html_url
stringlengths 48
51
| title
stringlengths 1
290
| comments
sequencelengths 0
30
| body
stringlengths 0
228k
⌀ | number
int64 2
7.08k
|
|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/5070 | Support default config name when no builder configs | [
"Thank you for creating this feature request, Albert.\r\n\r\nFor context this is the datatest where Albert has been helping me to switch to on-the-fly split config https://huggingface.co/datasets/HuggingFaceM4/cm4-synthetic-testing\r\n\r\nand the attempt to switch on-the-fly splits was here: https://huggingface.co/datasets/HuggingFaceM4/cm4-synthetic-testing/discussions/2/files\r\n\r\nbut which I had to revert since providing no split breaks at run time.\r\n"
] | **Is your feature request related to a problem? Please describe.**
As discussed with @stas00, we could support defining a default config name, even if no predefined allowed config names are set. That is, support `DEFAULT_CONFIG_NAME`, even when `BUILDER_CONFIGS` is not defined.
**Additional context**
In order to support creating configs on the fly **by name** (not using kwargs), the list of allowed builder configs `BUILDER_CONFIGS` must not be set.
However, if so, then `DEFAULT_CONFIG_NAME` is not supported.
| 5,070 |
https://github.com/huggingface/datasets/issues/5061 | `_pickle.PicklingError: logger cannot be pickled` in multiprocessing `map` | [
"This is maybe related to python 3.10, do you think you could try on 3.8 ?\r\n\r\nIn the meantime we'll keep improving the support for 3.10. Let me add a dedicated CI",
"I did some binary search and seems like the root cause is either `multiprocess` or `dill`. python 3.10 is fine. Specifically:\r\n- `multiprocess==0.70.12.2, dill==0.3.4`: works\r\n- `multiprocess==0.70.12.2, dill==0.3.5.1`: doesn't work\r\n- `multiprocess==0.70.13, dill==0.3.5.1`: doesn't work\r\n- `multiprocess==0.70.13, dill==0.3.4`: can't test, `multiprocess==0.70.13` requires `dill>=0.3.5.1`\r\n\r\nI will pin their versions on my end. I don't have enough knowledge of how python multiprocessing works to debug this, but ideally there could be a fix. It's also possible that I'm doing something wrong in my code, but again the `.name` of the logger that failed to pickle is `datasets.fingerprint`, which I'm not using directly.",
"Do you know which logger fails at being pickled ?",
"I'm not 100% sure how to figure it out -- the stack trace above doesn't clearly give me a place where I can print out who owns the logger, etc. I only found out its `.name` is `datasets.fingerprint` by printing right before\r\n```\r\n File \".../logging/__init__.py\", line 1774, in __reduce__\r\n raise pickle.PicklingError('logger cannot be pickled')\r\n```\r\nIf you have any idea on how to find it out, please let me know.",
"Ok I see, not sure why it triggers this error though, in `logging.py` the code is\r\n\r\nhttps://github.com/python/cpython/blob/c9da063e32725a66495e4047b8a5ed13e72d9e8e/Lib/logging/__init__.py#L1769-L1775\r\n\r\nand on my side it works on 3.10 with dill 0.3.5.1 and multiprocess 0.70.13\r\n```python\r\n>>> datasets.fingerprint.logger.__reduce__() \r\n(<function logging.getLogger(name=None)>, ('datasets.fingerprint',))\r\n```\r\nCould you try to run this code ?\r\n\r\nAre you in an environment where the loggers are instantiated differently ? Can you check the source code of `logging.Logger.__reduce__` in `\".../logging/__init__.py\", line 1774` ?",
"Closing due to inactivity."
] | ## Describe the bug
When I `map` with multiple processes, this error occurs. The `.name` of the `logger` that fails to pickle in the final line is `datasets.fingerprint`.
```
File "~/project/dataset.py", line 204, in <dictcomp>
split: dataset.map(
File ".../site-packages/datasets/arrow_dataset.py", line 2489, in map
transformed_shards[index] = async_result.get()
File ".../site-packages/multiprocess/pool.py", line 771, in get
raise self._value
File ".../site-packages/multiprocess/pool.py", line 537, in _handle_tasks
put(task)
File ".../site-packages/multiprocess/connection.py", line 214, in send
self._send_bytes(_ForkingPickler.dumps(obj))
File ".../site-packages/multiprocess/reduction.py", line 54, in dumps
cls(buf, protocol, *args, **kwds).dump(obj)
File ".../site-packages/dill/_dill.py", line 620, in dump
StockPickler.dump(self, obj)
File ".../pickle.py", line 487, in dump
self.save(obj)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../pickle.py", line 902, in save_tuple
save(element)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../site-packages/dill/_dill.py", line 1963, in save_function
_save_with_postproc(pickler, (_create_function, (
File ".../site-packages/dill/_dill.py", line 1140, in _save_with_postproc
pickler.save_reduce(*reduction, obj=obj)
File ".../pickle.py", line 717, in save_reduce
save(state)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../pickle.py", line 887, in save_tuple
save(element)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../site-packages/dill/_dill.py", line 1251, in save_module_dict
StockPickler.save_dict(pickler, obj)
File ".../pickle.py", line 972, in save_dict
self._batch_setitems(obj.items())
File ".../pickle.py", line 998, in _batch_setitems
save(v)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../site-packages/dill/_dill.py", line 1963, in save_function
_save_with_postproc(pickler, (_create_function, (
File ".../site-packages/dill/_dill.py", line 1140, in _save_with_postproc
pickler.save_reduce(*reduction, obj=obj)
File ".../pickle.py", line 717, in save_reduce
save(state)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../pickle.py", line 887, in save_tuple
save(element)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../site-packages/dill/_dill.py", line 1251, in save_module_dict
StockPickler.save_dict(pickler, obj)
File ".../pickle.py", line 972, in save_dict
self._batch_setitems(obj.items())
File ".../pickle.py", line 998, in _batch_setitems
save(v)
File ".../pickle.py", line 560, in save
f(self, obj) # Call unbound method with explicit self
File ".../site-packages/dill/_dill.py", line 1963, in save_function
_save_with_postproc(pickler, (_create_function, (
File ".../site-packages/dill/_dill.py", line 1154, in _save_with_postproc
pickler._batch_setitems(iter(source.items()))
File ".../pickle.py", line 998, in _batch_setitems
save(v)
File ".../pickle.py", line 578, in save
rv = reduce(self.proto)
File ".../logging/__init__.py", line 1774, in __reduce__
raise pickle.PicklingError('logger cannot be pickled')
_pickle.PicklingError: logger cannot be pickled
```
## Steps to reproduce the bug
Sorry I failed to have a minimal reproducible example, but the offending line on my end is
```python
dataset.map(
lambda examples: self.tokenize(examples), # this doesn't matter, lambda e: [1] * len(...) also breaks. In fact I'm pretty sure it breaks before executing this lambda
batched=True,
num_proc=4,
)
```
This does work when `num_proc=1`, so it's likely a multiprocessing thing.
## Expected results
`map` succeeds
## Actual results
The error trace above.
## Environment info
- `datasets` version: 1.16.1 and 2.5.1 both failed
- Platform: Ubuntu 20.04.4 LTS
- Python version: 3.10.4
- PyArrow version: 9.0.0
| 5,061 |
https://github.com/huggingface/datasets/issues/5060 | Unable to Use Custom Dataset Locally | [
"Hi ! I opened a PR in your repo to fix this :)\r\nhttps://huggingface.co/datasets/zpn/pubchem_selfies/discussions/7\r\n\r\nbasically you need to use `open` for streaming to work properly",
"Thank you so much for this! Naive question, is this a feature of `open` or have you all overloaded it to be able to read from a URL? Any links to code/documentation would be greatly appreciated, I'd love to learn more",
"`datasets` extends `open` in dataset scripts to work with URLs. The builtin `open` from python only works with local files.\r\n\r\nYou can find the extension here: https://github.com/huggingface/datasets/blob/6ad430ba0cdeeb601170f732d4bd977f5c04594d/src/datasets/download/streaming_download_manager.py#L435-L451\r\n\r\nI think we can create a docs section dedicated to streaming to explain how this works",
"Closing this one - feel free to reopen if you have more questions"
] | ## Describe the bug
I have uploaded a [dataset](https://huggingface.co/datasets/zpn/pubchem_selfies) and followed the instructions from the [dataset_loader](https://huggingface.co/docs/datasets/dataset_script#download-data-files-and-organize-splits) tutorial. In that tutorial, it says
```
If the data files live in the same folder or repository of the dataset script,
you can just pass the relative paths to the files instead of URLs.
```
Accordingly, I put the [relative path](https://huggingface.co/datasets/zpn/pubchem_selfies/blob/main/pubchem_selfies.py#L76) to the data to be used. I was able to test the dataset and generate the metadata locally with `datasets-cli test path/to/<your-dataset-loading-script> --save_infos --all_configs`
However, if I try to load the data using `load_dataset`, I get the following error
```
with gzip.open(filepath, mode="rt") as f:
File "/usr/local/Cellar/[email protected]/3.9.7_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/gzip.py", line 58, in open
binary_file = GzipFile(filename, gz_mode, compresslevel)
File "/usr/local/Cellar/[email protected]/3.9.7_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/gzip.py", line 173, in __init__
fileobj = self.myfileobj = builtins.open(filename, mode or 'rb')
FileNotFoundError: [Errno 2] No such file or directory: 'https://huggingface.co/datasets/zpn/pubchem_selfies/resolve/main/data/Compound_021000001_021500000/Compound_021000001_021500000_SELFIES.jsonl.gz'
```
## Steps to reproduce the bug
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset("zpn/pubchem_selfies", streaming=True)
>>> t = dataset["train"]
>>> for item in t:
...... print(item)
...... break
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/Users/zachnussbaum/env/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 723, in __iter__
for key, example in self._iter():
File "/Users/zachnussbaum/env/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 713, in _iter
yield from ex_iterable
File "/Users/zachnussbaum/env/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 113, in __iter__
yield from self.generate_examples_fn(**self.kwargs)
File "/Users/zachnussbaum/.cache/huggingface/modules/datasets_modules/datasets/zpn--pubchem_selfies/d2571f35996765aea70fd3f3f8e3882d59c401fb738615c79282e2eb1d9f7a25/pubchem_selfies.py", line 475, in _generate_examples
with gzip.open(filepath, mode="rt") as f:
File "/usr/local/Cellar/[email protected]/3.9.7_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/gzip.py", line 58, in open
binary_file = GzipFile(filename, gz_mode, compresslevel)
File "/usr/local/Cellar/[email protected]/3.9.7_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/gzip.py", line 173, in __init__
fileobj = self.myfileobj = builtins.open(filename, mode or 'rb')
FileNotFoundError: [Errno 2] No such file or directory: 'https://huggingface.co/datasets/zpn/pubchem_selfies/resolve/main/data/Compound_021000001_021500000/Compound_021000001_021500000_SELFIES.jsonl.gz'
````
```
## Expected results
A clear and concise description of the expected results.
## Actual results
Specify the actual results or traceback.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.5.1
- Platform: macOS-12.5.1-x86_64-i386-64bit
- Python version: 3.9.7
- PyArrow version: 9.0.0
- Pandas version: 1.5.0
| 5,060 |
https://github.com/huggingface/datasets/issues/5053 | Intermittent JSON parse error when streaming the Pile | [
"Maybe #2838 can help. In this PR we allow to skip bad chunks of JSON data to not crash the training\r\n\r\nDid you have warning messages before the error ?\r\n\r\nsomething like this maybe ?\r\n```\r\n03/24/2022 02:19:46 - WARNING - datasets.utils.streaming_download_manager - Got disconnected from remote data host. Retrying in 5sec [1/20]\r\n03/24/2022 02:20:01 - WARNING - datasets.utils.streaming_download_manager - Got disconnected from remote data host. Retrying in 5sec [2/20]\r\n03/24/2022 02:20:09 - ERROR - datasets.packaged_modules.json.json - Failed to read file 'gzip://file-000000000007.json::https://huggingface.co/datasets/lvwerra/codeparrot-clean-train/resolve/1d740acb9d09cf7a3307553323e2c677a6535407/file-000000000007.json.gz' with error <class 'pyarrow.lib.ArrowInvalid'>: JSON parse error: Invalid value. in row 0\r\n```",
"Ah, thanks! I did get errors like that. Sad that PR wasn't merged in! \r\n\r\nI'm currently just downloading 200GB of the Pile locally to avoid streaming (I have space and it's faster anyway), but that's really useful! I can probably apply the dumb patch of just commenting out the bits that raise the JSON Parse Error lol, based on your code - if I continue the loop should it be fine?",
"Yup you can get some inspiration from this PR. It simply ignores the bad chunks (a chunk is ~a few MBs of data).\r\nWe'll try to merge this PR soon"
] | ## Describe the bug
I have an intermittent error when streaming the Pile, where I get a JSON parse error which causes my program to crash.
This is intermittent - when I rerun the program with the same random seed it does not crash in the same way. The exact point this happens also varied - it happened to me 11B tokens and 4 days into a training run, and now just happened 2 minutes into one, but I can't reliably reproduce it.
I'm using a remote machine with 8 A6000 GPUs via runpod.io
## Expected results
I have a DataLoader which can iterate through the whole Pile
## Actual results
Stack trace:
```
Failed to read file 'zstd://12.jsonl::https://the-eye.eu/public/AI/pile/train/12.jsonl.zst' with error <class 'pyarrow.lib.ArrowInvalid'>: JSON parse error: Invalid value. in row 0
```
I'm currently using HuggingFace accelerate, which also gave me the following stack trace, but I've also experienced this problem intermittently when using DataParallel, so I don't think it's to do with parallelisation
```
Traceback (most recent call last):
File "ddp_script.py", line 1258, in <module>
main()
File "ddp_script.py", line 1143, in main
for c, batch in tqdm.tqdm(enumerate(data_iter)):
File "/opt/conda/lib/python3.7/site-packages/tqdm/std.py", line 1195, in __iter__
for obj in iterable:
File "/opt/conda/lib/python3.7/site-packages/accelerate/data_loader.py", line 503, in __iter__
next_batch, next_batch_info, next_skip = self._fetch_batches(main_iterator)
File "/opt/conda/lib/python3.7/site-packages/accelerate/data_loader.py", line 454, in _fetch_batches
broadcast_object_list(batch_info)
File "/opt/conda/lib/python3.7/site-packages/accelerate/utils/operations.py", line 333, in broadcast_object_list
torch.distributed.broadcast_object_list(object_list, src=from_process)
File "/opt/conda/lib/python3.7/site-packages/torch/distributed/distributed_c10d.py", line 1900, in broadcast_object_list
object_list[i] = _tensor_to_object(obj_view, obj_size)
File "/opt/conda/lib/python3.7/site-packages/torch/distributed/distributed_c10d.py", line 1571, in _tensor_to_object
return _unpickler(io.BytesIO(buf)).load()
_pickle.UnpicklingError: invalid load key, '@'.
```
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset(
cfg["dataset_name"], streaming=True, split="train")
dataset = dataset.remove_columns("meta")
dataset = dataset.map(tokenize_and_concatenate, batched=True)
dataset = dataset.with_format(type="torch")
train_data_loader = DataLoader(
dataset, batch_size=cfg["batch_size"], num_workers=3)
for batch in train_data_loader:
continue
```
`tokenize_and_concatenate` is a custom tokenization function I defined on the GPT-NeoX tokenizer to tokenize the text, separated by endoftext tokens, and reshape to have length batch_size, I don't think this is related to tokenization:
```
import numpy as np
import einops
import torch
def tokenize_and_concatenate(examples):
texts = examples["text"]
full_text = tokenizer.eos_token.join(texts)
div = 20
length = len(full_text) // div
text_list = [full_text[i * length: (i + 1) * length]
for i in range(div)]
tokens = tokenizer(text_list, return_tensors="np", padding=True)[
"input_ids"
].flatten()
tokens = tokens[tokens != tokenizer.pad_token_id]
n = len(tokens)
curr_batch_size = n // (seq_len - 1)
tokens = tokens[: (seq_len - 1) * curr_batch_size]
tokens = einops.rearrange(
tokens,
"(batch_size seq) -> batch_size seq",
batch_size=curr_batch_size,
seq=seq_len - 1,
)
prefix = np.ones((curr_batch_size, 1), dtype=np.int64) * \
tokenizer.bos_token_id
return {
"text": np.concatenate([prefix, tokens], axis=1)
}
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-5.4.0-105-generic-x86_64-with-debian-buster-sid
- Python version: 3.7.13
- PyArrow version: 9.0.0
- Pandas version: 1.3.5
ZStandard data:
Version: 0.18.0
Summary: Zstandard bindings for Python
Home-page: https://github.com/indygreg/python-zstandard
Author: Gregory Szorc
Author-email: [email protected]
License: BSD
Location: /opt/conda/lib/python3.7/site-packages
Requires:
Required-by: | 5,053 |
https://github.com/huggingface/datasets/issues/5050 | Restore saved format state in `load_from_disk` | [
"Hi, can I work on this?",
"Hi, sure! Let us know if you need some pointers/help."
] | Even though we save the `format` state in `save_to_disk`, we don't restore it in `load_from_disk`. We should fix that.
Reported here: https://discuss.huggingface.co/t/save-to-disk-loses-formatting-information/23815 | 5,050 |
https://github.com/huggingface/datasets/issues/5046 | Audiofolder creates empty Dataset if files same level as metadata | [
"Hi! Unfortunately, I can't reproduce this behavior. Instead, I get `ValueError: audio at 2063_fe9936e7-62b2-4e62-a276-acbd344480ce_1.wav doesn't have metadata in /audio-data/metadata.csv`, which can be fixed by removing the `./` from the file name.\r\n\r\n(Link to a Colab that tries to reproduce this behavior: https://colab.research.google.com/drive/1IhQzULYi0Van1xLrN_SddBX1JF7mLZZK?usp=sharing)",
"I think we can make the file name matching part more robust by replacing `file_name` with `os.path.normpath(file_name)`, to ignore \"./\" among other things, in these two places:\r\n* https://github.com/huggingface/datasets/blob/85cd129bde605cd9acacdff0d065fc02e39e09b1/src/datasets/packaged_modules/folder_based_builder/folder_based_builder.py#L319\r\n* https://github.com/huggingface/datasets/blob/85cd129bde605cd9acacdff0d065fc02e39e09b1/src/datasets/packaged_modules/folder_based_builder/folder_based_builder.py#L388",
"@mariosasko Some tests failed (see my PR). Any thoughts on that?",
"Yes, I mentioned the solution in my review.",
"I realized what I was doing wrong.\r\n\r\nThe documentation puts the files in a subfolder.\r\nOnce I have done that, it worked.\r\n\r\nBut l agree that this should be handled better if possible."
] | ## Describe the bug
When audio files are at the same level as the metadata (`metadata.csv` or `metadata.jsonl` ), the `load_dataset` returns a `DatasetDict` with no rows but the correct columns.
https://github.com/huggingface/datasets/blob/1ea4d091b7a4b83a85b2eeb8df65115d39af3766/docs/source/audio_dataset.mdx?plain=1#L88
## Steps to reproduce the bug
`metadata.csv`:
```csv
file_name,duration,transcription
./2063_fe9936e7-62b2-4e62-a276-acbd344480ce_1.wav,10.768,hello
```
```python
>>> audio_dataset = load_dataset("audiofolder", data_dir="/audio-data/")
>>> audio_dataset
DatasetDict({
train: Dataset({
features: ['audio', 'duration', 'transcription'],
num_rows: 0
})
validation: Dataset({
features: ['audio', 'duration', 'transcription'],
num_rows: 0
})
})
```
I've tried, with no success,:
- setting `split` to something else so I don't get a `DatasetDict`,
- removing the `./`,
- using `.jsonl`.
## Expected results
```
Dataset({
features: ['audio', 'duration', 'transcription'],
num_rows: 1
})
```
## Actual results
```
DatasetDict({
train: Dataset({
features: ['audio', 'duration', 'transcription'],
num_rows: 0
})
validation: Dataset({
features: ['audio', 'duration', 'transcription'],
num_rows: 0
})
})
```
## Environment info
- `datasets` version: 2.5.1
- Platform: Linux-5.13.0-1025-aws-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 9.0.0
- Pandas version: 1.5.0
| 5,046 |
https://github.com/huggingface/datasets/issues/5045 | Automatically revert to last successful commit to hub when a push_to_hub is interrupted | [
"Could you share the error you got please ? Maybe the full stack trace if you have it ?\r\n\r\nMaybe `push_to_hub` be implemented as a single commit @Wauplin ? This way if it fails, the repo is still at the previous (valid) state instead of ending-up in an invalid/incimplete state.",
"> Maybe push_to_hub be implemented as a single commit ? \r\n\r\nI think that would definitely be the way to go. Do you know the reasons why not implementing it like this in the first place ? I guess it is because of not been able to upload all at once with `huggingface_hub` but if there was another reason, please let me know.\r\nAbout pushing all at once, it seems to be a more and more requested feature. I have created this issue https://github.com/huggingface/huggingface_hub/issues/1085 recently but other discussions already happened in the past. The `moon-landing` team is working on it (cc @coyotte508). The `huggingface_hub` integration will come afterwards.\r\n\r\nFor now, maybe it's best to wait for a proper implementation instead of creating a temporary workaround :)\r\n",
"> I think that would definitely be the way to go. Do you know the reasons why not implementing it like this in the first place ? I guess it is because of not been able to upload all at once with huggingface_hub but if there was another reason, please let me know.\r\n\r\nIdeally we would want to upload the files iteratively - and then once everything is uploaded we proceed to commit. When we implemented `push_to_hub`, using `upload_file` for each shard was the only option.\r\n\r\nFor more context: for each shard to upload we do:\r\n1. load the arrow shard in memory\r\n2. convert to parquet\r\n3. upload\r\n\r\nSo to avoid OOM we need to upload the files iteratively.\r\n\r\n> For now, maybe it's best to wait for a proper implementation instead of creating a temporary workaround :)\r\n\r\nLet us know if we can help !",
"> Ideally we would want to upload the files iteratively - and then once everything is uploaded we proceed to commit. \r\n\r\nOh I see. So maybe this has to be done in an implementation specific to `datasets/` as it is not a very common case (upload a bunch of files on the fly).\r\n\r\nYou can maybe have a look at how `huggingface_hub` is implemented for LFS files (arrow shards are LFS anyway, right?).\r\nIn [`upload_lfs_files`](https://github.com/huggingface/huggingface_hub/blob/e28646c977fc9304a4c3576ce61ff07f9778950b/src/huggingface_hub/_commit_api.py#L164) LFS files are uploaded 1 by 1 (multithreaded) and then [the commit is pushed](https://github.com/huggingface/huggingface_hub/blob/e28646c977fc9304a4c3576ce61ff07f9778950b/src/huggingface_hub/hf_api.py#L1926) to the Hub once all files have been uploaded. This is pretty much what you need, right ?\r\n\r\nI can help you if you have questions how to do it in `datasets`. If that makes sense we could then move the implementation from `datasets` to `huggingface_hub` once it's mature. Next week I'm on holidays but feel free to start without my input.\r\n\r\n(also cc @coyotte508 and @SBrandeis who implemented LFS upload in `hfh`)",
"> Could you share the error you got please ? Maybe the full stack trace if you have it ?\r\n\r\nHere’s part of the stack trace, that I can reproduce at the moment from a photo I took (potential typos from OCR):\r\n```\r\nValueError\r\nTraceback (most recent call last)\r\n<ipython-input-4-274613b7d3f5> in <module>\r\nfrom datasets import load dataset\r\nds = load_dataset('jrahn/chessv6', use_auth_token-True)\r\n\r\n/us/local/1ib/python3.7/dist-packages/datasets/table.py in cast_table _to_schema (table, schema)\r\nLine 2005 raise ValueError()\r\n\r\nValueError: Couldn't cast \r\nfen: string \r\nmove: string \r\nres: string \r\neco: string \r\nmove_id: int64\r\nres_num: int64 to\r\n{ 'fen': Value(dtype='string', id=None), \r\n'move': Value(dtype=' string', id=None),\r\n'res': Value(dtype='string', id=None),\r\n'eco': Value(dtype='string', id=None), \r\n'hc': Value(dtype='string', id=None), \r\n'move_ id': Value(dtype='int64', id=None),\r\n'res_num': Value(dtype= 'int64' , id=None) }\r\nbecause column names don't match \r\n```\r\n\r\nThe column 'hc' was removed before the interrupted push_to_hub(). It appears in the column list in curly brackets but not in the column list above.\r\n\r\nLet me know, if I can be of any help."
] | **Is your feature request related to a problem? Please describe.**
I pushed a modification of a large dataset (remove a column) to the hub. The push was interrupted after some files were committed to the repo. This left the dataset to raise an error on load_dataset() (ValueError couldn’t cast … because column names don’t match). Only by specifying the previous (complete) commit as revision=commit_hash in load_data(), I was able to repair this and after a successful, complete push, the dataset loads without error again.
**Describe the solution you'd like**
Would it make sense to detect an incomplete push_to_hub() and automatically revert to the previous commit/revision?
**Describe alternatives you've considered**
Leave everything as is, the revision parameter in load_dataset() allows to manually fix this problem.
**Additional context**
Provide useful defaults
| 5,045 |
https://github.com/huggingface/datasets/issues/5044 | integrate `load_from_disk` into `load_dataset` | [
"I agree the situation is not ideal and it would be awesome to use `load_dataset` to reload a dataset saved locally !\r\n\r\nFor context:\r\n\r\n- `load_dataset` works in three steps: download the dataset, then prepare it as an arrow dataset, and finally return a memory mapped arrow dataset. In particular it creates a cache directory to store the arrow data and the subsequent cache files for `map`.\r\n\r\n- `load_from_disk` directly returns a memory mapped dataset from the arrow file (similar to `Dataset.from_file`). It doesn't create a cache diretory, instead all the subsequent `map` calls write in the same directory as the original data. \r\n\r\nIf we want to keep the download_and_prepare step for consistency, it would unnecessarily copy the arrow data into the datasets cache. On the other hand if we don't do this step, the cache directory doesn't exist which is inconsistent.\r\n\r\nI'm curious, what would you expect to happen in this situation ?",
"Thank you for the detailed breakdown, @lhoestq \r\n\r\n> I'm curious, what would you expect to happen in this situation ?\r\n\r\n1. the simplest solution is to add a flag to the dataset saved by `save_to_disk` and have `load_dataset` check that flag - if it's set simply switch control to `load_from_disk` behind the scenes. So `load_dataset` detects it's a local filesystem, looks inside to see whether it's something it can cache or whether it should use it directly as is and continues accordingly with one of the 2 dataset-type specific APIs.\r\n\r\n2. the more evolved solution is to look at a dataset produced by `save_to_disk` as a remote resource like hub. So the first time `load_dataset` sees it, it'll take a fingerprint and create a normal cached dataset. On subsequent uses it'll again discover it as a remote resource, validate that it has it cached via the fingerprint and serve as a normal dataset. \r\n\r\nAs you said the cons of approach 2 is that if the dataset is huge it'll make 2 copies on the same machine. So it's possible that both approaches can be integrated. Say if `save_to_disc(do_not_cache=True)` is passed it'll use solution 1, otherwise solution 2. or could even symlink the huge arrow files to the cache instead? or perhaps it's more intuitive to use `load_dataset(do_not_cache=True)` instead. So that one can choose whether to make a cached copy or not for the locally saved dataset. i.e. a simple at use point user control.\r\n\r\nSurely there are other ways to handle it, this is just one possibility.\r\n",
"I think the simplest is to always memory map the local file without copy, but still have a cached directory in the cache at `~/.cache/huggingface` instead of saving `map` results next to the original data.\r\n\r\nIn practice we can even use symlinks if it makes the implementation simpler",
"Yes, so that you always have the cached entry for any dataset, but the \"payload\" doesn't have to be physically in the cache if it's already on the local filesystem. As you said a symlink will do. ",
"Any updates?",
"We haven't had the bandwidth to implement this so far. Let me know if you'd be interested in contributing this feature :)",
"@lhoestq I can jump into that. What I don't like is having functions with many parameters input. Even though they are optional, it's always harder to reason about and test such cases.\r\nIf there are more features worth to work on, feel free to ping me. It's a lot of fun to help :smile: ",
"Thanks a lot for your help @mariusz-jachimowicz-83 :)\r\n\r\nI think as a first step we could implement an Arrow dataset builder to be able to load and stream Arrow datasets locally or from Hugging Face. Maybe something similar to the Parquet builder at [src/datasets/packaged_modules/parquet/parquet.py](https://github.com/huggingface/datasets/blob/main/src/datasets/packaged_modules/parquet/parquet.py) ?\r\n\r\nAnd we can deal with the disk space optimization as a second step. What do you think ?\r\n\r\n(this issue is also related to https://github.com/huggingface/datasets/issues/3035)",
"@lhoestq I made a PR based on suggestion https://github.com/huggingface/datasets/pull/5944. Could you please review it?",
"@lhoestq Let me know if you have further recommendations or anything that you would like to add but you don't have bandwith for. ",
"Any update on this issue? It makes existing scripts and examples fall flat when provided with a customized/preprocessed dataset saved to disk."
] | **Is your feature request related to a problem? Please describe.**
Is it possible to make `load_dataset` more universal similar to `from_pretrained` in `transformers` so that it can handle the hub, and the local path datasets of all supported types?
Currently one has to choose a different loader depending on how the dataset has been created.
e.g. this won't work:
```
$ git clone https://huggingface.co/datasets/severo/test-parquet
$ python -c 'from datasets import load_dataset; ds=load_dataset("test-parquet"); \
ds.save_to_disk("my_dataset"); load_dataset("my_dataset")'
[...]
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/load.py", line 1746, in load_dataset
builder_instance.download_and_prepare(
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/builder.py", line 793, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/builder.py", line 1277, in _prepare_split
writer.write_table(table)
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_writer.py", line 524, in write_table
pa_table = table_cast(pa_table, self._schema)
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/table.py", line 2005, in table_cast
return cast_table_to_schema(table, schema)
File "/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/table.py", line 1968, in cast_table_to_schema
raise ValueError(f"Couldn't cast\n{table.schema}\nto\n{features}\nbecause column names don't match")
ValueError: Couldn't cast
_data_files: list<item: struct<filename: string>>
child 0, item: struct<filename: string>
child 0, filename: string
```
both times the dataset is being loaded from disk. Why does it fail the second time?
Why can't `save_to_disk` generate a dataset that can be immediately loaded by `load_dataset`?
e.g. the simplest hack would be to have `save_to_disk` add some flag to the saved dataset, that tells `load_dataset` to internally call `load_from_disk`. like having `save_to_disk` create a `load_me_with_load_from_disk.txt` file ;) and `load_dataset` will support that feature from saved datasets from new `datasets` versions. The old ones will still need to use `load_from_disk` explicitly. Unless the flag is not needed and one can immediately tell by looking at the saved dataset that it was saved via `save_to_disk` and thus use `load_from_disk` internally.
The use-case is defining a simple API where the user only ever needs to pass a `dataset_name_or_path` and it will always just work. Currently one needs to manually add additional switches telling the system whether to use one loading method or the other which works but it's not smooth.
Thank you! | 5,044 |
https://github.com/huggingface/datasets/issues/5039 | Hendrycks Checksum | [
"Thanks for reporting, @DanielHesslow. We are fixing it. ",
"@albertvillanova thanks for taking care of this so quickly!",
"The dataset metadata is fixed. You can download it normally."
] | Hi,
The checksum for [hendrycks_test](https://huggingface.co/datasets/hendrycks_test) does not compare correctly, I guess it has been updated on the remote.
```
datasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://people.eecs.berkeley.edu/~hendrycks/data.tar']
```
| 5,039 |
https://github.com/huggingface/datasets/issues/5038 | `Dataset.unique` showing wrong output after filtering | [
"Hi! It seems like `flatten_indices` (called in `unique`) doesn't know how to handle empty indices mappings. I'm working on the fix.",
"Thanks, that was fast!"
] | ## Describe the bug
After filtering a dataset, and if no samples remain, `Dataset.unique` will return the unique values of the unfiltered dataset.
## Steps to reproduce the bug
```python
from datasets import Dataset
dataset = Dataset.from_dict({'id': [0]})
dataset = dataset.filter(lambda _: False)
print(dataset.unique('id'))
```
## Expected results
The above code should return an empty list since the dataset is empty.
## Actual results
```bash
[0]
```
## Environment info
- `datasets` version: 2.5.1
- Platform: Linux-5.18.19-100.fc35.x86_64-x86_64-with-glibc2.34
- Python version: 3.9.14
- PyArrow version: 7.0.0
- Pandas version: 1.3.5 | 5,038 |
https://github.com/huggingface/datasets/issues/5032 | new dataset type: single-label and multi-label video classification | [
"Hi ! You can in the `features` folder how we implemented the audio and image feature types.\r\n\r\nWe can have something similar to videos. What we need to decide:\r\n- the video loading library to use\r\n- the output format when a user accesses a video type object\r\n- what parameters a `Video()` feature type needs\r\n\r\nalso cc @nateraw who also took a look at what we can do for video",
"@lhoestq @nateraw is there any progress on adding video classification datasets? ",
"Hi ! I think we just missing which lib we're going to use to decode the videos + which parameters must go in the `Video` type",
"Hmm. `decord` could be nice but it's no longer maintained [it seems](https://github.com/dmlc/decord/issues/214). ",
"pytorchvideo uses [pyav](https://github.com/PyAV-Org/PyAV) as the default decoder: https://github.com/facebookresearch/pytorchvideo/blob/c8d23d8b7e597586a9e2d18f6ed31ad8aa379a7a/pytorchvideo/data/labeled_video_dataset.py#L37\r\n\r\nAlso it would be great if `optionally` audio can also be decoded from the video as in pytorchvideo: https://github.com/facebookresearch/pytorchvideo/blob/c8d23d8b7e597586a9e2d18f6ed31ad8aa379a7a/pytorchvideo/data/labeled_video_dataset.py#L35\r\n\r\nHere are the other decoders supported in pytorchvideo: https://github.com/facebookresearch/pytorchvideo/blob/c8d23d8b7e597586a9e2d18f6ed31ad8aa379a7a/pytorchvideo/data/encoded_video.py#L17\r\n",
"@sayakpaul I did do quite a bit of work on [this PR](https://github.com/huggingface/datasets/pull/4532) a while back to add a video feature. It's outdated, but uses my `encoded_video` [package](https://github.com/nateraw/encoded-video) under the hood, which is basically a wrapper around PyAV stolen from [pytorchvideo](https://github.com/facebookresearch/pytorchvideo/) that gets rid of the `torch` dependency. \r\n\r\nwould be really great to get something like this in...it's just a really tricky and time consuming feature to add. "
] | **Is your feature request related to a problem? Please describe.**
In my research, I am dealing with multi-modal (audio+text+frame sequence) video classification. It would be great if the datasets library supported generating multi-modal batches from a video dataset.
**Describe the solution you'd like**
Assume I have video files having single/multiple labels. I want to train a single/multi-label video classification model. I want datasets to support generating multi-modal batches (audio+frame sequence) from video files. Audio waveform and frame sequence can be extracted from each video clip then I can use any audio, image and video model from transformers library to extract features which will be fed into my model.
**Describe alternatives you've considered**
Currently, I am using https://github.com/facebookresearch/pytorchvideo dataloaders. There seems to be not much alternative.
**Additional context**
I am wiling to open a PR but don't know where to start.
| 5,032 |
https://github.com/huggingface/datasets/issues/5028 | passing parameters to the method passed to Dataset.from_generator() | [
"Hi! Yes, you can either use the `gen_kwargs` param in `Dataset.from_generator` (`ds = Dataset.from_generator(gen, gen_kwargs={\"param1\": val})`) or wrap the generator function with `functools.partial`\r\n(`ds = Dataset.from_generator(functools.partial(gen, param1=\"val\"))`) to pass custom parameters to it.\r\n"
] | Big thanks for providing dataset creation via a generator.
I want to ask whether there is any way that parameters can be passed to the method Dataset.from_generator() method, like as follows.
```
from datasets import Dataset
def gen(param1):
for idx in len(custom_dataset):
yield custom_dataset[idx] + param1
ds = Dataset.from_generator(gen(param1))
```
| 5,028 |
https://github.com/huggingface/datasets/issues/5025 | Custom Json Dataset Throwing Error when batch is False | [
"Hi! Our processors are meant to be used in `batched` mode, so if `batched` is `False`, you need to drop the batch dimension (the error message warns you that the array has an extra dimension meaning it's 4D instead of 3D) to avoid the error:\r\n```python\r\ndef prepare_examples(examples):\r\n #Some preporcessing for each image and text as all my data saved in cloud\r\n #For this reason I couldn't set the batch to True. \r\n encoding = processor(img_as_tensor, words, boxes=boxes, word_labels=labels,\r\n truncation=True, padding=\"max_length\", return_tensors=\"np\")\r\n # drop extra dim\r\n for k in encoding.items():\r\n encoding[k]=encoding[k][0]\r\n return encoding\r\n```",
"> Hi! Our processors are meant to be used in `batched` mode, so if `batched` is `False`, you need to drop the batch dimension (the error message warns you that the array has an extra dimension meaning it's 4D instead of 3D) to avoid the error:\r\n> \r\n> ```python\r\n> def prepare_examples(examples):\r\n> #Some preporcessing for each image and text as all my data saved in cloud\r\n> #For this reason I couldn't set the batch to True. \r\n> encoding = processor(img_as_tensor, words, boxes=boxes, word_labels=labels,\r\n> truncation=True, padding=\"max_length\", return_tensors=\"np\")\r\n> # drop extra dim\r\n> for k in encoding.items():\r\n> encoding[k]=encoding[k][0]\r\n> return encoding\r\n> ```\r\n\r\nThank you it did work\r\n\r\n```\r\nfor k,v in encoding.items():\r\n encoding[k]=encoding[k][0]\r\n```"
] | ## Describe the bug
A clear and concise description of what the bug is.
I tried to create my custom dataset using below code
```
from datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D
from torchvision import transforms
from transformers import AutoProcessor
# we'll use the Auto API here - it will load LayoutLMv3Processor behind the scenes,
# based on the checkpoint we provide from the hub
from datasets import load_dataset
def prepare_examples(examples):
#Some preporcessing for each image and text as all my data saved in cloud
#For this reason I couldn't set the batch to True.
encoding = processor(img_as_tensor, words, boxes=boxes, word_labels=labels,
truncation=True, padding="max_length")
# encoding['pixel_values']=np.array(encoding['pixel_values'])
return encoding
dataset = load_dataset("json", data_files='issues.jsonl')
processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)
features = dataset["train"].features
column_names = dataset["train"].column_names
# we need to define custom features for `set_format` (used later on) to work properly
features = Features({
'pixel_values': Array3D(dtype="float32", shape=(3, 224, 224)),
'input_ids': Sequence(feature=Value(dtype='int64')),
'attention_mask': Sequence(Value(dtype='int64')),
'bbox': Array2D(dtype="int64", shape=(512, 4)),
'labels': Sequence(feature=Value(dtype='int64')),
})
train_dataset = dataset["train"].map(
prepare_examples,
batched=False,
remove_columns=column_names,
features=features
)
```
It throws below error.
```
/opt/conda/lib/python3.7/site-packages/datasets/arrow_writer.py in __arrow_array__(self, type)
172 storage = to_pyarrow_listarray(data, pa_type)
--> 173 return pa.ExtensionArray.from_storage(pa_type, storage)
174
/opt/conda/lib/python3.7/site-packages/pyarrow/array.pxi in pyarrow.lib.ExtensionArray.from_storage()
TypeError: Incompatible storage type list<item: list<item: list<item: list<item: float>>>> for extension type extension<arrow.py_extension_type<Array3DExtensionType>>
```
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
```
rom datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D
from torchvision import transforms
from transformers import AutoProcessor
# we'll use the Auto API here - it will load LayoutLMv3Processor behind the scenes,
# based on the checkpoint we provide from the hub
from datasets import load_dataset
def prepare_examples(examples):
#Some preporcessing for each image and text as all my data saved in cloud
encoding = processor(img_as_tensor, words, boxes=boxes, word_labels=labels,
truncation=True, padding="max_length")
# encoding['pixel_values']=np.array(encoding['pixel_values'])
return encoding
dataset = load_dataset("json", data_files='issues.jsonl')
processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)
features = dataset["train"].features
column_names = dataset["train"].column_names
# we need to define custom features for `set_format` (used later on) to work properly
features = Features({
'pixel_values': Array3D(dtype="float32", shape=(3, 224, 224)),
'input_ids': Sequence(feature=Value(dtype='int64')),
'attention_mask': Sequence(Value(dtype='int64')),
'bbox': Array2D(dtype="int64", shape=(512, 4)),
'labels': Sequence(feature=Value(dtype='int64')),
})
train_dataset = dataset["train"].map(
prepare_examples,
batched=False,
remove_columns=column_names,
features=features
)
## Expected results
A clear and concise description of the expected results.
Expected would be similar to all the otherdatasets with no error.
## Actual results
Specify the actual results or traceback.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: Unix
- Python version: 3.9
- PyArrow version: 9.0.0
| 5,025 |
https://github.com/huggingface/datasets/issues/5023 | Text strings are split into lists of characters in xcsr dataset | [] | ## Describe the bug
Text strings are split into lists of characters.
Example for "X-CSQA-en":
```
{'id': 'd3845adc08414fda',
'lang': 'en',
'question': {'stem': ['T',
'h',
'e',
' ',
'd',
'e',
'n',
't',
'a',
'l',
' ',
'o',
'f',
'f',
'i',
'c',
'e',
' ',
'h',
'a',
'n',
'd',
'l',
'e',
'd',
' ',
'a',
' ',
'l',
'o',
't',
' ',
'o',
'f',
' ',
'p',
'a',
't',
'i',
'e',
'n',
't',
's',
' ',
'w',
'h',
'o',
' ',
'e',
'x',
'p',
'e',
'r',
'i',
'e',
'n',
'c',
'e',
'd',
' ',
't',
'r',
'a',
'u',
'm',
'a',
't',
'i',
'c',
' ',
'm',
'o',
'u',
't',
'h',
' ',
'i',
'n',
'j',
'u',
'r',
'y',
',',
' ',
'w',
'h',
'e',
'r',
'e',
' ',
'w',
'e',
'r',
'e',
' ',
't',
'h',
'e',
's',
'e',
' ',
'p',
'a',
't',
'i',
'e',
'n',
't',
's',
' ',
'c',
'o',
'm',
'i',
'n',
'g',
' ',
'f',
'r',
'o',
'm',
'?'],
'choices': [{'label': ['A'], 'text': ['t', 'o', 'w', 'n']},
{'label': ['B'], 'text': ['m', 'i', 'c', 'h', 'i', 'g', 'a', 'n']},
{'label': ['C'], 'text': ['h', 'o', 's', 'p', 'i', 't', 'a', 'l']},
{'label': ['D'], 'text': ['s', 'c', 'h', 'o', 'o', 'l', 's']},
{'label': ['E'],
'text': ['o',
'f',
'f',
'i',
'c',
'e',
' ',
'b',
'u',
'i',
'l',
'd',
'i',
'n',
'g']}]},
'answerKey': 'C'}
## Steps to reproduce the bug
```python
ds = load_dataset("datasets/xcsr", "X-CSQA-en", split="validation", streaming=True)
item = next(iter(ds))
item
```
## Expected results
```
{'id': 'd3845adc08414fda',
'lang': 'en',
'question': {'stem': 'The dental office handled a lot of patients who experienced traumatic mouth injury, where were these patients coming from?',
'choices': {'label': ['A', 'B', 'C', 'D', 'E'],
'text': ['town', 'michigan', 'hospital', 'schools', 'office building']}},
'answerKey': 'C'}
```
| 5,023 |
https://github.com/huggingface/datasets/issues/5021 | Split is inferred from filename and overrides metadata.jsonl | [
"Hi! What's the structure of your image folder? `datasets` by default tries to infer to what split each file belongs based on directory/file names. If it's OK to load all the images inside the `dataset` folder in the `train` split, you can do the following:\r\n```python\r\ndataset = load_dataset(\"imagefolder\", data_files=\"dataset/**\")\r\n```",
"Thanks! Specifying `data_files` worked for that case.\r\n\r\nI'm new to the library, so let me try rephrasing the issue. If there's no actual bug here, sorry for the trouble.\r\n\r\nI've uploaded an example [here](https://files.catbox.moe/nfj2pd.zip) with the following files: \r\n\r\n```\r\n.\r\n├── bug.py\r\n└── imagefolder\r\n ├── test\r\n │ ├── metadata.jsonl\r\n │ ├── dog.jpg\r\n │ └── personal trainer.jpg\r\n └── train\r\n ├── metadata.jsonl\r\n ├── cat.jpg\r\n └── testing center.jpg\r\n```\r\n\r\n`bug.py`\r\n```\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"imagefolder\")\r\n\r\nprint(dataset)\r\n# DatasetDict({\r\n# test: Dataset({\r\n# features: ['image', 'text'],\r\n# num_rows: 1\r\n# })\r\n# })\r\n\r\nfor split in dataset:\r\n print(\"Split:\", split)\r\n for n in dataset[split]:\r\n print(n['text'])\r\n\r\n\r\n# Split: test\r\n# testing center\r\n```\r\n\r\nAs far as I can tell, this conforms with the example given here: https://huggingface.co/docs/datasets/image_dataset#imagefolder. It appears to me that, even though `metadata.jsonl` is present, the inferred labels from the path are taking precedent. Does this sound like a bug/undocumented behavior?",
"This looks like a duplicate of https://github.com/huggingface/datasets/issues/4895 (the problem is explained in this comment: https://github.com/huggingface/datasets/issues/4895#issuecomment-1248269550).\r\n\r\nIn the meantime, you can do the following to fetch all the splits:\r\n```python\r\ndataset = load_dataset(\"imagefolder\", data_files={\"train\": \"imagefolder/train/**\", \"test\": \"imagefolder/test/**\"})\r\n```\r\n"
] | ## Describe the bug
Including the strings "test" or "train" anywhere in a filename causes `datasets` to infer the split and silently ignore all other files.
