html_url
stringlengths 48
51
| title
stringlengths 1
290
| comments
sequencelengths 0
30
| body
stringlengths 0
228k
β | number
int64 2
7.08k
|
|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/4612 | Release 2.3.0 broke custom iterable datasets | [
"Apparently, `fsspec` does not allow access to attribute-based modules anymore, such as `fsspec.async`.\r\n\r\nHowever, this is a fairly simple fix:\r\n- Change the import to: `from fsspec import asyn`;\r\n- Change line 18 to: `asyn.iothread[0] = None`;\r\n- Change line 19 to `asyn.loop[0] = None`.",
"Hi! I think it's easier to replace `import fsspec` with `import fsspec.asyn` and leave the rest unchanged. @gugarosa Are you interested in submitting a PR?",
"Perfect, it is even better!\r\n\r\nJust submitted the PR: #4630.\r\n\r\nThank you!"
] | ## Describe the bug
Trying to iterate examples from custom iterable dataset fails to bug introduced in `torch_iterable_dataset.py` since the release of 2.3.0.
## Steps to reproduce the bug
```python
next(iter(custom_iterable_dataset))
```
## Expected results
`next(iter(custom_iterable_dataset))` should return examples from the dataset
## Actual results
```
/usr/local/lib/python3.7/dist-packages/datasets/formatting/dataset_wrappers/torch_iterable_dataset.py in _set_fsspec_for_multiprocess()
16 See https://github.com/fsspec/gcsfs/issues/379
17 """
---> 18 fsspec.asyn.iothread[0] = None
19 fsspec.asyn.loop[0] = None
20
AttributeError: module 'fsspec' has no attribute 'asyn'
```
## Environment info
- `datasets` version: 2.3.0
- Platform: Linux-5.4.188+-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.13
- PyArrow version: 8.0.0
- Pandas version: 1.3.5
| 4,612 |
https://github.com/huggingface/datasets/issues/4610 | codeparrot/github-code failing to load | [
"I believe the issue is in `codeparrot/github-code`. `base_path` param is missing - https://huggingface.co/datasets/codeparrot/github-code/blob/main/github-code.py#L169\r\n\r\nFunction definition has changed.\r\nhttps://github.com/huggingface/datasets/blob/0e1c629cfb9f9ba124537ba294a0ec451584da5f/src/datasets/data_files.py#L547\r\n\r\n@mariosasko could you please confirm my finding? And are there any changes that need to be done from my side?",
"Good catch ! We recently did a breaking change in `get_patterns_in_dataset_repository`, I think we can revert it",
"> Good catch ! We recently did a breaking change in `get_patterns_in_dataset_repository`, I think we can revert it\n\nI can't wait for that releasee. Broke my application",
"This simple workaround should fix: https://huggingface.co/datasets/codeparrot/github-code/discussions/2\r\n\r\n`get_patterns_in_dataset_repository` can treat whether `base_path=None`, so we just need to make sure that codeparrot/github-code `_split_generators` calls with such an argument.",
"I am afraid your suggested change @gugarosa will break compatibility with older datasets versions that don't have `base_path` argument in `get_patterns_in_dataset_repository`, as a workaround while the issue gets resolved in `datasets` can you downgrade your datasets version to `<=2.1.0` ? \r\n@lvwerra do you think we should adapt the script to check the datasets version before calling `get_patterns_in_dataset_repository`?",
"Actually I think it's just simpler to fix it in the dataset itself, let me open a PR\r\n\r\nEDIT: PR opened here: https://huggingface.co/datasets/codeparrot/github-code/discussions/3",
"PR is merged, it's working now ! Closing this one :)",
"> I am afraid your suggested change @gugarosa will break compatibility with older datasets versions that don't have `base_path` argument in `get_patterns_in_dataset_repository`, as a workaround while the issue gets resolved in `datasets` can you downgrade your datasets version to `<=2.1.0` ?\r\n> @lvwerra do you think we should adapt the script to check the datasets version before calling `get_patterns_in_dataset_repository`?\r\n\r\nYou are definitely right, sorry about it. I always keep forgetting that we need to keep in mind users from past versions, my bad."