This behavior is documented for directory names but not filenames: https://huggingface.co/docs/datasets/image_dataset#imagefolder
## Steps to reproduce the bug
`metadata.jsonl`
```json
{"file_name": "photo of a cat.jpg", "text": "a photo of a cat"}
{"file_name": "photo of a dog.jpg", "text": "a photo of a dog"}
{"file_name": "photo of a train.jpg", "text": "a photo of a train"}
{"file_name": "photo of test tubes.jpg", "text": "a photo of test tubes"}
```
`bug.py`
```python
from datasets import load_dataset
dataset = load_dataset("dataset")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['image', 'text'],
# num_rows: 1
# })
# test: Dataset({
# features: ['image', 'text'],
# num_rows: 1
# })
# })
for split in dataset:
for n in dataset[split]:
print(n['text'])
# a photo of a train
# a photo of test tubes
```
## Expected results
One single dataset with all four images / a warning for unused files / documentation of this behavior
## Actual results
Only the images with "test" or "train" in the name are loaded
## Environment info
- `datasets` version: 2.5.1
- Platform: macOS-12.5.1-x86_64-i386-64bit
- Python version: 3.10.4
- PyArrow version: 9.0.0
- Pandas version: 1.5.0 | 5,021 |
https://github.com/huggingface/datasets/issues/5017 | xcsr: X-CSQA simply uses english for all alleged non-english data | [
"Thanks for reporting, @thesofakillers. Good catch. We are fixing this. "
] | ## Describe the bug
All the alleged non-english subcollections for the X-CSQA task in the [xcsr benchmark dataset ](https://huggingface.co/datasets/xcsr) seem to be copies of the english subcollection, rather than translations. This is in contrast to the data description:
> we automatically translate the original CSQA and CODAH datasets, which only have English versions, to 15 other languages, forming development and test sets for studying X-CSR
## Steps to reproduce the bug
```python
# let's say you want to load the french X-CSQA subcollection
french = datasets.load_dataset("xcsr", "X-CSQA-fr")
# for good measure, let's load english too
english = datasets.load_dataset("xcsr", "X-CSQA-en")
# let's inspect
"".join(english['test'][0]['question']['stem'])
# output: 'The people wanted to stop the parade, so what did they set up to thwart it?'
"".join(french['test'][0]['question']['stem'])
# output: 'The people wanted to stop the parade, so what did they set up to thwart it?'
# what? Why are they both in english?
# I've checked this for validation and train splits too, across many datapoints. It's all the same english dataset
# maybe i need to look better?
french['test'].unique('lang')
# output: ['en']
# no, it's all english
```
## Expected results
Accessing a subcollection in language X should return a subcollection containg samples in language X
## Actual results
Accessing a subcollection in language X returns a subcollection containing samples in English.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.5.1
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.3
| 5,017 |
https://github.com/huggingface/datasets/issues/5015 | Transfer dataset scripts to Hub | [
"Sounds good ! Can I help with anything ?"
] | Before merging:
- #4974
TODO:
- [x] Create label: ["dataset contribution"](https://github.com/huggingface/datasets/pulls?q=label%3A%22dataset+contribution%22)
- [x] Create project: [Datasets: Transfer datasets to Hub](https://github.com/orgs/huggingface/projects/22/)
- [x] PRs:
- [x] Add dataset: we should recommend transfer all additions of datasets to the Hub, under the appropriate namespace; no more additions of datasets on GitHub
- [x] Update dataset: in general, we should merge bug fixes; enhancements should be considered on a case-by-case basis, depending on whether there is a more suitable namespace on the Hub
- [ ] Issues
Finally:
- [x] #4974
Let me know what you think! :hugs: | 5,015 |
https://github.com/huggingface/datasets/issues/5014 | I need to read the custom dataset in conll format | [
"Hi! We don't currently have a builder for parsing custom `conll` datasets, but I guess we could add one as a packaged module (similarly to what [TFDS](https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/core/dataset_builders/conll/conll_dataset_builder.py) did). @lhoestq @albertvillanova WDYT?\r\n\r\nIn the meantime, you can use `Dataset.from_generator` to create a dataset as follows:\r\n```python\r\nfrom datasets import Dataset\r\n\r\n# 2009 version\r\nINPUT_COLUMNS = \"ID FORM LEMMA PLEMMA POS PPOS FEAT PFEAT HEAD PHEAD DEPREL PDEPREL\".split()\r\n\r\ndef read_conll(file):\r\n example = {col: [] for col in INPUT_COLUMNS}\r\n idx = 0\r\n with open(file) as f:\r\n for line in f:\r\n if line.startswith(\"-DOCSTART-\") or line == \"\\n\" or not line:\r\n if example[next(iter(example))]:\r\n yield idx, example\r\n idx += 1\r\n example = {col: [] for col in INPUT_COLUMNS}\r\n else:\r\n row_cols = line.split()\r\n for i, col in enumerate(example):\r\n example[col] = row_cols[i].rstrip()\r\n\r\n# (optional) pass custom features with `features=Features(...)`\r\ndset = Dataset.from_generator(read_conll, gen_kwargs={\"file\": \"path/to/conll/file\"}) \r\n``` ",
"I think we could add a dedicated builder if you think this format is general enough.",
"\r\n\r\n\r\n> I think we could add a dedicated builder if you think this format is general enough.\r\n\r\nI think its functions are incomplete. It should have to_ Conll and from_ There are two methods of conll."
] | I need to read the custom dataset in conll format
| 5,014 |
https://github.com/huggingface/datasets/issues/5013 | would huggingface like publish cpp binding for datasets package ? | [
"Hi ! Can you share more information about your use case ? How could it help you to have cpp bindings versus using the python libraries ?",
"> Hi ! Can you share more information about your use case ? How could it help you to have cpp bindings versus using the python libraries ?\r\n\r\nfor example ,the huggingface load_model() and load_dataset() can execute in cpp env",
"If it's a viable option for you, you can check [tch-rs](https://github.com/LaurentMazare/tch-rs) to load models in Rust. Regarding datasets, you can first download them in python and then use Arrow C++ or Rust to load them",
"If you are more adventurous, another option is to embed python calls inside c++ e.g. with `pybind11`.",
"> pybind11\r\n\r\nI think it is not the best solution"
] | HI:
I use cpp env libtorch, I like use hugggingface ,but huggingface not cpp binding, would you like publish cpp binding for it.
thanks | 5,013 |
https://github.com/huggingface/datasets/issues/5012 | Force JSON format regardless of file naming on S3 | [
"Hi ! Support for URIs like `s3://...` is not implemented yet in `data_files=`. You can use the HTTP URL instead if your data is public in the meantime",
"Hi,\r\nI want to make sure I understand this response. I have a set of files on S3 that are private for security reasons. Because they are not public files I cannot read those files (many are parquet) into my hf notebooks in Kaggle? That can't be correct, can it? ",
"Hi ! There is a discussion at https://github.com/huggingface/datasets/issues/5281\r\n\r\nUsing the latest `datasets` 2.11 you can try passing fsspec URLs to private buckets to `data_files` in `load_dataset()`. Though this is still experimental and undocumented, so feedback is welcome. You may not have the best experience though, since anything related to performance and caching hasn't been tested properly yet.",
"closing this one since data_files supports fsspec (still experimental/untested/undocumented for s3 though)"
] | I have a file on S3 created by Data Version Control, it looks like `s3://dvc/ac/badff5b134382a0f25248f1b45d7b2` but contains a json file. If I run
```python
dataset = load_dataset(
"json",
data_files='s3://dvc/ac/badff5b134382a0f25248f1b45d7b2'
)
```
It gives me
```
InvalidSchema: No connection adapters were found for 's3://dvc/ac/badff5b134382a0f25248f1b45d7b2'
```
However, I cannot go ahead and change the names of the s3 file. Is there a way to "force" load a S3 url with certain decoder (JSON, CSV, etc.) regardless of s3 URL naming? | 5,012 |
https://github.com/huggingface/datasets/issues/5011 | Audio: `encode_example` fails with IndexError | [
"Sorry bug on my part 😅 Closing "
] | ## Describe the bug
Loading the dataset [earnings-22](https://huggingface.co/datasets/sanchit-gandhi/earnings22_split) from the Hub yields an Index Error. I created this dataset locally and then pushed to hub at the specified URL. Thus, I expect the dataset should work out-of-the-box! Indeed, the dataset viewer functions correctly, and there were no issues when I had the dataset locally.
Don't think it's a sound file bug as the version matches what worked previously.
Update: the bug appeared for me on a GPU, mysteriously on a TPU I can't repro and it downloads correctly...
## Steps to reproduce the bug
```python
from datasets import load_dataset
earnings22 = load_dataset("sanchit-gandhi/earnings22_split")
```
## Expected results
```
>>> earnings22
DatasetDict({
validation: Dataset({
features: ['source_id', 'audio', 'segment_id', 'sentence', 'start_ts', 'end_ts', 'id'],
num_rows: 2650
})
train: Dataset({
features: ['source_id', 'audio', 'segment_id', 'sentence', 'start_ts', 'end_ts', 'id'],
num_rows: 52006
})
test: Dataset({
features: ['source_id', 'audio', 'segment_id', 'sentence', 'start_ts', 'end_ts', 'id'],
num_rows: 2735
})
})
```
## Actual results
```
Traceback (most recent call last):
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/arrow_dataset.py", line 2764, in _map_single
writer.write(example)
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/arrow_writer.py", line 451, in write
self.write_examples_on_file()
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/arrow_writer.py", line 409, in write_examples_on_file
self.write_batch(batch_examples=batch_examples)
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/arrow_writer.py", line 508, in write_batch
arrays.append(pa.array(typed_sequence))
File "pyarrow/array.pxi", line 231, in pyarrow.lib.array
File "pyarrow/array.pxi", line 110, in pyarrow.lib._handle_arrow_array_protocol
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/arrow_writer.py", line 197, in __arrow_array__
out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/table.py", line 1683, in wrapper
return func(array, *args, **kwargs)
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/table.py", line 1795, in cast_array_to_feature
return feature.cast_storage(array)
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/features/audio.py", line 190, in cast_storage
storage = pa.array([Audio().encode_example(x) if x is not None else None for x in storage.to_pylist()])
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/features/audio.py", line 190, in <listcomp>
storage = pa.array([Audio().encode_example(x) if x is not None else None for x in storage.to_pylist()])
File "/opt/conda/envs/hf/lib/python3.8/site-packages/datasets/features/audio.py", line 92, in encode_example
sf.write(buffer, value["array"], value["sampling_rate"], format="wav")
File "/opt/conda/envs/hf/lib/python3.8/site-packages/soundfile.py", line 313, in write
channels = data.shape[1]
IndexError: tuple index out of range
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-4.19.0-21-cloud-amd64-x86_64-with-glibc2.10
- Python version: 3.8.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.3
Plus:
- SoundFile version: 0.10.3.post1
cc @lhoestq @polinaeterna | 5,011 |
https://github.com/huggingface/datasets/issues/5009 | Error loading StonyBrookNLP/tellmewhy dataset from hub even though local copy loads correctly | [
"I think this is because some columns are mostly empty lists. In particular the train and validation splits only have empty lists for `val_ann`. Therefore the type inference doesn't know which type is inside (or it would have to scan the other splits first before knowing).\r\n\r\nYou can fix that by specifying the features types explicitly.\r\nThen you can save the feature types inside the dataset repository, so that you won't need to specify the features in subsequent calls:\r\n```python\r\nfrom datasets import load_dataset, Features, Sequence, Value\r\nfrom datasets.info import DatasetInfosDict\r\n\r\nfeatures = Features({\r\n 'narrative': Value('string'),\r\n 'question': Value('string'),\r\n 'original_sentence_for_question': Value('string'),\r\n 'narrative_lexical_overlap': Value('float64'),\r\n 'is_ques_answerable': Value('string'),\r\n 'answer': Value('string'),\r\n 'is_ques_answerable_annotator': Value('string'),\r\n 'original_narrative_form': Sequence(Value('string')),\r\n 'question_meta': Value('string'),\r\n 'helpful_sentences': Sequence(Value('int64')),\r\n 'human_eval': Value('bool'),\r\n 'val_ann': Sequence(Value('int64')),\r\n 'gram_ann': Sequence(Value('int64'))\r\n})\r\nds = load_dataset('StonyBrookNLP/tellmewhy', features=features)\r\nDatasetInfosDict({\"default\": ds[\"train\"].info}).write_to_directory(\"path/to/local/tellmewhy\")\r\n```\r\nand then after pushing the change to the dataset repository on the Hub, `load_dataset(\"StonyBrookNLP/tellmewhy\")` will work directly`",
"(Note that specifying explicit types will be made easier with https://github.com/huggingface/datasets/pull/4926)",
"`gram_ann` and `val_ann` are annotations that only exist for part of the test set. I wanted to keep all the columns consistent across all files, so I added them to train and validation as well. I'll check if removing them from those files is still compliant with this repo. Otherwise, I will do as you suggested. Thanks @lhoestq !",
"@lhoestq I followed the exact steps you described but it seems like I'm getting the same error unfortunately. Any other ideas? Thanks in advance",
"Hi ! If you move `dataset_infos.json` from `data/` to the root of your dataset repository if should work :)",
"I tried that and pushed to the [hub](https://huggingface.co/datasets/StonyBrookNLP/tellmewhy/tree/main). Now, there is a new error.\r\n```\r\n File \"/home/yklal95/tellmewhy/src/prepare_data.py\", line 67, in main\r\n dataset = load_dataset('StonyBrookNLP/tellmewhy')\r\n File \"/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/load.py\", line 1746, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/builder.py\", line 704, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/builder.py\", line 775, in _download_and_prepare\r\n verify_checksums(\r\n File \"/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/utils/info_utils.py\", line 33, in verify_checksums\r\n raise ExpectedMoreDownloadedFiles(str(set(expected_checksums) - set(recorded_checksums)))\r\ndatasets.utils.info_utils.ExpectedMoreDownloadedFiles: {'/home/yklal95/tellmewhy/data/test.json', '/home/yklal95/tellmewhy/data/validation.json', '/home/yklal95/tellmewhy/data/train.json'}\r\n```\r\nNo changes were made to any of the other files and they are still on the hub. Let me know if you have any ideas @lhoestq Thanks!",
"Oh I see - the code I gave you returns local paths instead of URLs to store metadata about files to download.\r\nI opened a PR in your repo here to remove this: https://huggingface.co/datasets/StonyBrookNLP/tellmewhy/discussions/1\r\nsorry for the inconvenience !",
"It works now! Thanks a lot @lhoestq "
] | ## Describe the bug
I have added a new dataset with the identifier `StonyBrookNLP/tellmewhy` to the hub. When I load the individual files using my local copy using `dataset = datasets.load_dataset("json", data_files="data/train.jsonl")`, it loads the dataset correctly. However, when I try to load it from the hub, I get an error (pasted below). Additionally, `dataset = datasets.load_dataset("json", data_dir="data/")` throws the same error.
## Steps to reproduce the bug
```python
dataset = datasets.load_dataset('StonyBrookNLP/tellmewhy')
```
## Expected results
Successfully load the `StonyBrookNLP/tellmewhy` dataset.
## Actual results
```
Using custom data configuration StonyBrookNLP--tellmewhy-82712924092694ff
Downloading and preparing dataset json/StonyBrookNLP--tellmewhy to /home/yklal95/.cache/huggingface/datasets/StonyBrookNLP___json/StonyBrookNLP--tellmewhy-82712924092694ff/0.0.0/a3e658c4731e59120d44081ac10bf85dc7e1388126b92338344ce9661907f253...
Downloading data files: 100%|██████████████████████████████| 3/3 [00:00<00:00, 957.46it/s]
Extracting data files: 100%|███████████████████████████████| 3/3 [00:00<00:00, 299.14it/s]
Traceback (most recent call last):
File "/home/yklal95/tmw-generalization/src/load_datasets.py", line 17, in <module>
main(args)
File "/home/yklal95/tmw-generalization/src/load_datasets.py", line 11, in main
dataset = datasets.load_dataset(args.dataset_name)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/load.py", line 1746, in load_dataset
builder_instance.download_and_prepare(
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/builder.py", line 793, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/builder.py", line 1277, in _prepare_split
writer.write_table(table)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/arrow_writer.py", line 524, in write_table
pa_table = table_cast(pa_table, self._schema)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 2005, in table_cast
return cast_table_to_schema(table, schema)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1969, in cast_table_to_schema
arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1969, in <listcomp>
arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1681, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1681, in <listcomp>
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1822, in cast_array_to_feature
casted_values = _c(array.values, feature.feature)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1683, in wrapper
return func(array, *args, **kwargs)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1853, in cast_array_to_feature
return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1683, in wrapper
return func(array, *args, **kwargs)
File "/home/yklal95/anaconda3/envs/tmw-generalization/lib/python3.9/site-packages/datasets/table.py", line 1761, in array_cast
raise TypeError(f"Couldn't cast array of type {array.type} to {pa_type}")
TypeError: Couldn't cast array of type int64 to null
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-4.15.0-121-generic-x86_64-with-glibc2.27
- Python version: 3.9.13
- PyArrow version: 9.0.0
- Pandas version: 1.5.0
| 5,009 |
https://github.com/huggingface/datasets/issues/5005 | Release 2.5.0 breaks transformers CI | [
"Shall we revert https://github.com/huggingface/datasets/pull/4971 @mariosasko ?\r\n\r\nAnd for consistency we can update IterableDataset.map later"
] | ## Describe the bug
As reported by @lhoestq:
> see https://app.circleci.com/pipelines/github/huggingface/transformers/47634/workflows/b491886b-e66e-4edb-af96-8b459e72aa25/jobs/564563
this is used here: [https://github.com/huggingface/transformers/blob/3b19c0317b6909e2d7f11b5053895ac55[…]torch/speech-pretraining/run_wav2vec2_pretraining_no_trainer.py](https://github.com/huggingface/transformers/blob/3b19c0317b6909e2d7f11b5053895ac55250e7da/examples/pytorch/speech-pretraining/run_wav2vec2_pretraining_no_trainer.py#L482-L488)
| 5,005 |
https://github.com/huggingface/datasets/issues/5002 | Dataset Viewer issue for loubnabnl/humaneval-x | [
"It's a bug! Thanks for reporting, I'm looking at it",
"Fixed."
] | ### Link
https://huggingface.co/datasets/loubnabnl/humaneval-x/viewer/
### Description
The dataset has subsets but the viewer gets stuck in the default subset even when I select another one (the data loading of the subsets works fine)
### Owner
Yes | 5,002 |
https://github.com/huggingface/datasets/issues/5000 | Dataset Viewer issue for asapp/slue | [
"<img width=\"519\" alt=\"Capture d’écran 2022-09-20 à 22 33 47\" src=\"https://user-images.githubusercontent.com/1676121/191358952-1220cb7d-745a-4203-a66b-3c707b25038f.png\">\r\n\r\n```\r\nNot found.\r\n\r\nError code: SplitsResponseNotFound\r\n```\r\n\r\nhttps://datasets-server.huggingface.co/splits?dataset=asapp/slue\r\n\r\n```json\r\n{\"error\":\"Not found.\"}\r\n```",
"I just launched a refresh. It's weird, I don't see any entry for this dataset in the cache, it's a bug on our side. In order to try to understand what happened, did you change the visibility status from private to public, by any chance?",
"The dataset is being refreshed, please retry later.\r\n\r\n<img width=\"802\" alt=\"Capture d’écran 2022-09-20 à 22 39 46\" src=\"https://user-images.githubusercontent.com/1676121/191360072-7cc86486-4e84-4b47-8f9a-4a69fe84a5ac.png\">\r\n",
"OK. We now have an issue because the dataset cannot be streamed, and the dataset viewer relies on it.\r\n\r\nMaybe @huggingface/datasets can help:\r\n\r\n```\r\nError code: StreamingRowsError\r\nException: NotImplementedError\r\nMessage: Extraction protocol for TAR archives like 'https://public-dataset-model-store.awsdev.asapp.com/users/sshon/public/slue/slue-voxpopuli_v0.2_blind.tar.gz' is not implemented in streaming mode. Please use `dl_manager.iter_archive` instead.\r\nTraceback: Traceback (most recent call last):\r\n File \"/src/services/worker/src/worker/responses/first_rows.py\", line 337, in get_first_rows_response\r\n rows = get_rows(dataset, config, split, streaming=True, rows_max_number=rows_max_number, hf_token=hf_token)\r\n File \"/src/services/worker/src/worker/utils.py\", line 123, in decorator\r\n return func(*args, **kwargs)\r\n File \"/src/services/worker/src/worker/responses/first_rows.py\", line 65, in get_rows\r\n ds = load_dataset(\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/load.py\", line 1739, in load_dataset\r\n return builder_instance.as_streaming_dataset(split=split)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1025, in as_streaming_dataset\r\n splits_generators = {sg.name: sg for sg in self._split_generators(dl_manager)}\r\n File \"/tmp/modules-cache/datasets_modules/datasets/asapp--slue/adaa0c78233e1a1df9c2f054e690ec5fc3eaf453bd76b80fe5cbe5728e55d9b1/slue.py\", line 189, in _split_generators\r\n dl_dir = dl_manager.download_and_extract(_DL_URLS[config_name])\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 944, in download_and_extract\r\n return self.extract(self.download(url_or_urls))\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 907, in extract\r\n urlpaths = map_nested(self._extract, path_or_paths, map_tuple=True)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 385, in map_nested\r\n return function(data_struct)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 912, in _extract\r\n protocol = _get_extraction_protocol(urlpath, use_auth_token=self.download_config.use_auth_token)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 390, in _get_extraction_protocol\r\n raise NotImplementedError(\r\n NotImplementedError: Extraction protocol for TAR archives like 'https://public-dataset-model-store.awsdev.asapp.com/users/sshon/public/slue/slue-voxpopuli_v0.2_blind.tar.gz' is not implemented in streaming mode. Please use `dl_manager.iter_archive` instead.\r\n```",
"Thanks @severo, \r\n\r\nDo I have to modify the python script to support streaming so that it can be previewed?\r\nIs there a document somewhere that I can follow?\r\n",
"Hi @fwu-asapp thanks for reporting, and thanks @severo for the investigation.\r\n\r\nAs explained by @severo, the preview requires that your dataset loading script supports streaming.\r\n\r\nThere are several options here:\r\n- the easiest would be to replace the source files, archived using ZIP instead TAR: the TAR format does not allow random access while streaming, but only sequential access; the ZIP files support streaming out of the box.\r\n- alternatively, to stream TAR archives you can use `dl_manager.iter_archive`: the only prerequisite is that your \"index\" files (.tsv) should have been archived before their corresponding audio files, so while iterating the content of the TAR archive, the metadata files appear first. I think this is the case for voxpopuli tar but not for voxceleb.\r\n- if your .tsv files were not archived before their corresponding audio files (I think this is the case for voxceleb), then you should extract the .tsv files and host them separately (you can host them on the same Hugging Face Hub).\r\n - you can take as example, e.g.: https://huggingface.co/datasets/vivos/blob/main/vivos.py\r\n\r\nAs an advanced approach, you can handle both streaming and non-streaming cases separately.\r\n- as for example: https://huggingface.co/datasets/librispeech_asr/blob/main/librispeech_asr.py or https://huggingface.co/datasets/google/fleurs/blob/main/fleurs.py\r\n\r\nSee related discussion:\r\n- https://github.com/huggingface/datasets/issues/4697#issuecomment-1191502492",
"Thanks @albertvillanova for your clarification. I'll talk to my collaborators to see if we can replace those files. Let me just close this issue for now.",
"FYI, after replacing the source files with the ZIP ones, the dataset viewer works well. Thanks again to @severo and @albertvillanova for your help!",
"Great! And thank you for sharing that interesting dataset!"
] | ### Link
https://huggingface.co/datasets/asapp/slue/viewer/
### Description
Hi,
I wonder how to get the dataset viewer of our slue dataset to work.
Best,
Felix
### Owner
Yes | 5,000 |
https://github.com/huggingface/datasets/issues/4996 | Dataset Viewer issue for Jean-Baptiste/wikiner_fr | [
"The script uses `Dataset.load_from_disk`, which as you can expect, doesn't work in streaming mode.\r\n\r\nIt would probably be more practical to load the dataset locally using `Dataset.load_from_disk` first and then `push_to_hub` to upload it in Parquet on the Hub",
"I've transferred this issue to the Hub repo: https://huggingface.co/datasets/Jean-Baptiste/wikiner_fr/discussions/3\r\n\r\nI'm closing this."
] | ### Link
https://huggingface.co/datasets/Jean-Baptiste/wikiner_fr
### Description
```
Error code: StreamingRowsError
Exception: FileNotFoundError
Message: [Errno 2] No such file or directory: 'zip:/data/train::https:/huggingface.co/datasets/Jean-Baptiste/wikiner_fr/resolve/main/data.zip/state.json'
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/responses/first_rows.py", line 337, in get_first_rows_response
rows = get_rows(dataset, config, split, streaming=True, rows_max_number=rows_max_number, hf_token=hf_token)
File "/src/services/worker/src/worker/utils.py", line 123, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/responses/first_rows.py", line 77, in get_rows
rows_plus_one = list(itertools.islice(ds, rows_max_number + 1))
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 718, in __iter__
for key, example in self._iter():
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 708, in _iter
yield from ex_iterable
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 112, in __iter__
yield from self.generate_examples_fn(**self.kwargs)
File "/tmp/modules-cache/datasets_modules/datasets/Jean-Baptiste--wikiner_fr/683a580ba6ec769d508f7dfc603a651667b0ed3817b1ae5bfd45f97cc024923f/wikiner_fr.py", line 165, in _generate_examples
dataset = Dataset.load_from_disk(filepath)
File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 1210, in load_from_disk
with open(Path(dataset_path, config.DATASET_STATE_JSON_FILENAME).as_posix(), encoding="utf-8") as state_file:
FileNotFoundError: [Errno 2] No such file or directory: 'zip:/data/train::https:/huggingface.co/datasets/Jean-Baptiste/wikiner_fr/resolve/main/data.zip/state.json'
```
Is it an error with the dataset script, or the data itself, @huggingface/datasets?
https://huggingface.co/datasets/Jean-Baptiste/wikiner_fr/tree/main
### Owner
No | 4,996 |
https://github.com/huggingface/datasets/issues/4995 | Get a specific Exception when the dataset has no data | [] | In the dataset viewer on the Hub (https://huggingface.co/datasets/glue/viewer), we would like (https://github.com/huggingface/moon-landing/issues/3882) to show a specific message when the repository lacks any data files.
In that case, instead of showing a complex traceback, we want to show a call to action to help the user upload data.
To do that, it would be very helpful to know for sure that the repository is missing any (supported) data files.
It could be done by raising a custom exception, for example, `NoDataError`. | 4,995 |
https://github.com/huggingface/datasets/issues/4994 | delete the hardcoded license list in `datasets` | [] | > Feel free to delete the license list in `datasets` [...]
>
> Also FYI in #4926 I also removed all the validation steps anyway (language, license, types etc.)
_Originally posted by @lhoestq in https://github.com/huggingface/datasets/issues/4930#issuecomment-1238401662_
> [...], in my opinion we can just delete this file from `datasets`, the validation is happening hub-side anyways now?
_Originally posted by @julien-c in https://github.com/huggingface/datasets/issues/4930#issuecomment-1238390659_ | 4,994 |
https://github.com/huggingface/datasets/issues/4990 | "no-token" is passed to `huggingface_hub` when token is `None` | [
"Hi @Wauplin, thanks for raising this potential issue.\r\n\r\nThe choice of passing `\"no-token\"` instead of `None` was made in this PR:\r\n- #4536 \r\n\r\nAccording to the PR description, the reason why it is passed is to avoid that `HfApi.dataset_info` uses the local token when no token should be used.",
"Hi @albertvillanova , thanks for finding the original issue :+1: \r\n\r\nAs of next release of `huggingface_hub`, the `token` argument will be deprecated in favor of the `use_auth_token` argument in `dataset_info` method. This change as been done by @SBrandeis in https://github.com/huggingface/huggingface_hub/pull/928. `use_auth_token` is a bit different and allow the case \"don't sent the cached token by default\".\r\n\r\nIf you want to strictly avoid sending the cached token from `datasets`, you can use:\r\n```py\r\n# token=token if token else \"no-token\", <- will fail because token is not valid\r\n\r\nuse_auth_token=token if token else False, # using the new `use_auth_token` parameter\r\n```\r\n\r\nAnd as a note, I am currently updating the \"don't send the cached token by default\"-rule to \"don't send the cached token on public repos by default but use it in private ones\" in https://github.com/huggingface/huggingface_hub/pull/1064. This will not change the fact that `use_auth_token=False` doesn't send the token at all.\r\n",
"What is current strategy in term of updating `huggingface_hub` version in `datasets` ? I don't want to break stuff in the next release so let's find a proper solution :) ",
"As soon as `token` is deprecated and hfh has a new release, we'll update `datasets` to use the new argument instead. Does it sound good to you ?",
"Perfect :ok_hand: ",
"Hi @Wauplin, thanks for the warning about the deprecation of `token` in favor of `use_auth_token`.\r\n\r\nIndeed, in datasets we use internally `use_auth_token`, which in this case was transformed to `token` to call `HfApi.dataset_info`:\r\nhttps://github.com/huggingface/datasets/blob/1a9385d7cc8a3241b44015145ef56a230fdadc51/src/datasets/load.py#L747\r\n\r\nTherefore, for the new hfh release, the fix will be trivial: we will pass directly `use_auth_token`.\r\n\r\nAs discussed during our meeting yesterday, due to the fact that at datasets we support multiple hfh versions, I think we should handle passing `token` or `use_auth_token` depending on the hfh version."
] | ## Describe the bug
In the 2 lines listed below, a token is passed to `huggingface_hub` to get information from a dataset. If no token is provided, a "no-token" string is passed. What is the purpose of it ? If no real, I would prefer if the `None` value could be sent directly to be handle by `huggingface_hub`. I feel that here it is working because we assume the token will never be validated.
https://github.com/huggingface/datasets/blob/5b23f58535f14cc4dd7649485bce1ccc836e7bca/src/datasets/load.py#L753
https://github.com/huggingface/datasets/blob/5b23f58535f14cc4dd7649485bce1ccc836e7bca/src/datasets/load.py#L1121
## Expected results
Pass `token=None` to `huggingface_hub`.
## Actual results
`token="no-token"` is passed.
## Environment info
`huggingface_hub v0.10.0dev` | 4,990 |
https://github.com/huggingface/datasets/issues/4989 | Running add_column() seems to corrupt existing sequence-type column info | [
"Nevermind, I was incorrect."
] | I have a dataset that contains a column ("foo") that is a sequence type of length 4. So when I run .to_pandas() on it, the resulting dataframe correctly contains 4 columns - foo_0, foo_1, foo_2, foo_3. So the 1st row of the dataframe might look like:
ds = load_dataset(...)
df = ds.to_pandas()
df:
foo_0 | foo_1 | foo_2 | foo_3
0.0 | 1.0 | 2.0 | 3.0
If I run .add_column("new_col", data) on the dataset, and then .to_pandas() on the resulting new dataset, the resulting dataframe contains only 2 columns - foo, new_col. The values in column foo are lists of length 4, the 4 elements that should have been split into separate columns. Dataframe 1st row would be:
ds = load_dataset(...)
new_ds = ds.add_column("new_col", data)
df = new_ds.to_pandas()
df:
foo | new_col
[0.0, 1.0, 2.0, 3.0] | new_val
I've explored the 2 datasets in a debugger and haven't noticed any changes to any attributes related to the foo column, but I can't determine why the dataframes are so different. | 4,989 |
https://github.com/huggingface/datasets/issues/4988 | Add `IterableDataset.from_generator` to the API | [
"#take",
"Thanks @hamid-vakilzadeh ! Let us know if you have some questions or if we can help",
"Thank you! I certainly will reach out if I need any help."
] | We've just added `Dataset.from_generator` to the API. It would also be cool to add `IterableDataset.from_generator` to support creating an iterable dataset from a generator.
cc @lhoestq | 4,988 |
https://github.com/huggingface/datasets/issues/4983 | How to convert torch.utils.data.Dataset to huggingface dataset? | [
"Hi! I think you can use the newly-added `from_generator` method for that:\r\n```python\r\nfrom datasets import Dataset\r\n\r\ndef gen():\r\n for idx in len(torch_dataset):\r\n yield torch_dataset[idx] # this has to be a dictionary\r\n ## or if it's an IterableDataset\r\n # for ex in torch_dataset:\r\n # yield ex\r\n\r\ndset = Dataset.from_generator(gen)\r\n```",
"Maybe `Dataset.from_list` can work as well no ?\r\n```python\r\nfrom datasets import Dataset\r\n\r\ndset = Dataset.from_list(torch_dataset)\r\n```",
"> ```python\r\n> from datasets import Dataset\r\n> \r\n> def gen():\r\n> for idx in len(torch_dataset):\r\n> yield torch_dataset[idx] # this has to be a dictionary\r\n> ## or if it's an IterableDataset\r\n> # for ex in torch_dataset:\r\n> # yield ex\r\n> \r\n> dset = Dataset.from_generator(gen)\r\n> ```\r\n\r\nI try to use `Dataset.from_generator()` method, and it returns an error:\r\n```bash\r\nAttributeError: type object 'Dataset' has no attribute 'from_generator'\r\n```\r\nAnd I think it maybe the version of my datasets package is out-of-date, so I update it\r\n```bash\r\npip install --upgrade datasets\r\n```\r\nBut after that, the code still return the above Error. ",
"> ```python\r\n> dset = Dataset.from_list(torch_dataset)\r\n> ```\r\n\r\nIt seems that Dataset also has no `from_list` method 😂\r\n```bash\r\nAttributeError: type object 'Dataset' has no attribute 'from_list'\r\n```",
"> I look through the huggingface dataset docs, and it seems that there is no offical support function to convert `torch.utils.data.Dataset` to huggingface dataset. However, there is a way to convert huggingface dataset to `torch.utils.data.Dataset`, like below:\r\n> \r\n> ```python\r\n> from datasets import Dataset\r\n> data = [[1, 2],[3, 4]]\r\n> ds = Dataset.from_dict({\"data\": data})\r\n> ds = ds.with_format(\"torch\")\r\n> ds[0]\r\n> ds[:2]\r\n> ```\r\n> \r\n> So is there something I miss, or there IS no function to convert `torch.utils.data.Dataset` to huggingface dataset. If so, is there any way to do this convert? Thanks.\r\n\r\nMy dummy code is like:\r\n```python\r\nimport os\r\nimport json\r\nfrom torch.utils import data\r\nimport datasets\r\n\r\ndef gen(torch_dataset):\r\n for idx in len(torch_dataset):\r\n yield torch_dataset[idx] # this has to be a dictionary\r\n\r\nclass MyDataset(data.Dataset):\r\n def __init__(self, path):\r\n self.dict = []\r\n for line in open(path, 'r', encoding='utf-8'):\r\n j_dict = json.loads(line)\r\n self.dict.append(j_dict['context'])\r\n \r\n def __getitem__(self, idx):\r\n return self.dict[idx]\r\n\r\n def __len__(self):\r\n return len(self.dict)\r\n\r\nroot_path = os.path.dirname(os.path.abspath(__file__))\r\npath = os.path.join(root_path, 'dataset', 'train.json')\r\ntorch_dataset = MyDataset(path)\r\n\r\ndit = []\r\nfor line in open(path, 'r', encoding='utf-8'):\r\n j_dict = json.loads(line)\r\n dit.append(j_dict['context'])\r\ndset1 = datasets.Dataset.from_list(dit)\r\nprint(dset1)\r\ndset2 = datasets.Dataset.from_generator(gen)\r\nprint(dset2)\r\n```",
"We're releasing `from_generator` and `from_list` today :)\r\nIn the meantime you can play with them by installing `datasets` from source",
"> We're releasing `from_generator` and `from_list` today :) In the meantime you can play with them by installing `datasets` from source\r\n\r\nThanks a lot for your work!",
"> > I look through the huggingface dataset docs, and it seems that there is no offical support function to convert `torch.utils.data.Dataset` to huggingface dataset. However, there is a way to convert huggingface dataset to `torch.utils.data.Dataset`, like below:\r\n> > ```python\r\n> > from datasets import Dataset\r\n> > data = [[1, 2],[3, 4]]\r\n> > ds = Dataset.from_dict({\"data\": data})\r\n> > ds = ds.with_format(\"torch\")\r\n> > ds[0]\r\n> > ds[:2]\r\n> > ```\r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > \r\n> > So is there something I miss, or there IS no function to convert `torch.utils.data.Dataset` to huggingface dataset. If so, is there any way to do this convert? Thanks.\r\n> \r\n> My dummy code is like:\r\n> \r\n> ```python\r\n> import os\r\n> import json\r\n> from torch.utils import data\r\n> import datasets\r\n> \r\n> def gen(torch_dataset):\r\n> for idx in len(torch_dataset):\r\n> yield torch_dataset[idx] # this has to be a dictionary\r\n> \r\n> class MyDataset(data.Dataset):\r\n> def __init__(self, path):\r\n> self.dict = []\r\n> for line in open(path, 'r', encoding='utf-8'):\r\n> j_dict = json.loads(line)\r\n> self.dict.append(j_dict['context'])\r\n> \r\n> def __getitem__(self, idx):\r\n> return self.dict[idx]\r\n> \r\n> def __len__(self):\r\n> return len(self.dict)\r\n> \r\n> root_path = os.path.dirname(os.path.abspath(__file__))\r\n> path = os.path.join(root_path, 'dataset', 'train.json')\r\n> torch_dataset = MyDataset(path)\r\n> \r\n> dit = []\r\n> for line in open(path, 'r', encoding='utf-8'):\r\n> j_dict = json.loads(line)\r\n> dit.append(j_dict['context'])\r\n> dset1 = datasets.Dataset.from_list(dit)\r\n> print(dset1)\r\n> dset2 = datasets.Dataset.from_generator(gen)\r\n> print(dset2)\r\n> ```\r\nHi, when I am using this code to build my own dataset, ` datasets.Dataset.from_generator(gen)` report `TypeError: cannot pickle generator object` whre MyDataset returns a dict like {'image': bytes, 'text': string}. How can I resolve this? Thanks a lot!",
"Hi ! Right now generator functions are expected to be picklable, so that `datasets` can hash it and use the hash to cache the resulting Dataset on disk. Maybe this can be improved.\r\n\r\nIn the meantime, can you check that you're not using unpickable objects. In your case it looks like you're using a generator object that is unpickable. It might come from an opened file, e.g. this doesn't work:\r\n```python\r\nwith open(...) as f:\r\n\r\n def gen():\r\n for x in f:\r\n yield json.loads(x)\r\n\r\n ds = Dataset.from_generator(gen)\r\n```\r\nbut this does work:\r\n```python\r\ndef gen():\r\n with open(...) as f:\r\n for x in f:\r\n yield json.loads(x)\r\n\r\nds = Dataset.from_generator(gen)\r\n```",
"> Hi ! Right now generator functions are expected to be picklable, so that `datasets` can hash it and use the hash to cache the resulting Dataset on disk. Maybe this can be improved.\r\n> \r\n> In the meantime, can you check that you're not using unpickable objects. In your case it looks like you're using a generator object that is unpickable. It might come from an opened file, e.g. this doesn't work:\r\n> \r\n> ```python\r\n> with open(...) as f:\r\n> \r\n> def gen():\r\n> for x in f:\r\n> yield json.loads(x)\r\n> \r\n> ds = Dataset.from_generator(gen)\r\n> ```\r\n> \r\n> but this does work:\r\n> \r\n> ```python\r\n> def gen():\r\n> with open(...) as f:\r\n> for x in f:\r\n> yield json.loads(x)\r\n> \r\n> ds = Dataset.from_generator(gen)\r\n> ```\r\n\r\nThanks a lot! That's the reason why I have encountered this issue. Sorry for bothering you again with another problem, since my dataset is large and I use IterableDataset.from_generator which has no attribute with_transform, how can I equip it with some customed preprocessings like Dataset.from_generator? Should I move the preprocessing to the my torch Dataset?",
"Iterable datasets are lazy: exactly like `with_transform` they apply processing on the fly when accessing the examples.\r\n\r\nTherefore you can use `my_iterable_dataset.map()` instead :)",
"@lhoestq thanks a lot and I have successfully made it work~",
"@lhoestq I am having a similar issue. Can you help me understand which kinds of generators are picklable? I previously thought that no generators are picklable so I'm intrigued to hear this.",
"Generator functions are generally picklable. E.g.\r\n```python\r\nimport dill as pickle\r\n\r\ndef generator_fn():\r\n for i in range(10):\r\n yield i\r\n\r\npickle.dumps(generator_fn)\r\n```\r\n\r\nhowever generators are not picklable\r\n```python\r\ngenerator = generator_fn()\r\npickle.dumps(generator)\r\n# TypeError: cannot pickle 'generator' object\r\n```\r\n\r\nThough it can happen that some generator functions are not recursively picklable if they use global objects that are not picklable:\r\n```python\r\ndef generator_fn_not_picklable():\r\n for i in generator:\r\n yield i\r\n\r\npickle.dumps(generator_fn_not_picklable, recurse=True)\r\n# TypeError: cannot pickle 'generator' object\r\n````",
"I'm trying to create an IterableDataset from a generator but I get this error:\r\n`PicklingError: Can't pickle <built-in function input>: it's not the same object as builtins.input`\r\n\r\nWhat can I do?"
] | I look through the huggingface dataset docs, and it seems that there is no offical support function to convert `torch.utils.data.Dataset` to huggingface dataset. However, there is a way to convert huggingface dataset to `torch.utils.data.Dataset`, like below:
```python
from datasets import Dataset
data = [[1, 2],[3, 4]]
ds = Dataset.from_dict({"data": data})
ds = ds.with_format("torch")
ds[0]
ds[:2]
```
So is there something I miss, or there IS no function to convert `torch.utils.data.Dataset` to huggingface dataset. If so, is there any way to do this convert?
Thanks. | 4,983 |
https://github.com/huggingface/datasets/issues/4982 | Create dataset_infos.json with VALIDATION and TEST splits | [
"@mariosasko could you help me with this issue? we've started the discussion from [here](https://github.com/huggingface/datasets/issues/4895#issuecomment-1248227130)",
"Hi again! Can you please pass the directory name containing the dataset script instead of the script name to `datasets-cli test`?",
"Yes, it worked! thanks a lot"
] | The problem is described in that [issue](https://github.com/huggingface/datasets/issues/4895#issuecomment-1247975569).
> When I try to create data_infos.json using datasets-cli test Peter.py --save_infos --all_configs I get an error:
> ValueError: Unknown split "test". Should be one of ['train'].
>
> The data_infos.json is created perfectly fine when I use only one split - datasets.Split.TRAIN
>
> You can find the code here: https://huggingface.co/datasets/sberbank-ai/Peter/tree/add_splits (add_splits branch)
I tried to clear the cache folder, than I got an another error. I run:
```
git clone https://huggingface.co/datasets/sberbank-ai/Peter
cd Peter
git checkout add_splits # switch to a add_splits branch
rm dataset_infos.json # remove local dataset_infos.json
rm -r ~/.cache/huggingface # remove cached dataset_infos.json
datasets-cli test Peter.py --save_infos --all_configs # trying to create new dataset_infos.json
```
The error message:
```
Using custom data configuration default
Testing builder 'default' (1/1)
Downloading and preparing dataset peter/default to /Users/kalinin/.cache/huggingface/datasets/peter/default/0.0.0/ef579519e140d6a40df2555996f26165f04c47557d7373709c8d7e7b4fd7465d...
Downloading data files: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 5160.63it/s]
Extracting data files: 0%| | 0/4 [00:00<?, ?it/s]Traceback (most recent call last):
File "/usr/local/bin/datasets-cli", line 8, in <module>
sys.exit(main())
File "/usr/local/lib/python3.9/site-packages/datasets/commands/datasets_cli.py", line 39, in main
service.run()
File "/usr/local/lib/python3.9/site-packages/datasets/commands/test.py", line 137, in run
builder.download_and_prepare(
File "/usr/local/lib/python3.9/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/usr/local/lib/python3.9/site-packages/datasets/builder.py", line 1227, in _download_and_prepare
super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File "/usr/local/lib/python3.9/site-packages/datasets/builder.py", line 771, in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
File "/Users/kalinin/.cache/huggingface/modules/datasets_modules/datasets/Peter/ef579519e140d6a40df2555996f26165f04c47557d7373709c8d7e7b4fd7465d/Peter.py", line 23, in _split_generators
data_files = dl_manager.download_and_extract(_URLS)
File "/usr/local/lib/python3.9/site-packages/datasets/download/download_manager.py", line 431, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/usr/local/lib/python3.9/site-packages/datasets/download/download_manager.py", line 403, in extract
extracted_paths = map_nested(
File "/usr/local/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 393, in map_nested
mapped = [
File "/usr/local/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 394, in <listcomp>
_single_map_nested((function, obj, types, None, True, None))
File "/usr/local/lib/python3.9/site-packages/datasets/utils/py_utils.py", line 330, in _single_map_nested
return function(data_struct)
File "/usr/local/lib/python3.9/site-packages/datasets/utils/file_utils.py", line 213, in cached_path
output_path = ExtractManager(cache_dir=download_config.cache_dir).extract(
File "/usr/local/lib/python3.9/site-packages/datasets/utils/extract.py", line 46, in extract
self.extractor.extract(input_path, output_path, extractor_format)
File "/usr/local/lib/python3.9/site-packages/datasets/utils/extract.py", line 263, in extract
with FileLock(lock_path):
File "/usr/local/lib/python3.9/site-packages/datasets/utils/filelock.py", line 399, in __init__
max_filename_length = os.statvfs(os.path.dirname(lock_file)).f_namemax
FileNotFoundError: [Errno 2] No such file or directory: ''
Exception ignored in: <function BaseFileLock.__del__ at 0x11caeec10>
Traceback (most recent call last):
File "/usr/local/lib/python3.9/site-packages/datasets/utils/filelock.py", line 328, in __del__
self.release(force=True)
File "/usr/local/lib/python3.9/site-packages/datasets/utils/filelock.py", line 303, in release
with self._thread_lock:
AttributeError: 'UnixFileLock' object has no attribute '_thread_lock'
Extracting data files: 0%| | 0/4 [00:00<?, ?it/s]
```
Can you help me please?
## Environment info
- `datasets` version: 2.4.0
- Platform: macOS-12.5.1-x86_64-i386-64bit
- Python version: 3.9.5
- PyArrow version: 9.0.0
- Pandas version: 1.2.4
| 4,982 |
https://github.com/huggingface/datasets/issues/4981 | Can't create a dataset with `float16` features | [
"Hi @dconathan, thanks for reporting.\r\n\r\nWe rely on Arrow as a backend, and as far as I know currently support for `float16` in Arrow is not fully implemented in Python (C++), hence the `ArrowNotImplementedError` you get.\r\n\r\nSee, e.g.: https://arrow.apache.org/docs/status.html?highlight=float16#data-types",
"Thanks for the link…. didn’t realize arrow didn’t support it yet. Should it be removed from https://huggingface.co/docs/datasets/v2.4.0/en/package_reference/main_classes#datasets.Value until Arrow supports it?",
"Yes, you are right: maybe we should either remove it from our docs or add a comment explaining the issue.\r\n\r\nThe thing is that in Arrow it is partially supported: you can create `float16` values, but you can't cast them from/to other types. And current implementation of `Value` always tries to perform a cast from `float64` to `float16`.",
"Maybe we can just add a note in the `Value` documentation ?",
"Would you accept a PR to fix this? @lhoestq Do you have an idea of how hard it would be to fix?",
"I think the issue comes mostly from pyarrow not supporting `float16` completely.\r\n\r\nFor example you stil can't cast from/to `float16`\r\n```python\r\nimport numpy as np\r\nimport pyarrow as pa\r\n\r\npa.array(range(5)).cast(pa.float16())\r\n# ArrowNotImplementedError: Unsupported cast from int64 to halffloat using function cast_half_float\r\npa.array(range(5), pa.float32()).cast(pa.float16())\r\n# ArrowNotImplementedError: Unsupported cast from float to halffloat using function cast_half_float\r\npa.array(range(5), pa.float16())\r\n# ArrowTypeError: Expected np.float16 instance\r\npa.array(np.arange(5, dtype=np.float16())).cast(pa.float32())\r\n# ArrowNotImplementedError: Unsupported cast from halffloat to float using function cast_float\r\n```",
"Hmm it seems like we can either:\r\n1. try to fix pyarrow upstream\r\n2. half-support float16 with some workaround to make sure we don't ever do casting internally\r\n"
] | ## Describe the bug
I can't create a dataset with `float16` features.
I understand from the traceback that this is a `pyarrow` error, but I don't see anywhere in the `datasets` documentation about how to successfully do this. Is it actually supported? I've tried older versions of `pyarrow` as well with the same exact error.
The bug seems to arise from `datasets` casting the values to `double` and then `pyarrow` doesn't know how to convert those back to `float16`... does that sound right? Is there a way to bypass this since it's not necessary in the `numpy` and `torch` cases?
Thanks!
## Steps to reproduce the bug
All of the following raise the following error with the same exact (as far as I can tell) traceback:
```python
ArrowNotImplementedError: Unsupported cast from double to halffloat using function cast_half_float
```
```python
from datasets import Dataset, Features, Value
Dataset.from_dict({"x": [0.0, 1.0, 2.0]}, features=Features(x=Value("float16")))
import numpy as np
Dataset.from_dict({"x": np.arange(3, dtype=np.float16)}, features=Features(x=Value("float16")))
import torch
Dataset.from_dict({"x": torch.arange(3).to(torch.float16)}, features=Features(x=Value("float16")))
```
## Expected results
A dataset with `float16` features is successfully created.
## Actual results
```python
---------------------------------------------------------------------------
ArrowNotImplementedError Traceback (most recent call last)
Cell In [14], line 1
----> 1 Dataset.from_dict({"x": [1.0, 2.0, 3.0]}, features=Features(x=Value("float16")))
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/arrow_dataset.py:870, in Dataset.from_dict(cls, mapping, features, info, split)
865 mapping = features.encode_batch(mapping)
866 mapping = {
867 col: OptimizedTypedSequence(data, type=features[col] if features is not None else None, col=col)
868 for col, data in mapping.items()
869 }
--> 870 pa_table = InMemoryTable.from_pydict(mapping=mapping)
871 if info.features is None:
872 info.features = Features({col: ts.get_inferred_type() for col, ts in mapping.items()})
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/table.py:750, in InMemoryTable.from_pydict(cls, *args, **kwargs)
734 @classmethod
735 def from_pydict(cls, *args, **kwargs):
736 """
737 Construct a Table from Arrow arrays or columns
738
(...)
748 :class:`datasets.table.Table`:
749 """
--> 750 return cls(pa.Table.from_pydict(*args, **kwargs))
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/table.pxi:3648, in pyarrow.lib.Table.from_pydict()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/table.pxi:5174, in pyarrow.lib._from_pydict()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/array.pxi:343, in pyarrow.lib.asarray()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/array.pxi:231, in pyarrow.lib.array()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/array.pxi:110, in pyarrow.lib._handle_arrow_array_protocol()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/arrow_writer.py:197, in TypedSequence.__arrow_array__(self, type)
192 # otherwise we can finally use the user's type
193 elif type is not None:
194 # We use cast_array_to_feature to support casting to custom types like Audio and Image
195 # Also, when trying type "string", we don't want to convert integers or floats to "string".
196 # We only do it if trying_type is False - since this is what the user asks for.
--> 197 out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
198 return out
199 except (TypeError, pa.lib.ArrowInvalid) as e: # handle type errors and overflows
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/table.py:1683, in _wrap_for_chunked_arrays.<locals>.wrapper(array, *args, **kwargs)
1681 return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
1682 else:
-> 1683 return func(array, *args, **kwargs)
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/table.py:1853, in cast_array_to_feature(array, feature, allow_number_to_str)
1851 return array_cast(array, get_nested_type(feature), allow_number_to_str=allow_number_to_str)
1852 elif not isinstance(feature, (Sequence, dict, list, tuple)):
-> 1853 return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
1854 raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{feature}")
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/table.py:1683, in _wrap_for_chunked_arrays.<locals>.wrapper(array, *args, **kwargs)
1681 return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
1682 else:
-> 1683 return func(array, *args, **kwargs)
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/datasets/table.py:1762, in array_cast(array, pa_type, allow_number_to_str)
1760 if pa.types.is_null(pa_type) and not pa.types.is_null(array.type):
1761 raise TypeError(f"Couldn't cast array of type {array.type} to {pa_type}")
-> 1762 return array.cast(pa_type)
1763 raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{pa_type}")
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/array.pxi:919, in pyarrow.lib.Array.cast()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/compute.py:389, in cast(arr, target_type, safe, options)
387 else:
388 options = CastOptions.safe(target_type)
--> 389 return call_function("cast", [arr], options)
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/_compute.pyx:560, in pyarrow._compute.call_function()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/_compute.pyx:355, in pyarrow._compute.Function.call()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/error.pxi:144, in pyarrow.lib.pyarrow_internal_check_status()
File ~/scratch/scratch-env-39/.venv/lib/python3.9/site-packages/pyarrow/error.pxi:121, in pyarrow.lib.check_status()
ArrowNotImplementedError: Unsupported cast from double to halffloat using function cast_half_float
```
## Environment info
- `datasets` version: 2.4.0
- Platform: macOS-12.5.1-arm64-arm-64bit
- Python version: 3.9.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.4
| 4,981 |
https://github.com/huggingface/datasets/issues/4980 | Make `pyarrow` optional | [
"The whole datasets library is pretty much a wrapper to pyarrow (just take a look at some of the source for a Dataset) https://github.com/huggingface/datasets/blob/51aef08ad7053c0bfe8f9a961207b26df15850d3/src/datasets/arrow_dataset.py#L639 \r\n\r\nI think removing the pyarrow dependency would involve a complete rewrite / a different library with minimal functionality (datasets-lite ?)",
"Thanks for the proposal, @KOLANICH. And also thanks for your answer, @dconathan.\r\n\r\nIndeed, we are using `pyarrow` as the backend for our datasets, in order to cache them and also allow memory-mapping (using datasets larger than your RAM memory).\r\n\r\nOne way to avoid using `pyarrow` could be loading the datasets in streaming mode, by passing `streaming=True` to `load_dataset`. This way you basically get a generator for the dataset; nothing is downloaded, nor cached. ",
"Thanks for the info. Could `datasets` then be made optional for `transformers` instead? I used `transformers` only to deal with pretrained models to deploy them (convert to ONNX, and then I use TVM), so I don't really need `pyarrow` and `datasets` by now.\r\n"
] | **Is your feature request related to a problem? Please describe.**
Is `pyarrow` really needed for every dataset?
**Describe the solution you'd like**
It is made optional.
**Describe alternatives you've considered**
Likely, no.
| 4,980 |
https://github.com/huggingface/datasets/issues/4977 | Providing dataset size | [
"Hi @sashavor, thanks for your suggestion.\r\n\r\nUntil now we have the CLI command \r\n```\r\ndatasets-cli test datasets/<your-dataset-folder> --save_infos --all_configs\r\n```\r\nthat generates the `dataset_infos.json` with the size of the downloaded dataset, among other information.\r\n\r\nWe are currently in the middle of removing those JSON files and putting their information directly in the header of the `README.md` (as YAML tags). Normally, the CLI command should continue working but saving its output to the dataset card instead. See:\r\n- #4926",
"Additionally, the download size can be inferred by doing HEAD requests to the files to be downloaded. And for files hosted on the hub you can even get the file sizes using the Hub API",
"Amazing @albertvillanova ! I think just having that information visible in the dataset info (without having to do any requests/additional coding) would be really useful :hugs: "
] | **Is your feature request related to a problem? Please describe.**
Especially for big datasets like [LAION](https://huggingface.co/datasets/laion/laion2B-en/), it's hard to know exactly the downloaded size (because there are many files and you don't have their exact size when downloaded).
**Describe the solution you'd like**
Auto-populating the downloaded dataset size on the dataset page would be really useful, including that of each split (when there are some).
**Describe alternatives you've considered**
People should be adding this to dataset cards, but I don't think that is systematically the case :slightly_smiling_face:
**Additional context**
Mentioned to @lhoestq
| 4,977 |
https://github.com/huggingface/datasets/issues/4976 | Hope to adapt Python3.9 as soon as possible | [
"Hi! `datasets` should work in Python 3.9. What kind of issue have you encountered?",
"There is this related issue already: https://github.com/huggingface/datasets/issues/4113\r\nAnd I guess we need a CI job for 3.9 ^^",
"Perhaps we should report this issue in the `filelock` repo?"
] | **Is your feature request related to a problem? Please describe.**
A clear and concise description of what the problem is.
**Describe the solution you'd like**
A clear and concise description of what you want to happen.
**Describe alternatives you've considered**
A clear and concise description of any alternative solutions or features you've considered.
**Additional context**
Add any other context about the feature request here.
| 4,976 |
https://github.com/huggingface/datasets/issues/4965 | [Apple M1] MemoryError: Cannot allocate write+execute memory for ffi.callback() | [
"Hi! This seems like a bug in `soundfile`. Could you please open an issue in their repo? `soundfile` works without any issues on my M1, so I'm not sure we can help.",
"Hi @mariosasko, can you share how you installed `soundfile` on your mac M1?",
"Hi @hoangtnm - I upgraded to python 3.10 and it fixed the problem for me. I was also running 3.8 on an M1 mac.",
"Same here, upgrade python didn't work for me \r\n\r\nMemoryError: Cannot allocate write+execute memory for ffi.callback()\r\n\r\nany idea?",
"This is a `soundfile` issue, so there isn't much we can do about it. Hopefully, it gets fixed soon.",
"> Hi @hoangtnm - I upgraded to python 3.10 and it fixed the problem for me. I was also running 3.8 on an M1 mac.\r\n\r\nit work for me too \r\n"
] | ## Describe the bug
I'm trying to run `cast_column("audio", Audio())` on Apple M1 Pro, but it seems that it doesn't work.
## Steps to reproduce the bug
```python
import datasets
dataset = load_dataset("csv", data_files="./train.csv")["train"]
dataset = dataset.map(lambda x: {"audio": str(DATA_DIR / "audio" / x["audio"])})
dataset = dataset.cast_column("audio", Audio())
dataset[0]
```
## Expected results
```
{'audio': {'bytes': None,
'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav'},
'english_transcription': 'I would like to set up a joint account with my partner',
'intent_class': 11,
'lang_id': 4,
'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
'transcription': 'I would like to set up a joint account with my partner'}
```
## Actual results
````---------------------------------------------------------------------------
MemoryError Traceback (most recent call last)
Input In [6], in <cell line: 1>()
----> 1 dataset[0]
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/arrow_dataset.py:2165, in Dataset.__getitem__(self, key)
2163 def __getitem__(self, key): # noqa: F811
2164 """Can be used to index columns (by string names) or rows (by integer index or iterable of indices or bools)."""
-> 2165 return self._getitem(
2166 key,
2167 )
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/arrow_dataset.py:2150, in Dataset._getitem(self, key, decoded, **kwargs)
2148 formatter = get_formatter(format_type, features=self.features, decoded=decoded, **format_kwargs)
2149 pa_subtable = query_table(self._data, key, indices=self._indices if self._indices is not None else None)
-> 2150 formatted_output = format_table(
2151 pa_subtable, key, formatter=formatter, format_columns=format_columns, output_all_columns=output_all_columns
2152 )
2153 return formatted_output
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/formatting/formatting.py:532, in format_table(table, key, formatter, format_columns, output_all_columns)
530 python_formatter = PythonFormatter(features=None)
531 if format_columns is None:
--> 532 return formatter(pa_table, query_type=query_type)
533 elif query_type == "column":
534 if key in format_columns:
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/formatting/formatting.py:281, in Formatter.__call__(self, pa_table, query_type)
279 def __call__(self, pa_table: pa.Table, query_type: str) -> Union[RowFormat, ColumnFormat, BatchFormat]:
280 if query_type == "row":
--> 281 return self.format_row(pa_table)
282 elif query_type == "column":
283 return self.format_column(pa_table)
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/formatting/formatting.py:312, in PythonFormatter.format_row(self, pa_table)
310 row = self.python_arrow_extractor().extract_row(pa_table)
311 if self.decoded:
--> 312 row = self.python_features_decoder.decode_row(row)
313 return row
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/formatting/formatting.py:221, in PythonFeaturesDecoder.decode_row(self, row)
220 def decode_row(self, row: dict) -> dict:
--> 221 return self.features.decode_example(row) if self.features else row
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/features/features.py:1647, in Features.decode_example(self, example, token_per_repo_id)
1634 def decode_example(self, example: dict, token_per_repo_id: Optional[Dict[str, Union[str, bool, None]]] = None):
1635 """Decode example with custom feature decoding.
1636
1637 Args:
(...)
1644 :obj:`dict[str, Any]`
1645 """
-> 1647 return {
1648 column_name: decode_nested_example(feature, value, token_per_repo_id=token_per_repo_id)
1649 if self._column_requires_decoding[column_name]
1650 else value
1651 for column_name, (feature, value) in zip_dict(
1652 {key: value for key, value in self.items() if key in example}, example
1653 )
1654 }
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/features/features.py:1648, in <dictcomp>(.0)
1634 def decode_example(self, example: dict, token_per_repo_id: Optional[Dict[str, Union[str, bool, None]]] = None):
1635 """Decode example with custom feature decoding.
1636
1637 Args:
(...)
1644 :obj:`dict[str, Any]`
1645 """
1647 return {
-> 1648 column_name: decode_nested_example(feature, value, token_per_repo_id=token_per_repo_id)
1649 if self._column_requires_decoding[column_name]
1650 else value
1651 for column_name, (feature, value) in zip_dict(
1652 {key: value for key, value in self.items() if key in example}, example
1653 )
1654 }
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/features/features.py:1260, in decode_nested_example(schema, obj, token_per_repo_id)
1257 # Object with special decoding:
1258 elif isinstance(schema, (Audio, Image)):
1259 # we pass the token to read and decode files from private repositories in streaming mode
-> 1260 return schema.decode_example(obj, token_per_repo_id=token_per_repo_id) if obj is not None else None
1261 return obj
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/features/audio.py:156, in Audio.decode_example(self, value, token_per_repo_id)
154 array, sampling_rate = self._decode_non_mp3_file_like(file)
155 else:
--> 156 array, sampling_rate = self._decode_non_mp3_path_like(path, token_per_repo_id=token_per_repo_id)
157 return {"path": path, "array": array, "sampling_rate": sampling_rate}
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/datasets/features/audio.py:257, in Audio._decode_non_mp3_path_like(self, path, format, token_per_repo_id)
254 use_auth_token = None
256 with xopen(path, "rb", use_auth_token=use_auth_token) as f:
--> 257 array, sampling_rate = librosa.load(f, sr=self.sampling_rate, mono=self.mono)
258 return array, sampling_rate
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/librosa/util/decorators.py:88, in deprecate_positional_args.<locals>._inner_deprecate_positional_args.<locals>.inner_f(*args, **kwargs)
86 extra_args = len(args) - len(all_args)
87 if extra_args <= 0:
---> 88 return f(*args, **kwargs)
90 # extra_args > 0
91 args_msg = [
92 "{}={}".format(name, arg)
93 for name, arg in zip(kwonly_args[:extra_args], args[-extra_args:])
94 ]
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/librosa/core/audio.py:164, in load(path, sr, mono, offset, duration, dtype, res_type)
161 else:
162 # Otherwise try soundfile first, and then fall back if necessary
163 try:
--> 164 y, sr_native = __soundfile_load(path, offset, duration, dtype)
166 except RuntimeError as exc:
167 # If soundfile failed, try audioread instead
168 if isinstance(path, (str, pathlib.PurePath)):
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/librosa/core/audio.py:195, in __soundfile_load(path, offset, duration, dtype)
192 context = path
193 else:
194 # Otherwise, create the soundfile object
--> 195 context = sf.SoundFile(path)
197 with context as sf_desc:
198 sr_native = sf_desc.samplerate
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/soundfile.py:629, in SoundFile.__init__(self, file, mode, samplerate, channels, subtype, endian, format, closefd)
626 self._mode = mode
627 self._info = _create_info_struct(file, mode, samplerate, channels,
628 format, subtype, endian)
--> 629 self._file = self._open(file, mode_int, closefd)
630 if set(mode).issuperset('r+') and self.seekable():
631 # Move write position to 0 (like in Python file objects)
632 self.seek(0)
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/soundfile.py:1179, in SoundFile._open(self, file, mode_int, closefd)
1177 file_ptr = _snd.sf_open_fd(file, mode_int, self._info, closefd)
1178 elif _has_virtual_io_attrs(file, mode_int):
-> 1179 file_ptr = _snd.sf_open_virtual(self._init_virtual_io(file),
1180 mode_int, self._info, _ffi.NULL)
1181 else:
1182 raise TypeError("Invalid file: {0!r}".format(self.name))
File ~/miniconda3/envs/rodan/lib/python3.8/site-packages/soundfile.py:1197, in SoundFile._init_virtual_io(self, file)
1194 def _init_virtual_io(self, file):
1195 """Initialize callback functions for sf_open_virtual()."""
1196 @_ffi.callback("sf_vio_get_filelen")
-> 1197 def vio_get_filelen(user_data):
1198 curr = file.tell()
1199 file.seek(0, SEEK_END)
MemoryError: Cannot allocate write+execute memory for ffi.callback(). You might be running on a system that prevents this. For more information, see https://cffi.readthedocs.io/en/latest/using.html#callbacks
```
## Environment info
- `datasets` version: 2.4.0
- Platform: macOS-12.5.1-arm64-arm-64bit
- Python version: 3.8.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.4 | 4,965 |
https://github.com/huggingface/datasets/issues/4964 | Column of arrays (2D+) are using unreasonably high memory | [
"note i have tried the same code with `datasets` version 2.4.0, the outcome is the very same as described above.",
"Seems related to issues #4623 and #4802 so it would appear this issue has been around for a few months.",
"Hi ! `Dataset.from_dict` keeps the data in memory. You can write on disk and reload them with\r\n```python\r\ndataset.save_to_disk(\"path/to/local\")\r\ndataset = load_from_disk(\"path/to/local\")\r\n```\r\nthis way you'll end up with a dataset loaded from your disk using memory mapping, and it won't fill up your RAM :)\r\n\r\nrelated to https://github.com/huggingface/datasets/issues/4861",
"@lhoestq thnx for getting back to me! i've tested the suggested method, but unfortunately the memory consumption is the very same:\r\n\r\n```\r\nfrom datasets import Dataset, Features, Array2D, Array3D, load_from_disk\r\nimport numpy as np\r\n\r\ncolumn_name = \"a\"\r\narray_shape = (64, 64, 3)\r\n\r\ndata = np.random.random((10000,) + array_shape)\r\ndataset = Dataset.from_dict({column_name: data}, features=Features({column_name: Array3D(shape=array_shape, dtype=\"float64\")}))\r\ndataset.save_to_disk(\"foo\")\r\n\r\nfoo_db = load_from_disk(\"foo\")\r\ncolum_value = foo_db[column_name]\r\n```\r\n\r\nthe very same happens when you create the dataset, but dont specify the feature type.\r\n\r\ni've tried running this on different envs (macOS, linux) and it's behaving the very same way.",
"When you call `colum_value = foo_db[column_name]`, you load the full column in memory.\r\n\r\nIf you want to avoid filling up your memory, you can access chunks of data instead\r\n```python\r\nembeddings = dataset[i:i + chunk_size][\"embeddings\"]\r\n```",
"@lhoestq yeah that's intentional, i.e. i really want to load the whole column into the memory. but as said above there's an unreasonable amount of overhead for the memory. the np array itself is using about 1G of memory:\r\n```\r\n>>> getsizeof(data)/1024/1024\r\n937.5001525878906\r\n```\r\nthat accessing of column above is using 10x memory compared to the original numpy array.",
"The dataset must be twice as big because we use regular arrow ListArray under the hood and not FixedSizeListArray. Basically we store unnecessary offsets.\r\n\r\nAnd this should affect performance as well. When we developed this, FixedSizeListArray still had some issues but they should be resolved on the PyArrow side now",
"A doubling would be fine. My very basic understanding of PyArrow is that using ListArray is probably related to the issue though. Using a multi-dimensional array in datasets is storing everything as strange nested 1d object arrays, which I imagine is creating the massive overhead.\r\n\r\nI think it should be a PyArrow Tensor, no?",
"PyArrow tensors are not part of the Arrow format AFAIK:\r\n\r\n> There is no direct support in the arrow columnar format to store Tensors as column values.\r\n\r\nsource: https://github.com/apache/arrow/issues/4802#issuecomment-508494694",
"That's... unfortunate. I didn't realize that."
] | ## Describe the bug
When trying to store `Array2D, Array3D, etc` as column values in a dataset, accessing that column (or creating depending on how you create it, see code below) will cause more than 10 fold of memory usage.
## Steps to reproduce the bug
```python
from datasets import Dataset, Features, Array2D, Array3D
import numpy as np
column_name = "a"
array_shape = (64, 64, 3)
data = np.random.random((10000,) + array_shape)
dataset = Dataset.from_dict({column_name: data}, features=Features({column_name: Array3D(shape=array_shape, dtype="float64")}))
```
the code above will use about 10Gb of RAM while constructing the `dataset` object.
The code below will use roughly the same amount of memory (and time) when trying to actually access the data itself of that column.
```python
from datasets import Dataset
import numpy as np
column_name = "a"
array_shape = (64, 64, 3)
data = np.random.random((10000,) + array_shape)
dataset = Dataset.from_dict({column_name: data})
dataset[column_name]
```
## Expected results
Some memory overhead, but not like as it is now and certainly not an overhead of such runtime that is currently happening.
## Actual results
Enormous memory- and runtime overhead.
## Environment info
- `datasets` version: 2.3.2
- Platform: macOS-12.5.1-arm64-arm-64bit
- Python version: 3.8.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.4 | 4,964 |
https://github.com/huggingface/datasets/issues/4963 | Dataset without script does not support regular JSON data file | [
"Hi @julien-c,\r\n\r\nOut of the box, we only support JSON lines (NDJSON) data files, but your data file is a regular JSON file. The reason is we use `pyarrow.json.read_json` and this only supports line-delimited JSON. "
] | ### Link
https://huggingface.co/datasets/julien-c/label-studio-my-dogs
### Description
<img width="1115" alt="image" src="https://user-images.githubusercontent.com/326577/189422048-7e9c390f-bea7-4521-a232-43f049ccbd1f.png">
### Owner
Yes | 4,963 |
https://github.com/huggingface/datasets/issues/4961 | fsspec 2022.8.2 breaks xopen in streaming mode | [
"loading `fsspec==2022.7.1` fixes this issue, setup.py would need to be changed to prevent users from using the latest version of fsspec.",
"Opened [PR](https://github.com/huggingface/datasets/pull/4962) to address this.",
"Hi @DCNemesis, thanks for reporting.\r\n\r\nThat was a temporary issue in `fsspec` releases 2022.8.0 and 2022.8.1. But they fixed it in their patch release 2022.8.2 (and yanked both previous versions). See:\r\n- https://github.com/huggingface/transformers/pull/18846\r\n\r\nAre you sure you have version 2022.8.2 installed?\r\n```shell\r\npip install -U fsspec\r\n```\r\n",
"@albertvillanova I was using a temporary Google Colab instance, but checking it again today it seems it was loading 2022.8.1 rather than 2022.8.2. It's surprising that colab is using the version that was replaced the same day it was released. Testing with 2022.8.2 did work. It appears Colab [will be fixing it](https://github.com/googlecolab/colabtools/issues/3055) on their end too. ",
"Thanks for the additional information.\r\n\r\nOnce we know 2022.8.2 works, I'm closing this issue. Feel free to reopen it if necessary.",
"Colab just upgraded their default `fsspec` version to 2022.8.2:\r\n- https://github.com/googlecolab/colabtools/issues/3055#issuecomment-1244019010"
] | ## Describe the bug
When fsspec 2022.8.2 is installed in your environment, xopen will prematurely close files, making streaming mode inoperable.
## Steps to reproduce the bug
```python
import datasets
data = datasets.load_dataset('MLCommons/ml_spoken_words', 'id_wav', split='train', streaming=True)
```
## Expected results
Dataset should load as iterator.
## Actual results
```
[/usr/local/lib/python3.7/dist-packages/datasets/load.py](https://localhost:8080/#) in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, **config_kwargs)
1737 # Return iterable dataset in case of streaming
1738 if streaming:
-> 1739 return builder_instance.as_streaming_dataset(split=split)
1740
1741 # Some datasets are already processed on the HF google storage
[/usr/local/lib/python3.7/dist-packages/datasets/builder.py](https://localhost:8080/#) in as_streaming_dataset(self, split, base_path)
1023 )
1024 self._check_manual_download(dl_manager)
-> 1025 splits_generators = {sg.name: sg for sg in self._split_generators(dl_manager)}
1026 # By default, return all splits
1027 if split is None:
[~/.cache/huggingface/modules/datasets_modules/datasets/MLCommons--ml_spoken_words/321ea853cf0a05abb7a2d7efea900692a3d8622af65a2f3ce98adb7800a5d57b/ml_spoken_words.py](https://localhost:8080/#) in _split_generators(self, dl_manager)
182 name=datasets.Split.TRAIN,
183 gen_kwargs={
--> 184 "audio_archives": [download_audio(split="train", lang=lang) for lang in self.config.languages],
185 "local_audio_archives_paths": [download_extract_audio(split="train", lang=lang) for lang in
186 self.config.languages] if not dl_manager.is_streaming else None,
[~/.cache/huggingface/modules/datasets_modules/datasets/MLCommons--ml_spoken_words/321ea853cf0a05abb7a2d7efea900692a3d8622af65a2f3ce98adb7800a5d57b/ml_spoken_words.py](https://localhost:8080/#) in <listcomp>(.0)
182 name=datasets.Split.TRAIN,
183 gen_kwargs={
--> 184 "audio_archives": [download_audio(split="train", lang=lang) for lang in self.config.languages],
185 "local_audio_archives_paths": [download_extract_audio(split="train", lang=lang) for lang in
186 self.config.languages] if not dl_manager.is_streaming else None,
[~/.cache/huggingface/modules/datasets_modules/datasets/MLCommons--ml_spoken_words/321ea853cf0a05abb7a2d7efea900692a3d8622af65a2f3ce98adb7800a5d57b/ml_spoken_words.py](https://localhost:8080/#) in _download_audio_archives(dl_manager, lang, format, split)
267 # for streaming case
268 def _download_audio_archives(dl_manager, lang, format, split):
--> 269 archives_paths = _download_audio_archives_paths(dl_manager, lang, format, split)
270 return [dl_manager.iter_archive(archive_path) for archive_path in archives_paths]
[~/.cache/huggingface/modules/datasets_modules/datasets/MLCommons--ml_spoken_words/321ea853cf0a05abb7a2d7efea900692a3d8622af65a2f3ce98adb7800a5d57b/ml_spoken_words.py](https://localhost:8080/#) in _download_audio_archives_paths(dl_manager, lang, format, split)
251 n_files_path = dl_manager.download(n_files_url)
252
--> 253 with open(n_files_path, "r", encoding="utf-8") as file:
254 n_files = int(file.read().strip()) # the file contains a number of archives
255
ValueError: I/O operation on closed file.
```
## Environment info
- `datasets` version: 2.4.0
- Platform: Linux-5.10.133+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.13
- PyArrow version: 6.0.1
- Pandas version: 1.3.5
| 4,961 |
https://github.com/huggingface/datasets/issues/4960 | BioASQ AttributeError: 'BuilderConfig' object has no attribute 'schema' | [
"Following worked:\r\n\r\n```\r\ndata_dir = \"/Users/dlituiev/repos/datasets/bioasq/\"\r\nbioasq_task_b = load_dataset(\"aps/bioasq_task_b\", data_dir=data_dir, name=\"bioasq_9b_source\")\r\n```\r\n\r\nWould maintainers be open to one of the following:\r\n- automating this with a latest default config (e.g. `bioasq_9b_source`); how can this be generalized to other datasets?\r\n- providing an actionable error message that lists available `name` values? I only got available `name` values once I've provided something there (`name=\"aps/bioasq_task_b\"`), before it would not even mention that it requires `name` argument",
"Hi ! In general the list of available configurations is prompted. I think this is an issue with this specific dataset.\r\n\r\nFeel free to open a new discussions at https://huggingface.co/datasets/aps/bioasq_task_b/discussions\r\n\r\ncc @apsdehal\r\n\r\nIn particular it sounds like the `BUILDER_CONFIG_CLASS= BigBioConfig ` class attribute is missing and the _info should account for schema being None and raise an error"
] | ## Describe the bug
I am trying to load a dataset from drive and running into an error.
## Steps to reproduce the bug
```python
data_dir = "/Users/dlituiev/repos/datasets/bioasq/BioASQ-training9b"
bioasq_task_b = load_dataset("aps/bioasq_task_b", data_dir=data_dir)
```
## Actual results
`AttributeError: 'BuilderConfig' object has no attribute 'schema'`
<details>
```
Using custom data configuration default-a1ca3e05be5abf2f
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Input In [8], in <cell line: 2>()
1 data_dir = "/Users/dlituiev/repos/datasets/bioasq/BioASQ-training9b"
----> 2 bioasq_task_b = load_dataset("aps/bioasq_task_b", data_dir=data_dir)
File ~/opt/anaconda3/envs/spacy3/lib/python3.10/site-packages/datasets/load.py:1723, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, **config_kwargs)
1720 ignore_verifications = ignore_verifications or save_infos
1722 # Create a dataset builder
-> 1723 builder_instance = load_dataset_builder(
1724 path=path,
1725 name=name,
1726 data_dir=data_dir,
1727 data_files=data_files,
1728 cache_dir=cache_dir,
1729 features=features,
1730 download_config=download_config,
1731 download_mode=download_mode,
1732 revision=revision,
1733 use_auth_token=use_auth_token,
1734 **config_kwargs,
1735 )
1737 # Return iterable dataset in case of streaming
1738 if streaming:
File ~/opt/anaconda3/envs/spacy3/lib/python3.10/site-packages/datasets/load.py:1526, in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, **config_kwargs)
1523 raise ValueError(error_msg)
1525 # Instantiate the dataset builder
-> 1526 builder_instance: DatasetBuilder = builder_cls(
1527 cache_dir=cache_dir,
1528 config_name=config_name,
1529 data_dir=data_dir,
1530 data_files=data_files,
1531 hash=hash,
1532 features=features,
1533 use_auth_token=use_auth_token,
1534 **builder_kwargs,
1535 **config_kwargs,
1536 )
1538 return builder_instance
File ~/opt/anaconda3/envs/spacy3/lib/python3.10/site-packages/datasets/builder.py:1154, in GeneratorBasedBuilder.__init__(self, writer_batch_size, *args, **kwargs)
1153 def __init__(self, *args, writer_batch_size=None, **kwargs):
-> 1154 super().__init__(*args, **kwargs)
1155 # Batch size used by the ArrowWriter
1156 # It defines the number of samples that are kept in memory before writing them
1157 # and also the length of the arrow chunks
1158 # None means that the ArrowWriter will use its default value
1159 self._writer_batch_size = writer_batch_size or self.DEFAULT_WRITER_BATCH_SIZE
File ~/opt/anaconda3/envs/spacy3/lib/python3.10/site-packages/datasets/builder.py:307, in DatasetBuilder.__init__(self, cache_dir, config_name, hash, base_path, info, features, use_auth_token, repo_id, data_files, data_dir, name, **config_kwargs)
305 if info is None:
306 info = self.get_exported_dataset_info()
--> 307 info.update(self._info())
308 info.builder_name = self.name
309 info.config_name = self.config.name
File ~/.cache/huggingface/modules/datasets_modules/datasets/aps--bioasq_task_b/3d54b1213f7e8001eef755af92877f9efa44161ee83c2a70d5d649defa95759e/bioasq_task_b.py:477, in BioasqTaskBDataset._info(self)
474 def _info(self):
475
476 # BioASQ Task B source schema
--> 477 if self.config.schema == "source":
478 features = datasets.Features(
479 {
480 "id": datasets.Value("string"),
(...)
504 }
505 )
506 # simplified schema for QA tasks
AttributeError: 'BuilderConfig' object has no attribute 'schema'
```
</details>
## Environment info
- `datasets` version: 2.4.0
- Platform: macOS-10.16-x86_64-i386-64bit
- Python version: 3.10.4
- PyArrow version: 9.0.0
- Pandas version: 1.4.3 | 4,960 |
https://github.com/huggingface/datasets/issues/4958 | ConnectionError: Couldn't reach https://raw.githubusercontent.com/huggingface/datasets/2.4.0/datasets/jsonl/jsonl.py | [
"I have solved this problem... The extension of the file should be `.json` not `.jsonl`"
] | Hi,
When I use load_dataset from local jsonl files, below error happens, and I type the link into the browser prompting me `404: Not Found`. I download the other `.py` files using the same method and it works. It seems that the server is missing the appropriate file, or it is a problem with the code version.
```
ConnectionError: Couldn't reach https://raw.githubusercontent.com/huggingface/datasets/2.3.0/datasets/jsonl/jsonl.py (ConnectionError(MaxRetryError("HTTPSConnectionPool(host='raw.githubusercontent.com', port=443): Max retries exceeded with url: /huggingface/datasets/2.3.0/datasets/jsonl/jsonl.py (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x2b08342004c0>: Failed to establish a new connection: [Errno 101] Network is unreachable'))")))
```
| 4,958 |
https://github.com/huggingface/datasets/issues/4955 | Raise a more precise error when the URL is unreachable in streaming mode | [] | See for example:
- https://github.com/huggingface/datasets/issues/3191
- https://github.com/huggingface/datasets/issues/3186
It would help provide clearer information on the Hub and help the dataset maintainer solve the issue by themselves quicker. Currently:
- https://huggingface.co/datasets/compguesswhat
<img width="1029" alt="Capture d’écran 2022-09-08 à 15 51 37" src="https://user-images.githubusercontent.com/1676121/189139946-6deffb91-f21b-4281-8825-a98026c69740.png">
- https://huggingface.co/datasets/nli_tr
<img width="1032" alt="Capture d’écran 2022-09-08 à 15 51 44" src="https://user-images.githubusercontent.com/1676121/189139963-d26490ed-ad23-48ea-9cfc-1ab9c4d08d0c.png">
cc @albertvillanova | 4,955 |
https://github.com/huggingface/datasets/issues/4953 | CI test of TensorFlow is failing | [] | ## Describe the bug
The following CI test fails: https://github.com/huggingface/datasets/runs/8246722693?check_suite_focus=true
```
FAILED tests/test_py_utils.py::TempSeedTest::test_tensorflow - AssertionError:
```
Details:
```
_________________________ TempSeedTest.test_tensorflow _________________________
[gw0] linux -- Python 3.7.13 /opt/hostedtoolcache/Python/3.7.13/x64/bin/python
self = <tests.test_py_utils.TempSeedTest testMethod=test_tensorflow>
@require_tf
def test_tensorflow(self):
import tensorflow as tf
from tensorflow.keras import layers
def gen_random_output():
model = layers.Dense(2)
x = tf.random.uniform((1, 3))
return model(x).numpy()
with temp_seed(42, set_tensorflow=True):
out1 = gen_random_output()
with temp_seed(42, set_tensorflow=True):
out2 = gen_random_output()
out3 = gen_random_output()
> np.testing.assert_equal(out1, out2)
E AssertionError:
E Arrays are not equal
E
E Mismatched elements: 2 / 2 (100%)
E Max absolute difference: 0.84619296
E Max relative difference: 16.083529
E x: array([[-0.793581, 0.333286]], dtype=float32)
E y: array([[0.052612, 0.539708]], dtype=float32)
tests/test_py_utils.py:149: AssertionError
```
| 4,953 |
https://github.com/huggingface/datasets/issues/4945 | Push to hub can push splits that do not respect the regex | [] | ## Describe the bug
The `push_to_hub` method can push splits that do not respect the regex check that is used for downloads. Therefore, splits may be pushed but never re-used, which can be painful if the split was done after runtime preprocessing.
## Steps to reproduce the bug
```python
>>> from datasets import Dataset, DatasetDict, load_dataset
>>> d = Dataset.from_dict({'x': [1,2,3], 'y': [1,2,3]})
>>> di = DatasetDict()
>>> di['identifier-with-column'] = d
>>> di.push_to_hub('open-source-metrics/test')
Pushing split identifier-with-column to the Hub.
Pushing dataset shards to the dataset hub: 100%|██████████| 1/1 [00:04<00:00, 4.40s/it]
```
Loading it afterwards:
```python
>>> load_dataset('open-source-metrics/test')
Downloading: 100%|██████████| 610/610 [00:00<00:00, 432kB/s]
Using custom data configuration open-source-metrics--test-28b63ec7cde80488
Downloading and preparing dataset None/None (download: 950 bytes, generated: 48 bytes, post-processed: Unknown size, total: 998 bytes) to /home/lysandre/.cache/huggingface/datasets/open-source-metrics___parquet/open-source-metrics--test-28b63ec7cde80488/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec...
Downloading data files: 0%| | 0/1 [00:00<?, ?it/s]
Downloading data: 100%|██████████| 950/950 [00:00<00:00, 1.01MB/s]
Downloading data files: 100%|██████████| 1/1 [00:01<00:00, 1.48s/it]
Extracting data files: 100%|██████████| 1/1 [00:00<00:00, 2291.97it/s]
Traceback (most recent call last):
File "/home/lysandre/.pyenv/versions/3.10.6/lib/python3.10/code.py", line 90, in runcode
exec(code, self.locals)
File "<input>", line 1, in <module>
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/load.py", line 1746, in load_dataset
builder_instance.download_and_prepare(
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/builder.py", line 771, in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/packaged_modules/parquet/parquet.py", line 48, in _split_generators
splits.append(datasets.SplitGenerator(name=split_name, gen_kwargs={"files": files}))
File "<string>", line 5, in __init__
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/splits.py", line 599, in __post_init__
NamedSplit(self.name) # check that it's a valid split name
File "/home/lysandre/Workspaces/python/Metrics/GitHub-Metrics/.env/lib/python3.10/site-packages/datasets/splits.py", line 346, in __init__
raise ValueError(f"Split name should match '{_split_re}' but got '{split_name}'.")
ValueError: Split name should match '^\w+(\.\w+)*$' but got 'identifier-with-column'.
```
## Expected results
I would expect `push_to_hub` to stop me in my tracks if trying to upload a split that will not be working afterwards.
## Actual results
See above
## Environment info
- `datasets` version: 2.4.0
- Platform: Linux-5.15.64-1-lts-x86_64-with-glibc2.36
- Python version: 3.10.6
- PyArrow version: 9.0.0
- Pandas version: 1.4.4
| 4,945 |
https://github.com/huggingface/datasets/issues/4944 | larger dataset, larger GPU memory in the training phase? Is that correct? | [
"does the trainer save it in GPU? sooo curious... how to fix it",
"It's my bad. didn't limit the input length"
] | from datasets import set_caching_enabled
set_caching_enabled(False)
for ds_name in ["squad","newsqa","nqopen","narrativeqa"]:
train_ds = load_from_disk("../../../dall/downstream/processedproqa/{}-train.hf".format(ds_name))
break
train_ds = concatenate_datasets([train_ds,train_ds,train_ds,train_ds]) #operation 1
trainer = QuestionAnsweringTrainer( #huggingface trainer
model=model,
args=training_args,
train_dataset=train_ds,
eval_dataset= None,
eval_examples=None,
answer_column_name=answer_column,
dataset_name="squad",
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
)
with operation 1, the GPU memory increases from 16G to 23G | 4,944 |
https://github.com/huggingface/datasets/issues/4942 | Trec Dataset has incorrect labels | [
"Thanks for reporting, @wmpauli. \r\n\r\nIndeed we recently fixed this issue:\r\n- #4801 \r\n\r\nThe fix will be accessible after our next library release. In the meantime, you can have it by passing `revision=\"main\"` to `load_dataset`."