] | ## Describe the bug
codeparrot/github-code fails to load with a `TypeError: get_patterns_in_dataset_repository() missing 1 required positional argument: 'base_path'`
## Steps to reproduce the bug
```python
from datasets import load_dataset
```
## Expected results
loaded dataset object
## Actual results
```python
[3]: dataset = load_dataset("codeparrot/github-code")
No config specified, defaulting to: github-code/all-all
Downloading and preparing dataset github-code/all-all to /home/bebr/.cache/huggingface/datasets/codeparrot___github-code/all-all/0.0.0/a55513bc0f81db773f9896c7aac225af0cff5b323bb9d2f68124f0a8cc3fb817...
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
Input In [3], in <cell line: 1>()
----> 1 dataset = load_dataset("codeparrot/github-code")
File ~/miniconda3/envs/fastapi-kube/lib/python3.10/site-packages/datasets/load.py:1679, 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)
1676 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES
1678 # Download and prepare data
-> 1679 builder_instance.download_and_prepare(
1680 download_config=download_config,
1681 download_mode=download_mode,
1682 ignore_verifications=ignore_verifications,
1683 try_from_hf_gcs=try_from_hf_gcs,
1684 use_auth_token=use_auth_token,
1685 )
1687 # Build dataset for splits
1688 keep_in_memory = (
1689 keep_in_memory if keep_in_memory is not None else is_small_dataset(builder_instance.info.dataset_size)
1690 )
File ~/miniconda3/envs/fastapi-kube/lib/python3.10/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 ~/miniconda3/envs/fastapi-kube/lib/python3.10/site-packages/datasets/builder.py:1221, in GeneratorBasedBuilder._download_and_prepare(self, dl_manager, verify_infos)
1220 def _download_and_prepare(self, dl_manager, verify_infos):
-> 1221 super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File ~/miniconda3/envs/fastapi-kube/lib/python3.10/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/codeparrot--github-code/a55513bc0f81db773f9896c7aac225af0cff5b323bb9d2f68124f0a8cc3fb817/github-code.py:169, in GithubCode._split_generators(self, dl_manager)
162 def _split_generators(self, dl_manager):
164 hfh_dataset_info = HfApi(datasets.config.HF_ENDPOINT).dataset_info(
165 _REPO_NAME,
166 timeout=100.0,
167 )
--> 169 patterns = datasets.data_files.get_patterns_in_dataset_repository(hfh_dataset_info)
170 data_files = datasets.data_files.DataFilesDict.from_hf_repo(
171 patterns,
172 dataset_info=hfh_dataset_info,
173 )
175 files = dl_manager.download_and_extract(data_files["train"])
TypeError: get_patterns_in_dataset_repository() missing 1 required positional argument: 'base_path'
```
## Environment info
- `datasets` version: 2.3.2
- Platform: Linux-5.18.7-arch1-1-x86_64-with-glibc2.35
- Python version: 3.10.5
- PyArrow version: 8.0.0
- Pandas version: 1.4.2 | 4,610 |
https://github.com/huggingface/datasets/issues/4609 | librispeech dataset has to download whole subset when specifing the split to use | [
"Hi! You can use streaming to fetch only a subset of the data:\r\n```python\r\nraw_dataset = load_dataset(\"librispeech_asr\", \"clean\", split=\"train.100\", streaming=True)\r\n```\r\nAlso, we plan to make it possible to download a particular split in the non-streaming mode, but this task is not easy due to how our dataset scripts are structured.",
"Hi,\r\n\r\nThat's a great help. Thank you very much."