] | ## Describe the bug
Both coarse and fine labels seem to be out of line.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = "trec"
raw_datasets = load_dataset(dataset)
df = pd.DataFrame(raw_datasets["test"])
df.head()
```
## Expected results
text (string) | coarse_label (class label) | fine_label (class label)
-- | -- | --
How far is it from Denver to Aspen ? | 5 (NUM) | 40 (NUM:dist)
What county is Modesto , California in ? | 4 (LOC) | 32 (LOC:city)
Who was Galileo ? | 3 (HUM) | 31 (HUM:desc)
What is an atom ? | 2 (DESC) | 24 (DESC:def)
When did Hawaii become a state ? | 5 (NUM) | 39 (NUM:date)
## Actual results
index | label-coarse |label-fine | text
-- |-- | -- | --
0 | 4 | 40 | How far is it from Denver to Aspen ?
1 | 5 | 21 | What county is Modesto , California in ?
2 | 3 | 12 | Who was Galileo ?
3 | 0 | 7 | What is an atom ?
4 | 4 | 8 | When did Hawaii become a state ?
## Environment info
- `datasets` version: 2.4.0
- Platform: Linux-5.4.0-1086-azure-x86_64-with-glibc2.27
- Python version: 3.9.13
- PyArrow version: 8.0.0
- Pandas version: 1.4.3
| 4,942 |
https://github.com/huggingface/datasets/issues/4936 | vivos (Vietnamese speech corpus) dataset not accessible | [
"If you need an example of a small audio datasets, I just created few hours ago a speech dataset with only 300MB of compressed audio files https://huggingface.co/datasets/indonesian-nlp/librivox-indonesia. It works also with streaming (@albertvillanova helped me adding this functionality) :-)",
"@cahya-wirawan omg this is awesome!! thank you! ",
"We have contacted the authors to ask them."
] | ## Describe the bug
VIVOS data is not accessible anymore, neither of these links work (at least from France):
* https://ailab.hcmus.edu.vn/assets/vivos.tar.gz (data)
* https://ailab.hcmus.edu.vn/vivos (dataset page)
Therefore `load_dataset` doesn't work.
## Steps to reproduce the bug
```python
ds = load_dataset("vivos")
```
## Expected results
dataset loaded
## Actual results
```
ConnectionError: Couldn't reach https://ailab.hcmus.edu.vn/assets/vivos.tar.gz (ConnectionError(MaxRetryError("HTTPSConnectionPool(host='ailab.hcmus.edu.vn', port=443): Max retries exceeded with url: /assets/vivos.tar.gz (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f9d8a27d190>: Failed to establish a new connection: [Errno -5] No address associated with hostname'))")))
```
Will try to contact the authors, as we wanted to use Vivos as an example in documentation on how to create scripts for audio datasets (https://github.com/huggingface/datasets/pull/4872), because it's small and straightforward and uses tar archives. | 4,936 |
https://github.com/huggingface/datasets/issues/4935 | Dataset Viewer issue for ubuntu_dialogs_corpus | [
"The dataset maintainers (https://huggingface.co/datasets/ubuntu_dialogs_corpus) decided to forbid the dataset from being downloaded automatically (https://huggingface.co/docs/datasets/v2.4.0/en/loading#manual-download), and the dataset viewer respects this.\r\nWe will try to improve the error display though. Thanks for reporting."
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,935 |
https://github.com/huggingface/datasets/issues/4934 | Dataset Viewer issue for indonesian-nlp/librivox-indonesia | [
"The error is not related to the dataset viewer. I'm having a look...",
"Thanks @albertvillanova for checking the issue. Actually, I can use the dataset like following:\r\n```\r\n>>> from datasets import load_dataset\r\n>>> ds=load_dataset(\"indonesian-nlp/librivox-indonesia\")\r\nNo config specified, defaulting to: librivox-indonesia/all\r\nReusing dataset librivox-indonesia (/root/.cache/huggingface/datasets/indonesian-nlp___librivox-indonesia/all/1.0.0/9a934a42bfb53dc103003d191618443b8a786bea2bd7bb0bc2d9454b8494521e)\r\n100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 500.87it/s]\r\n>>> ds\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['path', 'language', 'reader', 'sentence', 'audio'],\r\n num_rows: 7815\r\n })\r\n})\r\n>>> ds[\"train\"][0]\r\n{'path': '/root/.cache/huggingface/datasets/downloads/extracted/c8ead52370fa28feb64643ea9d05cd7d820192dc8a1700d665ec45ec7624f5a3/librivox-indonesia/sundanese/universal-declaration-of-human-rights/human_rights_un_sun_brc_0000.mp3', 'language': 'sun', 'reader': '3174', 'sentence': 'pernyataan umum ngeunaan hak hak asasi manusa sakabeh manusa', 'audio': {'path': '/root/.cache/huggingface/datasets/downloads/extracted/c8ead52370fa28feb64643ea9d05cd7d820192dc8a1700d665ec45ec7624f5a3/librivox-indonesia/sundanese/universal-declaration-of-human-rights/human_rights_un_sun_brc_0000.mp3', 'array': array([ 0. , 0. , 0. , ..., -0.02419001,\r\n -0.01957154, -0.01502833], dtype=float32), 'sampling_rate': 44100}}\r\n\r\n```\r\nIt would be just nice if I also can see it using dataset viewer.",
"Yes, the issue arises when streaming (that is used by the viewer): your script does not support streaming and to support it in this case there are some subtleties that we are explaining better in our docs in a work-in progress pull request:\r\n- #4872\r\n\r\nJust note that when streaming, `local_extracted_archive` is None, and this code line generates the error:\r\n```python\r\nfilepath = local_extracted_archive + \"/librivox-indonesia/audio_transcription.csv\"\r\n```\r\n\r\nFor a proper implementation, you could have a look at: https://huggingface.co/datasets/common_voice/blob/main/common_voice.py\r\n\r\nYou can test your script locally by passing `streaming=True` to `load_dataset`:\r\n```python\r\nds = load_dataset(\"indonesian-nlp/librivox-indonesia\", split=\"train\", streaming=True); item = next(iter(ds)); item\r\n```",
"Great, I will have a look and update the script. Thanks.",
"Hi @albertvillanova , I just add the streaming functionality and it works in the first try :-) Thanks a lot!",
"Awesome!!! :hugs: "
] | ### Link
https://huggingface.co/datasets/indonesian-nlp/librivox-indonesia
### Description
I created a new speech dataset https://huggingface.co/datasets/indonesian-nlp/librivox-indonesia, but the dataset preview doesn't work with following error message:
```
Server error
Status code: 400
Exception: TypeError
Message: unsupported operand type(s) for +: 'NoneType' and 'str'
```
Please help, I am not sure what the problem here is. Thanks a lot.
### Owner
Yes | 4,934 |
https://github.com/huggingface/datasets/issues/4933 | Dataset/DatasetDict.filter() cannot have `batched=True` due to `mask` (numpy array?) being non-iterable. | [
"Hi ! When `batched=True`, you filter function must take a batch as input, and return a list of booleans.\r\n\r\nIn your case, something like\r\n```python\r\nfrom datasets import load_dataset\r\n\r\n\r\nds_mc4_ja = load_dataset(\"mc4\", \"ja\") # This will take 6+ hours... perhaps test it with a toy dataset instead?\r\nds_mc4_ja_2020 = ds_mc4_ja.filter(\r\n lambda batch: [timestamp[:4] == \"2020\" for timestamp in batch[\"timestamp\"]],\r\n batched=True,\r\n)\r\n```\r\n\r\nLet me know if it helps !",
"> Hi ! When `batched=True`, you filter function must take a batch as input, and return a list of booleans.\r\n> [...]\r\n> Let me know if it helps !\r\n\r\nHi @lhoestq,\r\n\r\nAh, my bad, I totally forgot that part...\r\nSorry for the trouble and thank you for the kind help!"
] | ## Describe the bug
`Dataset/DatasetDict.filter()` cannot have `batched=True` due to `mask` (numpy array?) being non-iterable.
## Steps to reproduce the bug
(In a python 3.7.12 env, I've tried 2.4.0 and 2.3.2 with both `pyarraw==9.0.0` and `pyarrow==8.0.0`.)
```python
from datasets import load_dataset
ds_mc4_ja = load_dataset("mc4", "ja") # This will take 6+ hours... perhaps test it with a toy dataset instead?
ds_mc4_ja_2020 = ds_mc4_ja.filter(
lambda example: example["timestamp"][:4] == "2020",
batched=True,
)
```
## Expected results
No error
## Actual results
```python
---------------------------------------------------------------------------
RemoteTraceback Traceback (most recent call last)
RemoteTraceback:
"""
Traceback (most recent call last):
File "/opt/conda/lib/python3.7/site-packages/multiprocess/pool.py", line 121, in worker
result = (True, func(*args, **kwds))
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 557, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 524, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/opt/conda/lib/python3.7/site-packages/datasets/fingerprint.py", line 480, in wrapper
out = func(self, *args, **kwargs)
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2779, in _map_single
offset=offset,
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2655, in apply_function_on_filtered_inputs
processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 2347, in decorated
result = f(decorated_item, *args, **kwargs)
File "/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 4946, in get_indices_from_mask_function
indices_array = [i for i, to_keep in zip(indices, mask) if to_keep]
TypeError: zip argument #2 must support iteration
"""
The above exception was the direct cause of the following exception:
TypeError Traceback (most recent call last)
/tmp/ipykernel_51348/2345782281.py in <module>
7 batched=True,
8 # batch_size=10_000,
----> 9 num_proc=111,
10 )
11 # ds_mc4_ja_clean_2020 = ds_mc4_ja.filter(
/opt/conda/lib/python3.7/site-packages/datasets/dataset_dict.py in filter(self, function, with_indices, input_columns, batched, batch_size, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, fn_kwargs, num_proc, desc)
878 desc=desc,
879 )
--> 880 for k, dataset in self.items()
881 }
882 )
/opt/conda/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
878 desc=desc,
879 )
--> 880 for k, dataset in self.items()
881 }
882 )
/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
522 }
523 # apply actual function
--> 524 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
525 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
526 # re-apply format to the output
/opt/conda/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
478 # Call actual function
479
--> 480 out = func(self, *args, **kwargs)
481
482 # Update fingerprint of in-place transforms + update in-place history of transforms
/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py in filter(self, function, with_indices, input_columns, batched, batch_size, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
2920 new_fingerprint=new_fingerprint,
2921 input_columns=input_columns,
-> 2922 desc=desc,
2923 )
2924 new_dataset = copy.deepcopy(self)
/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, with_rank, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
2498
2499 for index, async_result in results.items():
-> 2500 transformed_shards[index] = async_result.get()
2501
2502 assert (
/opt/conda/lib/python3.7/site-packages/multiprocess/pool.py in get(self, timeout)
655 return self._value
656 else:
--> 657 raise self._value
658
659 def _set(self, i, obj):
TypeError: zip argument #2 must support iteration
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-4.19.0-21-cloud-amd64-x86_64-with-debian-10.12
- Python version: 3.7.12
- PyArrow version: 9.0.0
- Pandas version: 1.3.5
(I've tried 2.4.0 and 2.3.2 with both `pyarraw==9.0.0` and `pyarrow==8.0.0`.) | 4,933 |
https://github.com/huggingface/datasets/issues/4932 | Dataset Viewer issue for bigscience-biomedical/biosses | [
"Possibly not related to the dataset viewer in itself. cc @huggingface/datasets.\r\n\r\nIn particular, I think that the import of bigbiohub is not working here: https://huggingface.co/datasets/bigscience-biomedical/biosses/blob/main/biosses.py#L29 (requires a relative path?)\r\n\r\n```python\r\n>>> from datasets import get_dataset_config_names\r\n>>> get_dataset_config_names('bigscience-biomedical/biosses')\r\nDownloading builder script: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8.00k/8.00k [00:00<00:00, 7.47MB/s]\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 289, in get_dataset_config_names\r\n dataset_module = dataset_module_factory(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/load.py\", line 1247, in dataset_module_factory\r\n raise e1 from None\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/load.py\", line 1220, in dataset_module_factory\r\n return HubDatasetModuleFactoryWithScript(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/load.py\", line 931, in get_module\r\n local_imports = _download_additional_modules(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/load.py\", line 215, in _download_additional_modules\r\n raise ImportError(\r\nImportError: To be able to use bigscience-biomedical/biosses, you need to install the following dependency: bigbiohub.\r\nPlease install it using 'pip install bigbiohub' for instance'\r\n```",
"Opened a PR here to (hopefully) fix the dataset script: https://huggingface.co/datasets/bigscience-biomedical/biosses/discussions/1/files",
"thanks for taking a look @severo . agree this isn't related to dataset viewer (sorry just clicked on the auto issue creator). also thanks @lhoestq , I see the format to use for relative imports. was a bit confused b/c it seems to be working here \r\n\r\nhttps://huggingface.co/datasets/bigscience-biomedical/scitail/blob/main/scitail.py#L31\r\n\r\nI'll try this PR a see what happens. ",
"closing as I think the issue is relative imports and attempting to read json files directly in the repo (thanks again @lhoestq ) "
] | ### Link
https://huggingface.co/datasets/bigscience-biomedical/biosses
### Description
I've just been working on adding the dataset loader script to this dataset and working with the relative imports. I'm not sure how to interpret the error below (show where the dataset preview used to be) .
```
Status code: 400
Exception: ModuleNotFoundError
Message: No module named 'datasets_modules.datasets.bigscience-biomedical--biosses.ddbd5893bf6c2f4db06f407665eaeac619520ba41f69d94ead28f7cc5b674056.bigbiohub'
```
### Owner
Yes | 4,932 |
https://github.com/huggingface/datasets/issues/4924 | Concatenate_datasets loads everything into RAM | [] | ## Describe the bug
When loading the datasets seperately and saving them on disk, I want to concatenate them. But `concatenate_datasets` is filling up my RAM and the process gets killed. Is there a way to prevent this from happening or is this intended behaviour? Thanks in advance
## Steps to reproduce the bug
```python
gcs = gcsfs.GCSFileSystem(project='project')
datasets = [load_from_disk(f'path/to/slice/of/data/{i}', fs=gcs, keep_in_memory=False) for i in range(10)]
dataset = concatenate_datasets(datasets)
```
## Expected results
A concatenated dataset which is stored on my disk.
## Actual results
Concatenated dataset gets loaded into RAM and overflows it which gets the process killed.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-4.19.0-21-cloud-amd64-x86_64-with-glibc2.10
- Python version: 3.8.13
- PyArrow version: 8.0.1
- Pandas version: 1.4.3 | 4,924 |
https://github.com/huggingface/datasets/issues/4922 | I/O error on Google Colab in streaming mode | [] | ## Describe the bug
When trying to load a streaming dataset in Google Colab the loading fails with an I/O error
## Steps to reproduce the bug
```python
import datasets
from datasets import load_dataset
hf_ds = load_dataset(path='wmt19', name='cs-en', streaming=True, split=datasets.Split.VALIDATION)
list(hf_ds.take(5))
```
## Expected results
It should load five data points
## Actual results
```
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
[<ipython-input-13-7b5b8b1e7e58>](https://localhost:8080/#) in <module>
2 from datasets import load_dataset
3 hf_ds = load_dataset(path='wmt19', name='cs-en', streaming=True, split=datasets.Split.VALIDATION)
----> 4 list(hf_ds.take(5))
6 frames
[/usr/local/lib/python3.7/dist-packages/datasets/iterable_dataset.py](https://localhost:8080/#) in __iter__(self)
716
717 def __iter__(self):
--> 718 for key, example in self._iter():
719 if self.features:
720 # `IterableDataset` automatically fills missing columns with None.
[/usr/local/lib/python3.7/dist-packages/datasets/iterable_dataset.py](https://localhost:8080/#) in _iter(self)
706 else:
707 ex_iterable = self._ex_iterable
--> 708 yield from ex_iterable
709
710 def _iter_shard(self, shard_idx: int):
[/usr/local/lib/python3.7/dist-packages/datasets/iterable_dataset.py](https://localhost:8080/#) in __iter__(self)
582
583 def __iter__(self):
--> 584 yield from islice(self.ex_iterable, self.n)
585
586 def shuffle_data_sources(self, generator: np.random.Generator) -> "TakeExamplesIterable":
[/usr/local/lib/python3.7/dist-packages/datasets/iterable_dataset.py](https://localhost:8080/#) in __iter__(self)
110
111 def __iter__(self):
--> 112 yield from self.generate_examples_fn(**self.kwargs)
113
114 def shuffle_data_sources(self, generator: np.random.Generator) -> "ExamplesIterable":
[~/.cache/huggingface/modules/datasets_modules/datasets/wmt19/aeadcbe9f1cbf9969e603239d33d3e43670cf250c1158edf74f5f6e74d4f21d0/wmt_utils.py](https://localhost:8080/#) in _generate_examples(self, split_subsets, extraction_map, with_translation)
845 raise ValueError("Invalid number of files: %d" % len(files))
846
--> 847 for sub_key, ex in sub_generator(*sub_generator_args):
848 if not all(ex.values()):
849 continue
[~/.cache/huggingface/modules/datasets_modules/datasets/wmt19/aeadcbe9f1cbf9969e603239d33d3e43670cf250c1158edf74f5f6e74d4f21d0/wmt_utils.py](https://localhost:8080/#) in _parse_parallel_sentences(f1, f2, filename1, filename2)
923 l2_sentences, l2 = parse_file(f2_i, filename2)
924
--> 925 for line_id, (s1, s2) in enumerate(zip(l1_sentences, l2_sentences)):
926 key = f"{f_id}/{line_id}"
927 yield key, {l1: s1, l2: s2}
[~/.cache/huggingface/modules/datasets_modules/datasets/wmt19/aeadcbe9f1cbf9969e603239d33d3e43670cf250c1158edf74f5f6e74d4f21d0/wmt_utils.py](https://localhost:8080/#) in gen()
895
896 def gen():
--> 897 with open(path, encoding="utf-8") as f:
898 for line in f:
899 seg_match = re.match(seg_re, line)
ValueError: I/O operation on closed file.
```
## Environment info
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 2.4.0
- Platform: Linux-5.4.188+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.13
- PyArrow version: 9.0.0. (the same error happened with PyArrow version 6.0.0)
- Pandas version: 1.3.5
| 4,922 |
https://github.com/huggingface/datasets/issues/4920 | Unable to load local tsv files through load_dataset method | [
"Hi @DataNoob0723,\r\n\r\nUnder the hood, we use `pandas` to load CSV/TSV files. Therefore, you should use \"csv\" and pass `sep=\"\\t\"`, as explained in our docs: https://huggingface.co/docs/datasets/v2.4.0/en/package_reference/loading_methods#from-files\r\n```python\r\nds = load_dataset('csv', sep=\"\\t\", data_files=data_files)\r\n``` "
] | ## Describe the bug
Unable to load local tsv files through load_dataset method.
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
data_files = {
'train': 'train.tsv',
'test': 'test.tsv'
}
raw_datasets = load_dataset('tsv', data_files=data_files)
## Expected results
I am pretty sure the data files exist in the current directory. The above code should load them as Datasets, but threw exceptions.
## Actual results
---------------------------------------------------------------------------
FileNotFoundError Traceback (most recent call last)
[<ipython-input-9-24207899c1af>](https://localhost:8080/#) in <module>
----> 1 raw_datasets = load_dataset('tsv', data_files='train.tsv')
2 frames
[/usr/local/lib/python3.7/dist-packages/datasets/load.py](https://localhost:8080/#) in dataset_module_factory(path, revision, download_config, download_mode, dynamic_modules_path, data_dir, data_files, **download_kwargs)
1244 f"Couldn't find a dataset script at {relative_to_absolute_path(combined_path)} or any data file in the same directory. "
1245 f"Couldn't find '{path}' on the Hugging Face Hub either: {type(e1).__name__}: {e1}"
-> 1246 ) from None
1247 raise e1 from None
1248 else:
FileNotFoundError: Couldn't find a dataset script at /content/tsv/tsv.py or any data file in the same directory. Couldn't find 'tsv' on the Hugging Face Hub either: FileNotFoundError: Couldn't find file at https://raw.githubusercontent.com/huggingface/datasets/main/datasets/tsv/tsv.py
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-5.4.188+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.13
- PyArrow version: 6.0.1
- Pandas version: 1.3.5
| 4,920 |
https://github.com/huggingface/datasets/issues/4918 | Dataset Viewer issue for pysentimiento/spanish-targeted-sentiment-headlines | [
"Thanks for reporting, it's fixed now (I refreshed it manually). It's a known issue; we hope it will be fixed permanently in a few days.\r\n\r\n<img width=\"1508\" alt=\"Capture d’écran 2022-09-05 à 18 31 22\" src=\"https://user-images.githubusercontent.com/1676121/188489762-0ed86a7e-dfb3-46e8-a125-43b815a2c6f4.png\">\r\n",
"Thanks @severo! "
] | ### Link
https://huggingface.co/datasets/pysentimiento/spanish-targeted-sentiment-headlines
### Description
After moving the dataset from my user (`finiteautomata`) to the `pysentimiento` organization, the dataset viewer says that it doesn't exist.
### Owner
_No response_ | 4,918 |
https://github.com/huggingface/datasets/issues/4917 | Keys mismatch: make error message more informative | [
"Good idea ! I think this can be improved in `Features.reorder_fields_as()` indeed at\r\n\r\nhttps://github.com/huggingface/datasets/blob/7feeb5648a63b6135a8259dedc3b1e19185ee4c7/src/datasets/features/features.py#L1739-L1740\r\n\r\nIs it something you would be interested in contributing ?",
"Is this open to work on? I'd love to take on this as my first issue.",
"Hi @daspartho I’ve opened a PR #4919 \r\nI don’t think there’s much left to do",
"ok : )"
] | **Is your feature request related to a problem? Please describe.**
When loading a dataset from disk with a defect in its `dataset_info.json` describing its features (I don’t know when/why/how this happens but it deserves its own issue), you will get an error message like:
`ValueError: Keys mismatch: between {'bar': Value(dtype='int64', id=None)} and {'foo': Value(dtype='int64', id=None)}`
Which is fine when you have only a few features like in the example but it gets very hard to read when you have a lot of features in your dataset.
**Describe the solution you'd like**
The error message should give the difference between the features (what keys are in A but missing in B and vice-versa). It should also tell which keys are inferred from `dataset.arrow` and which come from `dataset_info.json`.
Willing to help :)
| 4,917 |
https://github.com/huggingface/datasets/issues/4916 | Apache Beam unable to write the downloaded wikipedia dataset | [
"See:\r\n- #4915"
] | ## Describe the bug
Hi, I am currently trying to download wikipedia dataset using
load_dataset("wikipedia", language="aa", date="20220401", split="train",beam_runner='DirectRunner'). However, I end up in getting filenotfound error. I get this error for any language I try to download. It downloads the file but while saving it in hugging face cache it fails to write. This happens for any available date of any language in wikipedia dump. I had raised another issue earlier #4915 but probably was not that clear and the solution provider misunderstood my problem. Hence raising one more issue. Any help is appreciated.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("wikipedia", language="aa", date="20220401", split="train",beam_runner='DirectRunner')
```
## Expected results
to load the dataset
## Actual results
I am pasting the error trace here:
Downloading builder script: 35.9kB [00:00, ?B/s]
Downloading metadata: 30.4kB [00:00, 1.94MB/s]
Using custom data configuration 20220401.aa-date=20220401,language=aa
Downloading and preparing dataset wikipedia/20220401.aa to C:\Users\Shilpa.cache\huggingface\datasets\wikipedia\20220401.aa-date=20220401,language=aa\2.0.0\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559...
Downloading data: 100%|████████████████████████████████████████████████████████████| 11.1k/11.1k [00:00<00:00, 712kB/s]
Downloading data files: 100%|████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.82s/it]
Extracting data files: 100%|█████████████████████████████████████████████████████████████████████| 1/1 [00:00<?, ?it/s]
Downloading data: 100%|███████████████████████████████████████████████████████████| 35.6k/35.6k [00:00<00:00, 84.3kB/s]
Downloading data files: 100%|████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.93s/it]
Traceback (most recent call last):
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1571, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "G:\Python3.7\lib\site-packages\apache_beam\io\iobase.py", line 1193, in process
self.writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 202, in open_writer
return FileBasedSinkWriter(self, writer_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 419, in init
self.temp_handle = self.sink.open(temp_shard_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\parquetio.py", line 553, in open
self._file_handle = super().open(temp_path)
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 139, in open
temp_path, self.mime_type, self.compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filesystems.py", line 224, in create
return filesystem.create(path, mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 163, in create
return self._path_open(path, 'wb', mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 140, in _path_open
raw_file = io.open(path, mode)
FileNotFoundError: [Errno 2] No such file or directory: 'C:\Users\Shilpa\.cache\huggingface\datasets\wikipedia\20220401.aa-date=20220401,language=aa\2.0.0\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559.incomplete\beam-temp-wikipedia-train-880233e8287e11edaf9d3ca067f2714e\20a05238-6106-4420-a713-4eca6dd5959a.wikipedia-train'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "G:/abc/temp.py", line 32, in
beam_runner='DirectRunner')
File "G:\Python3.7\lib\site-packages\datasets\load.py", line 1751, in load_dataset
use_auth_token=use_auth_token,
File "G:\Python3.7\lib\site-packages\datasets\builder.py", line 705, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "G:\Python3.7\lib\site-packages\datasets\builder.py", line 1394, in _download_and_prepare
pipeline_results = pipeline.run()
File "G:\Python3.7\lib\site-packages\apache_beam\pipeline.py", line 574, in run
return self.runner.run_pipeline(self, self._options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\direct\direct_runner.py", line 131, in run_pipeline
return runner.run_pipeline(pipeline, options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 201, in run_pipeline
options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 212, in run_via_runner_api
return self.run_stages(stage_context, stages)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 443, in run_stages
runner_execution_context, bundle_context_manager, bundle_input)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 776, in _execute_bundle
bundle_manager))
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 1000, in _run_bundle
data_input, data_output, input_timers, expected_timer_output)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 1309, in process_bundle
result_future = self._worker_handler.control_conn.push(process_bundle_req)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\worker_handlers.py", line 380, in push
response = self.worker.do_instruction(request)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\sdk_worker.py", line 598, in do_instruction
getattr(request, request_type), request.instruction_id)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\sdk_worker.py", line 635, in process_bundle
bundle_processor.process_bundle(instruction_id))
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\bundle_processor.py", line 1004, in process_bundle
element.data)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\bundle_processor.py", line 227, in process_encoded
self.output(decoded_value)
File "apache_beam\runners\worker\operations.py", line 526, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam\runners\worker\operations.py", line 528, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam\runners\worker\operations.py", line 237, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 324, in apache_beam.runners.worker.operations.GeneralPurposeConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 905, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1507, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1571, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "G:\Python3.7\lib\site-packages\apache_beam\io\iobase.py", line 1193, in process
self.writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 202, in open_writer
return FileBasedSinkWriter(self, writer_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 419, in init
self.temp_handle = self.sink.open(temp_shard_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\parquetio.py", line 553, in open
self._file_handle = super().open(temp_path)
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 139, in open
temp_path, self.mime_type, self.compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filesystems.py", line 224, in create
return filesystem.create(path, mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 163, in create
return self._path_open(path, 'wb', mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 140, in _path_open
raw_file = io.open(path, mode)
RuntimeError: FileNotFoundError: [Errno 2] No such file or directory: 'C:\Users\Shilpa\.cache\huggingface\datasets\wikipedia\20220401.aa-date=20220401,language=aa\2.0.0\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559.incomplete\beam-temp-wikipedia-train-880233e8287e11edaf9d3ca067f2714e\20a05238-6106-4420-a713-4eca6dd5959a.wikipedia-train' [while running 'train/Save to parquet/Write/WriteImpl/WriteBundles']
## Environment info
Python: 3.7.6
Windows 10 Pro
datasets :2.4.0
apache_beam: 2.41.0
mwparserfromhell: 0.6.4 | 4,916 |
https://github.com/huggingface/datasets/issues/4915 | FileNotFoundError while downloading wikipedia dataset for any language | [
"Hi @Shilpac20,\r\n\r\nAs explained in the Wikipedia dataset card: https://huggingface.co/datasets/wikipedia\r\n> You can find the full list of languages and dates [here](https://dumps.wikimedia.org/backup-index.html).\r\n\r\nThis means that, before passing a specific date, you should first make sure it is available online, as Wikimedia only keeps last X months (depending on the size of the corresponding language dump)): e.g. to see which dates \"aa\" Wikipedia is available online, see https://dumps.wikimedia.org/aawiki/ (as of today 2022-08-31, the available dates are from [20220401](https://dumps.wikimedia.org/aawiki/20220401/) to [20220820](https://dumps.wikimedia.org/aawiki/20220820/)).",
"Hi, the date that I have specified \"20220401\" is available for the language \"aa\". The error persists for any other available dates as present in https://dumps.wikimedia.org/aawiki/. The error is mainly due to apache beam not able to write the downloaded files. Any help on this?",
"I see, sorry, I misread your issue.\r\n\r\nWe are investigating this.",
"I am struggling with basically the same issue. I am trying to download the German Wikipedia dump.\r\n\r\nAs per the [documentation](https://huggingface.co/datasets/wikipedia), `\"20220301.de\"` should be available as a pre-processed dataset.\r\n\r\nIssuing the command mentioned in the documentation cited above\r\n\r\n from datasets import load_dataset\r\n load_dataset(\"wikipedia\", \"20220301.de\")\r\n\r\nraises the following `FileNotFound` error\r\n\r\n FileNotFoundError: Couldn't find file at https://dumps.wikimedia.org/dewiki/20220301/dumpstatus.json\r\n\r\nUsing the ([undocumented](https://huggingface.co/docs/datasets/v1.2.1/package_reference/loading_methods.html#datasets.load_dataset)?) call to `load_dataset()` with `language` and `date` parameters\r\n\r\n load_dataset(\"wikipedia\", language=\"de\", date=\"20220301\", beam_runner=\"DirectRunner\")\r\n\r\nproduces the same error.\r\n\r\nEDIT: as I am using `datasets` v2.7.1, I should be looking at [that version's documentation](https://huggingface.co/docs/datasets/v2.7.1/en/package_reference/loading_methods#datasets.load_dataset)! It is mentioned there, that additional `kwargs` are \"passed to the [BuilderConfig](https://huggingface.co/docs/datasets/v2.7.1/en/package_reference/builder_classes#datasets.BuilderConfig) and used in the [DatasetBuilder](https://huggingface.co/docs/datasets/v2.7.1/en/package_reference/builder_classes#datasets.DatasetBuilder)\". So I guess that is how `language` and `date` are used.\r\n\r\nAs I can see a folder `20221130` on `https://dumps.wikimedia.org/dewiki/`, I also tried\r\n\r\n from datasets import load_dataset\r\n load_dataset(\"wikipedia\", \"20221130.de\")\r\n\r\nwhich throws another error:\r\n\r\n ValueError: BuilderConfig 20221120.de not found. Available: ['20220301.aa', ... '20220301.de', ...\r\n\r\nbasically telling me that the dataset I originally requested (`'20220301.de'`) is available...\r\n\r\nIt seems that `load_dataset` is not handling the vanishing older dumps for Wikipedia correctly?",
"I am able to start downloading the dataset when trying anything with the recent dumps for 20221201. But obviously, those are the big wiki dumps and I need the smaller preloaded version.\r\n\r\nI am now getting some error when the files show up in my cache but it will say FileNotFoundError at the end of the download for some reason. The cache directory to the datasets\\wikipedia\\date.bn\\ had something in it, then when the error came up it disappeared. \r\n\r\nIt is easy to test with the langauge \"bn\" because the amount of files is low.\r\n\r\ndataset = load_dataset('wikipedia', date=\"20221201\", language=\"bn\", split='train', beam_runner='DirectRunner')"
] | ## Describe the bug
Hi, I am currently trying to download wikipedia dataset using
load_dataset("wikipedia", language="aa", date="20220401", split="train",beam_runner='DirectRunner'). However, I end up in getting filenotfound error. I get this error for any language I try to download.
Environment:
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("wikipedia", language="aa", date="20220401", split="train",beam_runner='DirectRunner')
```
## Expected results
to load the dataset
## Actual results
I am pasting the error trace here:
Downloading builder script: 35.9kB [00:00, ?B/s]
Downloading metadata: 30.4kB [00:00, 1.94MB/s]
Using custom data configuration 20220401.aa-date=20220401,language=aa
Downloading and preparing dataset wikipedia/20220401.aa to C:\Users\Shilpa\.cache\huggingface\datasets\wikipedia\20220401.aa-date=20220401,language=aa\2.0.0\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559...
Downloading data: 100%|████████████████████████████████████████████████████████████| 11.1k/11.1k [00:00<00:00, 712kB/s]
Downloading data files: 100%|████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.82s/it]
Extracting data files: 100%|█████████████████████████████████████████████████████████████████████| 1/1 [00:00<?, ?it/s]
Downloading data: 100%|███████████████████████████████████████████████████████████| 35.6k/35.6k [00:00<00:00, 84.3kB/s]
Downloading data files: 100%|████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.93s/it]
Traceback (most recent call last):
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1571, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "G:\Python3.7\lib\site-packages\apache_beam\io\iobase.py", line 1193, in process
self.writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 202, in open_writer
return FileBasedSinkWriter(self, writer_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 419, in __init__
self.temp_handle = self.sink.open(temp_shard_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\parquetio.py", line 553, in open
self._file_handle = super().open(temp_path)
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 139, in open
temp_path, self.mime_type, self.compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filesystems.py", line 224, in create
return filesystem.create(path, mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 163, in create
return self._path_open(path, 'wb', mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 140, in _path_open
raw_file = io.open(path, mode)
FileNotFoundError: [Errno 2] No such file or directory: 'C:\\Users\\Shilpa\\.cache\\huggingface\\datasets\\wikipedia\\20220401.aa-date=20220401,language=aa\\2.0.0\\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559.incomplete\\beam-temp-wikipedia-train-880233e8287e11edaf9d3ca067f2714e\\20a05238-6106-4420-a713-4eca6dd5959a.wikipedia-train'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "G:/abc/temp.py", line 32, in <module>
beam_runner='DirectRunner')
File "G:\Python3.7\lib\site-packages\datasets\load.py", line 1751, in load_dataset
use_auth_token=use_auth_token,
File "G:\Python3.7\lib\site-packages\datasets\builder.py", line 705, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "G:\Python3.7\lib\site-packages\datasets\builder.py", line 1394, in _download_and_prepare
pipeline_results = pipeline.run()
File "G:\Python3.7\lib\site-packages\apache_beam\pipeline.py", line 574, in run
return self.runner.run_pipeline(self, self._options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\direct\direct_runner.py", line 131, in run_pipeline
return runner.run_pipeline(pipeline, options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 201, in run_pipeline
options)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 212, in run_via_runner_api
return self.run_stages(stage_context, stages)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 443, in run_stages
runner_execution_context, bundle_context_manager, bundle_input)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 776, in _execute_bundle
bundle_manager))
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 1000, in _run_bundle
data_input, data_output, input_timers, expected_timer_output)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\fn_runner.py", line 1309, in process_bundle
result_future = self._worker_handler.control_conn.push(process_bundle_req)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\portability\fn_api_runner\worker_handlers.py", line 380, in push
response = self.worker.do_instruction(request)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\sdk_worker.py", line 598, in do_instruction
getattr(request, request_type), request.instruction_id)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\sdk_worker.py", line 635, in process_bundle
bundle_processor.process_bundle(instruction_id))
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\bundle_processor.py", line 1004, in process_bundle
element.data)
File "G:\Python3.7\lib\site-packages\apache_beam\runners\worker\bundle_processor.py", line 227, in process_encoded
self.output(decoded_value)
File "apache_beam\runners\worker\operations.py", line 526, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam\runners\worker\operations.py", line 528, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam\runners\worker\operations.py", line 237, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 324, in apache_beam.runners.worker.operations.GeneralPurposeConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 905, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 623, in apache_beam.runners.common.SimpleInvoker.invoke_process
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1491, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1581, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "apache_beam\runners\common.py", line 1694, in apache_beam.runners.common._OutputHandler._write_value_to_tag
File "apache_beam\runners\worker\operations.py", line 240, in apache_beam.runners.worker.operations.SingletonElementConsumerSet.receive
File "apache_beam\runners\worker\operations.py", line 907, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\worker\operations.py", line 908, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam\runners\common.py", line 1419, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 1507, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam\runners\common.py", line 1417, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam\runners\common.py", line 837, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam\runners\common.py", line 981, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "apache_beam\runners\common.py", line 1571, in apache_beam.runners.common._OutputHandler.handle_process_outputs
File "G:\Python3.7\lib\site-packages\apache_beam\io\iobase.py", line 1193, in process
self.writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 202, in open_writer
return FileBasedSinkWriter(self, writer_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 419, in __init__
self.temp_handle = self.sink.open(temp_shard_path)
File "G:\Python3.7\lib\site-packages\apache_beam\io\parquetio.py", line 553, in open
self._file_handle = super().open(temp_path)
File "G:\Python3.7\lib\site-packages\apache_beam\options\value_provider.py", line 193, in _f
return fnc(self, *args, **kwargs)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filebasedsink.py", line 139, in open
temp_path, self.mime_type, self.compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\filesystems.py", line 224, in create
return filesystem.create(path, mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 163, in create
return self._path_open(path, 'wb', mime_type, compression_type)
File "G:\Python3.7\lib\site-packages\apache_beam\io\localfilesystem.py", line 140, in _path_open
raw_file = io.open(path, mode)
RuntimeError: FileNotFoundError: [Errno 2] No such file or directory: 'C:\\Users\\Shilpa\\.cache\\huggingface\\datasets\\wikipedia\\20220401.aa-date=20220401,language=aa\\2.0.0\\aa542ed919df55cc5d3347f42dd4521d05ca68751f50dbc32bae2a7f1e167559.incomplete\\beam-temp-wikipedia-train-880233e8287e11edaf9d3ca067f2714e\\20a05238-6106-4420-a713-4eca6dd5959a.wikipedia-train' [while running 'train/Save to parquet/Write/WriteImpl/WriteBundles']
## Environment info
Python: 3.7.6
Windows 10 Pro
datasets :2.4.0
apache_beam: 2.41.0
mwparserfromhell: 0.6.4
| 4,915 |
https://github.com/huggingface/datasets/issues/4912 | datasets map() handles all data at a stroke and takes long time | [
"Hi ! Interesting question ;)\r\n\r\n> Which is better? Process in map() or in data-collator\r\n\r\nAs you said, both can be used in practice: map() if you want to preprocess before training, or a data-collator (or the equivalent `dataset.set_transform`) if you want to preprocess on-the-fly during training. Both options are great and really depend on your case.\r\n\r\nTo choose between the two, here are IMO the main caveats of each approach:\r\n- if your preprocessing takes too much CPU for example, using a data-collator may slow down your training and your GPUs may not work at full speed\r\n- on the other hand, map() may take a lot of time and disk space to run if your dataset is too big.\r\n\r\n> Why huggingface advises map() function? There should be some advantages to using map()\r\n\r\nTo get the best throughput when training a model, it is often recommended to preprocess your dataset before training. Note that preprocessing may include other steps before tokenization such as data filtering, cleaning, chunking etc. which are often done before training.",
"Thanks for your clear explanation @lhoestq ! \r\n> * if your preprocessing takes too much CPU for example, using a data-collator may slow down your training and your GPUs may not work at full speed\r\n> * on the other hand, map() may take a lot of time and disk space to run if your dataset is too big.\r\n\r\nI really agree with you. There should be some trade-off between processing before and during the train loop.\r\nBesides, I find `map()` function can cache the results once it has been executed. Very useful!",
"I'm closing this issue if you don't mind, feel free to reopen if needed ;)",
"@lhoestq How to preprocess on-the-fly during training？my data is about 1w hours, when I use map to preprocess, and It's not finished yet， but all disk space(2T) is full.",
"Hi ! You can do that using `set_transform`, see https://huggingface.co/docs/datasets/process#format-transform for more info :)",
"unfortunately , it not work.",
"Could you share more details ?"
] | **1. Background**
Huggingface datasets package advises using `map()` to process data in batches. In the example code on pretraining masked language model, they use `map()` to tokenize all data at a stroke before the train loop.
The corresponding code:
```
with accelerator.main_process_first():
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on every text in dataset"
)
```
**2. The problem**
Thus, when I try the same pertaining code with a much larger corpus, it takes quite a long time to tokenize.
Also, we can choose to tokenize data in `data-collator`. In this way, the program only tokenizes one batch in the next training step and avoids getting stuck in tokenization.
**3. My question**
As described above, my questions are:
* **Which is better? Process in `map()` or in `data-collator`**
* **Why huggingface advises `map()` function?** There should be some advantages to using `map()`
Thanks for your answers! | 4,912 |
https://github.com/huggingface/datasets/issues/4911 | [Tests] Ensure `datasets` supports renamed repositories | [
"You could also switch to using `huggingface_hub` more directly, where such a guarantee is already tested =)\r\n\r\ncc @Wauplin "
] | On https://hf.co/datasets you can rename a dataset (or sometimes move it to another user/org). The website handles redirections correctly and AFAIK `datasets` does as well.
However it would be nice to have an integration test to make sure we don't break support for renamed datasets.
To implement this we can use the /api/repos/move endpoint on hub-ci to rename/move a repo (it is documented at https://huggingface.co/docs/hub/api) | 4,911 |
https://github.com/huggingface/datasets/issues/4910 | Identical keywords in build_kwargs and config_kwargs lead to TypeError in load_dataset_builder() | [
"I am getting similar error - `TypeError: type object got multiple values for keyword argument 'name'` while following this [tutorial](https://huggingface.co/docs/datasets/dataset_script#create-a-dataset-loading-script). I am getting this error with the `dataset-cli test` command.\r\n\r\n`datasets` version: 2.4.0",
"In my case, this was happening because I defined multiple `BuilderConfig` for multiple types, but didn't had all the data files that are requierd by those configs. \r\n\r\nI think this is different than the original issue by @bablf .",
"Hi ! I think this can be fixed by letting the config_kwargs take over the builder kwargs here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/7feeb5648a63b6135a8259dedc3b1e19185ee4c7/src/datasets/load.py#L1533-L1534\r\n\r\nmaybe something like this ?\r\n```python\r\n **{**builder_kwargs, **config_kwargs}\r\n```\r\n\r\nLet me know if you'd like to contribute and fix this bug, so I can assign you :)\r\n\r\n> In my case, this was happening because I defined multiple BuilderConfig for multiple types, but didn't had all the data files that are requierd by those configs.\r\n> \r\n> I think this is different than the original issue by @bablf .\r\n\r\nFeel free to to open an new issue, I'd be happy to help\r\n",
"@lhoestq Yeah, I want to, please assign.",
"Cool thank you ! Let me know if you have questions or if I can help",
"@lhoestq On second thoughts, I think this might be expected behavior; although a better error message might help.\r\n\r\nReasoning: Given n configs, if no data file is provided for any config, then it should be an error. Then why it should not be the case if out of n configs, for some data files are provided but not for others. Also, I was using `--all_configs` flag with `dataset-cli test`.",
"Ok I see - maybe we should check the values of builder_kwargs raise an error if any key in config_kwargs tries to overwrite it ? The builder kwargs are determined from the builder's type and location (in some cases it forces the base_path, data_files and config name for example)"
] | ## Describe the bug
In `load_dataset_builder()`, `build_kwargs` and `config_kwargs` can contain the same keywords leading to a TypeError("type object got multiple values for keyword argument "xyz").
I ran into this problem with the keyword: `base_path`. It might happen with other kwargs as well. I think a quickfix would be
```python
builder_cls = import_main_class(dataset_module.module_path)
builder_kwargs = dataset_module.builder_kwargs
data_files = builder_kwargs.pop("data_files", data_files)
config_name = builder_kwargs.pop("config_name", name)
hash = builder_kwargs.pop("hash")
base_path = builder_kwargs.pop("base_path")
```
and then pass base_path into `builder_cls`.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("rotten_tomatoes", base_path="./sample_data")
```
## Expected results
The docs state: `**config_kwargs` — Keyword arguments to be passed to the [BuilderConfig](https://huggingface.co/docs/datasets/v2.4.0/en/package_reference/builder_classes#datasets.BuilderConfig) and used in the [DatasetBuilder](https://huggingface.co/docs/datasets/v2.4.0/en/package_reference/builder_classes#datasets.DatasetBuilder).
So I would expect to be able to pass the base_path into `load_dataset()`.
## Actual results
TypeError("type object got multiple values for keyword argument "base_path").
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: macOS-12.5-arm64-arm-64bit
- Python version: 3.8.9
- PyArrow version: 9.0.0
| 4,910 |
https://github.com/huggingface/datasets/issues/4907 | None Type error for swda datasets | [
"Thanks for reporting @hannan72 ! I couldn't reproduce the error on my side, can you share the full stack trace please ?",
"Thanks a lot for your response @lhoestq \r\nThe problem is solved accidentally today and I don't know exactly why it was happened yesterday.\r\nThe issue can be closed.",
"Ok, let us know if you encounter the issue again ;)"
] | ## Describe the bug
I got `'NoneType' object is not callable` error while calling the swda datasets.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset("swda")
```
## Expected results
Run without error
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Python version: 3.8.10
| 4,907 |
https://github.com/huggingface/datasets/issues/4906 | Can't import datasets AttributeError: partially initialized module 'datasets' has no attribute 'utils' (most likely due to a circular import) | [
"Thanks for reporting, @OPterminator.\r\n\r\nHowever, we are not able to reproduce this issue.\r\n\r\nThere might be 2 reasons why you get this exception:\r\n- Either the name of your local Python file: if it is called `datasets.py` this could generate a circular import when trying to import the Hugging Face `datasets` library.\r\n - You could try to rename it and run it again.\r\n- Another cause could be the simultaneous use of the packages `nlp` and `datasets`. Please note that we renamed the Hugging Face `nlp` library to `datasets` more than 2 years ago: they are 2 versions of the same library.\r\n - Please try to update your script and use only `datasets` (`nlp` name is no longer in use and is out of date).",
"i am also facing this issue\r\n\r\n\r\n```\r\n----> 1 import datasets\r\n 3 dataset = datasets.load_dataset(\"ucberkeley-dlab/measuring-hate-speech\", \"binary\")\r\n 4 df = dataset[\"train\"].to_pandas()\r\n\r\nFile ~/.pyenv/versions/3.10.9/lib/python3.10/site-packages/datasets/__init__.py:52\r\n 50 from .fingerprint import disable_caching, enable_caching, is_caching_enabled, set_caching_enabled\r\n 51 from .info import DatasetInfo, MetricInfo\r\n---> 52 from .inspect import (\r\n 53 get_dataset_config_info,\r\n 54 get_dataset_config_names,\r\n 55 get_dataset_infos,\r\n 56 get_dataset_split_names,\r\n 57 inspect_dataset,\r\n 58 inspect_metric,\r\n 59 list_datasets,\r\n 60 list_metrics,\r\n 61 )\r\n 62 from .iterable_dataset import IterableDataset\r\n 63 from .load import load_dataset, load_dataset_builder, load_from_disk, load_metric\r\n\r\nFile ~/.pyenv/versions/3.10.9/lib/python3.10/site-packages/datasets/inspect.py:30\r\n 28 from .download.streaming_download_manager import StreamingDownloadManager\r\n...\r\n---> 16 logger = datasets.utils.logging.get_logger(__name__)\r\n 19 if datasets.config.PYARROW_VERSION.major >= 7:\r\n 21 def pa_table_to_pylist(table):\r\n```",
"I am facing the same question. And this happens when i installing `evaluate` package while `jupyter notebook` running. I'm not sure if the error occured because of trying to import the package installed when the notebook is running. Surpringly when i stop the notebook and rerun, the issue has been solved itself. Hope this will be helpful : )",
"I also got this error.\r\nIt helped me to find the python process and kill it, then restart the kernel and the error disappeared.",
"> I also got this error. It helped me to find the python process and kill it, then restart the kernel and the error disappeared.\r\n\r\nYes!",
"> I am facing the same question. And this happens when i installing `evaluate` package while `jupyter notebook` running. I'm not sure if the error occured because of trying to import the package installed when the notebook is running. Surpringly when i stop the notebook and rerun, the issue has been solved itself. Hope this will be helpful : )\r\n\r\nThank you! :)"
] | ## Describe the bug
A clear and concise description of what the bug is.
Not able to import datasets
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
import os
os.environ["WANDB_API_KEY"] = "0" ## to silence warning
import numpy as np
import random
import sklearn
import matplotlib.pyplot as plt
import pandas as pd
import sys
import tensorflow as tf
import plotly.express as px
import transformers
import tokenizers
import nlp as nlp
import utils
import datasets
```
## Expected results
A clear and concise description of the expected results.
import should work normal
## Actual results
Specify the actual results or traceback.
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-21-b3b5b0b62103> in <module>
13 import nlp as nlp
14 import utils
---> 15 import datasets
~\anaconda3\lib\site-packages\datasets\__init__.py in <module>
44 from .fingerprint import disable_caching, enable_caching, is_caching_enabled, set_caching_enabled
45 from .info import DatasetInfo, MetricInfo
---> 46 from .inspect import (
47 get_dataset_config_info,
48 get_dataset_config_names,
~\anaconda3\lib\site-packages\datasets\inspect.py in <module>
28 from .download.streaming_download_manager import StreamingDownloadManager
29 from .info import DatasetInfo
---> 30 from .load import dataset_module_factory, import_main_class, load_dataset_builder, metric_module_factory
31 from .utils.file_utils import relative_to_absolute_path
32 from .utils.logging import get_logger
~\anaconda3\lib\site-packages\datasets\load.py in <module>
53 from .iterable_dataset import IterableDataset
54 from .metric import Metric
---> 55 from .packaged_modules import (
56 _EXTENSION_TO_MODULE,
57 _MODULE_SUPPORTS_METADATA,
~\anaconda3\lib\site-packages\datasets\packaged_modules\__init__.py in <module>
4 from typing import List
5
----> 6 from .csv import csv
7 from .imagefolder import imagefolder
8 from .json import json
~\anaconda3\lib\site-packages\datasets\packaged_modules\csv\csv.py in <module>
13
14
---> 15 logger = datasets.utils.logging.get_logger(__name__)
16
17 _PANDAS_READ_CSV_NO_DEFAULT_PARAMETERS = ["names", "prefix"]
AttributeError: partially initialized module 'datasets' has no attribute 'utils' (most likely due to a circular import)
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
Copy-and-paste the text below in your GitHub issue.
- `datasets` version: 2.4.0
- Platform: Windows-10-10.0.22000-SP0
- Python version: 3.8.8
- PyArrow version: 9.0.0
- Pandas version: 1.2.4
| 4,906 |
https://github.com/huggingface/datasets/issues/4902 | Name the default config `default` | [
"Addressed in #5331."
] | Currently, if a dataset has no configuration, a default configuration is created from the dataset name.
For example, for a dataset loaded from the hub repository, such as https://huggingface.co/datasets/user/dataset (repo id is `user/dataset`), the default configuration will be `user--dataset`.
It might be easier to handle to set it to `default`, or another reserved word. | 4,902 |
https://github.com/huggingface/datasets/issues/4900 | Dataset Viewer issue for asaxena1990/Dummy_dataset | [
"Seems to be linked to the use of the undocumented `_resolve_features` method in the dataset viewer backend:\r\n\r\n```\r\n>>> from datasets import load_dataset\r\n>>> dataset = load_dataset(\"asaxena1990/Dummy_dataset\", name=\"asaxena1990--Dummy_dataset\", split=\"train\", streaming=True)\r\nUsing custom data configuration asaxena1990--Dummy_dataset-4a704ed7e5627563\r\n>>> dataset._resolve_features()\r\nFailed to read file 'https://huggingface.co/datasets/asaxena1990/Dummy_dataset/resolve/06885879a8bdd767d2d27695484fc6c83244617a/dummy_dataset_train.json' with error <class 'pyarrow.lib.ArrowInvalid'>: JSON parse error: Column() changed from object to array in row 0\r\nTraceback (most recent call last):\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py\", line 109, in _generate_tables\r\n pa_table = paj.read_json(\r\n File \"pyarrow/_json.pyx\", line 246, in pyarrow._json.read_json\r\n File \"pyarrow/error.pxi\", line 143, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 99, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to array in row 0\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 1261, in _resolve_features\r\n features = _infer_features_from_batch(self._head())\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 686, in _head\r\n return _examples_to_batch([x for key, x in islice(self._iter(), n)])\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 686, in <listcomp>\r\n return _examples_to_batch([x for key, x in islice(self._iter(), n)])\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 708, in _iter\r\n yield from ex_iterable\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 112, in __iter__\r\n yield from self.generate_examples_fn(**self.kwargs)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 651, in wrapper\r\n for key, table in generate_tables_fn(**kwargs):\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py\", line 137, in _generate_tables\r\n f\"This JSON file contain the following fields: {str(list(dataset.keys()))}. \"\r\nAttributeError: 'list' object has no attribute 'keys'\r\n```\r\n\r\nPinging @huggingface/datasets",
"Hi ! JSON files containing a list of object are not supported yet, you can use JSON Lines files instead in the meantime\r\n```json\r\n{\"text\": \"can I know this?\", \"intent\": \"Know\", \"type\": \"Test\"}\r\n{\"text\": \"can I know this?\", \"intent\": \"Know\", \"type\": \"Test\"}\r\n...\r\n```",
"A JSON list of objects is supported as of version 2.5.0."
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,900 |
https://github.com/huggingface/datasets/issues/4898 | Dataset Viewer issue for timit_asr | [
"Yes, the dataset viewer is based on `datasets`, and the following does not work:\r\n\r\n```\r\n>>> from datasets import get_dataset_split_names\r\n>>> get_dataset_split_names('timit_asr')\r\nDownloading builder script: 7.48kB [00:00, 6.69MB/s]\r\nTraceback (most recent call last):\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 354, in get_dataset_config_info\r\n for split_generator in builder._split_generators(\r\n File \"/home/slesage/.cache/huggingface/modules/datasets_modules/datasets/timit_asr/43f9448dd5db58e95ee48a277f466481b151f112ea53e27f8173784da9254fb2/timit_asr.py\", line 117, in _split_generators\r\n data_dir = os.path.abspath(os.path.expanduser(dl_manager.manual_dir))\r\n File \"/home/slesage/.pyenv/versions/3.9.6/lib/python3.9/posixpath.py\", line 231, in expanduser\r\n path = os.fspath(path)\r\nTypeError: expected str, bytes or os.PathLike object, not NoneType\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 404, in get_dataset_split_names\r\n info = get_dataset_config_info(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 359, in get_dataset_config_info\r\n raise SplitsNotFoundError(\"The split names could not be parsed from the dataset config.\") from err\r\ndatasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.\r\n```\r\n\r\ncc @huggingface/datasets ",
"Due to license restriction, this dataset needs manual downloading of the original data.\r\n\r\nThis information is in the dataset card: https://huggingface.co/datasets/timit_asr\r\n> The dataset needs to be downloaded manually from https://catalog.ldc.upenn.edu/LDC93S1",
"Maybe a better error message for datasets that need manual downloading? @severo \r\n\r\nMaybe we can raise a specific excpetion as done from `load_dataset`...",
"Yes, ideally something like https://github.com/huggingface/datasets/blob/main/src/datasets/builder.py#L81\r\n",
"The preview is now disabled (and a descriptive warning is displayed) for datasets requiring manual download. See:\r\n\r\n\r\n"
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,898 |
https://github.com/huggingface/datasets/issues/4897 | datasets generate large arrow file | [
"Hi ! The cache files are the results of all the transforms you applied to the dataset using `map` for example.\r\nDid you run a transform that could potentially blow up the size of the dataset ?",
"@lhoestq,\r\nI don't remember, but I can't imagine what kind of transform may generate data that grow over 200 times in size. \r\nI think maybe it doesn' matter, it's just cache after all."
] | Checking the large file in disk, and found the large cache file in the cifar10 data directory:
As we know, the size of cifar10 dataset is ~130MB, but the cache file has almost 30GB size, there may be some problems here. | 4,897 |
https://github.com/huggingface/datasets/issues/4895 | load_dataset method returns Unknown split "validation" even if this dir exists | [
"I don't know the main problem but it looks like, it is ignoring the last directory in your case. So, create a directory called 'zzz' in the same folder as train, validation and test. if it doesn't work, create a directory called \"aaa\". It worked for me.\r\n",
"@SamSamhuns could you please try to load it with the current main-branch version of `datasets`? I suppose the problem is that it tries to get splits names from filenames in this case, ignoring directories names, but `val` wasn't in keywords at that time, but it was fixed recently in this PR https://github.com/huggingface/datasets/pull/4844. ",
"I have a similar problem.\r\nWhen I try to create `data_infos.json` using `datasets-cli test Peter.py --save_infos --all_configs` I get an error:\r\n`ValueError: Unknown split \"test\". Should be one of ['train'].`\r\n\r\nThe `data_infos.json` is created perfectly fine when I use only one split - `datasets.Split.TRAIN`\r\n\r\n@polinaeterna Could you help here please?\r\n\r\nYou can find the code here: https://huggingface.co/datasets/sberbank-ai/Peter/tree/add_splits (add_splits branch)",
"@skalinin It seems the `dataset_infos.json` of your dataset is missing the info on the test split (and `datasets-cli` doesn't ignore the cached infos at the moment, which is a known bug), so your issue is not related to this one. I think you can fix your issue by deleting all the cached `dataset_infos.json` (in the local repo and in `~/.cache/huggingface/modules`) before running the `datasets-cli test` command. Let us know if that doesn't help, and I can try to generate it myself.",
"This code indeed behaves as expected on `main`. But suppose the `val_234.png` is renamed to some other value not containing one of [these](https://github.com/huggingface/datasets/blob/38c8c725f3996ff1ff03f6fd461aa6d645321034/src/datasets/data_files.py#L31) keywords, in that case, this issue becomes relevant again because the real cause of it is the order in which we check the predefined split patterns to assign data files to each split - first we assign data files based on filenames, and only if this fails meaning not a single split found (`val` is not recognized here in the older versions of `datasets`, which results in an empty `validation` split), do we assign based on directory names.\r\n\r\n@polinaeterna @lhoestq Perhaps one way to fix this would be to swap the [order](https://github.com/huggingface/datasets/blob/38c8c725f3996ff1ff03f6fd461aa6d645321034/src/datasets/data_files.py#L78-L79) of the patterns if `data_dir` is specified (or if `load_dataset(data_dir)` is called)? ",
"> @polinaeterna @lhoestq Perhaps one way to fix this would be to swap the [order](https://github.com/huggingface/datasets/blob/38c8c725f3996ff1ff03f6fd461aa6d645321034/src/datasets/data_files.py#L78-L79) of the patterns if data_dir is specified (or if load_dataset(data_dir) is called)?\r\n\r\nyes that makes sense !",
"Looks like the `val/validation` dir name issue is fixed with the current main-branch version of the `datasets` repository. \r\n\r\n> @polinaeterna @lhoestq Perhaps one way to fix this would be to swap the [order](https://github.com/huggingface/datasets/blob/38c8c725f3996ff1ff03f6fd461aa6d645321034/src/datasets/data_files.py#L78-L79) of the patterns if data_dir is specified (or if load_dataset(data_dir) is called)?\r\n\r\nI agree with this as well. I would expect higher precedence to the directory name over the file name. Right now if I place a single file named `train_00001.jpg` under the `validation` directory, `load_dataset` cannot find the validation split.",
"Thanks for the reply\r\n\r\nI've created a separate [issue](https://github.com/huggingface/datasets/issues/4982#issue-1375604693) for my problem.",
"> @polinaeterna @lhoestq Perhaps one way to fix this would be to swap the [order](https://github.com/huggingface/datasets/blob/38c8c725f3996ff1ff03f6fd461aa6d645321034/src/datasets/data_files.py#L78-L79) of the patterns if data_dir is specified (or if load_dataset(data_dir) is called)?\r\n\r\nSounds good to me! opened a PR: https://github.com/huggingface/datasets/pull/4985",
"Hi there @polinaeterna @mariosasko ! I have installed 5.2.3.dev0, which should have this fix. Unfortunately, I am still getting the error:\r\n`ValueError: Unknown split \"validation\". Should be one of ['train'].` When I call `load_dataset(\"csv\", data_files=files, split=split)`\r\n\r\nAny help would be greatly appreciated!",
"hi @shaneacton ! could you please show your dataset structure?",
"Hi there @polinaeterna . My local CSV files are stored as follows:\r\nbinding:\r\n---------- tune.csv\r\n---------- public_data:\r\n--------------------------- train.csv\r\n\r\n`self.list_shards(split)` sucessfully finds the relevant data files",
"@shaneacton do you have `validation.csv`/`val.csv`/`valid.csv`/`dev.csv` file in your data folder? I can't find it in the structure you provided",
"@polinaeterna no, does the name of the split need to match the name of the file exactly?\r\n\r\nBut my train file is not actually named 'train.py' its called 'XXXXXXXXX_train_XXXXXXXX.csv'\r\nAnd the code works fine for train, but fails for validation.\r\nDoes the file name need to _contain_ the split name?",
"@shaneacton what files do you expect to be included in \"validation\" split? yes, you should somehow indicate that a file belongs to a certain split - either by including split name in a filename or by putting it into a folder with split name, you can also check out [this documentation page](https://huggingface.co/docs/datasets/main/en/repository_structure) :)\r\nby default all the data goes to a single `train` split",
"@polinaeterna I have specified my train/test/tune files via the `split_to_filepattern` argument when initialising my `FileDataSource` class. This is how `list_shards` is able to find the right files.\r\nAfter your last message, I have tried renaminig my data files to simply `train.csv` and `validation.csv`, however I am still getting the same error: `Unknown split \"validation\". Should be one of ['train']`",
"@polinaeterna I have solved the issue. The solution was to call:\r\n`load_dataset(\"csv\", data_files={split: files}, split=split)`",
"For me it resolved by adding the verification_mode param:\r\n```\r\n imdb_ds = load_dataset(\r\n \"imdb\", verification_mode=\"no_checks\"\r\n )\r\n imdb_ds\r\n```"
] | ## Describe the bug
The `datasets.load_dataset` returns a `ValueError: Unknown split "validation". Should be one of ['train', 'test'].` when running `load_dataset(local_data_dir_path, split="validation")` even if the `validation` sub-directory exists in the local data path.
The data directories are as follows and attached to this issue:
```
test_data1
|_ train
|_ 1012.png
|_ metadata.jsonl
...
|_ test
...
|_ validation
|_ 234.png
|_ metadata.jsonl
...
test_data2
|_ train
|_ train_1012.png
|_ metadata.jsonl
...
|_ test
...
|_ validation
|_ val_234.png
|_ metadata.jsonl
...
```
They contain the same image files and `metadata.jsonl` but the images in `test_data2` have the split names prepended i.e.
`train_1012.png, val_234.png` and the images in `test_data1` do not have the split names prepended to the image names i.e. `1012.png, 234.png`
I actually saw in another issue `val` was not recognized as a split name but here I would expect the files to take the split from the parent directory name i.e. val should become part of the validation split?
## Steps to reproduce the bug
```python
import datasets
datasets.logging.set_verbosity_error()
from datasets import load_dataset, get_dataset_split_names
# the following only finds train, validation and test splits correctly
path = "./test_data1"
print("######################", get_dataset_split_names(path), "######################")
dataset_list = []
for spt in ["train", "test", "validation"]:
dataset = load_dataset(path, split=spt)
dataset_list.append(dataset)
# the following only finds train and test splits
path = "./test_data2"
print("######################", get_dataset_split_names(path), "######################")
dataset_list = []
for spt in ["train", "test", "validation"]:
dataset = load_dataset(path, split=spt)
dataset_list.append(dataset)
```
## Expected results
```
###################### ['train', 'test', 'validation'] ######################
###################### ['train', 'test', 'validation'] ######################
```
## Actual results
```
Traceback (most recent call last):
File "test_data_loader.py", line 11, in <module>
dataset = load_dataset(path, split=spt)
File "/home/venv/lib/python3.8/site-packages/datasets/load.py", line 1758, in load_dataset
ds = builder_instance.as_dataset(split=split, ignore_verifications=ignore_verifications, in_memory=keep_in_memory)
File "/home/venv/lib/python3.8/site-packages/datasets/builder.py", line 893, in as_dataset
datasets = map_nested(
File "/home/venv/lib/python3.8/site-packages/datasets/utils/py_utils.py", line 385, in map_nested
return function(data_struct)
File "/home/venv/lib/python3.8/site-packages/datasets/builder.py", line 924, in _build_single_dataset
ds = self._as_dataset(
File "/home/venv/lib/python3.8/site-packages/datasets/builder.py", line 993, in _as_dataset
dataset_kwargs = ArrowReader(self._cache_dir, self.info).read(
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 211, in read
files = self.get_file_instructions(name, instructions, split_infos)
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 184, in get_file_instructions
file_instructions = make_file_instructions(
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 107, in make_file_instructions
absolute_instructions = instruction.to_absolute(name2len)
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 616, in to_absolute
return [_rel_to_abs_instr(rel_instr, name2len) for rel_instr in self._relative_instructions]
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 616, in <listcomp>
return [_rel_to_abs_instr(rel_instr, name2len) for rel_instr in self._relative_instructions]
File "/home/venv/lib/python3.8/site-packages/datasets/arrow_reader.py", line 433, in _rel_to_abs_instr
raise ValueError(f'Unknown split "{split}". Should be one of {list(name2len)}.')
ValueError: Unknown split "validation". Should be one of ['train', 'test'].
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: Linux Ubuntu 18.04
- Python version: 3.8.12
- PyArrow version: 9.0.0
Data files
[test_data1.zip](https://github.com/huggingface/datasets/files/9424463/test_data1.zip)
[test_data2.zip](https://github.com/huggingface/datasets/files/9424468/test_data2.zip)
| 4,895 |
https://github.com/huggingface/datasets/issues/4893 | Oversampling strategy for iterable datasets in `interleave_datasets` | [
"Hi @lhoestq,\r\nI plunged into the code and it should be manageable for me to work on it!\r\n#take\r\n\r\nAlso, setting `d1`, `d2` and `d3` as you did raised a `SyntaxError: 'yield' inside list comprehension` for me, on Python 3.8.10.\r\nThe following snippet works for me though:\r\n```\r\nd1 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [0, 1, 2]])), {}))\r\nd2 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [10, 11, 12, 13]])), {}))\r\nd3 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [20, 21, 22, 23, 24]])), {}))\r\n```\r\n\r\n",
"Great @ylacombe thanks ! I'm assigning you this issue",
"Hi @ylacombe :) Is there anything I can do to help ? Feel free to ping me if you have any question :)",
"Hi @lhoestq,\r\n\r\nI actually have already wrote the code last time [on this commit](https://github.com/ylacombe/datasets/commit/84769db97facc78a33ec53f7b1b395951e1804df) but I still have to change the docs and write some tests though. I'm working on it.\r\n\r\nHowever, I still your advice on one matter. \r\nIn #4831, when using a `Dataset` list with probabilities, I had change the original behavior so that it stops as soon as one or all datasets are out of samples. By nature, this behavior can't be applied with an `IterableDataset` because one only knows an iterable dataset is out of sample when receiving a StopIteration error after calling the iterator once again. \r\nTo sum up, as it is right know, the behavior is not consistent with an `IterableDataset` list or a `Dataset` list, when using probabilities.\r\nTo be honest, I think that the current behavior with a `Dataset` list is desirable and avoid having too many samples, so I would recommand keeping that as it is, but I can understand the desire to have the same behavior for both classes. \r\nWhat do you think ? Please let me know if you need more details.\r\n\r\n\r\nEDIT:\r\nHere is an example:\r\n```\r\n>>> from tests.test_iterable_dataset import *\r\n>>> d1 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [0, 1, 2]])), {}))\r\n>>> d2 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [10, 11, 12, 13]])), {}))\r\n>>> d3 = IterableDataset(ExamplesIterable((lambda: (yield from [(i, {\"a\": i}) for i in [20, 21, 22, 23, 24]])), {}))\r\n>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)\r\n>>> [x[\"a\"] for x in dataset]\r\n[10, 0, 11, 1, 2, 20, 12, 13]\r\n>>> from tests.test_arrow_dataset import *\r\n>>> d1 = Dataset.from_dict({\"a\": [0, 1, 2]})\r\n>>> d2 = Dataset.from_dict({\"a\": [10, 11, 12]})\r\n>>> d3 = Dataset.from_dict({\"a\": [20, 21, 22]})\r\n>>> interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)[\"a\"]\r\n[10, 0, 11, 1, 2]\r\n[10, 0, 11, 1, 2]\r\n```\r\n ",
"Hi ! Awesome :) \r\n\r\nMaybe you can pre-load the next sample to know if the dataset is empty or not ?\r\nThis way it should be possible to have the same behavior for `IterableDataset`",
"Hi @ylacombe let us know if we can help with anything :)",
"Hi @lhoestq, I've finally made some advances in the matter. I've modified the `IterableDataset` behavior so that it aligns with the `Dataset` behavior as we have discussed. The documentation has been dealt with too. \r\nIt works as expected on my examples. However I'm having trouble figuring out how to test `interleave_datasets` on `test_iterable_datasets.py` as I have never worked with pytest. Could you help me on that or give me some indications? \r\n",
"Thanks @ylacombe :)\r\n\r\nUsing the `pytest` command, you can run all the functions in a python file that start with \"test_*\" and make sure they return not errors:\r\n```\r\npytest tests/test_iterable_dataset.py\r\n```\r\n\r\nIn our case it can be nice to define a `test_interleave_datasets_with_oversampling` function. This function can contain the code example that we mentioned earlier in this github issue to make sure it works as expected.",
"Resolved via #5036."
] | In https://github.com/huggingface/datasets/pull/4831 @ylacombe added an oversampling strategy for `interleave_datasets`. However right now it doesn't work for datasets loaded using `load_dataset(..., streaming=True)`, which are `IterableDataset` objects.
It would be nice to expand `interleave_datasets` for iterable datasets as well to support this oversampling strategy
```python
>>> from datasets.iterable_dataset import IterableDataset, ExamplesIterable
>>> d1 = IterableDataset(ExamplesIterable(lambda: [(yield i, {"a": i}) for i in [0, 1, 2]], {}))
>>> d2 = IterableDataset(ExamplesIterable(lambda: [(yield i, {"a": i}) for i in [10, 11, 12, 13]], {}))
>>> d3 = IterableDataset(ExamplesIterable(lambda: [(yield i, {"a": i}) for i in [20, 21, 22, 23, 24]], {}))
>>> dataset = interleave_datasets([d1, d2, d3]) # is supported
>>> [x["a"] for x in dataset]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted") # is not supported yet
>>> [x["a"] for x in dataset]
[0, 10, 20, 1, 11, 21, 2, 12, 22, 0, 13, 23, 1, 0, 24]
```
This can be implemented by adding the strategy to both `CyclingMultiSourcesExamplesIterable` and `RandomlyCyclingMultiSourcesExamplesIterable` used in `_interleave_iterable_datasets` in `iterable_dataset.py`
I would be happy to share some guidance if anyone would like to give it a shot :) | 4,893 |
https://github.com/huggingface/datasets/issues/4889 | torchaudio 11.0 yields different results than torchaudio 12.1 when loading MP3 | [
"Maybe we can just pass this along to torchaudio @lhoestq @albertvillanova ? It be great if you could investigate if the errors lies in datasets or in torchaudio.",
"torchaudio did a change in [0.12](https://github.com/pytorch/audio/releases/tag/v0.12.0) on MP3 decoding (which affects common voice):\r\n> MP3 decoding is now handled by FFmpeg in sox_io backend. (https://github.com/pytorch/audio/pull/2419, https://github.com/pytorch/audio/pull/2428)\r\n> - FFmpeg is now used as fallback in sox_io backend, and now MP3 decoding is handled by FFmpeg. To load MP3 audio with torchaudio.load, please install a compatible version of FFmpeg (Version 4 when using an official binary distribution).\r\n> - Note that, whereas the previous MP3 decoding scheme pads the output audio, the new scheme does not. As a consequence, the new version returns shorter audio tensors.",
"Do we have a solution for this now? Should we just upgrade to `torchaudio 0.12.0` then? ",
"`datasets` supports `torchaudio` 0.12 if you have an environment that supports reading MP3 with `torchaudio`, i.e. if you have `ffmpeg>=4`",
"Closing as we no longer use `torchaudio` for decoding."
] | ## Describe the bug
When loading Common Voice with torchaudio 0.11.0 the results are different to 0.12.1 which leads to problems in transformers see: https://github.com/huggingface/transformers/pull/18749
## Steps to reproduce the bug
If you run the following code once with `torchaudio==0.11.0+cu102` and `torchaudio==0.12.1+cu102` you can see that the tensors differ. This is a pretty big breaking change and makes some integration tests fail in Transformers.
```python
#!/usr/bin/env python3
from datasets import load_dataset
import datasets
import numpy as np
import torch
import torchaudio
print("torch vesion", torch.__version__)
print("torchaudio vesion", torchaudio.__version__)
save_audio = True
load_audios = False
if save_audio:
ds = load_dataset("common_voice", "en", split="train", streaming=True)
ds = ds.cast_column("audio", datasets.Audio(sampling_rate=16_000))
ds_iter = iter(ds)
sample = next(ds_iter)
np.save(f"audio_sample_{torch.__version__}", sample["audio"]["array"])
print(sample["audio"]["array"])
if load_audios:
array_torch_11 = np.load("/home/patrick/audio_sample_1.11.0+cu102.npy")
print("Array 11 Shape", array_torch_11.shape)
print("Array 11 abs sum", np.sum(np.abs(array_torch_11)))
array_torch_12 = np.load("/home/patrick/audio_sample_1.12.1+cu102.npy")
print("Array 12 Shape", array_torch_12.shape)
print("Array 12 abs sum", np.sum(np.abs(array_torch_12)))
```
Having saved the tensors the print output yields:
```
torch vesion 1.12.1+cu102
torchaudio vesion 0.12.1+cu102
Array 11 Shape (122880,)
Array 11 abs sum 1396.4988
Array 12 Shape (123264,)
Array 12 abs sum 1396.5193
```
## Expected results
torchaudio 11.0 and 12.1 should yield same results.
## Actual results
See above.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.1.1.dev0
- Platform: Linux-5.18.10-76051810-generic-x86_64-with-glibc2.34
- Python version: 3.9.7
- PyArrow version: 6.0.1
- Pandas version: 1.4.2
| 4,889 |
https://github.com/huggingface/datasets/issues/4888 | Dataset Viewer issue for subjqa | [
"It's a bug in the viewer, thanks for reporting it. We're hoping to update to a new version in the next few days which should fix it.",
"Fixed \r\n\r\nhttps://huggingface.co/datasets/subjqa\r\n\r\n<img width=\"1040\" alt=\"Capture d’écran 2022-09-08 à 10 23 26\" src=\"https://user-images.githubusercontent.com/1676121/189073210-2a57ff88-8bb1-44bd-851e-0e75473cea3f.png\">\r\n"
] | ### Link
https://huggingface.co/datasets/subjqa
### Description
Getting the following error for this dataset:
```
Status code: 500
Exception: Status500Error
Message: 2 or more items returned, instead of 1
```
Not sure what's causing it though 🤔
### Owner
Yes | 4,888 |
https://github.com/huggingface/datasets/issues/4886 | Loading huggan/CelebA-HQ throws pyarrow.lib.ArrowInvalid | [
"Hi! IIRC one of the files in this dataset is corrupted due to https://github.com/huggingface/datasets/pull/4081 (fixed now).\r\n\r\n@NielsRogge Could you please re-generate and re-push this dataset (or I can do it if you share the generation script)?",
"Could you put something in place to catch these problems? I'm seeing this on another dataset consistently too and I guess I can't fix it in code?",
"Hey,\r\n\r\nYes the notebook I used to upload this dataset can be found here: https://colab.research.google.com/drive/141LJCcM2XyqprPY83nIQ-Zk3BbxWeahq?usp=sharing.\r\n\r\nIf you have time to regenerate the dataset, would be great.",
"Sorry, maybe I wasn't clear enough that it's a different dataset `laion2B-multi-joined-translated-to-en`. I think there should be checks in the upload, tests on the server, or validation after download (hashes) to catch these problems.\r\n\r\nLots of bandwidth wasted otherwise! /cc @mariosasko",
"Yes @alexjc sorry was more a reply to @JeanKaddour.\r\n\r\nAnd indeed it'd be great to have additional checks to avoid these errors. ",
"cc @severo since such checks should probably be implemented on the datasets-server side.",
"Hi,\r\n\r\nIt seems the problem is still persist. I have encountered the exact same problem using just 2 line of code above. \r\n\r\nThe error code is as follows:\r\n\r\n```\r\n發生例外狀況: DatasetGenerationError\r\nAn error occurred while generating the dataset\r\npyarrow.lib.ArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\n File \"/code/ddpm_learn/train.py\", line 65, in <module>\r\n dataset = load_dataset(\"huggan/CelebA-HQ\", cache_dir=\"./CelebA-HQ\"\r\ndatasets.builder.DatasetGenerationError: An error occurred while generating the dataset\r\n```",
"Yes for the moment refer to the notebook linked above if you want to create a HF dataset yourself",
"Hi @NielsRogge ,\r\nI can help to push the dataset to the cloud. However, I cannot locate the situation so far. I wonder if \r\n1. the downloaded files so far has corruption s.t. the file cannot generate properly, or\r\n2. the downloaded files has no bug, the bug is caused by buggy upload program so that I can use what I have just downloaded to re-upload to cloud\r\n\r\nThank, \r\nAllan"
] | ## Describe the bug
Loading huggan/CelebA-HQ throws pyarrow.lib.ArrowInvalid
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('huggan/CelebA-HQ')
```
## Expected results
See https://colab.research.google.com/drive/141LJCcM2XyqprPY83nIQ-Zk3BbxWeahq?usp=sharing#scrollTo=N3ml_7f8kzDd
## Actual results
```
File "/home/jean/projects/cold_diffusion/celebA.py", line 4, in <module>
dataset = load_dataset('huggan/CelebA-HQ')
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/datasets/load.py", line 1793, in load_dataset
builder_instance.download_and_prepare(
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/datasets/builder.py", line 793, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/datasets/builder.py", line 1274, in _prepare_split
for key, table in logging.tqdm(
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/tqdm/std.py", line 1195, in __iter__
for obj in iterable:
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/datasets/packaged_modules/parquet/parquet.py", line 67, in _generate_tables
parquet_file = pq.ParquetFile(f)
File "/home/jean/miniconda3/envs/seq/lib/python3.10/site-packages/pyarrow/parquet/__init__.py", line 286, in __init__
self.reader.open(
File "pyarrow/_parquet.pyx", line 1227, in pyarrow._parquet.ParquetReader.open
File "pyarrow/error.pxi", line 100, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: datasets-2.4.1.dev0
- Platform: Ubuntu 18.04
- Python version: 3.10
- PyArrow version: pyarrow 9.0.0
| 4,886 |
https://github.com/huggingface/datasets/issues/4885 | Create dataset from list of dicts | [
"Hi @sanderland, thanks for your enhancement proposal.\r\n\r\nI agree with you that this would be useful.\r\n\r\nPlease note that under the hood, we use PyArrow tables as backend:\r\n- The implementation of `Dataset.from_dict` uses the PyArrow `Table.from_pydict`\r\n\r\nTherefore, I would suggest:\r\n- Implementing `Dataset.from_list` using the PyArrow `Table.from_pylist`\r\n\r\nWhat do you think?\r\nLet's see if other people have other suggestions...",
"Thanks for the quick and positive reply @albertvillanova! \r\n`from_list` seems sensible. Have opened a PR so we can discuss details there.",
"Resolved via #4890."
] | I often find myself with data from a variety of sources, and a list of dicts is very common among these.
However, converting this to a Dataset is a little awkward, requiring either
```Dataset.from_pandas(pd.DataFrame(formatted_training_data))```
Which can error out on some more exotic values as 2-d arrays for reasons that are not entirely clear
> ArrowInvalid: ('Can only convert 1-dimensional array values', 'Conversion failed for column labels with type object')
Alternatively:
```Dataset.from_dict({k: [s[k] for s in formatted_training_data] for k in formatted_training_data[0].keys()})```
Which works, but is a little ugly.
**Describe the solution you'd like**
Either `.from_dict` accepting a list of dicts, or a `.from_records` function accepting such.
I am happy to PR this, just wanted to check you are happy to accept this I haven't missed something obvious, and which of the solutions would be prefered.