] | ## Describe the bug
librispeech dataset has to download whole subset when specifing the split to use
## Steps to reproduce the bug
see below
# Sample code to reproduce the bug
```
!pip install datasets
from datasets import load_dataset
raw_dataset = load_dataset("librispeech_asr", "clean", split="train.100")
```
## Expected results
The split "train.clean.100" is downloaded.
## Actual results
All four splits in "clean" subset is downloaded.
## Environment info
- `datasets` version: 2.3.2
- 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
| 4,609 |
https://github.com/huggingface/datasets/issues/4606 | evaluation result changes after `datasets` version change | [
"Hi! The GH/no-namespace datasets versioning is synced with the version of the `datasets` lib, which means that the `wikiann` script was modified between the two compared versions. In this scenario, you can ensure reproducibility by pinning the script version, which is done by passing `revision=\"x.y.z\"` (e.g. `revision=\"2.2.0\"`) to `load_dataset.`\r\n"
] | ## Describe the bug
evaluation result changes after `datasets` version change
## Steps to reproduce the bug
1. Train a model on WikiAnn
2. reload the ckpt -> test accuracy becomes same as eval accuracy
3. such behavior is gone after downgrading `datasets`
https://colab.research.google.com/drive/1kYz7-aZRGdayaq-gDTt30tyEgsKlpYOw?usp=sharing
## Expected results
evaluation result shouldn't change before/after `datasets` version changes
## Actual results
evaluation result changes before/after `datasets` version changes
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.2
- Platform: colab
- Python version: 3.7.13
- PyArrow version: 6.0.1
Q. How could the evaluation result change before/after `datasets` version changes? | 4,606 |
https://github.com/huggingface/datasets/issues/4605 | Dataset Viewer issue for boris/gis_filtered | [
"Yes, this dataset is \"gated\": you first have to go to https://huggingface.co/datasets/boris/gis_filtered and click \"Access repository\" (if you accept to share your contact information with the repository authors).",
"I already did that, it returns error when using streaming",
"Oh, sorry, I misread. Looking at it. Maybe @huggingface/datasets or @SBrandeis ",
"I could reproduce the error, even though I provided my token and accepted the gate form. It looks like an error from `datasets`",
"This is indeed a bug in `datasets`. Parquet datasets in gated/private repositories can't be streamed properly, which caused the viewer to fail. I opened a PR at https://github.com/huggingface/datasets/pull/4608"
] | ### Link
https://huggingface.co/datasets/boris/gis_filtered/viewer/boris--gis_filtered/train
### Description
When I try to access this from the website I get this error:
Status code: 400
Exception: ClientResponseError
Message: 401, message='Unauthorized', url=URL('https://huggingface.co/datasets/boris/gis_filtered/resolve/80b805053ce61d4eb487b6b8d9095d775c2c466e/data/train/0000.parquet')
If I try to load with code I also get the same issue:
```python
dataset2_train=load_dataset("boris/gis_filtered", use_auth_token=os.environ["HF_TOKEN"],split="train",streaming=True)
dataset2_validation=load_dataset("boris/gis_filtered", use_auth_token=os.environ["HF_TOKEN"], split="validation",streaming=True)
```
### Owner
No | 4,605 |
https://github.com/huggingface/datasets/issues/4603 | CI fails recurrently and randomly on Windows | [] | As reported by @lhoestq,
The windows CI is currently flaky: some dependencies like `aiobotocore`, `multiprocess` and `seqeval` sometimes fail to install.
In particular it seems that building the wheels fail. Here is an example of logs:
```
Building wheel for seqeval (setup.py): started
Running command 'C:\tools\miniconda3\envs\py37\python.exe' -u -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'C:\\Users\\circleci\\AppData\\Local\\Temp\\pip-install-h55pfgbv\\seqeval_d6cdb9d23ff6490b98b6c4bcaecb516e\\setup.py'"'"'; __file__='"'"'C:\\Users\\circleci\\AppData\\Local\\Temp\\pip-install-h55pfgbv\\seqeval_d6cdb9d23ff6490b98b6c4bcaecb516e\\setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' bdist_wheel -d 'C:\Users\circleci\AppData\Local\Temp\pip-wheel-x3cc8ym6'
No parent package detected, impossible to derive `name`
running bdist_wheel
running build
running build_py
package init file 'seqeval\__init__.py' not found (or not a regular file)
package init file 'seqeval\metrics\__init__.py' not found (or not a regular file)
C:\tools\miniconda3\envs\py37\lib\site-packages\setuptools\command\install.py:37: SetuptoolsDeprecationWarning: setup.py install is deprecated. Use build and pip and other standards-based tools.