| 4,885 |
https://github.com/huggingface/datasets/issues/4883 | With dataloader RSS memory consumed by HF datasets monotonically increases | [
"Are you sure there is a leak? How can I see it? You shared the script but not the output which you believe should indicate a leak.\r\n\r\nI modified your reproduction script to print only once per try as your original was printing too much info and you absolutely must add `gc.collect()` when doing any memory measurements, since python's GC is scheduled so you might be measuring the wrong thing. This gives us:\r\n\r\n```\r\nimport psutil\r\nimport os\r\nimport gc\r\nfrom transformers import BertTokenizer\r\nfrom datasets import load_dataset\r\nfrom torch.utils.data import DataLoader\r\n\r\nBATCH_SIZE = 32\r\nNUM_TRIES = 100\r\n\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\ndef transform(x):\r\n x.update(tokenizer(x[\"text\"], return_tensors=\"pt\", max_length=64, padding=\"max_length\", truncation=True))\r\n x.pop(\"text\")\r\n x.pop(\"label\")\r\n return x\r\ndataset = load_dataset(\"imdb\", split=\"train\")\r\ndataset.set_transform(transform)\r\ntrain_loader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)\r\n\r\nmem_before = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n\r\ncount = 0\r\nwhile count < NUM_TRIES:\r\n for idx, batch in enumerate(train_loader): pass\r\n gc.collect()\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(count, mem_after - mem_before)\r\n count += 1\r\n```\r\n\r\nNow running it:\r\n\r\n```\r\n$ python dl-leak.py \r\nReusing dataset imdb (/home/stas/.cache/huggingface/datasets/imdb/plain_text/1.0.0/2fdd8b9bcadd6e7055e742a706876ba43f19faee861df134affd7a3f60fc38a1)\r\n0 4.43359375\r\n1 4.4453125\r\n2 4.44921875\r\n3 4.44921875\r\n4 4.4609375\r\n5 4.46484375\r\n6 4.46484375\r\n7 4.46484375\r\n8 4.46484375\r\n9 4.46484375\r\n10 4.46484375\r\n11 4.46484375\r\n12 4.46484375\r\n13 4.46484375\r\n14 4.46484375\r\n15 4.46484375\r\n16 4.46484375\r\n```\r\n\r\nIt's normal that at the beginning there is a small growth in memory usage, but after 5 cycles it gets steady.",
"Unless of course you're referring the memory growth during the first try. Is that what you're referring to? And since your ds is small it's hard to see the growth - could it be just because some records are longer and it needs to allocate more memory for those?\r\n\r\nThough while experimenting with this I have observed a peculiar thing, if I concatenate 2 datasets, I don't see any growth at all. But that's probably because the program allocated additional peak RSS memory to concatenate and then is re-using the memory\r\n\r\nI basically tried to see if I make the dataset much longer, I'd expect not to see any memory growth once the 780 records of the imdb ds have been processed once.",
"It is hard to say if it is directly reproducible in this setup. Maybe it is specific to the images stored in the CM4 case which cause a memory leak. I am still running your script and seeing if I can reproduce that particular leak in this case.",
"I was able to reproduce the leak with:\r\n\r\n```\r\nimport psutil\r\nimport os\r\nimport gc\r\nfrom datasets import load_from_disk\r\nimport time\r\n\r\nDATASET_PATH = \"/hf/m4-master/data/cm4/cm4-10000-v0.1\"\r\n\r\ndataset = load_from_disk(DATASET_PATH)\r\n\r\n# truncate to a tiny dataset\r\ndataset = dataset.select(range(1000))\r\n\r\nprint(f\"dataset: {len(dataset)} records\")\r\n\r\nmem_before = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\nfor idx, rec in enumerate(dataset):\r\n if idx % 100 == 0:\r\n gc.collect()\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB\")\r\n```\r\nYou need to adjust the DATASET_PATH record.\r\n\r\nwhich you get from\r\n\r\n```\r\ngsutil -m cp \"gs://hf-science-m4/cm4/cm4-10000-v0.1/dataset.arrow\" \"gs://hf-science-m4/cm4/cm4-10000-v0.1/dataset_info.json\" \"gs://hf-science-m4/cm4/cm4-10000-v0.1/state.json\" .\r\n```\r\n(I assume the hf folks have the perms) - it's a smallish dataset (10k)\r\n\r\nthen you run:\r\n```\r\n$ python ds.py\r\ndataset: 1000 records\r\n 0 1.0156MB\r\n 100 126.3906MB\r\n 200 142.8906MB\r\n 300 168.5586MB\r\n 400 218.3867MB\r\n 500 230.7070MB\r\n 600 238.9570MB\r\n 700 263.3789MB\r\n 800 288.1289MB\r\n 900 300.5039MB\r\n```\r\n\r\nyou should be able to see the leak ",
"This issue has nothing to do with `PIL`'s decoder. I removed it and the problem is still there.\r\n\r\nI then traced this leak to this single call: `pa_table.to_pydict()` here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/08a7b389cdd6fb49264a72aa8ccfc49a233494b6/src/datasets/formatting/formatting.py#L138-L140\r\n\r\nI can make it leak much faster by modifying that code to repeat `pa_table.to_pydict()` many times in a row. It shouldn't have that impact:\r\n\r\n```\r\nclass PythonArrowExtractor(BaseArrowExtractor[dict, list, dict]):\r\n def extract_row(self, pa_table: pa.Table) -> dict:\r\n x = [pa_table.to_pydict() for x in range(200)]\r\n return _unnest(pa_table.to_pydict())\r\n```\r\n\r\n@lhoestq - do you know what might be happening inside `pa_table.to_pydict()`, as this is in the `pyarrow` domain. Perhaps you know someone to tag from that project?\r\n\r\nProbably next need to remove `datasets` from the equation and make a reproducible case with just `pyarrow` directly.\r\n\r\nThe problem already happens with `pyarrow==6.0.0` or later (minimum for current `datasets`)\r\n\r\nI'm also trying to dig in with `objgraph` to see if there are any circular references which prevent objects from being freed, but no luck there so far. And I'm pretty sure `to_pydict` is not a python code, so the problem is likely to happen somewhere outside of python's GC.",
"This appears to be the same issue I think: https://github.com/huggingface/datasets/issues/4528\r\nI dug into the repro code there and it's the same behavior with the same leak, but it's a pure nlp dataset and thus much faster to work with. \r\n",
"I went all the way back to `pyarrow==1.0.0` and `datasets==1.12.0` and the problem is still there. How is it even possible that it wasn't noticed all this time. \r\n\r\nCould it be that the leak is in some 3rd party component `pyarrow` relies on? as while downgrading I have only downgraded the above 2 packages.\r\n",
"Also found this warning \r\n\r\n> Be careful: if you don't pass the ArrowArray struct to a consumer,\r\n> array memory will leak. This is a low-level function intended for\r\n> expert users.\r\n\r\nsee: https://github.com/apache/arrow/blob/99b57e84277f24e8ec1ddadbb11ef8b4f43c8c89/python/pyarrow/table.pxi#L2515-L2517\r\n\r\nperhaps something triggers this condition?\r\n\r\nI have no idea if it's related - this is just something that came up during my research.",
"Does it crash with OOM at some point? If it doesn't, it isn't a leak, just agressive caching or a custom allocator that doesn't like to give memory back (not uncommon). #4528 looks like it hits a steady state.\r\n\r\nI believe the underlying arrow libs use a custom C allocator. Some of those are designed not to give back to OS, but keep heap memory for themselves to re-use (hitting up the OS involves more expensive mutex locks, contention, etc). The greedy behaviour can be undesirable though. There are likely flags to change the allocator behaviour, and one could likely build without any custom allocators (or use a different one).",
"> Does it crash with OOM at some point?\r\n\r\nIn the original setup where we noticed this problem, it was indeed ending in an OOM",
"> https://github.com/huggingface/datasets/issues/4528 looks like it hits a steady state.\r\n\r\n@rwightman in the plot I shared, the steady state comes from the `time.sleep(100)` I added in the end of the script, to showcase that even the garbage collector couldn't free that allocated memory.\r\n",
"Could this be related to this discussion about a potential memory leak in pyarrow: https://issues.apache.org/jira/browse/ARROW-11007 ?\r\n\r\n(Note: I've tried `import pyarrow; pyarrow.jemalloc_set_decay_ms(0)` and the memory leak is still happening on your toy example)",
"> @lhoestq - do you know what might be happening inside pa_table.to_pydict(), as this is in the pyarrow domain. Perhaps you know someone to tag from that project?\r\n\r\n`to_pydict` calls `to_pylist` on each column (i.e. on each PyArrow Array). Then it iterates on the array and calls `as_py` on each element. The `as_py` implementation depends on the data type. For strings I think it simply gets the buffer that contains the binary string data that is defined in C++\r\n\r\nThe Arrow team is pretty responsive at [email protected] if it can help\r\n\r\n> Probably next need to remove datasets from the equation and make a reproducible case with just pyarrow directly.\r\n\r\nThat would be ideal indeed. Would be happy to help on this, can you give me access to the bucket so I can try with your data ?",
"> That would be ideal indeed. Would be happy to help on this, can you give me access to the bucket so I can try with your data ?\r\n\r\nI added you to the bucket @lhoestq ",
"It looks like an issue with memory mapping:\r\n- the amount of memory used in the end corresponds to the size of the dataset\r\n- setting `keep_in_memory=True` in `load_from_disk` loads the dataset in RAM, and **doesn't cause any memory leak**",
"Here is a code to reproduce this issue using only PyArrow and a dummy arrow file:\r\n```python\r\nimport psutil\r\nimport os\r\nimport gc\r\nimport pyarrow as pa\r\nimport time\r\n\r\nARROW_PATH = \"tmp.arrow\"\r\n\r\nif not os.path.exists(ARROW_PATH):\r\n arr = pa.array([b\"a\" * (200 * 1024)] * 1000) # ~200MB\r\n table = pa.table({\"a\": arr})\r\n\r\n with open(ARROW_PATH, \"wb\") as f:\r\n writer = pa.RecordBatchStreamWriter(f, schema=table.schema)\r\n writer.write_table(table)\r\n writer.close()\r\n\r\n\r\ndef memory_mapped_arrow_table_from_file(filename: str) -> pa.Table:\r\n memory_mapped_stream = pa.memory_map(filename)\r\n opened_stream = pa.ipc.open_stream(memory_mapped_stream)\r\n pa_table = opened_stream.read_all()\r\n return pa_table\r\n\r\n\r\ntable = memory_mapped_arrow_table_from_file(ARROW_PATH)\r\narr = table[0]\r\n\r\nmem_before = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\nfor idx, x in enumerate(arr):\r\n if idx % 100 == 0:\r\n gc.collect()\r\n time.sleep(0.1)\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB\")\r\n```\r\nprints\r\n```\r\n 0 0.2500MB\r\n 100 19.8008MB\r\n 200 39.3320MB\r\n 300 58.8633MB\r\n 400 78.3945MB\r\n 500 97.9258MB\r\n 600 117.4570MB\r\n 700 136.9883MB\r\n 800 156.5195MB\r\n 900 176.0508MB\r\n```\r\nNote that this example simply iterates over the `pyarrow.lib.BinaryScalar` objects in the array. Running `.as_py()` is not needed to experience the memory issue.",
"@lhoestq that does indeed increase in memory, but if you iterate over array again after the first time, or re-open and remap the same file (repeat `table = memory_mapped_arrow_table_from_file(ARROW_PATH)`) before re-iterating, it doesn't move pas 195MB.... it would appear another step is needed to continue consuming memory past that.. hmmm\r\n\r\nAre the pa_tables held on to anywhere after they are iterated in the real code?\r\n\r\nin my hack, if you do a bunch cut & paste and then change the arr name for each iter \r\n\r\n```\r\ntable = memory_mapped_arrow_table_from_file(ARROW_PATH)\r\narr = table[0]\r\n\r\nfor idx, x in enumerate(arr):\r\n if idx % 100 == 0:\r\n gc.collect()\r\n time.sleep(0.1)\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB\")\r\n\r\ntable = memory_mapped_arrow_table_from_file(ARROW_PATH)\r\narr1 = table[0]\r\n\r\nfor idx, x in enumerate(arr1):\r\n if idx % 100 == 0:\r\n gc.collect()\r\n time.sleep(0.1)\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB\")\r\n\r\ntable = memory_mapped_arrow_table_from_file(ARROW_PATH)\r\narr2 = table[0]\r\n\r\nfor idx, x in enumerate(arr2):\r\n if idx % 100 == 0:\r\n gc.collect()\r\n time.sleep(0.1)\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB\")\r\n```\r\n\r\nit leaks, if all arr are the same name (so prev one gets cleaned up) it does not and goes back to 0, anything that could be holding onto a reference of an intermediary equivalent like arr in the real use case?\r\n\r\n\r\n\r\n",
"Yes, we have already established here https://github.com/huggingface/datasets/issues/4883#issuecomment-1232063891 that when one iterates over the whole dataset multiple times, it consumes a bit more memory in the next few repetitions and then remains steady. \r\n\r\nWhich means that when a new iterator is created over the same dataset, all the memory from the previous iterator is re-used.\r\n\r\nSo the leak happens primarily when the iterator is \"drained\" the first time. which tells me that either a circular reference is created somewhere which only gets released when the iterator is destroyed, or there is some global variable that keeps piling up the memory and doesn't release it in time.\r\n\r\nAlso I noticed some `__del__` methods which won't destroy objects automatically and there is usually a warning against using it https://stackoverflow.com/a/1481512/9201239\r\n\r\nThere are also some `weakref`s in the code which too may lead to leaks or weird problems at times.\r\n",
"@stas00 my point was, I'm not convinced @lhoestq last example illustrates the leak, but rather the differences between memory mapping and in memory usage patterns. If you destroy arr, memory map impl goes back to 0 each iteration. The amount of memory that 'looks' like it is leaked in first pass differes quite a bit between memory mapped vs in memory, but the underlying issue likely a circular reference, or reference(s) which were not cleaned up that would impact either case, but likely much more visible with mmap.",
"Thank you for clarifying, Ross. \r\n\r\nI think we agree that it's almost certain that the `datasets` iterator traps some inner variable that prevents object freeing, since if we create the iterator multiple times (and drain it) after a few runs no new memory is allocated. We could try to dig in more with `objgraph` - my main concern is if the problem happens somewhere outside of python, (i.e. in pyarrow cpp implementation) in which case it'd be much more difficult to trace. \r\n\r\nI wish there was a way on linux to tell the program to free no longer used memory at will.",
"FWIW, I revisted some code I had in the works to use HF datasets w/ timm train & val scripts. There is no leak there across multipe epochs. It uses the defaults. \r\n\r\nIt's worth noting that with imagenet `keep_in_memory=True` isn't even an option because the train arrow file is ~140GB and my local memory is less. The virtual address space reflects mmap (> 150GB) and doesn't increase over epochs that I noticed. I have some perf issues to bring up wrt to the current setup, but that's a separate and lower prio discussion to have elsewhere...",
"# Notes \r\n\r\nAfter reading many issues and trying many things here is the summary of my learning\r\n\r\nI'm now using @lhoestq repro case as it's pyarrow-isolated: https://github.com/huggingface/datasets/issues/4883#issuecomment-1242034985\r\n\r\n\r\n## 1. pyarrow memory backends\r\n\r\nit has 3 backends, I tried them all with the same results\r\n\r\n```\r\npa.set_memory_pool(pa.jemalloc_memory_pool())\r\npa.set_memory_pool(pa.mimalloc_memory_pool())\r\npa.set_memory_pool(pa.system_memory_pool())\r\n```\r\n\r\n## 2. quick release\r\n\r\nThe `jemalloc` backend supports quick release\r\n\r\n```\r\npa.jemalloc_set_decay_ms(0)\r\n```\r\n\r\nit doesn't make any difference in this case\r\n\r\n## 3. actual memory allocations\r\n\r\nthis is a useful tracer for PA memory allocators\r\n```\r\npa.log_memory_allocations(enable=True)\r\n```\r\n\r\nit nicely reports memory allocations and releases when the arrow file is created the first time.\r\n\r\nbut when we then try to do `enumerate(arr)` this logger reports 0 allocations.\r\n\r\nThis summary also reports no allocations when the script run the second time (post file creation):\r\n```\r\nmem_pool = pa.default_memory_pool()\r\nprint(f\"PyArrow mem pool info: {mem_pool.backend_name} backend, {mem_pool.bytes_allocated()} allocated, \"\r\n f\"{mem_pool.max_memory()} max allocated, \")\r\n\r\nprint(f\"PyArrow total allocated bytes: {pa.total_allocated_bytes()}\")\r\n```\r\n\r\nHowever it's easy to see by using `tracemalloc` which only measures python's memory allocations that it's PA that leaks, since `tracemalloc` shows fixed memory\r\n\r\n(this is bolted on top of the original repro script)\r\n\r\n```\r\nimport tracemalloc\r\ntracemalloc.start()\r\n\r\n[...]\r\nfor idx, x in enumerate(arr):\r\n if idx % 10 == 0:\r\n gc.collect()\r\n time.sleep(0.1)\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\n mem_use = pa.total_allocated_bytes() - start_use\r\n mem_peak = pool.max_memory() - start_peak_use\r\n\r\n second_size, second_peak = tracemalloc.get_traced_memory()\r\n mem_diff = (second_size - first_size) / 2**20\r\n mem_peak_diff = (second_peak - first_peak) / 2**20\r\n\r\n # pa.jemalloc_memory_pool().release_unused()\r\n # pa.mimalloc_memory_pool().release_unused()\r\n # pa.system_memory_pool().release_unused()\r\n\r\n print(f\"{idx:4d} {mem_after - mem_before:12.4f}MB {mem_diff:12.4f} {mem_peak_diff:12.4f} {memory_mapped_stream.size()/2**20:4.4}MB {mem_use/2**20:4.4}MB {mem_peak/2**20:4.4}MB\")\r\n\r\n```\r\n\r\ngives:\r\n\r\n```\r\n 0 5.4258MB 0.0110 0.0201 195.3MB 0.0MB 0.0MB\r\n 10 25.3672MB 0.0112 0.0202 195.3MB 0.0MB 0.0MB\r\n 20 45.9336MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 30 62.4336MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 40 83.0586MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 50 103.6836MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 60 124.3086MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 70 140.8086MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 80 161.4336MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n 90 182.0586MB 0.0112 0.0203 195.3MB 0.0MB 0.0MB\r\n```\r\n\r\nthe 3rd and 4th columns are `tracemalloc`'s report.\r\n\r\nthe 5th column is the size of mmaped stream - fixed.\r\n\r\nthe last 2 are the PA's malloc reports - you can see it's totally fixed and 0.\r\n\r\nSo what gives? PA's memory allocator says nothing was allocated and we can see python doesn't allocate any memory either.\r\n\r\nAs someone suggested in one of the PA issues that **IPC/GRPC could be the issue.** Any suggestions on how debug this one?\r\n\r\nThe main issue is that one can't step through with a python debugger as `arr` is an opaque cpp object binded to python.\r\n\r\nPlease see the next comment for a possible answer.\r\n\r\n# ref-count\r\n\r\nI also traced reference counts and they are all fixed using either `sys.getrefcount(x)` or `len(gc.get_referrers(x))`\r\n\r\nso it's not the python object\r\n\r\n# Important related discussions\r\n\r\nhttps://issues.apache.org/jira/browse/ARROW-11007 - looks very similar to our issue\r\nin particular this part of the report:\r\nhttps://issues.apache.org/jira/browse/ARROW-11007?focusedCommentId=17279642&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17279642\r\n",
"# There is no leak, just badly communicated linux RSS memory usage stats\r\n\r\nNext, lets revisit @rwightman's suggestion that there is actually no leak.\r\n\r\nAfter all - we are using mmap which **will try to map** the file to RAM as much as it can and then page out if there is no memory. i.e. MMAP is only fast if you have a lot of CPU RAM.\r\n\r\nSo let's do it:\r\n\r\n# Memory mapping OOM test\r\n\r\nWe first quickly start a cgroups-controlled shell which will instantly kill any program that consumes more than 1GB of memory:\r\n\r\n```\r\n$ systemd-run --user --scope -p MemoryHigh=1G -p MemoryMax=1G -p MemorySwapMax=1G --setenv=\"MEMLIMIT=1GB\" bash\r\n```\r\n\r\nLet's check that it indeed does so. Let's change @lhoestq's script to allocate a 10GB arrow file:\r\n\r\n```\r\n$ python -c 'import pyarrow as pa; pa.array([b\"a\" * (2000 * 1024)] * 5000)'\r\nKilled\r\n```\r\noops, that didn't work, as we tried to allocate 10GB when only 1GB is allowed. This is what we want!\r\n\r\nLet's do a sanity check - can we allocate 0.1GB?\r\n```\r\npython -c 'import pyarrow as pa; pa.array([b\"a\" * (2000 * 1024)] * 50)'\r\n```\r\n\r\nYes. So the limited shell does the right thing. It let's allocate `< 1GB` of RSS RAM.\r\n\r\nNext let's go back to @lhoestq's script but with 10GB arrow file.\r\n\r\nwe change his repro script https://github.com/huggingface/datasets/issues/4883#issuecomment-1242034985 to 50x larger file\r\n```\r\n arr = pa.array([b\"a\" * (2000 * 1024)] * 5000) # ~10000MB\r\n```\r\nwe first have to run into a normal unlimited shell so that we don't get killed (as the script allocates 10GB)\r\n\r\nlet's run the script now in the 1GB-limited shell while running a monitor:\r\n\r\n```\r\n$ htop -F python -s M_RESIDENT -u `whoami`\r\n```\r\n\r\nso we have 2 sources of RSS info just in case.\r\n\r\n```\r\n$ python pyar.py\r\n 0 4.3516MB 0.0103 0.0194 9.766e+03MB 0.0MB 0.0MB\r\n 10 24.3008MB 0.0104 0.0195 9.766e+03MB 0.0MB 0.0MB\r\n[...]\r\n4980 9730.3672MB 0.0108 0.0199 9.766e+03MB 0.0MB 0.0MB\r\n4990 9750.9922MB 0.0108 0.0199 9.766e+03MB 0.0MB 0.0MB\r\nPyArrow mem pool info: jemalloc backend, 0 allocated, 0 max allocated,\r\nPyArrow total allocated bytes: 0\r\n```\r\n\r\nBut wait, it reported 10GB RSS both in `htop` and in our log!\r\n\r\nSo that means it never allocated 10GB otherwise it'd have been killed.\r\n\r\n**Which tells us that there is no leak whatsoever** and this is just a really difficult situation where MMAPPED memory is reported as part of RSS which it probably shouldn't. As now we have no way how to measure real memory usage.\r\n\r\nI also attached the script with all the different things I have tried in it, so it should be easy to turn them on/off if you want to reproduce any of my findings.\r\n\r\n[pyar.txt](https://github.com/huggingface/datasets/files/9539430/pyar.txt)\r\n\r\njust rename it to `pyra.py` as gh doesn't let attaching scripts...\r\n\r\n(I have to remember to exit that special mem-limited shell or else I won't be able to do anything serious there.)\r\n\r\n",
"The original leak in the multi-modal code is very likely something else. But of course now it'd be very difficult to trace it using mmap.\r\n\r\nI think to debug we have to set `keep_in_memory=True` in `load_from_disk` to load the small dataset in RAM, so there will be no mmap misleading reporting component and then continue searching for another source of a leak.",
"To add to what @stas00 found, I'm gonna leave some links to where I believe the confusion came from in pyarrow's APIs, for future reference:\r\n* In the section where they talk about [efficiently writing and reading arrow data](https://arrow.apache.org/docs/dev/python/ipc.html?#efficiently-writing-and-reading-arrow-data), they give an example of how \r\n\r\n> Arrow can directly reference the data mapped from disk and avoid having to allocate its own memory. \r\n\r\nAnd where their example shows 0 RSS memory allocation, the way we used to measure RSS shows 39.6719MB allocated. Here's the script to reproduce:\r\n```\r\nimport psutil\r\nimport os\r\nimport gc\r\nimport pyarrow as pa\r\nimport time\r\nimport sys\r\n\r\n# gc.set_debug(gc.DEBUG_LEAK)\r\n# gc.set_threshold(0,0,0)\r\n\r\n#pa.set_memory_pool(pa.mimalloc_memory_pool())\r\n#pa.set_memory_pool(pa.system_memory_pool())\r\n\r\nimport tracemalloc\r\n\r\n#pa.jemalloc_set_decay_ms(0)\r\n# pa.log_memory_allocations(enable=True)\r\n\r\nBATCH_SIZE = 10000\r\nNUM_BATCHES = 1000\r\nschema = pa.schema([pa.field('nums', pa.int32())])\r\nwith pa.OSFile('bigfile.arrow', 'wb') as sink:\r\n with pa.ipc.new_file(sink, schema) as writer:\r\n for row in range(NUM_BATCHES):\r\n batch = pa.record_batch([pa.array(range(BATCH_SIZE), type=pa.int32())], schema)\r\n writer.write(batch)\r\n\r\nstart_use = pa.total_allocated_bytes()\r\npool = pa.default_memory_pool()\r\nstart_peak_use = pool.max_memory()\r\ntracemalloc.start()\r\nfirst_size, first_peak = tracemalloc.get_traced_memory()\r\nmem_before = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\n\r\n\r\n# with pa.OSFile('bigfile.arrow', 'rb') as source:\r\n# loaded_array = pa.ipc.open_file(source).read_all()\r\n\r\nwith pa.memory_map('bigfile.arrow', 'rb') as source:\r\n loaded_array = pa.ipc.open_file(source).read_all()\r\n\r\n\r\nprint(\"LEN:\", len(loaded_array))\r\nprint(\"RSS: {}MB\".format(pa.total_allocated_bytes() >> 20))\r\n\r\ngc.collect()\r\ntime.sleep(0.1)\r\nmem_after = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\nmem_use = pa.total_allocated_bytes() - start_use\r\nmem_peak = pool.max_memory() - start_peak_use\r\nsecond_size, second_peak = tracemalloc.get_traced_memory()\r\nmem_diff = (second_size - first_size) / 2**20\r\nmem_peak_diff = (second_peak - first_peak) / 2**20\r\n\r\nidx = 0\r\nprint(f\"{idx:4d} {mem_after - mem_before:12.4f}MB {mem_diff:12.4f} {mem_peak_diff:12.4f} {mem_use/2**20:4.4}MB {mem_peak/2**20:4.4}MB\")\r\n```\r\ngives:\r\n```\r\n\r\nLEN: 10000000\r\nRSS: 0MB\r\n 0 39.6719MB 0.0132 0.0529 0.0MB 0.0MB\r\n```\r\nWhich again just proves that we uncorrectly measure RSS, in the case of MMAPPED memory\r\n\r\n\r\n* [The recommended way to do memory profiling from Arrow's docs](https://arrow.apache.org/docs/dev/cpp/memory.html#memory-profiling)\r\n",
"@lhoestq, I have been working on a detailed article that shows that MMAP doesn't leak and it's mostly ready. I will share when it's ready.\r\n\r\nThe issue is that we still need to be able to debug memory leaks by turning MMAP off.\r\n\r\nBut, once I tried to show the user that using `load_dataset(... keep_in_memory=True)` is the way to debug an actual memory leak - guess I what I discovered? A potential actual leak.\r\n\r\nHere is the repro:\r\n\r\n```\r\n$ cat ds-mmap.py\r\nfrom datasets import load_dataset\r\nimport gc\r\nimport os\r\nimport psutil\r\n\r\nproc = psutil.Process(os.getpid())\r\ndef mem_read():\r\n gc.collect()\r\n return proc.memory_info().rss / 2**20\r\n\r\ndataset = load_dataset(\"wmt19\", 'cs-en', keep_in_memory=True, streaming=False)['train']\r\n\r\nprint(f\"{'idx':>6} {'RSS':>10} {'Δ RSS':>15}\")\r\nstep = 20000\r\nfor i in range(0, 10*step, step):\r\n mem_before = mem_read()\r\n _ = dataset[i:i+step]\r\n mem_after = mem_read()\r\n print(f\"{i:6d} {mem_after:12.4f}MB {mem_after - mem_before:12.4f}MB \")\r\n```\r\n\r\n```\r\npython ds-io.py\r\nReusing dataset wmt19 (/home/stas/.cache/huggingface/datasets/wmt19/cs-en/1.0.0/c3db1bf4240362ed1ef4673b354f468d70aac66d4e67d45f536d493a0840f0d3)\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 5.66it/s]\r\n idx RSS Δ RSS\r\n 0 1398.4609MB 3.5195MB\r\n 20000 1398.5742MB 0.1133MB\r\n 40000 1398.6016MB 0.0273MB\r\n 60000 1398.6016MB 0.0000MB\r\n 80000 1398.6016MB 0.0000MB\r\n100000 1398.6328MB 0.0312MB\r\n120000 1398.6953MB 0.0625MB\r\n140000 1398.6953MB 0.0000MB\r\n160000 1398.7500MB 0.0547MB\r\n180000 1398.7500MB 0.0000MB\r\n```",
"as I suggested on slack perhaps it was due to dataset records length variation, so with your help I wrote another repro with synthetic records which are all identical - which should remove my hypothese from the equation and we should expect 0 incremental growth as we iterate over the datasets. But alas this is not the case. There is a tiny but definite leak-like behavior.\r\n\r\nHere is the new repro:\r\n\r\n```\r\n$ cat ds-synthetic-no-mmap.py\r\nfrom datasets import load_from_disk, Dataset\r\nimport gc\r\nimport sys\r\nimport os\r\nimport psutil\r\n\r\nproc = psutil.Process(os.getpid())\r\ndef mem_read():\r\n gc.collect()\r\n return proc.memory_info().rss / 2**20\r\n\r\nDS_PATH = \"synthetic-ds\"\r\nif not os.path.exists(DS_PATH):\r\n records = 1_000_000\r\n print(\"Creating a synthetic dataset\")\r\n row = dict(foo=[dict(a='a'*500, b='b'*1000)])\r\n ds = Dataset.from_dict({k: [v] * records for k, v in row.items()})\r\n ds.save_to_disk(DS_PATH)\r\n print(\"Done. Please restart the program\")\r\n sys.exit()\r\n\r\ndataset = load_from_disk(DS_PATH, keep_in_memory=True)\r\nprint(f\"Dataset len={len(dataset)}\")\r\n\r\nprint(f\"{'idx':>8} {'RSS':>10} {'Δ RSS':>15}\")\r\nmem_start = 0\r\nstep = 25_000\r\nwarmup_iterations = 4\r\nfor idx, i in enumerate(range(0, len(dataset), step)):\r\n if idx == warmup_iterations: # skip the first few iterations while things get set up\r\n mem_start = mem_read()\r\n mem_before = mem_read()\r\n _ = dataset[i:i+step]\r\n mem_after = mem_read()\r\n print(f\"{i:8d} {mem_after:12.4f}MB {mem_after - mem_before:12.4f}MB\")\r\nmem_end = mem_read()\r\n\r\nprint(f\"Total diff: {mem_end - mem_start:12.4f}MB (after {warmup_iterations} warmup iterations)\")\r\n```\r\n\r\nand the run:\r\n\r\n```\r\n$ python ds-synthetic-no-mmap.py\r\nDataset len=1000000\r\n idx RSS Δ RSS\r\n 0 1601.9258MB 47.9688MB\r\n 25000 1641.6289MB 39.7031MB\r\n 50000 1641.8594MB 0.2305MB\r\n 75000 1642.1289MB 0.2695MB\r\n 100000 1642.1289MB 0.0000MB\r\n 125000 1642.3789MB 0.2500MB\r\n 150000 1642.3789MB 0.0000MB\r\n 175000 1642.6289MB 0.2500MB\r\n 200000 1642.6289MB 0.0000MB\r\n 225000 1642.8789MB 0.2500MB\r\n 250000 1642.8828MB 0.0039MB\r\n 275000 1643.1328MB 0.2500MB\r\n 300000 1643.1328MB 0.0000MB\r\n 325000 1643.3828MB 0.2500MB\r\n 350000 1643.3828MB 0.0000MB\r\n 375000 1643.6328MB 0.2500MB\r\n 400000 1643.6328MB 0.0000MB\r\n 425000 1643.8828MB 0.2500MB\r\n 450000 1643.8828MB 0.0000MB\r\n 475000 1644.1328MB 0.2500MB\r\n 500000 1644.1328MB 0.0000MB\r\n 525000 1644.3828MB 0.2500MB\r\n 550000 1644.3828MB 0.0000MB\r\n 575000 1644.6328MB 0.2500MB\r\n 600000 1644.6328MB 0.0000MB\r\n 625000 1644.8828MB 0.2500MB\r\n 650000 1644.8828MB 0.0000MB\r\n 675000 1645.1328MB 0.2500MB\r\n 700000 1645.1328MB 0.0000MB\r\n 725000 1645.3828MB 0.2500MB\r\n 750000 1645.3828MB 0.0000MB\r\n 775000 1645.6328MB 0.2500MB\r\n 800000 1645.6328MB 0.0000MB\r\n 825000 1645.8828MB 0.2500MB\r\n 850000 1645.8828MB 0.0000MB\r\n 875000 1646.1328MB 0.2500MB\r\n 900000 1646.1328MB 0.0000MB\r\n 925000 1646.3828MB 0.2500MB\r\n 950000 1646.3828MB 0.0000MB\r\n 975000 1646.6328MB 0.2500MB\r\nTotal diff: 4.5039MB (after 4 warmup iterations)\r\n```\r\nso I'm still not sure why we get this.\r\n\r\nAs you can see I started skipping the first few iterations where memory isn't stable yet. As the actual diff is much larger if we count all iterations.\r\n\r\nWhat do you think?",
"@stas00 my 2 cents from having looked at a LOT of memory leaks over the years, esp in Python, .3% memory increase over that many iterations of something is difficult to say with certainty it is a leak. \r\n\r\nAlso, just looking at RSS makes it hard to analyze leaks. RSS can stay near constant while you are leaking. RSS is paged in mem, if you have a big leak your RSS might not increase much (leaked mem tends not to get used again so often paged out) while your virtual page allocation could be going through the roof...",
"yes, that's true, but unless the leak is big, I'm yet to find another measurement tool.\r\n\r\nTo prove your point here is a very simple IO in a loop program that also reads the same line all over again:\r\n\r\n```\r\n$ cat mmap-no-leak-debug.py\r\nimport gc\r\nimport mmap\r\nimport os\r\nimport psutil\r\nimport sys\r\n\r\nproc = psutil.Process(os.getpid())\r\n\r\nPATH = \"./tmp.txt\"\r\n\r\ndef mem_read():\r\n gc.collect()\r\n return proc.memory_info().rss / 2**20\r\n\r\n# create a large data file with a few long lines\r\nif not os.path.exists(PATH):\r\n with open(PATH, \"w\") as fh:\r\n s = 'a'* 2**27 + \"\\n\" # 128MB\r\n # write ~2GB file\r\n for i in range(16):\r\n fh.write(s)\r\n\r\nprint(f\"{'idx':>4} {'RSS':>10} {'Δ RSS':>12} {'Δ accumulated':>10}\")\r\n\r\ntotal_read = 0\r\ncontent = ''\r\nmem_after = mem_before_acc = mem_after_acc = mem_before = proc.memory_info().rss / 2**20\r\nprint(f\"{0:4d} {mem_after:10.2f}MB {mem_after - 0:10.2f}MB {0:10.2f}MB\")\r\n\r\nmmap_mode = True if \"--mmap\" in sys.argv else False\r\n\r\nwith open(PATH, \"r\") as fh:\r\n\r\n if mmap_mode:\r\n mm = mmap.mmap(fh.fileno(), 0, access=mmap.ACCESS_READ)\r\n\r\n idx = 0\r\n while True:\r\n idx += 1\r\n mem_before = mem_read()\r\n line = mm.readline() if mmap_mode else fh.readline()\r\n if not line:\r\n break\r\n\r\n #total_read += len(line)\r\n\r\n if \"--accumulate\" in sys.argv:\r\n mem_before_acc = mem_read()\r\n content += str(line)\r\n mem_after_acc = mem_read()\r\n\r\n mem_after = mem_read()\r\n\r\n print(f\"{idx:4d} {mem_after:10.2f}MB {mem_after - mem_before:10.2f}MB {mem_after_acc - mem_before_acc:10.2f}MB\")\r\n```\r\n\r\nit has some other instrumentations to do mmap and accumulate data, but let's ignore that for now.\r\n\r\nHere it is running in a simple non-mmap IO:\r\n\r\n```\r\n$ python mmap-no-leak-debug.py\r\n idx RSS Δ RSS Δ accumulated\r\n 0 12.43MB 12.43MB 0.00MB\r\n 1 269.72MB 257.29MB 0.00MB\r\n 2 269.73MB 0.02MB 0.00MB\r\n 3 269.73MB 0.00MB 0.00MB\r\n 4 269.74MB 0.01MB 0.00MB\r\n 5 269.74MB 0.00MB 0.00MB\r\n 6 269.75MB 0.01MB 0.00MB\r\n 7 269.75MB 0.00MB 0.00MB\r\n 8 269.76MB 0.01MB 0.00MB\r\n 9 269.76MB 0.00MB 0.00MB\r\n 10 269.77MB 0.01MB 0.00MB\r\n 11 269.77MB 0.00MB 0.00MB\r\n 12 269.77MB 0.00MB 0.00MB\r\n 13 269.77MB 0.00MB 0.00MB\r\n 14 269.77MB 0.00MB 0.00MB\r\n 15 269.77MB 0.00MB 0.00MB\r\n 16 146.02MB -123.75MB 0.00MB\r\n```\r\n\r\nas you can see even this super-simplistic program that just performs `readline()` slightly increases in RSS over iterations.\r\n\r\nIf you have a better tool for measurement other than RSS, I'm all ears.",
"@stas00 if you aren't using memory maps, you should be able to clearly see the increase in the virtual mem for the process as well. Even then, it could still be challenging to determine if it's leak vs fragmentation due to problematic allocation patterns (not uncommon with Python). Using a better mem allocator like tcmalloc via LD_PRELOAD hooks could reduce impact of fragmentation across both Python and c libs. Not sure that plays nice with any allocator that arrow might use itself though. "
] | ## Describe the bug
When the HF datasets is used in conjunction with PyTorch Dataloader, the RSS memory of the process keeps on increasing when it should stay constant.
## Steps to reproduce the bug
Run and observe the output of this snippet which logs RSS memory.
```python
import psutil
import os
from transformers import BertTokenizer
from datasets import load_dataset
from torch.utils.data import DataLoader
BATCH_SIZE = 32
NUM_TRIES = 10
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
def transform(x):
x.update(tokenizer(x["text"], return_tensors="pt", max_length=64, padding="max_length", truncation=True))
x.pop("text")
x.pop("label")
return x
dataset = load_dataset("imdb", split="train")
dataset.set_transform(transform)
train_loader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)
mem_before = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)
count = 0
while count < NUM_TRIES:
for idx, batch in enumerate(train_loader):
mem_after = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)
print(count, idx, mem_after - mem_before)
count += 1
```
## Expected results
Memory should not increase after initial setup and loading of the dataset
## Actual results
Memory continuously increases as can be seen in the log.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.2
- Platform: Linux-4.19.0-21-cloud-amd64-x86_64-with-glibc2.10
- Python version: 3.8.13
- PyArrow version: 7.0.0
| 4,883 |
https://github.com/huggingface/datasets/issues/4881 | Language names and language codes: connecting to a big database (rather than slow enrichment of custom list) | [
"Thanks for opening this discussion, @alexis-michaud.\r\n\r\nAs the language validation procedure is shared with other Hugging Face projects, I'm tagging them as well.\r\n\r\nCC: @huggingface/moon-landing ",
"on the Hub side, there is not fine grained validation we just check that `language:` contains an array of lowercase strings between 2 and 3 characters long =)\r\n\r\nand for `language_bcp47:` we just check it's an array of strings.\r\n\r\nThe only page where we have a hardcoded list of languages is https://huggingface.co/languages and I've been thinking of hooking that page on an external database of languages (so any suggestion is super interesting), but it's not used for validation.\r\n\r\nThat being said, in `datasets` this file https://github.com/huggingface/datasets/blob/main/src/datasets/utils/resources/languages.json is not really used no? Or just in the tagging tool? What about just removing it?\r\n\r\nalso cc'ing @lbourdois who's been active and helpful on those subjects in the past!",
"PS @alexis-michaud is there a DB of language codes you would recommend? That would contain all `ISO 639-1, 639-2 or 639-3 codes` and be kept up to date, and ideally that would be accessible as a Node.js npm package?\r\n\r\ncc @albertvillanova too",
"> PS @alexis-michaud is there a DB of language codes you would recommend? That would contain all `ISO 639-1, 639-2 or 639-3 codes` and be kept up to date, and ideally that would be accessible as a Node.js npm package?\r\n> \r\n> cc @albertvillanova too\r\n\r\nMany thanks for your answer! \r\n\r\nThe Glottolog database is kept up to date, and has information on the closest ISO code for each Glottocode. So providing a clean table with equivalences sounds (to me) like something perfectly reasonable to expect from their team. \r\nTo what extent would [pyglottolog](https://github.com/glottolog/pyglottolog) fit the bill / do the job? (API documentation [here](https://pyglottolog.readthedocs.io/en/latest/index.html)) I'm reaching my technical limitations here: I can't assess the distance between what they offer and what the HF team needs. \r\nI have opened an Issue in [their repo](https://github.com/glottolog/glottolog-cldf/issues/13). \r\n\r\nVery interested to see where it goes from there.",
"I just tried pyglottolog to generate a file with all the current IDs (first column).\r\n\r\n`glottolog languoids` inside the `glottolog` repository.\r\n\r\n[glottolog-languoids-v4.6-10-g5c66eec874.csv](https://github.com/huggingface/datasets/files/9417456/glottolog-languoids-v4.6-10-g5c66eec874.csv)\r\n\r\n",
"Greetings @alexis-michaud and others,\r\nI think perhaps a standards-based approach here would help everyone out both at the technical and social layers of technical innovations. \r\n\r\nLet me say a few things: \r\n1. there are multiple kinds of assets in AI that should have associated language codes. \r\n * AI Training Data sets\r\n * AI models\r\n * AI outputs\r\nThese are all distinct components which should be tagged for the language and encoding methods they operate on or enhance. For example, an AI based cross-language tool from French to English (UK variety) still needs to consider if it is operating on oral language speech or written text. This is where [IANA language sub-tags](https://www.iana.org/assignments/language-subtag-registry/language-subtag-registry) come in and are so important. I link to the official source. If one wants to use middleware such as a python package or npm package to manage strings then please make sure those packages are updating codes as they are being revised. I see that @julien-c mentioned BCP-47. BCP-47 is the current standard for language tagging. Following it will make the resources you create more findable and let future users better understand or expect any biases which may have been introduced in the different AI based products.\r\n2. BCP-47 is a technical read. However, you will notice that it identifies when to use an ISO 639-1, ISO 639-2, or ISO 639-3. code. This is important for interoperability with many systems. If you are using library systems then you should likely just stick with ISO 639-3 codes.\r\n3. If you are going to use Glottolog codes use them after an `-x-` tag in the BCP-47 format to maintain BCP-47 validity. \r\n4. You should source ISO 639-3 codes directly from the [ISO 639-3 registrar](https://iso639-3.sil.org/code_tables/639/data) as these codes are updated annually, usually in February or March. ISO 639-3 codes have multiple classes: `Active`, `Deprecated`, and `Unassigned`. This means that string length checking is not a sufficient strategy for validation.\r\n5. The names of smaller languages often change depending on the language used to describe them. The [ISO639-2 documentation](https://www.loc.gov/standards/iso639-2/php/code_list.php) has a list of language names for languages with smaller populations for languages in which descriptions about these languages are often written. For example, ISO 639-2's documentation contains the names of languages as they are used in French, German, and English. ISO 639-2 rarely is updated as it is now tied to ISO 639-3's evolution and modern systems should just use ISO 639-3, but these additional names of languages in other languages may not appear in the ISO 369-3 tables.\r\n6. Glottolog codes are also updated at least annually. Usually sometime after ISO 639-3 updates.\r\n7. Please, if the material is in a written mode, please indicate which script is used unless the IANA field has a `suppress script` value. Please use the script tag that BCP-47 calls for from [ISO 15924](https://unicode.org/iso15924/iso15924-codes.html). This also updates at least annually. \r\n8. Another great place to look for language names is the [Unicode CLDR database for locales](https://cldr.unicode.org/translation/displaynames/languagelocale-names). These ought to be congruent with ISO 639-3 but, sometimes CLDR has additional references to languages (such as the french name for a language) which is not contained in ISO 639-2 or ISO 639-3.\r\n9. Wikidata for language names is not always a great source of authoritative information. Language names are asymmetrical. Many times they are contrived because there is no actual name for the language in the language referring... e.g. French doesn't have a name for every language in the world, often they say something like: the language of 'x' people. — English does the same. When a language name standard does not have the best name for a language the best way to handle that is to make a change request with the standards registrar. Keeping track of the source list and the version of your source list for your language codes is very important. \r\n10. Finally, It would be a great service to technologist, minority language communities, and linguists if for all resources of the three types mentioned in number 1 above you added a record to [OLAC](http://www.language-archives.org/). — I can help you with that. OLAC is a search interface for language resources.\r\n",
"Hi everybody!\r\n\r\nAbout the point:\r\n> also cc'ing @lbourdois who's been active and helpful on those subjects in the past!\r\n\r\nDiscussions on the need to improve the Hub's tagging system (applying to both datasets and models) can be found in the following discussion: https://github.com/huggingface/hub-docs/issues/193\r\nOnce this system has been redone and satisfies the identified needs, a redesign of the [Languages page](https://huggingface.co/languages) would also be relevant: https://github.com/huggingface/hub-docs/issues/194. \r\nI invite you to read them. But as a quick summary, the exchanges were oriented towards the ISO standard (the first HF system was based on it and it is generally the standard indicated in AI/DL papers) by favouring ISO 639-1 if it exists, and fallback to ISO 639-2 or ISO 639-3 if it doesn't. In addition, it is possible to add BCP-47 tags to consider existing varieties/regionalisms within a language (https://huggingface.co/datasets/AmazonScience/massive/discussions/1). If a language does not belong to either of these two standards, then a request should be made to the HF team to add it manually.\r\n\r\n\r\nTo return to the present discussion, thank you for the various databases and methodologies you mention. It makes a big difference to have linguists in the loop 🚀.\r\n\r\nI have a couple of questions where I think an expert perspective would be appreciated:\r\n- Do you think it's possible to easily handle tags that have been deprecated potentially for decades?\r\nFor example (I'm taking the case of Hebrew but this has happened for other languages) I tagged Google models with the \"iw\" [tag](https://huggingface.co/models?language=iw&sort=downloads) because I based it on what the authors gave in their [paper](https://arxiv.org/pdf/2010.11934.pdf) see table 6 page 12). It turns out that this ISO tag has in fact been deprecated since 1989 in favour of the \"he\" tag. It would therefore be necessary to have a verification that transforms the old tags into the most recent ones.\r\n\r\n- When you look up a language on Wikipedia, it usually shows, in addition to the ISO standard, the codes in the Glottolog (which you have already mentioned), [ELP](https://www.endangeredlanguages.com/?hl=en) and [Linguasphere](http://www.linguasphere.info/jr/index.php?l1=home&l2=welcome) databases. Would you have any opinion about these two other databases?\r\n\r\n- On the Hub, there is the following dataset where French people speak in English: https://huggingface.co/datasets/Datatang/French_Speaking_English_Speech_Data_by_Mobile_Phone \r\nIs there a database to take this case into account? I have not found any code in the Glottolog database. If based on an IETF BCP-47 standard, I would tend to tag the dataset with \"en-fr\" but would this be something accepted by linguists?\r\nBased on the first post in this thread that there are about 8000 languages, if one considers that a given language can be pronounced by a speaker of the other 7999, that would theoretically make about 64 million BCP-47 language1-language2 codes existing. And even much more if we consider regionalists with language1_regionalism_x-language2_regionalism_y. I guess there is no such database.\r\n\r\n- Are there any databases that take into account all the existing sign languages in the world?\r\nIt would be nice to have them included on the Hub.\r\n\r\n- Is there an international classification of languages?\r\nA bit like the [International Classification of Diseases](https://en.wikipedia.org/wiki/International_Classification_of_Diseases) in medicine, which is established by the WHO and used as a reference throughout the world. The idea would be to have a precise number of languages to which we would then have to assign a unique tag in order to find them later. \r\n\r\n- Finally for the CNRS team, when can we expect to see all the datasets of [Pangloss](https://pangloss.cnrs.fr/) on HF? 👀 And I don't know if you have a way to help to add also the datasets of [CoCoON](https://cocoon.huma-num.fr/exist/crdo/).",
"> I invite you to read them. But as a quick summary, the exchanges were oriented towards the ISO standard (the first HF system was based on it and it is generally the standard indicated in AI/DL papers) by favouring ISO 639-1 if it exists, and fallback to ISO 639-2 or ISO 639-3 if it doesn't. In addition, it is possible to add BCP-47 tags to consider existing varieties/regionalisms within a language (https://huggingface.co/datasets/AmazonScience/massive/discussions/1). If a language does not belong to either of these two standards, then a request should be made to the HF team to add it manually.\r\n\r\nOne comment on this fall back system (which generally follows the BCP-47 process). ISO 639-2 has some codes which refer to a language ambiguously. For example, I believe code `ara` is used for arabic. In some contexts arabic is considered a single language, however, Egyptian Arabic is quite different from Moroccan Arabic, which are both considered separate languages. These ambiguous codes are valid ISO 639-3 codes but they have a special status. They are called `macro codes`. They exist inside the ISO 639-3 standard to provide absolute fallback compatibility between ISO 639-2 and ISO 639-3. However, when considering AI and MT applications with language data, the unforeseen potential applications and the potential for bias using macro codes should be avoided for new applications of language tags to resources. For historical cases where it is not clear what resources were used to create the AI tools or datasets then I understand the use of ambiguous tag uses. So for clarity in language tagging I suggest:\r\n\r\n1. Strictly following BCP-47\r\n2. Whenever possible avoid the use of macro tags in the ISO 639-3 standard. These are BCP-47 valid, but could introduce biases in the application of their use in society. (Generally there are more specific tags available to use in the ISO 639-3 standard.)",
"> * Are there any databases that take into account all the existing sign languages in the world?\r\n> It would be nice to have them included on the Hub.\r\n\r\nSign Languages present an interesting case. As I understand the situation. The identification of sign languages has been identified as a component of their endangerment. Some sign languages do exist in ISO 639-3. For further discussion on the issue I refer readers to the following publications: \r\n\r\n* https://doi.org/10.3390/languages7010049\r\n* https://www.academia.edu/35870983/The_ethics_of_of_language_identification_and_ISO_639\r\n\r\nOne way to be BCP-47 compliant and identify a sign language which is not identified in any of the BCP-47 referenced standards is to use the ISO 639-3 code for undetermined language `und` and then apply a custom suffix indicator (as explained in BCP-47) `-x-` and a custom code, such as the ones used in https://doi.org/10.3390/languages7010049",
"> * Is there an international classification of languages?\r\n> A bit like the [International Classification of Diseases](https://en.wikipedia.org/wiki/International_Classification_of_Diseases) in medicine, which is established by the WHO and used as a reference throughout the world. The idea would be to have a precise number of languages to which we would then have to assign a unique tag in order to find them later.\r\n\r\nYes that would be the function of ISO 639-3. It is the reference standard for languages. It includes a code and its name and the status of the code. Many technical metadata standards for file and computer interoperability reference it, many technical library metadata standards reference it. Some linguists use it. Many governments reference it. \r\n\r\nIndexing diseases are different from indexing languages in several ways, one way is that diseases are the impact of a pathogen not the pathogen itself. If we take COVID-19 as an example, there are many varieties of the pathogen but broadly speaking there is only one disease — with many symptoms.\r\n\r\n",
">* When you look up a language on Wikipedia, it usually shows, in addition to the ISO standard, the codes in the Glottolog (which you have already mentioned), [ELP](https://www.endangeredlanguages.com/?hl=en) and [Linguasphere](http://www.linguasphere.info/jr/index.php?l1=home&l2=welcome) databases. Would you have any opinion about these two other databases?\r\n\r\nWhile these do appear on wikipedia, I don't know of any information system which uses these codes. I do know that glottolog did import ELP data at one time and its database does contain ELP data I'm not sure if Glottolog regularly ingests new versions of ELP data. I suspect that the use of Linguasphere data may be relevant to users of wikidata as a linked data attribute but I haven't heard of any linked data projects using Linguasphere data for analysis or product development. My impression is that it is fairly unused.",
"> * Do you think it's possible to easily handle tags that have been deprecated potentially for decades?\r\n>For example (I'm taking the case of Hebrew but this has happened for other languages) I [tag](https://huggingface.co/models?language=iw&sort=downloads)ged Google models with the \"iw\" tag because I based it on what the authors gave in their [paper](https://arxiv.org/pdf/2010.11934.pdf) see table 6 page 12). It turns out that this ISO tag has in fact been deprecated since 1989 in favour of the \"he\" tag. It would therefore be necessary to have a verification that transforms the old tags into the most recent ones.\r\n\r\nYes. You can parse the IANA file linked to above (it is regularly updated). All deprecated tags are marked as such in that file. The new prefered tag if there is one, is indicated. ISO 639-3 also indicates a code's status but their list is relevant only codes within their domain (ISO 639-3).",