setuptools.SetuptoolsDeprecationWarning,
installing to build\bdist.win-amd64\wheel
running install
running install_lib
warning: install_lib: 'build\lib' does not exist -- no Python modules to install
running install_egg_info
running egg_info
creating UNKNOWN.egg-info
writing UNKNOWN.egg-info\PKG-INFO
writing dependency_links to UNKNOWN.egg-info\dependency_links.txt
writing top-level names to UNKNOWN.egg-info\top_level.txt
writing manifest file 'UNKNOWN.egg-info\SOURCES.txt'
reading manifest file 'UNKNOWN.egg-info\SOURCES.txt'
writing manifest file 'UNKNOWN.egg-info\SOURCES.txt'
Copying UNKNOWN.egg-info to build\bdist.win-amd64\wheel\.\UNKNOWN-0.0.0-py3.7.egg-info
running install_scripts
creating build\bdist.win-amd64\wheel\UNKNOWN-0.0.0.dist-info\WHEEL
creating 'C:\Users\circleci\AppData\Local\Temp\pip-wheel-x3cc8ym6\UNKNOWN-0.0.0-py3-none-any.whl' and adding 'build\bdist.win-amd64\wheel' to it
adding 'UNKNOWN-0.0.0.dist-info/METADATA'
adding 'UNKNOWN-0.0.0.dist-info/WHEEL'
adding 'UNKNOWN-0.0.0.dist-info/top_level.txt'
adding 'UNKNOWN-0.0.0.dist-info/RECORD'
removing build\bdist.win-amd64\wheel
Building wheel for seqeval (setup.py): finished with status 'done'
Created wheel for seqeval: filename=UNKNOWN-0.0.0-py3-none-any.whl size=963 sha256=67eb93a6e1ff4796c5882a13f9fa25bb0d3d103796e2525f9cecf3b2ef26d4b1
Stored in directory: c:\users\circleci\appdata\local\pip\cache\wheels\05\96\ee\7cac4e74f3b19e3158dce26a20a1c86b3533c43ec72a549fd7
WARNING: Built wheel for seqeval is invalid: Wheel has unexpected file name: expected 'seqeval', got 'UNKNOWN'
``` | 4,603 |
https://github.com/huggingface/datasets/issues/4597 | Streaming issue for financial_phrasebank | [
"cc @huggingface/datasets: it seems like https://www.researchgate.net/ is flaky for datasets hosting (I put the \"hosted-on-google-drive\" tag since it's the same kind of issue I think)",
"Let's see if their license allows hosting their data on the Hub.",
"License is Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported (CC BY-NC-SA 3.0).\r\n\r\nWe can host their data on the Hub."
] | ### Link
https://huggingface.co/datasets/financial_phrasebank/viewer/sentences_allagree/train
### Description
As reported by a community member using [AutoTrain Evaluate](https://huggingface.co/spaces/autoevaluate/model-evaluator/discussions/5#62bc217436d0e5d316a768f0), there seems to be a problem streaming this dataset:
```
Server error
Status code: 400
Exception: Exception
Message: Give up after 5 attempts with ConnectionError
```
### Owner
No | 4,597 |
https://github.com/huggingface/datasets/issues/4596 | Dataset Viewer issue for universal_dependencies | [
"Thanks, looking at it!",
"Finally fixed! We updated the dataset viewer and it fixed the issue.\r\n\r\nhttps://huggingface.co/datasets/universal_dependencies/viewer/aqz_tudet/train\r\n\r\n<img width=\"1561\" alt=\"Capture dβeΜcran 2022-09-07 aΜ 13 29 18\" src=\"https://user-images.githubusercontent.com/1676121/188867795-4f7dd438-d4f2-46cd-8a92-20a37fb2d6bc.png\">\r\n"
] | ### Link
https://huggingface.co/datasets/universal_dependencies
### Description
invalid json response body at https://datasets-server.huggingface.co/splits?dataset=universal_dependencies reason: Unexpected token I in JSON at position 0
### Owner
_No response_ | 4,596 |
https://github.com/huggingface/datasets/issues/4595 | Dataset Viewer issue with False positive PII redaction | [
"The value is in the data, it's not an issue with the \"dataset-viewer\".\r\n\r\n<img width=\"1161\" alt=\"Capture dβeΜcran 2022-06-29 aΜ 10 25 51\" src=\"https://user-images.githubusercontent.com/1676121/176389325-4d2a9a7f-1583-45b8-aa7a-960ffaa6a36a.png\">\r\n\r\n Maybe open a PR: https://huggingface.co/datasets/cakiki/rosetta-code/discussions\r\n",
"This was indeed a scraping issue which I assumed was a display issue; sorry about that!"
] | ### Link
https://huggingface.co/datasets/cakiki/rosetta-code
### Description
Hello, I just noticed an entry being redacted that shouldn't have been:
`RootMeanSquare@Range[10]` is being displayed as `[email protected][10]`
### Owner
_No response_ | 4,595 |
https://github.com/huggingface/datasets/issues/4594 | load_from_disk suggests incorrect fix when used to load DatasetDict | [] | Edit: Please feel free to remove this issue. The problem was not the error message but the fact that the DatasetDict.load_from_disk does not support loading nested splits, i.e. if one of the splits is itself a DatasetDict. If nesting splits is an antipattern, perhaps the load_from_disk function can throw a warning indicating that? | 4,594 |
https://github.com/huggingface/datasets/issues/4592 | Issue with jalFaizy/detect_chess_pieces when running datasets-cli test | [
"Hi @faizankshaikh\r\n\r\nPlease note that we have recently launched the Community feature, specifically targeted to create Discussions (about issues/questions/asking-for-help) on each Dataset on the Hub:\r\n- Blog post: https://huggingface.co/blog/community-update\r\n- Docs: https://huggingface.co/docs/hub/repositories-pull-requests-discussions\r\n\r\nThe Discussion tab for your \"jalFaizy/detect_chess_pieces\" dataset is here: https://huggingface.co/datasets/jalFaizy/detect_chess_pieces/discussions\r\nYou can use it to ask for help by pinging the Datasets maintainers: see our docs here: https://huggingface.co/docs/datasets/master/en/share#ask-for-a-help-and-reviews\r\n\r\nI'm transferring this discussion to your Discussion tab and trying to address it: https://huggingface.co/datasets/jalFaizy/detect_chess_pieces/discussions/1",
"Thank you @albertvillanova , I will keep that in mind.\r\n\r\nJust a quick note - I posted the issue on Github because the dataset viewer suggested me to \"open an issue for direct support\". Maybe it can be updated with your suggestion\r\n\r\n\r\n\r\n\r\n",
"Thank you pointing this out: yes, definitely, we should fix the error message. We are working on this."