"> * On the Hub, there is the following dataset where French people speak in English: https://huggingface.co/datasets/Datatang/French_Speaking_English_Speech_Data_by_Mobile_Phone\r\nIs there a database to take this case into account? I have not found any code in the Glottolog database. If based on an IETF BCP-47 standard, I would tend to tag the dataset with \"en-fr\" but would this be something accepted by linguists?\r\n\r\nI would interpret `en-fr` as english as spoken in France. `fr`in this position refers to the geo-political entity not a second language. I see no reason that other linguists should have a different option after having read BCP-47 and understood how it works.\r\n\r\nThe functional goal here is to tag a language resource as being produced by nonnative speakers, while tagging both languages. There are several problems here. The first is that BCP-47 has no way explicit way to do this. One could use the sub code `x-` with a private use code to indicate a second language and infer some meaning as to that language's role. However, there is another problem here which complexifies the situation greatly... how do we know that those english speakers (in France, or from France, or who were native French speakers) were not speaking their third or fourth language rather than their second language. So to conceptualize a sub-tag which indicates the first language of a speech act for speakers in a second (or other) language would need to be carefully crafted. It might then be proposed to the appropriate authorities. For example three sub-tags exist.\r\n\r\nThere are three registered sub-tags out of a BCP-47 allowed 35. These are `x-`, `u-`, and `t-`. `u-` and `t-` are defined in [RFC6067 ](https://www.rfc-editor.org/rfc/rfc6067)and [RFC6497](https://www.rfc-editor.org/rfc/rfc6497) . For more information see the [Unicode CLDR documentation](https://cldr.unicode.org/index/bcp47-extension) where it says: \r\n\r\n\r\n>[IETF BCP 47 ](http://www.google.com/url?q=http%3A%2F%2Ftools.ietf.org%2Fhtml%2Fbcp47&sa=D&sntz=1&usg=AOvVaw1DoMN1IBGg-JHgECBvdW1t)[Tags for Identifying Languages](http://www.google.com/url?q=http%3A%2F%2Ftools.ietf.org%2Fhtml%2Fbcp47&sa=D&sntz=1&usg=AOvVaw1DoMN1IBGg-JHgECBvdW1t) defines the language identifiers (tags) used on the Internet and in many standards. It has an extension mechanism that allows additional information to be included. The Unicode Consortium is the maintainer of the extension ‘u’ for Locale Extensions, as described in [rfc6067](https://www.google.com/url?q=https%3A%2F%2Ftools.ietf.org%2Fhtml%2Frfc6067&sa=D&sntz=1&usg=AOvVaw0gGWi0EjHfy1WId8k8oKAi), and the extension 't' for Transformed Content, as described in [rfc6497](https://www.google.com/url?q=https%3A%2F%2Ftools.ietf.org%2Fhtml%2Frfc6497&sa=D&sntz=1&usg=AOvVaw0w-OUsFX1PtaKYIq31P64I).\r\n>\r\n>The subtags available for use in the 'u' extension provide language tag extensions that provide for additional information needed for identifying locales. The 'u' subtags consist of a set of keys and associated values (types). For example, a locale identifier for British English with numeric collation has the following form: en-GB-u-kn-true\r\n>\r\n>The subtags available for use in the 't' extension provide language tag extensions that provide for additional information needed for identifying transformed content, or a request to transform content in a certain way. For example, the language tag \"ja-Kana-t-it\" can be used as a content tag indicates Japanese Katakana transformed from Italian. It can also be used as a request for a given transformation.\r\n>\r\n>For more details on the valid subtags for these extensions, their syntax, and their meanings, see LDML Section 3.7 [Unicode BCP 47 Extension Data](http://www.google.com/url?q=http%3A%2F%2Fwww.unicode.org%2Freports%2Ftr35%2F%23Locale_Extension_Key_and_Type_Data&sa=D&sntz=1&usg=AOvVaw0lMthb9KbTJtoOd5mvv3Ha).",
"Hi @lbourdois ! Many thanks for the detailed information.\r\n\r\n> Discussions on the need to improve the Hub's tagging system (applying to both datasets and models) can be found in the following discussion: [huggingface/hub-docs#193](https://github.com/huggingface/hub-docs/issues/193) \r\nFascinating topic! To me, the following suggestion has a lot of appeal:\r\n\"if consider that it was necessary to create an ISO 639-3 because ISO 639-1 was deficient, it would be to do the reverse and thus convert the tags from ISO 639-1 to ISO 639-2 or 3 (https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes or https://iso639-3.sil.org/code_tables/639/data).\"\r\n\r\nYes, ISO 639-1 is unsuitable because it has so few codes: less than 200. To address linguistic diversity in 'unrestricted mode', a list of all languages is wanted. \r\n\r\nThe idea of letting people use their favourite nomenclature and automatically adding the ISO 639-3 three-letter code as a tag is appealing. Thus all the HF datasets would have three-letter language tags (handy for basic search), alongside the authors' preferred tags and language names (including Glottolog tags as well as ISO 639-{1, 2}, and all other options allowed by BCP-47). \r\n\r\nRetaining the authors' original tags and language names would be best. \r\n* For language names: some people favour one name over another and it is important to respect their choice. In the case of Yongning Na: alternative names include 'Mosuo', 'Narua', 'Eastern Naxi'... and the names carry implications: people have been reported to come to blows about the use of the term 'Mosuo'. \r\n* For language tags: Glottocodes can be more fine-grained than Ethnologue (ISO 639-3), and some colleagues feel strongly about those. \r\n\r\nThus there would be a BCP-47 tag (sounds like a solid technical choice, though not 'passer-by-friendly': requiring some expertise to interpret) **plus** an ISO 639-3 tag that could be grabbed easily, and (last but not least) language names spelled out in full. Searches would be easier. No information would be lost. \r\n\r\nAre industry practices so conservative that many people are happy with two-letter codes, and consider ISO 639-3 three-letter codes an unnecessary complication? That would be a pity, since there are so many advantages to using longer lists. (Somewhat like the transition to Unicode: sooo much better!) But maybe that conservative attitude _is_ widespread, and it would then need to be taken into account. In which case, one could consider offering two-letter codes as a search option. Internally, the search engine would look up the corresponding 3-letter codes, and produce the search results accordingly. \r\n\r\nNow to the other questions:\r\n\r\n> * Do you think it's possible to easily handle tags that have been deprecated potentially for decades?\r\n> For example (I'm taking the case of Hebrew but this has happened for other languages) I tagged Google models with the \"iw\" [tag](https://huggingface.co/models?language=iw&sort=downloads) because I based it on what the authors gave in their [paper](https://arxiv.org/pdf/2010.11934.pdf) see table 6 page 12). It turns out that this ISO tag has in fact been deprecated since 1989 in favour of the \"he\" tag. It would therefore be necessary to have a verification that transforms the old tags into the most recent ones.\r\nI guess that the above suggestion takes care of this case. The original tag (in this example, \"iw\") is retained (facilitating cross-reference with the published paper, and respecting the real: the way the dataset was originally tagged). This old tag goes into the `BCP-47` field of the dataset, which can handle quirks & oddities like this one. And a new tag is added in the `ISO 639-3` field: the 3-letter code \"heb\". \r\n\r\n> * When you look up a language on Wikipedia, it usually shows, in addition to the ISO standard, the codes in the Glottolog (which you have already mentioned), [ELP](https://www.endangeredlanguages.com/?hl=en) and [Linguasphere](http://www.linguasphere.info/jr/index.php?l1=home&l2=welcome) databases. Would you have any opinion about these two other databases?\r\n\r\nI'm afraid I never heard about Linguasphere. The [online register for Linguasphere (PDF)](http://www.linguasphere.info/jr/pdf/index/LS_index_n-n.pdf) seems to be from 1999-2000. It seems that the level of interoperability is not very high right now. (By contrast, Glottolog has [pyglottolog](https://github.com/glottolog/pyglottolog) and in my experience contacts flow well.) \r\n\r\nThe Endangered Languages Project is something Google started but initially did not 'push' very strongly, it seems. Just airing an opinion on the public Internet, it seems that the project is now solidly rooted at University of Hawaiʻi at Mānoa. It seems that they do not generate codes of their own. They refer to ISO 639-3 (Ethnologue) as a code authority when applicable, and otherwise provide comments in so many words, such as that language L currently lacks an Ethnologue code of its own (example [here](https://www.endangeredlanguages.com/lang/10624)). \r\n\r\n> * On the Hub, there is the following dataset where French people speak in English: https://huggingface.co/datasets/Datatang/French_Speaking_English_Speech_Data_by_Mobile_Phone\r\n> Is there a database to take this case into account? I have not found any code in the Glottolog database. If based on an IETF BCP-47 standard, I would tend to tag the dataset with \"en-fr\" but would this be something accepted by linguists?\r\n> Based on the first post in this thread that there are about 8000 languages, if one considers that a given language can be pronounced by a speaker of the other 7999, that would theoretically make about 64 million BCP-47 language1-language2 codes existing. And even much more if we consider regionalists with language1_regionalism_x-language2_regionalism_y. I guess there is no such database.\r\n\r\nYes, you noted the difficulty here: that there are so many possible situations. Eventually, each dataset would required descriptors of its own. @BenjaminGalliot points out that, in addition to specifying the speakers' native languages, the degree of language proficiency would also be relevant. How many years did the speakers spend in which area? Talking which languages? In what chronological order? Etc. The complexity defies encoding. The purpose of language codes is to allow for searches that group resources into sets that make sense. Additional information is very important, but would seem to be a matter for 'comments' fields. \r\n\r\n> * Is there an international classification of languages?\r\n> A bit like the [International Classification of Diseases](https://en.wikipedia.org/wiki/International_Classification_of_Diseases) in medicine, which is established by the WHO and used as a reference throughout the world. The idea would be to have a precise number of languages to which we would then have to assign a unique tag in order to find them later.\r\n\r\nAs I understand, Ethnologue and Glottolog both try to do that, each in its own way. The simile with diseases seems interesting, to some extent: in both cases it's about human classification of phenomena that have complexity (though some diseases are simpler than others, whereas all languages have much complexity, in different ways).\r\n\r\n> * Finally, when can we expect to see all the datasets of [Pangloss](https://pangloss.cnrs.fr/) on HF? eyes And I don't know if you have a way to help to add also the datasets of [CoCoON](https://cocoon.huma-num.fr/exist/crdo/).\r\n\r\nThree concerns: (i) Technical specifications: we have not yet received feedback on the Japhug and Na datasets in HF. There may be technical considerations that we have not yet thought of and that would need to be taken into account before 'bulk upload'. (ii) Would there be a way to automate the process? The way @BenjaminGalliot did it for Japhug and Na, there was a manual component involved, and doing it by hand for all 200 datasets would not be an ideal workflow, given that the metadata are all clearly arranged. (iii) Some datasets are currently under a 'No derivatives' CreativeCommons license. We could go back to the depositors and argue that the 'No derivatives' mention were best omitted (see [here a similar argument about publications](https://creativecommons.org/2020/04/21/academic-publications-under-no-derivatives-licenses-is-misguided/)): again, we'd want to be sure about the way forward before we set the process into motion.\r\n\r\nOur hope would be that some colleagues try out the [OutilsPangloss](https://gitlab.com/lacito/outilspangloss) download tool, assemble datasets from Pangloss/Cocoon as they wish, then deposit them to HF.",
"> The idea of letting people use their favourite nomenclature and automatically adding the ISO 639-3 three-letter code as a tag is appealing. Thus all the HF datasets would have three-letter language tags (handy for basic search), alongside the authors' preferred tags and language names (including Glottolog tags as well as ISO 639-{1, 2}, and all other options allowed by BCP-47).\r\n> \r\n> Retaining the authors' original tags and language names would be best.\r\n> \r\n> * For language names: some people favour one name over another and it is important to respect their choice. In the case of Yongning Na: alternative names include 'Mosuo', 'Narua', 'Eastern Naxi'... and the names carry implications: people have been reported to come to blows about the use of the term 'Mosuo'.\r\n> * For language tags: Glottocodes can be more fine-grained than Ethnologue (ISO 639-3), and some colleagues feel strongly about those.\r\n> \r\n> Thus there would be a BCP-47 tag (sounds like a solid technical choice, though not 'passer-by-friendly': requiring some expertise to interpret) **plus** an ISO 639-3 tag that could be grabbed easily, and (last but not least) language names spelled out in full. Searches would be easier. No information would be lost.\r\n\r\n@alexis-michaud raises an excellent point. Language Resource users have varying search habits (or approaches). This includes cases where two or more language names refer to a single language. A search utility/interface needs to be flexible and able to present results from various kinds of input in the search process. This could be like how the terms French/Français/Franzosisch (en/fr/de) are names for the same language or it could be a variety of the following: autoglottonyms (how the speakers of the language refer to their language), or exoglottonyms (how others refer to the language). Additionally, in web based searches I have also needed to implement diacritic sensitive and insensitive logic so that users can type with or without diacritics and not have results unnecessarily excluded. \r\n\r\nDepending on how detailed of a search problem HF seeks to solve. It may be better to off load complex search to search engines like OLAC which aggregate a lot of language resources. — as I mentioned above I can assist with the informatics on creating an OLAC feed.\r\n\r\nAbstracting search logic from actual metadata may prove a useful way to lower the technical debt overhead. Technical tools and library standards use ISO and BCP-47 Standards. So, from a bibliographic metadata perspective this seems to be the way forward with the widest set of use cases. ",
"To get a visual idea of these first exchanges, I coded a Streamlit app that I put online on Spaces: https://huggingface.co/spaces/lbourdois/Language-tags-demo. \r\nThe code is in Python so I don't know if it can be used by HF who seems to need something in Node.js but it serves as a proof of concept. The advantage is also that you can directly test ideas by enter things in a search bar and see what comes up. \r\n\r\nThis application is divided into 3 points:\r\n- The first is to enter a language in natural language to get its code which can then be filled in the YAML file of the README.MD files of the HF datasets or models in order to be referenced and found by everyone.\r\nIn practice, enter the language (e.g: `English`) you are interested in to get its associated tag (e.g: `en`). You can enter several languages by separating them with a comma (e.g `French,English,German`). You will be given priority to the ISO 639-3 code if it exists otherwise the Glottocode or the BCP47 code (for varieties in particular). If none of these codes are available, it links to a page where the user can contact HF to request to add this tag. \r\nIf you enter a BCP47 code, it must be entered as follows: `Language(Territory)`, for example `French(Canada)`. Attention! If you enter a BCP-47 language, it must be entered first, otherwise the plant code will be displayed. I have to fix this problem but I am moving to a new place, I don't have an internet connection when I want and I prefer to push this first version so that you can already test things now and not have to wait days or weeks.\r\nThis point is intended to simulate the user's side of the equation, which wonders which tag he should fill in for his language.\r\n\r\n\r\n- The second is to enter a language code to obtain the name of the language in natural language.\r\nIn practice, enter the tag (ISO 639-1/2/3, Glottolog or BCP-47) you are interested in (e.g: `fra`) to get its associated language (e.g: French). You can enter several languages by separating them with a comma (e.g `fra,eng,deu`). Attention! If you enter a BCP-47 code, it must be entered first, otherwise the plant code will be displayed. Same as the other bug above (it's actually the same one).\r\nThis point is intended to simulate the side of HF that for a given tag must return the correct language.\r\n\r\n\r\n\r\nTo code these two points, I tested two approaches. \r\n\r\n1. The first one (internal DB in the app) consists in querying a database that HF would have locally at their place. To create this database, I merged the ISO 639 database (https://iso639-3.sil.org/sites/iso639-3/files/downloads/iso-639-3.tab) and the Glottolog database (https://glottolog.org/meta/downloads). The result of this merge is visible in the 3rd point of the application qui is an overview of the database.\r\nIn the image below, on line 1 of the database, we can see that the Glottocode database gives an ISO 639-3 code (column ISO639P3code) but not the ISO 639 database (column 639-3). Do you have an explanation for this phenomenon?\r\n\r\n\r\n\r\nFor BCP 47 codes of the type `fr-CA`, I have retrieved the ISO-3166 alpha 1 codes of the territories (https://www.iso.org/iso-3166-country-codes.html).\r\nIn practice, what I do is if we enter `fr-CA` is that the letters before the `-` refer to a language in the `Name` column for a `639-1` == `fr` (`639-3` for `fra` or `fre`) in the base of my image above. Then I look at the letters after the `-` which refers to a territory. It comes out `French (Canada)`. I used https://cldr.unicode.org/translation/displaynames/languagelocale-name-patterns for the pattern that came up.\r\n\r\n\r\n2. The second approach (with langcodes lib in the app) consists in using the Python `langcodes` library (https://github.com/rspeer/langcodes) which offers a lot of features in ready-made functions. It manages for example deprecated codes, the validity of an entered code, gives languages from code in the language of your choice (by default in English, but also autoglottonyms), etc. I invite you to read the README of the library. The only negative point is that it hasn't been updated for 10 months so basing your tag system on an external tool that isn't necessarily up to date can cause problems in the long run. But it is certainly an interesting source.\r\n\r\nFinally, I have added some information on the number of people speaking/reading the language(s) searched (figures provided by langcodes which are based on those given by ISO). This is not relevant for our topic but it could be figures that could be added as information on the https://huggingface.co/languages page. \r\n\r\n\r\n\r\nWhat could be done to improve the app if I have time:\r\n- Write the text for the app's homepage to describe what it does. This could serve as a basis for a documentation that I think will be necessary to add somewhere on the HF website to explain how the language tagging system works.\r\n- Deal with the bug mentioned above\r\n- Integrate ISO 3166-1 alpha 2 territories (https://www.iso.org/obp/ui#iso:pub:PUB500001:en)? They offer a finer granularity than ISO 3166-1 alpha 1 which is limited to the country level, but they are very administrative (for French, ISO 3166-1 alpha 2 gives us the \"départements\" for example).\r\n- Add autoglottonyms? (I only handle English language names for the moment)\r\n- For each language indicate to which family it belongs, in practice this could help to make data augmentation, but especially to classify the languages and find them more easily on the page https://huggingface.co/languages.",
"Very impressive! Using the prompt 'Japhug' (a language name), the app finds the intended language:\r\n\r\n\r\nA first question: based on the Glottocode, would it be possible to grab the closest ISO639-3 code? In case there is no match for the exact language variety, one needs to explore the higher-level groupings, level by level. For this language (Japhug), the information provided in the extracted CSV file (`glottolog-languoids-v4.6.csv`) is: \r\n`sino1245/burm1265/naqi1236/qian1263/rgya1241/core1262/jiar1240` \r\nOne need not look further than the first higher-level grouping, [`jiar1240`](https://glottolog.org/resource/languoid/id/jiar1240), to get an ISO639-3 code, namely `jya`.\r\n\r\nThus users searching by language names would get ISO639-3 (often less fine-grained than Glottolog) as a bonus.\r\nIt might be possible to ask the Glottolog team to provide this piece of information as part of an export from their database.",
"> on line 1 of the database, we can see that the Glottocode database gives an ISO 639-3 code (column ISO639P3code) but not the ISO 639 database (column 639-3). Do you have an explanation for this phenomenon?\r\n\r\nThat is because the language name 'Aewa' is not found in the Ethnologue (ISO 639-3) export that you are using. [This export in table form](https://iso639-3.sil.org/sites/iso639-3/files/downloads/iso-639-3.tab) only has one reference name (`Ref_Name`). For the language at issue, it is not 'Aewa' but ['Awishira'](https://www.ethnologue.com/language/ash).\r\n\r\nBy contrast, the language on line 0 of the database is called 'Abinomn' by both Ethnologue and Glottolog, and accordingly, columns `ISO639P3code` and `639-3` both contain the ISO 639-3 code, `bsa`.\r\n \r\nThe full Ethnologue database records alternate names for each language, and I'd bet that 'Aewa' is recorded among alternate names for the 'Ashiwira' language. I can't check because the full Ethnologue database is paywalled. \r\n\r\n\r\n[Glottolog](https://glottolog.org/resource/languoid/id/abis1238) does provide the corresponding ISO 639-3 code for 'Aewa', `ash`, which is an exact match (it refers to the same variety as Glottolog `abis1238`).\r\nIn this specific case, Glottolog provides all the relevant information. I'd say that Glottolog can be trusted for all the codes they provide, including ISO 639-3 codes: they only include them when the match is good. \r\n\r\nSee previous comment about the cases where there is no exact match between Glottolog and ISO 639-3 (suggested workaround: look at a higher-level grouping to get an ISO 639-3 code).",
"I will add these two points to my TODO list.\r\n- Since Glottolog can be trust, I will add a condition to the code that if there is no ISO 639-3 code in the \"official\" database (https://iso639-3.sil.org/sites/iso639-3/files/downloads/iso-639-3.tab), look for it in the \"ISO639P3code\" column of Glottolog.\r\n- For the point of adding the closest ISO 639-3 code for a Glottolog code, what convention should be adopted for the output? Just the ISO 639-3 code, or the ISO 639-3 code - Glottolog code, or the ISO 639-3 code - language name?\r\nTo use the example of `Japhug` , should it be just `jya`, or `jya-japh1234` or `jya-Japhug`?",
"> * Integrate ISO 3166-1 alpha 2 territories (https://www.iso.org/obp/ui#iso:pub:PUB500001:en)? They offer a finer granularity than ISO 3166-1 alpha 1 which is limited to the country level, but they are very administrative (for French, ISO 3166-1 alpha 2 gives us the \"départements\" for example).\r\n\r\nI'm concerned with this sort of exploration. Not because I am against innovation. In fact this is an interesting thought exercise. However, to explore this thought further creates cognitive dissidence between BCP-47 authorized codes and other code sets which are not BP-47 compliant. For that reason, I think adding additional codes is a waste of time both for HF devs and for future users who get a confusing idea about language tagging. ",
"Good job for the application!\r\n\r\n> On the Hub, there is the following dataset where French people speak in English: https://huggingface.co/datasets/Datatang/French_Speaking_English_Speech_Data_by_Mobile_Phone\r\n Is there a database to take this case into account? I have not found any code in the Glottolog database. If based on an IETF BCP-47 standard, I would tend to tag the dataset with \"en-fr\" but would this be something accepted by linguists?\r\n Based on the first post in this thread that there are about 8000 languages, if one considers that a given language can be pronounced by a speaker of the other 7999, that would theoretically make about 64 million BCP-47 language1-language2 codes existing. And even much more if we consider regionalists with language1_regionalism_x-language2_regionalism_y. I guess there is no such database.\r\n\r\n> Yes, you noted the difficulty here: that there are so many possible situations. Eventually, each dataset would required descriptors of its own. @BenjaminGalliot points out that, in addition to specifying the speakers' native languages, the degree of language proficiency would also be relevant. How many years did the speakers spend in which area? Talking which languages? In what chronological order? Etc. The complexity defies encoding. The purpose of language codes is to allow for searches that group resources into sets that make sense. Additional information is very important, but would seem to be a matter for 'comments' fields.\r\n\r\nTo briefly complete what I said on this subject in a private discussion group, there is a lot of (meta)data associated with each element of a corpus (which language level, according to which criteria, knowing that even among native speakers there are differences, some of which may go beyond what seems obvious to us from a linguistic point of view, such as socio-professional category, life history, environment in the broad sense, etc.), which can be placed in ad-hoc columns, or more freely in a comment/note column. And it is the role of the researcher (in this case a linguist, in all likelihood) to do analyses (statistics...) to determine the relevant data, including criteria that may justify separating different languages (in the broad sense), making separate corpora, etc. Putting this information in the language code is in my opinion doing the job in the opposite and wrong direction, as well as bringing other problems, like where to stop in the list of multidimensional criteria to be integrated, so in my opinion, here, the minimum is the best (the important thing is in my opinion to have well-documented data, globally, by sub-corpus or by line)...\r\n\r\n> If you are going to use Glottolog codes use them after an -x- tag in the BCP-47 format to maintain BCP-47 validity.\r\n\r\nYes, for the current corpora, I have written:\r\n\r\n```\r\nlanguage:\r\n- jya\r\n- nru\r\nlanguage_bcp47:\r\n- x-japh1234\r\n- x-yong1288\r\n```\r\n\r\n> * Add autoglottonyms? (I only handle English language names for the moment)\r\n\r\nAutoglossonyms are useful (I use them prior to other glossonyms), but I'm not sure there is an easy way to retrieve them. We can find some of them in the \"Alternative Names\" panel of Glottolog, but even if we have an API to retrieve them easily, their associated language code will often not be the one we are in (hence the need to do several cycles to find one, which might not be the right one...). Maybe this problem needs more investigation...\r\n\r\n> For the point of adding the closest ISO 639-3 code for a Glottolog code, what convention should be adopted for the output? Just the ISO 639-3 code, or the ISO 639-3 code - Glottolog code, or the ISO 639-3 code - language name?\r\nTo use the example of Japhug , should it be just jya, or jya-japh1234 or jya-Japhug?\r\n\r\nI strongly insist not to add **a** language name after the code, it would restart a spiral of problems, notably the choice of the language in question:\r\n* the autoglossonym: in my opinion the best choice, but you have to know it…\r\n* the English name: iniquitous,\r\n* the name in the administratively/politically dominant language of the target language if it is relevant (strictly localized without overlapping, for example): iniquitous and tendentious (and in a way a special case of the previous one)...\r\n* etc.\r\n",
"> To get a visual idea of these first exchanges, I coded a Streamlit app that I put online on Spaces: https://huggingface.co/spaces/lbourdois/Language-tags-demo.\r\n> The code is in Python so I don't know if it can be used by HF who seems to need something in Node.js but it serves as a proof of concept. The advantage is also that you can directly test ideas by enter things in a search bar and see what comes up.\r\n\r\nThis is really great. You're doing a fantastic job. I love watching the creative process evolve. It is exciting. Let me provide some links to some search interfaces for further inspiration. I always find it helpful to know how others have approached a problem when figuring out my approach. I will link to three examples Glottolog, r12a's language sub-tag chooser, and the FLEx project builder wizard. The first two are online, but the last one is in an application which must be downloaded and works only on windows or linux. I have placed some notes on each of the screenshots.\r\n\r\n* **[Glottolog](https://glottolog.org/)** | [Search Query](https://glottolog.org/glottolog?name=en&namequerytype=part&multilingual=on#2/20.9/150.0) \r\n\r\n\r\n\r\n\r\n\r\n* **[r12a language sub-tag chooser](https://r12a.github.io/app-subtags/)** | [Code on github](https://github.com/r12a/app-subtags)\r\n\r\n\r\n\r\n\r\n* **FLEx Language Chooser** | [application page](https://software.sil.org/fieldworks/)\r\n\r\n\r\n",
"> In practice, what I do is if we enter `fr-CA` is that the letters before the `-` refer to a language in the `Name` column for a `639-1` == `fr` (`639-3` for `fra` or `fre`) in the base of my image above. Then I look at the letters after the `-` which refers to a territory. It comes out `French (Canada)`. I used https://cldr.unicode.org/translation/displaynames/languagelocale-name-patterns for the pattern that came up.\r\n\r\nWhat you are doing is looking at the algorithm for Locale generation rather than BCP-47's original documentation. I'm not sure there are difference, there might be. I know that locale IDs generally follow BCP-47 But I think there are some differences such as the use of `_` vs. `-`. ",
"> A first question: based on the Glottocode, would it be possible to grab the closest ISO639-3 code? In case there is no match for the exact language variety, one needs to explore the higher-level groupings, level by level. For this language (Japhug), the information provided in the extracted CSV file (`glottolog-languoids-v4.6.csv`) is: `sino1245/burm1265/naqi1236/qian1263/rgya1241/core1262/jiar1240` One need not look further than the first higher-level grouping, [`jiar1240`](https://glottolog.org/resource/languoid/id/jiar1240), to get an ISO639-3 code, namely `jya`.\r\n> \r\n> Thus users searching by language names would get ISO639-3 (often less fine-grained than Glottolog) as a bonus. It might be possible to ask the Glottolog team to provide this piece of information as part of an export from their database.\r\n\r\nThis is logical, but the fine grained assertions are not the same. That is just because they are in a hierarchical structure today doesn't mean they will be tomorrow. In some cases the glottolog is clearly referring to sub-language variants which will never receive full language status, whereas in other cases glottolog is referencing to unequal entities one or more of which should be a language. Many of the codes in glottolog have no associated documentation indicating what sort of speech variety they are. ",
"@lbourdois \r\n> * Since Glottolog can be trust, I will add a condition to the code that if there is no ISO 639-3 code in the \"official\" database (https://iso639-3.sil.org/sites/iso639-3/files/downloads/iso-639-3.tab), look for it in the \"ISO639P3code\" column of Glottolog.\r\n\r\nI'm confused here... if there is no ISO639-3 code in the official database from the registrar, why would you look for an \"unofficial\" code from someone else? What is the use case here?",
"> For the point of adding the closest ISO 639-3 code for a Glottolog code, what convention should be adopted for the output? Just the ISO 639-3 code, or the ISO 639-3 code - Glottolog code, or the ISO 639-3 code - language name?\r\nTo use the example of Japhug , should it be just jya, or jya-japh1234 or jya-Japhug?\r\n\r\n(answer edited in view of [Benjamin Galliot's comment](https://github.com/huggingface/datasets/issues/4881#issuecomment-1237420600) \r\nEasy part of the answer first: jya-Japhug is out, because, as @BenjaminGalliot pointed out above, mixing language names with language codes will make trouble. For Japhug, `jya-Japhug` looks rather good: the pair looks nice, the one (`jya`) packed together, the other (`Japhug`) good and complete while still pretty compact. But think about languages like 'Yongning Na' or 'Yucatán Maya': a code with a space in the middle, like `nru-Yongning Na`, is really unsightly and unwieldy, not?\r\n\r\nSome [principles for language naming in English](http://hdl.handle.net/10125/24725) have been put forward, with some linguistic arguments, but always supposing that such standardization is desirable, actual standardization of language names in English may well never happen.\r\n\r\nAs for `jya-japh1234`: again, at first sight it seems cute, combining two fierce competitors (Ethnologue and Glottolog) into something that gets the best of both worlds. \r\nBut @HughP has a point: _adding additional codes is a waste of time both for HF devs and for future users who get a confusing idea about language tagging_ Strong wording, for an important comment: better stick with BCP 47. \r\n\r\nSo the solution pointed out by Benjamin, from Frances Gillis-Webber and Sabine Tittel, looks attractive: \r\njya-x-japh1234\r\n\r\nOn the other hand, if the idea for HF Datasets is simply to add the closest ISO 639-3 code for a Glottolog code, maybe it could be provided simply in three letters: providing the 'raw' ISO 639-3 code `jya`. Availability of 'straight' ISO 639-3 codes could save trouble for some users, and those who want more detail could look at the rest of the metadata and general information associated with datasets.",
"The problem seems to have already been raised here: https://drops.dagstuhl.de/opus/volltexte/2019/10368/pdf/OASIcs-LDK-2019-4.pdf\r\n\r\nAn example can be seen here :\r\n\r\n> 3.1.2 The use of privateuse sub-tag\r\nIn light of unambiguous language codes being available for the two Khoisan varieties, we\r\npropose to combine the ISO 639-3 code for the parent language N‖ng, i.e., ‘ngh’, with the\r\nprivateuse sub-tag ‘x-’ and the respective Glottocodes stated above.\r\nThe language tags for N|uu and ‖’Au can then be defined accordingly:\r\nN|uu: ngh-x-nuuu1242\r\n‖’Au: ngh-x-auni1243\r\n\r\nBy the way, while searching for this, I came across this application: https://huggingface.co/spaces/cdleong/langcode-search",
"> > * Since Glottolog can be trust, I will add a condition to the code that if there is no ISO 639-3 code in the \"official\" database (https://iso639-3.sil.org/sites/iso639-3/files/downloads/iso-639-3.tab), look for it in the \"ISO639P3code\" column of Glottolog.\r\n> \r\n> I'm confused here... if there is no ISO639-3 code in the official database from the registrar, why would you look for an \"unofficial\" code from someone else? What is the use case here?\r\n\r\nHi @HughP, I'm happy to clear what confusion may exist here :innocent: Here is the use case. \r\nGuillaume Jacques (@rgyalrong) put together a sizeable corpus of the Japhug language. It is up on HF Datasets ([here](https://huggingface.co/datasets/Lacito/pangloss/viewer/japh1234)) as well as on Zenodo. \r\n\r\nZenodo is an all-purpose repository without adequate domain-specific metadata (\"[métadonnées métier](https://www.cines.fr/archivage/des-expertises/les-metadonnees/metadonnees-metier/)\"), and the deposits in there are not easy to locate. The Zenodo deposit is intended for a highly specific user case: someone reads about the dataset in a paper, goes to the address on Zenodo and grabs the dataset at one go. \r\n\r\nHF Datasets, on the other hand, allows users to look around among corpora. The Japhug corpus needs proper tagging so that HF Datasets users can find out about it. \r\nJaphug has an entry of its own in Glottolog, whereas it lacks an entry of its own in Ethnologue. Hence the practical usefulness of Glottolog. Ethnologue pools together, under the code `jya`, three different languages (Japhug, Tshobdun `tsho1240` and Zbu `zbua1234`). \r\n\r\nI hope that this helps.",
"> By the way, while searching for this, I came across this application: https://huggingface.co/spaces/cdleong/langcode-search\r\n\r\nReally relevant Space, so tagging its author @cdleong, just in case!",
"@cdleong A one-stop shop for language codes: terrific!\r\nHow do you feel about the use of Glottocodes? When searching the language names 'Japhug' and 'Yongning Na' (real examples, related to a HF Datasets deposit & various research projects), the relevant Glottocodes are retrieved, and that is great (and not that easy, notably with the space in the middle of 'Yongning Na'). But this positive result is 'hidden' in the results page. Specifically: \r\n\r\n- for Japhug: when searching by language name ('Japhug'), the result in big print is 'Failure', even though there is an available Glottocode (at bottom).\r\n\r\nWhen searching by Glottocode (japh1234), same outcome: 'Result: failure!' (even though this _is_ the right Glottocode\r\nWhen searching by x-japh1234 (Glottocode encapsulated in BCP 47 syntax), one gets the message \r\n\r\n> ''x-japh1234' parses meaningfully as a language tag according to IANA\"\r\n\r\nbut there is paradoxically no link provided to Glottolog: the 'Glottolog' part of the results page is empty\r\n\r\n\r\n- Yongning Na: the correct code is identified (yong1288) but instead of foregrounding this exact match, the first result that comes up is a completely different language, called 'Yong'. \r\n\r\nTrying to formulate a conclusion (admittedly, this note is not based on intensive testing, it is just feedback on initial contact): from a user perspective, it seems that the tool could make more extensive use of Glottolog. `langcode-search` does a great job querying Glottolog, why not make more extensive use of that information? (including: to arrive at the nearest ISO 639-3 code)"
] | **The problem:**
Language diversity is an important dimension of the diversity of datasets. To find one's way around datasets, being able to search by language name and by standardized codes appears crucial.
Currently the list of language codes is [here](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/resources/languages.json), right? At about 1,500 entries, it is roughly at 1/4th of the world's diversity of extant languages. (Probably less, as the list of 1,418 contains variants that are linguistically very close: 108 varieties of English, for instance.)
Looking forward to ever increasing coverage, how will the list of language names and language codes improve over time?
Enrichment of the custom list by HFT contributors (like [here](https://github.com/huggingface/datasets/pull/4880)) has several issues:
* progress is likely to be slow:
(input required from reviewers, etc.)
* the more contributors, the less consistency can be expected among contributions. No need to elaborate on how much confusion is likely to ensue as datasets accumulate.
* there is no information on which language relates with which: no encoding of the special closeness between the languages of the Northwestern Germanic branch (English+Dutch+German etc.), for instance. Information on phylogenetic closeness can be relevant to run experiments on transfer of technology from one language to its close relatives.
**A solution that seems desirable:**
Connecting to an established database that (i) aims at full coverage of the world's languages and (ii) has information on higher-level groupings, alternative names, etc.
It takes a lot of hard work to do such databases. Two important initiatives are [Ethnologue](https://www.ethnologue.com/) (ISO standard) and [Glottolog](https://glottolog.org/). Both have pros and cons. Glottolog contains references to Ethnologue identifiers, so adopting Glottolog entails getting the advantages of both sets of language codes.
Both seem technically accessible & 'developer-friendly'. Glottolog has a [GitHub repo](https://github.com/glottolog/glottolog). For Ethnologue, harvesting tools have been devised (see [here](https://github.com/lyy1994/ethnologue); I did not try it out).
In case a conversation with linguists seemed in order here, I'd be happy to participate ('pro bono', of course), & to rustle up more colleagues as useful, to help this useful development happen.
With appreciation of HFT, | 4,881 |
https://github.com/huggingface/datasets/issues/4878 | [not really a bug] `identical_ok` is deprecated in huggingface-hub's `upload_file` | [
"Resolved via https://github.com/huggingface/datasets/pull/4937."
] | In the huggingface-hub dependency, the `identical_ok` argument has no effect in `upload_file` (and it will be removed soon)
See
https://github.com/huggingface/huggingface_hub/blob/43499582b19df1ed081a5b2bd7a364e9cacdc91d/src/huggingface_hub/hf_api.py#L2164-L2169
It's used here:
https://github.com/huggingface/datasets/blob/fcfcc951a73efbc677f9def9a8707d0af93d5890/src/datasets/dataset_dict.py#L1373-L1381
https://github.com/huggingface/datasets/blob/fdcb8b144ce3ef241410281e125bd03e87b8caa1/src/datasets/arrow_dataset.py#L4354-L4362
https://github.com/huggingface/datasets/blob/fdcb8b144ce3ef241410281e125bd03e87b8caa1/src/datasets/arrow_dataset.py#L4197-L4213
We should remove it.
Maybe the third code sample has an unexpected behavior since it uses the non-default value `identical_ok = False`, but the argument is ignored. | 4,878 |
https://github.com/huggingface/datasets/issues/4876 | Move DatasetInfo from `datasets_infos.json` to the YAML tags in `README.md` | [
"also @osanseviero @Pierrci @SBrandeis potentially",
"Love this in principle 🚀 \r\n\r\nLet's keep in mind users might rely on `dataset_infos.json` already.\r\n\r\nI'm not convinced by the two-syntax solution, wouldn't it be simpler to have only one syntax with a `default` config for datasets with only one config? ie, always having the `configs` field. This makes parsing the metadata easier IMO.\r\n\r\nMight also be good to wrap the tags under a `datasets_info` tag as follows:\r\n\r\n```yaml\r\ndescription: ...\r\ncitation: ...\r\ndataset_infos:\r\n download_size: 35142551\r\n dataset_size: 89789763\r\n version: 1.0.0\r\n configs:\r\n - ...\r\n[...]\r\n```\r\n\r\nLet's also keep in mind that extracting YAML metadata from a markdown readme is a bit more fastidious for users than just parsing a JSON file.",
"> Let's keep in mind users might rely on dataset_infos.json already.\r\n\r\nYea we'll full full backward compatibility\r\n\r\n> Let's also keep in mind that extracting YAML metadata from a markdown readme is a bit more fastidious for users than just parsing a JSON file.\r\n\r\nThe main things that may use or ingest these data IMO are:\r\n- users in the UI or IDE\r\n- `datasets` to populate `DatasetInfo` python object\r\n- moon landing which is already parsing YAML\r\n\r\nAm I missing something ? If not I think it's ok to use YAML\r\n\r\n> Might also be good to wrap the tags under a datasets_info tag as follows:\r\n\r\nMaybe one single syntax like this then ?\r\n```yaml\r\ndataset_infos:\r\n- config: unlabeled\r\n download_size: 35142551\r\n dataset_size: 89789763\r\n version: 1.0.0\r\n splits:\r\n - name: train\r\n num_examples: 10000\r\n features:\r\n - name: text\r\n dtype: string\r\n- config: labeled\r\n download_size: 35142551\r\n dataset_size: 89789763\r\n version: 1.0.0\r\n splits:\r\n - name: train\r\n num_examples: 100\r\n features:\r\n - name: text\r\n dtype: string\r\n - name: label\r\n dtype: ClassLabel\r\n names:\r\n - negative\r\n - positive\r\n```\r\nand when you have only one config\r\n```yaml\r\ndataset_infos:\r\n- config: default\r\n splits:\r\n - name: train\r\n num_examples: 10000\r\n features:\r\n - name: text\r\n dtype: string\r\n```",
"love the idea, and the trend in general to move more things (like tasks) to a single place (YAML).\r\n\r\nalso, if you browse files on a dataset's page (in \"Files and versions\"), raw `README.md` files looks nice and readable, while `.json` files are just one long line that users need to scroll. \r\n\r\n> Let's also keep in mind that extracting YAML metadata from a markdown readme is a bit more fastidious for users than just parsing a JSON file.\r\n\r\ndo users often parse `datasets_infos.json` file themselves? ",
"> do users often parse datasets_infos.json file themselves?\r\n\r\nNot AFAIK, but I'm sure there should be a few users.\r\nUsers that access these info via the `DatasetInfo` from `datasets` won't see the change though e.g.\r\n```python\r\n>> from datasets import get_datasets_infos\r\n>>> get_datasets_infos(\"squad\")\r\n{'plain_text': DatasetInfo(description='Stanford Question Answering Dataset...\r\n```",
"> Maybe one single syntax like this then ?\r\n\r\nLGTM!\r\n\r\n> The main things that may use or ingest these data IMO are:\r\n> - users in the UI or IDE\r\n> - datasets to populate DatasetInfo python object\r\n> - moon landing which is already parsing YAML\r\n\r\nFair point!\r\n\r\nHaving dataset info in the README's YAML is great for API / `huggingface_hub` consumers as well as it will be inserted in the `cardData` field out of the box 🔥 \r\n",
"Very supportive of this!\r\n\r\nNesting an array of configs inside `dataset_infos: ` sounds good to me. One small tweak is that `config: default` can be optional for the default config (which can be the first one by convention)\r\n\r\nWe'll be able to implement metadata validation on the Hub side so we ensure that those metadata are always in the right format (maybe for @coyotte508 ? cc @Pierrci). From a quick glance the `features` might be the harder part to validate here, any doc will be welcome.\r\n\r\n### Other high-level points:\r\n- as we move from mostly academic datasets to *all* datasets (which include the data inside the repos), my intuition is that more and more datasets (Hub-stored) are going to be **single-config**\r\n- similarly, less and less datasets will have a loading script, **just the data + some metadata**\r\n- to lower the barrier to entry to contribution, in the long term users shouldn't need to compute/update this data via a command line. It could be filled automatically on the Hub through a \"bot\" inside Discussions & Pull requests for instance.",
"re: `config: default`\r\n\r\nNote also that the default config is not named `default`, afaiu, but create from the repo name, eg: https://huggingface.co/datasets/nbtpj/bionlp2021SAS default config is `nbtpj--bionlp2021SAS` (which is awful)",
"> Note also that the default config is not named default, afaiu, but create from the repo name, eg: https://huggingface.co/datasets/nbtpj/bionlp2021SAS default config is nbtpj--bionlp2021SAS (which is awful)\r\n\r\nWe can change this to `default` I think or something else",
"> From a quick glance the features might be the harder part to validate here, any doc will be welcome.\r\n\r\nI dug into features validation, see:\r\n\r\n- the OpenAPI spec: https://github.com/huggingface/datasets-server/blob/main/chart/static-files/openapi.json#L460-L697\r\n- the node.js code: https://github.com/huggingface/moon-landing/blob/upgrade-datasets-server-client/server/lib/datasets/FeatureType.ts",
"> We can change this to default I think or something else\r\n\r\nI created https://github.com/huggingface/datasets/issues/4902 to discuss that",
"> Note also that the default config is not named `default`, afaiu, but create from the repo name\r\n\r\nin case of single-config you can even hide the config name from the UI IMO\r\n\r\n> I dug into features validation, see: the OpenAPI spec\r\n\r\nin moon-landing we use [Joi](https://joi.dev/api/) to validate metadata so we would need to generate from Joi code from the OpenAPI spec (or from somewhere else) but I guess that's doable – or just rewrite it manually, as it won't change often",
"I remember there was an ongoing discussion on this topic:\r\n- #3507\r\n\r\nI recall some of the concerns raised on that discussion:\r\n- @lhoestq: Tensorflow Datasets catalog includes a community catalog where you can find and use HF datasets. They are using the exported dataset_infos.json files from github to get the metadata: [#3507 (comment)](https://github.com/huggingface/datasets/issues/3507#issuecomment-1056997627)\r\n- @severo: [#3507 (comment)](https://github.com/huggingface/datasets/issues/3507#issuecomment-1042779776)\r\n - the metadata header might be very long, before reaching the start of the README/dataset card. \r\n - It also somewhat prevents including large strings like the checksums\r\n - two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file. \r\n- @severo: the future \"datasets server\" could be in charge of generating the dataset-info.json file: [#3507 (comment)](https://github.com/huggingface/datasets/issues/3507#issuecomment-1033752157)",
"Thanks for bringing these points up !\r\n\r\n> @lhoestq: Tensorflow Datasets catalog includes a community catalog where you can find and use HF datasets. They are using the exported dataset_infos.json files from github to get the metadata: https://github.com/huggingface/datasets/issues/3507#issuecomment-1056997627\r\n\r\nThe TFDS implementation is not super advanced, so it's ok IMO as long as we don't break all the dataset scripts. Note that users can still use `to_tf_dataset`.\r\n\r\nWe had a chance to discuss the two nexts points with @julien-c as well:\r\n\r\n> @severo: https://github.com/huggingface/datasets/issues/3507#issuecomment-1042779776\r\nthe metadata header might be very long, before reaching the start of the README/dataset card.\r\n\r\nIf we don't add the checksums we should be fine. We can also set a maximum number of supported configs in the README to keep it readable.\r\n\r\n> @severo: the future \"datasets server\" could be in charge of generating the dataset-info.json file: https://github.com/huggingface/datasets/issues/3507#issuecomment-1033752157\r\n\r\nI guess the \"HF Hub actions\" could open PRs to do the same in the YAML directly\r\n",
"Thanks for linking that similar discussion for context, @albertvillanova!"