] | ### Link
https://huggingface.co/datasets/jalFaizy/detect_chess_pieces
### Description
I am trying to write a appropriate data loader for [a custom dataset](https://huggingface.co/datasets/jalFaizy/detect_chess_pieces) using [this script](https://huggingface.co/datasets/jalFaizy/detect_chess_pieces/blob/main/detect_chess_pieces.py)
When I run the command
`$ datasets-cli test "D:\workspace\HF\detect_chess_pieces" --save_infos --all_configs`
It gives the following error
```
Using custom data configuration default
Traceback (most recent call last):
File "c:\users\faiza\anaconda3\lib\runpy.py", line 194, in _run_module_as_main
return _run_code(code, main_globals, None,
File "c:\users\faiza\anaconda3\lib\runpy.py", line 87, in _run_code
exec(code, run_globals)
File "C:\Users\faiza\anaconda3\Scripts\datasets-cli.exe\__main__.py", line 7, in <module>
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\commands\datasets_cli.py", line 39, in main
service.run()
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\commands\test.py", line 132, in run
for j, builder in enumerate(get_builders()):
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\commands\test.py", line 125, in get_builders
yield builder_cls(
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\builder.py", line 1148, in __init__
super().__init__(*args, **kwargs)
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\builder.py", line 306, in __init__
info = self.get_exported_dataset_info()
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\builder.py", line 405, in get_exported_dataset_info
return self.get_all_exported_dataset_infos().get(self.config.name, DatasetInfo())
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\builder.py", line 390, in get_all_exported_dataset_infos
return DatasetInfosDict.from_directory(cls.get_imported_module_dir())
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\info.py", line 309, in from_directory
dataset_infos_dict = {
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\info.py", line 310, in <dictcomp>
config_name: DatasetInfo.from_dict(dataset_info_dict)
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\info.py", line 272, in from_dict
return cls(**{k: v for k, v in dataset_info_dict.items() if k in field_names})
File "<string>", line 20, in __init__
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\info.py", line 160, in __post_init__
templates = [
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\info.py", line 161, in <listcomp>
template if isinstance(template, TaskTemplate) else task_template_from_dict(template)
File "c:\users\faiza\anaconda3\lib\site-packages\datasets\tasks\__init__.py", line 43, in task_template_from_dict
return template.from_dict(task_template_dict)
AttributeError: 'NoneType' object has no attribute 'from_dict'
```
My assumption is that there is some kind of issue in how the "task_templates" are read, because even if I keep them as None, or not include the argument at all, the same error occurs
### Owner
Yes | 4,592 |
https://github.com/huggingface/datasets/issues/4591 | Can't push Images to hub with manual Dataset | [
"Hi, thanks for reporting! This issue stems from the changes introduced in https://github.com/huggingface/datasets/pull/4282 (cc @lhoestq), in which list casts are ignored if they don't change the list type (required to preserve `null` values). And `push_to_hub` does a special cast to embed external image files but doesn't change the types, hence the failure."
] | ## Describe the bug
If I create a dataset including an 'Image' feature manually, when pushing to hub decoded images are not pushed,
instead it looks for image where image local path is/used to be.
This doesn't (at least didn't used to) happen with imagefolder. I want to build dataset manually because it is complicated.
This happens even though the dataset is looking like decoded images:
and I use `embed_external_files=True` while `push_to_hub` (same with false)
## Steps to reproduce the bug
```python
from PIL import Image
from datasets import Image as ImageFeature
from datasets import Features,Dataset
#manually create dataset
feats=Features(
{
"images": [ImageFeature()], #same even if explicitly ImageFeature(decode=True)
"input_image": ImageFeature(),
}
)
test_data={"images":[[Image.open("test.jpg"),Image.open("test.jpg"),Image.open("test.jpg")]], "input_image":[Image.open("test.jpg")]}
test_dataset=Dataset.from_dict(test_data,features=feats)
print(test_dataset)
test_dataset.push_to_hub("ceyda/image_test_public",private=False,token="",embed_external_files=True)
# clear cache rm -r ~/.cache/huggingface
# remove "test.jpg" # remove to see that it is looking for image on the local path
test_dataset=load_dataset("ceyda/image_test_public",use_auth_token="")
print(test_dataset)
print(test_dataset['train'][0])
```
## Expected results
should be able to push image bytes if dataset has `Image(decode=True)`
## Actual results
errors because it is trying to decode file from the non existing local path.