] | Currently there are two places to find metadata for datasets:
- datasets_infos.json, which contains **per dataset config**
- description
- citation
- license
- splits and sizes
- checksums of the data files
- feature types
- and more
- YAML tags, which contain
- license
- language
- train-eval-index
- and more
It would be nice to have a single place instead. We can rely on the YAML tags more than the JSON file for consistency with models. And it would all be indexed by our back-end directly, which is nice to have.
One way would be to move everything to the YAML tags except the checksums (there can be tens of thousands of them). The description/citation is already in the dataset card so we probably don't need to have them in the YAML card, it would be redundant.
Here is an example for SQuAD
```yaml
download_size: 35142551
dataset_size: 89789763
version: 1.0.0
splits:
- name: train
num_examples: 87599
num_bytes: 79317110
- name: validation
num_examples: 10570
num_bytes: 10472653
features:
- name: id
dtype: string
- name: title
dtype: string
- name: context
dtype: string
- name: question
dtype: string
- name: answers
struct:
- name: text
list:
dtype: string
- name: answer_start
list:
dtype: int32
```
Since there is only one configuration for SQuAD, this structure is ok. For datasets with several configs we can see in a second step, but IMO it would be ok to have these fields per config using another syntax
```yaml
configs:
- config: unlabeled
splits:
- name: train
num_examples: 10000
features:
- name: text
dtype: string
- config: labeled
splits:
- name: train
num_examples: 100
features:
- name: text
dtype: string
- name: label
dtype: ClassLabel
names:
- negative
- positive
```
So in the end you could specify a YAML tag either at the top level (for all configs) or per config in the `configs` field
Alternatively we could keep config specific stuff in the `dataset_infos.json` as it it today
Not sure yet what's the best approach here but cc @julien-c @mariosasko @albertvillanova @polinaeterna for feedback :) | 4,876 |
https://github.com/huggingface/datasets/issues/4875 | `_resolve_features` ignores the token | [
"Hi ! Your HF_ENDPOINT seems wrong because of the extra \"/\"\r\n```diff\r\n- os.environ[\"HF_ENDPOINT\"] = \"https://hub-ci.huggingface.co/\"\r\n+ os.environ[\"HF_ENDPOINT\"] = \"https://hub-ci.huggingface.co\"\r\n```\r\n\r\ncan you try again without the extra \"/\" ?",
"Oh, yes, sorry, but it's not the issue.\r\n\r\nIn my code, I set `HF_ENDPOINT=https://hub-ci.huggingface.co`. I added `os.environ[\"HF_ENDPOINT\"] = \"https://hub-ci.huggingface.co/\"` afterward just to indicate that we had to have this env var and made a mistake there",
"I can't reproduce on my side. I tried using a private dataset repo with a CSV file on hub-ci\r\n\r\nWhat's your version of `huggingface_hub` ?",
"I can't reproduce either... Not sure what has occurred, very sorry to have made you lost your time on that ",
"I got something similar in https://github.com/huggingface/datasets-server/pull/608. Look how changing the order of the tests (https://github.com/huggingface/datasets-server/pull/608/commits/2c50fe833323de3dfdc76c5cd68639279e0887f8) change the result, which means that something has a side-effect:\r\n- https://github.com/huggingface/datasets-server/actions/runs/3264636253/jobs/5365612918 works\r\n- https://github.com/huggingface/datasets-server/actions/runs/3264651839/jobs/5365654924 does not work\r\n\r\nI still couldn't reproduce it with a simpler script... ",
"The issue happens because `extend_module_for_streaming` can't be used several times on packaged builders like `csv` to apply a new auth token. Indeed `extend_module_for_streaming` only applies authentication once, and on subsequent calls does nothing:\r\n\r\nhttps://github.com/huggingface/datasets/blob/07b7c38d9e9c72c74b02524c432ca64d0d3738f4/src/datasets/streaming.py#L62-L64\r\n\r\nThis behavior exists because the authenticatoin wrapper only supports one token. This is an issue for packaged builders which can be used to load several datasets, so it may require several tokens.\r\n\r\nThis can be fixed by storing a dict `token_per_repo_id` instead of `use_auth_token` in the authentication wrapper, and by making it possible to update the authentication wrapper with a new token",
"I fixed the datasets-server CI with: https://github.com/huggingface/datasets-server/pull/608\r\n\r\nSee https://github.com/huggingface/datasets-server/actions/runs/3265359326/jobs/5367445018\r\n\r\nThanks for the help @lhoestq !",
"> This can be fixed by storing a dict token_per_repo_id instead of use_auth_token in the authentication wrapper, and by making it possible to update the authentication wrapper with a new token\r\n\r\nIf I call the module on the same repo twice: first with authentication, then without authentication, would the second call use authentication anyway? It sounds like a bug: the argument passed to the function would be silently ignored.",
"Yes exactly, this is a known bug",
"And do you think this bug could be solved as well when fixing this issue?",
"yes definitely !"
] | ## Describe the bug
When calling [`_resolve_features()`](https://github.com/huggingface/datasets/blob/54b532a8a2f5353fdb0207578162153f7b2da2ec/src/datasets/iterable_dataset.py#L1255) on a gated dataset, ie. a dataset which requires a token to be loaded, the token seems to be ignored even if it has been provided to `load_dataset` before.
## Steps to reproduce the bug
```python
import os
os.environ["HF_ENDPOINT"] = "https://hub-ci.huggingface.co/"
hf_token = "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"
from datasets import load_dataset
# public
dataset_name = "__DUMMY_DATASETS_SERVER_USER__/repo_csv_data-16612654226756"
config_name = "__DUMMY_DATASETS_SERVER_USER__--repo_csv_data-16612654226756"
split_name = "train"
iterable_dataset = load_dataset(
dataset_name,
name=config_name,
split=split_name,
streaming=True,
use_auth_token=hf_token,
)
iterable_dataset = iterable_dataset._resolve_features()
print(iterable_dataset.features)
# gated
dataset_name = "__DUMMY_DATASETS_SERVER_USER__/repo_csv_data-16612654317644"
config_name = "__DUMMY_DATASETS_SERVER_USER__--repo_csv_data-16612654317644"
split_name = "train"
iterable_dataset = load_dataset(
dataset_name,
name=config_name,
split=split_name,
streaming=True,
use_auth_token=hf_token,
)
try:
iterable_dataset = iterable_dataset._resolve_features()
except FileNotFoundError as e:
print("FAILS")
```
## Expected results
I expect to have the same result on a public dataset and on a gated (or private) dataset, if the token has been provided.
## Actual results
An exception is thrown on gated datasets.
## Environment info
- `datasets` version: 2.4.0
- Platform: Linux-5.15.0-1017-aws-x86_64-with-glibc2.35
- Python version: 3.9.6
- PyArrow version: 7.0.0
- Pandas version: 1.4.2 | 4,875 |
https://github.com/huggingface/datasets/issues/4873 | Multiple dataloader memory error | [
"Hi!\r\n\r\n200+ data loaders is a lot. Have you tried to reduce the number of datasets by concatenating/interleaving the ones with the same structure/task (the API is `{concatenate_datasets/interleave_datasets}([dset1, ..., dset_N])`)?",
"Hi @mariosasko, thank you for your reply. I tried pre-concatenating different datasets into one, but one key need is to keep each batch the same data type. Considering that the concatenate-then-segment operation for prefetched samples may span across different data types after concatenating/interleaving (cuz different data sources are mixed), any solution to remain the same data source for each batch?",
"@cyk1337 have you found any solutions to it?\r\n@mariosasko I tried with interleave_datasets to sample batches from two large datasets (wikipedia alike) and it results in out-of-memory error during data loading (16gpus, >1TB physical memory). Do you have any idea about it?"
] | For the use of multiple datasets and tasks, we use around more than 200+ dataloaders, then pass it into `dataloader1, dataloader2, ..., dataloader200=accelerate.prepare(dataloader1, dataloader2, ..., dataloader200)`
It causes the memory error when generating batches. Any solutions to it?
```bash
File "/home/xxx/my_code/src/utils/data_utils.py", line 54, in generate_batch
x = next(iterator)
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/accelerate/data_loader.py", line 301, in __iter__
for batch in super().__iter__():
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
data = self._next_data()
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 475, in _next_data
data = self._dataset_fetcher.fetch(index) # may raise StopIteration
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 28, in fetch
data.append(next(self.dataset_iter))
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/accelerate/data_loader.py", line 249, in __iter__
for element in self.dataset:
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 503, in __iter__
for key, example in self._iter():
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 500, in _iter
yield from ex_iterable
File "/home/xxx/anaconda3/envs/pt1.7/lib/python3.7/site-packages/datasets/iterable_dataset.py", line 231, in __iter__
new_key = "_".join(str(key) for key in keys)
MemoryError
``` | 4,873 |
https://github.com/huggingface/datasets/issues/4865 | Dataset Viewer issue for MoritzLaurer/multilingual_nli | [
"Thanks for reporting @MoritzLaurer.\r\n\r\nCurrently, the dataset preview is working properly: https://huggingface.co/datasets/MoritzLaurer/multilingual_nli\r\n\r\nPlease note that when a dataset is modified, it might take some time until the preview is completely updated.\r\n\r\n@severo might it be worth adding a clearer error message, something like \"The preview is updating, please retry later\"?",
"Thanks for your response. You are right, its now working well. I had waited for 30 min or so and refreshed several times and thought there was some other error. Yeah, a different error message sounds like a good idea to avoid confusion. ",
"I'm closing this issue then.",
"> @severo might it be worth adding a clearer error message, something like \"The preview is updating, please retry later\"?\r\n\r\nYes, it's a known issue, and we're about to ship a better version"
] | ### Link
_No response_
### Description
I've just uploaded a new dataset to the hub and the viewer does not work for some reason, see here: https://huggingface.co/datasets/MoritzLaurer/multilingual_nli
It displays the error:
```
Status code: 400
Exception: Status400Error
Message: The dataset does not exist.
```
Weirdly enough the dataviewer works for an earlier version of the same dataset. The only difference is that it is smaller, but I'm not aware of other changes I have made: https://huggingface.co/datasets/MoritzLaurer/multilingual_nli_test
Do you know why the dataviewer is not working?
### Owner
_No response_ | 4,865 |
https://github.com/huggingface/datasets/issues/4864 | Allow pathlib PoxisPath in Dataset.read_json | [
"This same error will occur using `ds = datasets.load_dataset('json', data_files=['test.jsonl'])`",
"@cccntu I want to make a quick fix for this, but I am struggling to find where the json dataset builder is. Do you know?",
"@vvvm23 I think you mean think:\r\n```python\r\nds = datasets.load_dataset('json', data_files=[Path('test.jsonl')])\r\n```\r\nAnd the place you want to modify is here:\r\n```\r\nutils/file_utils.py:64, in is_remote_url(url_or_filename)\r\n 63 def is_remote_url(url_or_filename: str) -> bool:\r\n---> 64 parsed = urlparse(url_or_filename)\r\n 65 return parsed.scheme in (\"http\", \"https\", \"s3\", \"gs\", \"hdfs\", \"ftp\")\r\n```\r\n\r\nProbably just need to check first if `url_or_filename` is [PathLike](https://docs.python.org/3/library/os.html#os.PathLike) and return False early.\r\n\r\nBtw, I tried installing from main, and ran my code above and got a different error. Probably because the API have changed.\r\n`AttributeError: module 'datasets' has no attribute 'read_json'`\r\n",
"> @vvvm23 I think you mean think:\r\n\r\nYou are correct, thanks!\r\n\r\n> Probably just need to check first if url_or_filename is [PathLike](https://docs.python.org/3/library/os.html#os.PathLike) and return False early.\r\n\r\nIs PathLike sufficient, or should I check the file exists here? Or is that handled later?",
"I think here we just want to avoid passing Path to urlparse. A simpler solution is to add a str() call and convert the input to string before passing to the next step. No need to check anything.",
"Above PR should do your first suggestion. Hope that works for you, as I am going on holiday and won't be able to change much :wink: "
] | **Is your feature request related to a problem? Please describe.**
```
from pathlib import Path
from datasets import Dataset
ds = Dataset.read_json(Path('data.json'))
```
causes an error
```
AttributeError: 'PosixPath' object has no attribute 'decode'
```
**Describe the solution you'd like**
It should be able to accept PosixPath and read the json from inside. | 4,864 |
https://github.com/huggingface/datasets/issues/4863 | TFDS wiki_dialog dataset to Huggingface dataset | [
"@albertvillanova any help ? The linked dataset is in beam format which is similar to wikipedia dataset in huggingface that you scripted..",
"Nvm, I was able to port it to huggingface datasets, will upload to the hub soon",
"https://huggingface.co/datasets/djaym7/wiki_dialog",
"Thanks for the addition, @djaym7."
] | ## Adding a Dataset
- **Name:** *Wiki_dialog*
- **Description: https://github.com/google-research/dialog-inpainting#:~:text=JSON%20object%2C%20for-,example,-%3A
- **Paper: https://arxiv.org/abs/2205.09073
- **Data: https://github.com/google-research/dialog-inpainting
- **Motivation:** *Research and Development on biggest corpus of dialog data*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/main/ADD_NEW_DATASET.md).
| 4,863 |
https://github.com/huggingface/datasets/issues/4862 | Got "AttributeError: 'xPath' object has no attribute 'read'" when loading an excel dataset with my own code | [
"What's more, the downloaded data is actually a folder instead of an excel file.",
"Hi hi, instead of using `download_and_extract` function, I only use `download` function: `base_dir = Path(dl_manager.download(urls))`. It turns out that the code works for `datasets==2.2.2`, however, it doesn't work with `datasets==2.4.0`. ",
"Hi @yana-xuyan, thanks for reporting.\r\n\r\nIndeed you already found the answer: an Excel file should be just downloaded and not downloaded-and-extracted.\r\n\r\nThe reason why is that if you call also extract, our library will try to infer the compression format (and extract it). And Excel files are viewed as ZIP files and extracted as so (into a directory). This is because the Office Open XML is indeed a zipped file under the hood): https://en.wikipedia.org/wiki/Office_Open_XML\r\n> Office Open XML (also informally known as OOXML) is a **zipped**, XML-based file format\r\n```python\r\nimport zipfile\r\n\r\nzipfile.is_zipfile(\"filename.xlsx\")\r\n```\r\nreturns `True`.",
"Hi @albertvillanova, thank you for your reply! Do you have any clue on why the same error still exists with `datasets==2.4.0` even after I don't extract the downloaded file? FYI, if I downgrade to `datasets==2.2.2`, the code works fine.",
"I guess this has to do with the cache: you should remove the previously-wrongly generated directory from the cache; otherwise `datasets` tries to re-use it."
] | ## Describe the bug
A clear and concise description of what the bug is.
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
# The dataset function is as follows：
from pathlib import Path
from typing import Dict, List, Tuple
import datasets
import pandas as pd
_CITATION = """\
"""
_DATASETNAME = "jadi_ide"
_DESCRIPTION = """\
"""
_HOMEPAGE = ""
_LICENSE = "Unknown"
_URLS = {
_DATASETNAME: "https://github.com/fathanick/Javanese-Dialect-Identification-from-Twitter-Data/raw/main/Update 16K_Dataset.xlsx",
}
_SOURCE_VERSION = "1.0.0"
class JaDi_Ide(datasets.GeneratorBasedBuilder):
SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
BUILDER_CONFIGS = [
NusantaraConfig(
name="jadi_ide_source",
version=SOURCE_VERSION,
description="JaDi-Ide source schema",
schema="source",
subset_id="jadi_ide",
),
]
DEFAULT_CONFIG_NAME = "source"
def _info(self) -> datasets.DatasetInfo:
if self.config.schema == "source":
features = datasets.Features(
{
"id": datasets.Value("string"),
"text": datasets.Value("string"),
"label": datasets.Value("string")
}
)
return datasets.DatasetInfo(
description=_DESCRIPTION,
features=features,
homepage=_HOMEPAGE,
license=_LICENSE,
citation=_CITATION,
)
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
"""Returns SplitGenerators."""
# Dataset does not have predetermined split, putting all as TRAIN
urls = _URLS[_DATASETNAME]
base_dir = Path(dl_manager.download_and_extract(urls))
data_files = {"train": base_dir}
return [
datasets.SplitGenerator(
name=datasets.Split.TRAIN,
gen_kwargs={
"filepath": data_files["train"],
"split": "train",
},
),
]
def _generate_examples(self, filepath: Path, split: str) -> Tuple[int, Dict]:
"""Yields examples as (key, example) tuples."""
df = pd.read_excel(filepath, engine='openpyxl')
df.columns = ["id", "text", "label"]
if self.config.schema == "source":
for row in df.itertuples():
ex = {
"id": str(row.id),
"text": row.text,
"label": row.label,
}
yield row.id, ex
```
## Expected results
Expecting to load the dataset smoothly.
## Actual results
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/load.py", line 1751, in load_dataset
use_auth_token=use_auth_token,
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/builder.py", line 705, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/builder.py", line 1227, in _download_and_prepare
super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/builder.py", line 793, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/builder.py", line 1216, in _prepare_split
desc=f"Generating {split_info.name} split",
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/tqdm/std.py", line 1195, in __iter__
for obj in iterable:
File "/home/xuyan/.cache/huggingface/modules/datasets_modules/datasets/jadi_ide/7a539f2b6f726defea8fbe36ceda17bae66c370f6d6c418e3a08d760ebef7519/jadi_ide.py", line 107, in _generate_examples
df = pd.read_excel(filepath, engine='openpyxl')
File "/home/xuyan/anaconda3/lib/python3.7/site-packages/datasets/download/streaming_download_manager.py", line 701, in xpandas_read_excel
return pd.read_excel(BytesIO(filepath_or_buffer.read()), **kwargs)
AttributeError: 'xPath' object has no attribute 'read'
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-4.15.0-142-generic-x86_64-with-debian-stretch-sid
- Python version: 3.7.4
- PyArrow version: 9.0.0
- Pandas version: 0.25.1
| 4,862 |
https://github.com/huggingface/datasets/issues/4861 | Using disk for memory with the method `from_dict` | [
"This issue was also causing an OOM in @nateraw 's workflow and shows again that behavior is confusing - we should definitely switch to using the disk IMO"
] | **Is your feature request related to a problem? Please describe.**
I start with an empty dataset. In a loop, at each iteration, I create a new dataset with the method `from_dict` (based on some data I load) and I concatenate this new dataset with the one at the previous iteration. After some iterations, I have an OOM error.
**Describe the solution you'd like**
The method `from_dict` loads the data in RAM. It could be good to add an option to use the disk instead.
**Describe alternatives you've considered**
To solve the problem, I have to do an intermediate step where I save the new datasets at each iteration with `save_to_disk`. Once it's done, I open them all and concatenate them.
| 4,861 |
https://github.com/huggingface/datasets/issues/4859 | can't install using conda on Windows 10 | [] | ## Describe the bug
I wanted to install using conda or Anaconda navigator. That didn't work, so I had to install using pip.
## Steps to reproduce the bug
conda install -c huggingface -c conda-forge datasets
## Expected results
Should have indicated successful installation.
## Actual results
Solving environment: failed with initial frozen solve. Retrying with flexible solve.
Solving environment: failed with repodata from current_repodata.json, will retry with next repodata source.
... took forever, so I cancelled it with ctrl-c
## Environment info
- `datasets` version: 2.4.0 # after installing with pip
- Platform: Windows-10-10.0.19044-SP0
- Python version: 3.9.12
- PyArrow version: 9.0.0
- Pandas version: 1.4.2
- conda version: 4.13.0
conda info
active environment : base
active env location : G:\anaconda2022
shell level : 1
user config file : C:\Users\michael\.condarc
populated config files : C:\Users\michael\.condarc
conda version : 4.13.0
conda-build version : 3.21.8
python version : 3.9.12.final.0
virtual packages : __cuda=11.1=0
__win=0=0
__archspec=1=x86_64
base environment : G:\anaconda2022 (writable)
conda av data dir : G:\anaconda2022\etc\conda
conda av metadata url : None
channel URLs : https://conda.anaconda.org/pytorch/win-64
https://conda.anaconda.org/pytorch/noarch
https://conda.anaconda.org/huggingface/win-64
https://conda.anaconda.org/huggingface/noarch
https://conda.anaconda.org/conda-forge/win-64
https://conda.anaconda.org/conda-forge/noarch
https://conda.anaconda.org/anaconda-fusion/win-64
https://conda.anaconda.org/anaconda-fusion/noarch
https://repo.anaconda.com/pkgs/main/win-64
https://repo.anaconda.com/pkgs/main/noarch
https://repo.anaconda.com/pkgs/r/win-64
https://repo.anaconda.com/pkgs/r/noarch
https://repo.anaconda.com/pkgs/msys2/win-64
https://repo.anaconda.com/pkgs/msys2/noarch
package cache : G:\anaconda2022\pkgs
C:\Users\michael\.conda\pkgs
C:\Users\michael\AppData\Local\conda\conda\pkgs
envs directories : G:\anaconda2022\envs
C:\Users\michael\.conda\envs
C:\Users\michael\AppData\Local\conda\conda\envs
platform : win-64
user-agent : conda/4.13.0 requests/2.27.1 CPython/3.9.12 Windows/10 Windows/10.0.19044
administrator : False
netrc file : None
offline mode : False
| 4,859 |
https://github.com/huggingface/datasets/issues/4858 | map() function removes columns when input_columns is not None | [
"Hi! Thanks for reporting! This looks like a bug. I've just opened a PR with the fix.",
"Awesome! Thank you. I'll close the issue once the PR gets merged. :-)",
"I guess we should reopen after the revert by:\r\n- #5006"
] | ## Describe the bug
The map function, removes features from the dataset that are not present in the _input_columns_ list of columns, despite the columns being removed not mentioned in the _remove_columns_ argument.
## Steps to reproduce the bug
```python
from datasets import Dataset
ds = Dataset.from_dict({"a" : [1,2,3],"b" : [0,1,0], "c" : [2,4,5]})
def double(x,y):
x = x*2
y = y*2
return {"d" : x, "e" : y}
ds.map(double, input_columns=["a","c"])
```
## Expected results
```
Dataset({
features: ['a', 'b', 'c', 'd', 'e'],
num_rows: 3
})
```
## Actual results
```
Dataset({
features: ['a', 'c', 'd', 'e'],
num_rows: 3
})
```
In this specific example feature **b** should not be removed.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: linux (colab)
- Python version: 3.7.13
- PyArrow version: 6.0.1
| 4,858 |
https://github.com/huggingface/datasets/issues/4857 | No preprocessed wikipedia is working on huggingface/datasets | [
"Thanks for reporting @aninrusimha.\r\n\r\nPlease, note that the preprocessed datasets are still available, as described in the dataset card, e.g.: https://huggingface.co/datasets/wikipedia\r\n```python\r\nds = load_dataset(\"wikipedia\", \"20220301.en\")\r\n``` ",
"This is working now, but I was getting an error a few days ago when running an existing script. Unfortunately I did not do a proper bug report, but for some reason I was unable to load the dataset due to a request being made to the wikimedia website. However, its working now. Thanks for the reply!"
] | ## Describe the bug
20220301 wikipedia dump has been deprecated, so now there is no working wikipedia dump on huggingface
https://huggingface.co/datasets/wikipedia
https://dumps.wikimedia.org/enwiki/
| 4,857 |
https://github.com/huggingface/datasets/issues/4856 | file missing when load_dataset with openwebtext on windows | [
"I have tried to extract ```0015896-b1054262f7da52a0518521e29c8e352c.txt``` from ```17ecf461bfccd469a1fbc264ccb03731f8606eea7b3e2e8b86e13d18040bf5b3/urlsf_subset00-16_data.xz``` with 7-zip\r\nand put the file into cache_path ```F://huggingface/datasets/downloads/extracted/0901d27f43b7e9ac0577da0d0061c8c632ba0b70ecd1b4bfb21562d9b7486faa```\r\nthere is still raise the same error and I find the file was removed from cache_path after I run the run_mlm.py with ```python run_mlm.py --model_type roberta --tokenizer_name roberta-base --dataset_name openwebtext --per_device_train_batch_size 8 --per_device_eval_batch_size 8 --do_train --do_eval --output_dir F:/model/roberta-base```."
] | ## Describe the bug
0015896-b1054262f7da52a0518521e29c8e352c.txt is missing when I run run_mlm.py with openwebtext. I check the cache_path and can not find 0015896-b1054262f7da52a0518521e29c8e352c.txt. but I can find this file in the 17ecf461bfccd469a1fbc264ccb03731f8606eea7b3e2e8b86e13d18040bf5b3/urlsf_subset00-16_data.xz with 7-zip.
## Steps to reproduce the bug
```sh
python run_mlm.py --model_type roberta --tokenizer_name roberta-base --dataset_name openwebtext --per_device_train_batch_size 8 --per_device_eval_batch_size 8 --do_train --do_eval --output_dir F:/model/roberta-base
```
or
```python
from datasets import load_dataset
load_dataset("openwebtext", None, cache_dir=None, use_auth_token=None)
```
## Expected results
Loading is successful
## Actual results
Traceback (most recent call last):
File "D:\Python\v3.8.5\lib\site-packages\datasets\builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "D:\Python\v3.8.5\lib\site-packages\datasets\builder.py", line 1227, in _download_and_prepare
super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File "D:\Python\v3.8.5\lib\site-packages\datasets\builder.py", line 795, in _download_and_prepare
raise OSError(
OSError: Cannot find data file.
Original error:
[Errno 22] Invalid argument: 'F://huggingface/datasets/downloads/extracted/0901d27f43b7e9ac0577da0d0061c8c632ba0b70ecd1b4bfb21562d9b7486faa/0015896-b1054262f7da52a0518521e29c8e352c.txt'
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: windows
- Python version: 3.8.5
- PyArrow version: 9.0.0
| 4,856 |
https://github.com/huggingface/datasets/issues/4855 | Dataset Viewer issue for super_glue | [
"Thanks for reporting @wzsxxa.\r\n\r\nHowever the \"super_glue\" dataset is rendered properly by the Dataset preview: https://huggingface.co/datasets/super_glue"
] | ### Link
https://huggingface.co/datasets/super_glue
### Description
can't view super_glue dataset on the web page
### Owner
_No response_ | 4,855 |
https://github.com/huggingface/datasets/issues/4852 | Bug in multilingual_with_para config of exams dataset and checksums error | [
"Hi @albertvillanova. Unfortunately I still get this error. Is this because the merge has yet to be released? Is there a way to track the release?",
"Hi @thesofakillers, yes you are right: the fix will be available after next release (it was planned for today; Monday at the latest).\r\n\r\nIn the meantime, you can use the version of the `exams` on our main branch by passing `revision` to `load_dataset`:\r\n```python\r\nds = load_dataset(\"exams\", revision=\"main\")\r\n```"
] | ## Describe the bug
There is a bug for "multilingual_with_para" config in exams dataset:
```python
ds = load_dataset("./datasets/exams", split="train")
```
raises:
```
KeyError: 'choices'
```
Moreover, there is a NonMatchingChecksumError:
```
NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://github.com/mhardalov/exams-qa/raw/main/data/exams/multilingual/with_paragraphs/train_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/multilingual/with_paragraphs/dev_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/multilingual/with_paragraphs/test_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/test_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_bg_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_bg_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_hr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_hr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_hu_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_hu_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_it_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_it_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_mk_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_mk_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_pl_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_pl_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_pt_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_pt_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_sq_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_sq_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_sr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_sr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_tr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_tr_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/train_vi_with_para.jsonl.tar.gz', 'https://github.com/mhardalov/exams-qa/raw/main/data/exams/cross-lingual/with_paragraphs/dev_vi_with_para.jsonl.tar.gz']
```
CC: @thesofakillers | 4,852 |
https://github.com/huggingface/datasets/issues/4840 | Dataset Viewer issue for darragh/demo_data_raw3 | [
"do you have an idea of why it can occur @huggingface/datasets? The dataset consists of a single parquet file.",
"Thanks for reporting @severo.\r\n\r\nI'm not able to reproduce that error. I get instead:\r\n```\r\nFileNotFoundError: [Errno 2] No such file or directory: 'orix/data/ChiSig/唐合乐-9-3.jpg'\r\n```\r\n\r\nWhich pyarrow version are you using? Mine is 6.0.1. ",
"OK, I get now your error when not streaming.",
"OK!\r\n\r\nIf it's useful, the pyarrow version is 7.0.0:\r\n\r\nhttps://github.com/huggingface/datasets-server/blob/487c39d87998f8d5a35972f1027d6c8e588e622d/services/worker/poetry.lock#L1537-L1543",
"Apparently, there is something weird with that Parquet file: its schema is:\r\n```\r\nimages: extension<arrow.py_extension_type<pyarrow.lib.UnknownExtensionType>>\r\n```\r\n\r\nI have forced a right schema:\r\n```python\r\nfrom datasets import Features, Image, load_dataset\r\n\r\nfeatures = Features({\"images\": Image()})\r\nds = datasets.load_dataset(\"parquet\", split=\"train\", data_files=\"train-00000-of-00001.parquet\", features=features)\r\n```\r\nand then recreated a new Parquet file:\r\n```python\r\nds.to_parquet(\"train.parquet\")\r\n```\r\n\r\nNow this Parquet file has the right schema:\r\n```\r\nimages: struct<bytes: binary, path: string>\r\n child 0, bytes: binary\r\n child 1, path: string\r\n```\r\nand can be loaded normally:\r\n```python\r\nIn [26]: ds = load_dataset(\"parquet\", split=\"train\", data_files=\"dataset.parquet\")\r\nn [27]: ds\r\nOut[27]: \r\nDataset({\r\n features: ['images'],\r\n num_rows: 20\r\n})\r\n```"
] | ### Link
https://huggingface.co/datasets/darragh/demo_data_raw3
### Description
```
Exception: ValueError
Message: Arrow type extension<arrow.py_extension_type<pyarrow.lib.UnknownExtensionType>> does not have a datasets dtype equivalent.
```
reported by @NielsRogge
### Owner
No | 4,840 |
https://github.com/huggingface/datasets/issues/4839 | ImageFolder dataset builder does not read the validation data set if it is named as "val" | [
"#take"
] | **Is your feature request related to a problem? Please describe.**
Currently, the `'imagefolder'` data set builder in [`load_dataset()`](https://github.com/huggingface/datasets/blob/2.4.0/src/datasets/load.py#L1541] ) only [supports](https://github.com/huggingface/datasets/blob/6c609a322da994de149b2c938f19439bca99408e/src/datasets/data_files.py#L31) the following names as the validation data set directory name: `["validation", "valid", "dev"]`. When the validation directory is named as `'val'`, the Data set will not have a validation split. I expected this to be a trivial task but ended up spending a lot of time before knowing that only the above names are supported.
Here's a minimal example of `val` not being recognized:
```python
import os
import numpy as np
import cv2
from datasets import load_dataset
# creating a dummy data set with the following structure:
# ROOT
# | -- train
# | ---- class_1
# | ---- class_2
# | -- val
# | ---- class_1
# | ---- class_2
ROOT = "data"
for which in ["train", "val"]:
for class_name in ["class_1", "class_2"]:
dir_name = os.path.join(ROOT, which, class_name)
if not os.path.exists(dir_name):
os.makedirs(dir_name)
for i in range(10):
cv2.imwrite(
os.path.join(dir_name, f"{i}.png"),
np.random.random((224, 224))
)
# trying to create a data set
dataset = load_dataset(
"imagefolder",
data_dir=ROOT
)
>> dataset
DatasetDict({
train: Dataset({
features: ['image', 'label'],
num_rows: 20
})
})
# ^ note how the dataset only has a 'train' subset
```
**Describe the solution you'd like**
The suggestion is to include `"val"` to [that list ](https://github.com/huggingface/datasets/blob/6c609a322da994de149b2c938f19439bca99408e/src/datasets/data_files.py#L31) as that's a commonly used phrase to name the validation directory.
Also, In the documentation, explicitly mention that only such directory names are supported as train/val/test directories to avoid confusion.
**Describe alternatives you've considered**
In the documentation, explicitly mention that only such directory names are supported as train/val/test directories without adding `val` to the above list.
**Additional context**
A question asked in the forum: [
Loading an imagenet-style image dataset with train/val directories](https://discuss.huggingface.co/t/loading-an-imagenet-style-image-dataset-with-train-val-directories/21554) | 4,839 |
https://github.com/huggingface/datasets/issues/4836 | Is it possible to pass multiple links to a split in load script? | [] | **Is your feature request related to a problem? Please describe.**
I wanted to use a python loading script in hugging face datasets that use different sources of text (it's somehow a compilation of multiple datasets + my own dataset) based on how `load_dataset` [works](https://huggingface.co/docs/datasets/loading) I assumed I could do something like bellow in my loading script:
```python
...
_URL = "MY_DATASET_URL/resolve/main/data/"
_URLS = {
"train": [
"FIRST_URL_TO.txt",
_URL + "train-00000-of-00001-676bfebbc8742592.parquet"
]
}
...
```
but when loading the dataset it raises the following error:
```python
File ~/.local/lib/python3.8/site-packages/datasets/builder.py:704, in DatasetBuilder.download_and_prepare(self, download_config, download_mode, ignore_verifications, try_from_hf_gcs, dl_manager, base_path, use_auth_token, **download_and_prepare_kwargs)
702 logger.warning("HF google storage unreachable. Downloading and preparing it from source")
703 if not downloaded_from_gcs:
--> 704 self._download_and_prepare(
705 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
...
668 if isinstance(a, str):
669 # Force-cast str subclasses to str (issue #21127)
670 parts.append(str(a))
TypeError: expected str, bytes or os.PathLike object, not list
```
**Describe the solution you'd like**
I believe since it's possible for `load_dataset` to get list of URLs instead of just a URL for `train` split it can be possible here too.
**Describe alternatives you've considered**
An alternative solution would be to download all needed datasets locally and `push_to_hub` them all, but since the datasets I'm talking about are huge it's not among my options.
**Additional context**
I think loading `text` beside the `parquet` is completely a different issue but I believe I can figure it out by proposing a config for my dataset to load each entry of `_URLS['train']` separately either by `load_dataset("text", ...` or `load_dataset("parquet", ...`.
| 4,836 |
https://github.com/huggingface/datasets/issues/4829 | Misalignment between card tag validation and docs | [
"(Note that the doc is aligned with the hub validation rules, and the \"ground truth\" is the hub validation rules given that they apply to all datasets, not just the canonical ones)",
"Instead of our own implementation, we now use `huggingface_hub`'s `DatasetCardData`, which has the correct type hint, so I think we can close this issue."
] | ## Describe the bug
As pointed out in other issue: https://github.com/huggingface/datasets/pull/4827#discussion_r943536284
the validation of the dataset card tags is not aligned with its documentation: e.g.
- implementation: `license: List[str]`
- docs: `license: Union[str, List[str]]`
They should be aligned.
CC: @julien-c
| 4,829 |
https://github.com/huggingface/datasets/issues/4820 | Terminating: fork() called from a process already using GNU OpenMP, this is unsafe. | [
"Fixed by installing either resampy<3 or resampy>=4"
] | Hi, when i try to run prepare_dataset function in [fine tuning ASR tutorial 4](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_tuning_Wav2Vec2_for_English_ASR.ipynb) , i got this error.
I got this error
Terminating: fork() called from a process already using GNU OpenMP, this is unsafe.
There is no other logs available, so i have no clue what is the cause of it.
```
def prepare_dataset(batch):
audio = batch["path"]
# batched output is "un-batched"
batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
batch["input_length"] = len(batch["input_values"])
with processor.as_target_processor():
batch["labels"] = processor(batch["text"]).input_ids
return batch
data = data.map(prepare_dataset, remove_columns=data.column_names["train"],
num_proc=4)
```
Specify the actual results or traceback.
There is no traceback except
`Terminating: fork() called from a process already using GNU OpenMP, this is unsafe.`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.4.0
- Platform: Linux-5.15.0-43-generic-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 9.0.0
- Pandas version: 1.4.3
| 4,820 |
https://github.com/huggingface/datasets/issues/4817 | Outdated Link for mkqa Dataset | [
"Thanks for reporting @liaeh, we are investigating this. "
] | ## Describe the bug
The URL used to download the mkqa dataset is outdated. It seems the URL to download the dataset is currently https://github.com/apple/ml-mkqa/blob/main/dataset/mkqa.jsonl.gz instead of https://github.com/apple/ml-mkqa/raw/master/dataset/mkqa.jsonl.gz (master branch has been renamed to main).
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset("mkqa")
```
## Expected results
downloads the dataset
## Actual results
```python
Downloading builder script:
4.79k/? [00:00<00:00, 201kB/s]
Downloading metadata:
13.2k/? [00:00<00:00, 504kB/s]
Downloading and preparing dataset mkqa/mkqa (download: 11.35 MiB, generated: 34.29 MiB, post-processed: Unknown size, total: 45.65 MiB) to /home/lhr/.cache/huggingface/datasets/mkqa/mkqa/1.0.0/5401489c674c81257cf563417aaaa5de2c7e26a1090ce9b10eb0404f10003d4d...
Downloading data files: 0%
0/1 [00:00<?, ?it/s]
---------------------------------------------------------------------------
FileNotFoundError Traceback (most recent call last)
Input In [3], in <cell line: 3>()
1 from datasets import load_dataset
----> 3 dataset = load_dataset("mkqa")
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/load.py:1746, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, **config_kwargs)
1743 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES
1745 # Download and prepare data
-> 1746 builder_instance.download_and_prepare(
1747 download_config=download_config,
1748 download_mode=download_mode,
1749 ignore_verifications=ignore_verifications,
1750 try_from_hf_gcs=try_from_hf_gcs,
1751 use_auth_token=use_auth_token,
1752 )
1754 # Build dataset for splits
1755 keep_in_memory = (
1756 keep_in_memory if keep_in_memory is not None else is_small_dataset(builder_instance.info.dataset_size)
1757 )
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/builder.py:704, in DatasetBuilder.download_and_prepare(self, download_config, download_mode, ignore_verifications, try_from_hf_gcs, dl_manager, base_path, use_auth_token, **download_and_prepare_kwargs)
702 logger.warning("HF google storage unreachable. Downloading and preparing it from source")
703 if not downloaded_from_gcs:
--> 704 self._download_and_prepare(
705 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
706 )
707 # Sync info
708 self.info.dataset_size = sum(split.num_bytes for split in self.info.splits.values())
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/builder.py:1227, in GeneratorBasedBuilder._download_and_prepare(self, dl_manager, verify_infos)
1226 def _download_and_prepare(self, dl_manager, verify_infos):
-> 1227 super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/builder.py:771, in DatasetBuilder._download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)
769 split_dict = SplitDict(dataset_name=self.name)
770 split_generators_kwargs = self._make_split_generators_kwargs(prepare_split_kwargs)
--> 771 split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
773 # Checksums verification
774 if verify_infos and dl_manager.record_checksums:
File ~/.cache/huggingface/modules/datasets_modules/datasets/mkqa/5401489c674c81257cf563417aaaa5de2c7e26a1090ce9b10eb0404f10003d4d/mkqa.py:130, in Mkqa._split_generators(self, dl_manager)
128 # download and extract URLs
129 urls_to_download = _URLS
--> 130 downloaded_files = dl_manager.download_and_extract(urls_to_download)
132 return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_files["train"]})]
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/download/download_manager.py:431, in DownloadManager.download_and_extract(self, url_or_urls)
415 def download_and_extract(self, url_or_urls):
416 """Download and extract given url_or_urls.
417
418 Is roughly equivalent to:
(...)
429 extracted_path(s): `str`, extracted paths of given URL(s).
430 """
--> 431 return self.extract(self.download(url_or_urls))
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/download/download_manager.py:309, in DownloadManager.download(self, url_or_urls)
306 download_func = partial(self._download, download_config=download_config)
308 start_time = datetime.now()
--> 309 downloaded_path_or_paths = map_nested(
310 download_func,
311 url_or_urls,
312 map_tuple=True,
313 num_proc=download_config.num_proc,
314 disable_tqdm=not is_progress_bar_enabled(),
315 desc="Downloading data files",
316 )
317 duration = datetime.now() - start_time
318 logger.info(f"Downloading took {duration.total_seconds() // 60} min")
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/utils/py_utils.py:393, in map_nested(function, data_struct, dict_only, map_list, map_tuple, map_numpy, num_proc, types, disable_tqdm, desc)
391 num_proc = 1
392 if num_proc <= 1 or len(iterable) <= num_proc:
--> 393 mapped = [
394 _single_map_nested((function, obj, types, None, True, None))
395 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)
396 ]
397 else:
398 split_kwds = [] # We organize the splits ourselve (contiguous splits)
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/utils/py_utils.py:394, in <listcomp>(.0)
391 num_proc = 1
392 if num_proc <= 1 or len(iterable) <= num_proc:
393 mapped = [
--> 394 _single_map_nested((function, obj, types, None, True, None))
395 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)
396 ]
397 else:
398 split_kwds = [] # We organize the splits ourselve (contiguous splits)
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/utils/py_utils.py:330, in _single_map_nested(args)
328 # Singleton first to spare some computation
329 if not isinstance(data_struct, dict) and not isinstance(data_struct, types):
--> 330 return function(data_struct)
332 # Reduce logging to keep things readable in multiprocessing with tqdm
333 if rank is not None and logging.get_verbosity() < logging.WARNING:
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/download/download_manager.py:335, in DownloadManager._download(self, url_or_filename, download_config)
332 if is_relative_path(url_or_filename):
333 # append the relative path to the base_path
334 url_or_filename = url_or_path_join(self._base_path, url_or_filename)
--> 335 return cached_path(url_or_filename, download_config=download_config)
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/utils/file_utils.py:185, in cached_path(url_or_filename, download_config, **download_kwargs)
181 url_or_filename = str(url_or_filename)
183 if is_remote_url(url_or_filename):
184 # URL, so get it from the cache (downloading if necessary)
--> 185 output_path = get_from_cache(
186 url_or_filename,
187 cache_dir=cache_dir,
188 force_download=download_config.force_download,
189 proxies=download_config.proxies,
190 resume_download=download_config.resume_download,
191 user_agent=download_config.user_agent,
192 local_files_only=download_config.local_files_only,
193 use_etag=download_config.use_etag,
194 max_retries=download_config.max_retries,
195 use_auth_token=download_config.use_auth_token,
196 ignore_url_params=download_config.ignore_url_params,
197 download_desc=download_config.download_desc,
198 )
199 elif os.path.exists(url_or_filename):
200 # File, and it exists.
201 output_path = url_or_filename
File ~/repos/punc-cap/venv/lib/python3.9/site-packages/datasets/utils/file_utils.py:530, in get_from_cache(url, cache_dir, force_download, proxies, etag_timeout, resume_download, user_agent, local_files_only, use_etag, max_retries, use_auth_token, ignore_url_params, download_desc)
525 raise FileNotFoundError(
526 f"Cannot find the requested files in the cached path at {cache_path} and outgoing traffic has been"
527 " disabled. To enable file online look-ups, set 'local_files_only' to False."
528 )
529 elif response is not None and response.status_code == 404:
--> 530 raise FileNotFoundError(f"Couldn't find file at {url}")
531 _raise_if_offline_mode_is_enabled(f"Tried to reach {url}")
532 if head_error is not None:
FileNotFoundError: Couldn't find file at https://github.com/apple/ml-mkqa/raw/master/dataset/mkqa.jsonl.gz
```
## Environment info
- `datasets` version: 2.4.0
- Platform: Linux-5.13.0-40-generic-x86_64-with-glibc2.31
- Python version: 3.9.7
- PyArrow version: 9.0.0
- Pandas version: 1.4.2
| 4,817 |
https://github.com/huggingface/datasets/issues/4815 | Outdated loading script for OPUS ParaCrawl dataset | [] | ## Describe the bug
Our loading script for OPUS ParaCrawl loads its 7.1 version. Current existing version is 9.
| 4,815 |
https://github.com/huggingface/datasets/issues/4814 | Support CSV as metadata file format in AudioFolder/ImageFolder | [] | Requested here: https://discuss.huggingface.co/t/how-to-structure-an-image-dataset-repo-using-the-image-folder-approach/21004. CSV is also used in AutoTrain for specifying metadata in image datasets. | 4,814 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.