```
----> print(test_dataset['train'][0])
File ~/.local/lib/python3.8/site-packages/datasets/arrow_dataset.py:2154, in Dataset.__getitem__(self, key)
2152 def __getitem__(self, key): # noqa: F811
2153 """Can be used to index columns (by string names) or rows (by integer index or iterable of indices or bools)."""
-> 2154 return self._getitem(
2155 key,
2156 )
File ~/.local/lib/python3.8/site-packages/datasets/arrow_dataset.py:2139, in Dataset._getitem(self, key, decoded, **kwargs)
2137 formatter = get_formatter(format_type, features=self.features, decoded=decoded, **format_kwargs)
2138 pa_subtable = query_table(self._data, key, indices=self._indices if self._indices is not None else None)
-> 2139 formatted_output = format_table(
2140 pa_subtable, key, formatter=formatter, format_columns=format_columns, output_all_columns=output_all_columns
2141 )
2142 return formatted_output
File ~/.local/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:
...
-> 3068 fp = builtins.open(filename, "rb")
3069 exclusive_fp = True
3071 try:
FileNotFoundError: [Errno 2] No such file or directory: 'test.jpg'
```
## Environment info
- `datasets` version: 2.3.2
- Platform: Linux-5.4.0-1074-azure-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,591 |
https://github.com/huggingface/datasets/issues/4589 | Permission denied: '/home/.cache' when load_dataset with local script | [] | null | 4,589 |
https://github.com/huggingface/datasets/issues/4581 | Dataset Viewer issue for pn_summary | [
"linked to https://github.com/huggingface/datasets/issues/4580#issuecomment-1168373066?",
"Note that I refreshed twice this dataset, and I still have (another) error on one of the splits\r\n\r\n```\r\nStatus code: 400\r\nException: ClientResponseError\r\nMessage: 403, message='Forbidden', url=URL('https://doc-14-4c-docs.googleusercontent.com/docs/securesc/ha0ro937gcuc7l7deffksulhg5h7mbp1/pgotjmcuh77q0lk7p44rparfrhv459kp/1656403650000/11771870722949762109/*/16OgJ_OrfzUF_i3ftLjFn9kpcyoi7UJeO?e=download')\r\n```\r\n\r\nLike the three splits are processed in parallel by the workers, I imagine that the Google hosting is rate-limiting us.\r\n\r\ncc @albertvillanova \r\n\r\n",
"Exactly, Google Drive bans our loading scripts.\r\n\r\nWhen possible, we should host somewhere else."
] | ### Link
https://huggingface.co/datasets/pn_summary/viewer/1.0.0/validation
### Description
Getting an index error on the `validation` and `test` splits:
```
Server error
Status code: 400
Exception: IndexError
Message: list index out of range
```
### Owner
No | 4,581 |
https://github.com/huggingface/datasets/issues/4580 | Dataset Viewer issue for multi_news | [
"Thanks for reporting, @lewtun.\r\n\r\nI forced the refreshing of the preview and it worked OK for train and validation splits.\r\n\r\nI guess the error has to do with the data files being hosted at Google Drive: this gives errors when requested automatically using scripts.\r\nWe should host them to fix the error. Let's see if the license allows that.",
"I guess we can host the data: https://github.com/Alex-Fabbri/Multi-News/blob/master/LICENSE.txt"
] | ### Link
https://huggingface.co/datasets/multi_news
### Description
Not sure what the index error is referring to here:
```
Status code: 400
Exception: IndexError
Message: list index out of range
```
### Owner
No | 4,580 |
https://github.com/huggingface/datasets/issues/4578 | [Multi Configs] Use directories to differentiate between subsets/configurations | [
"I want to be able to create folders in a model.",
"How to set new split names, instead of train/test/validation? For example, I have a local dataset, consists of several subsets, named \"A\", \"B\", and \"C\". How can I create a huggingface dataset, with splits A/B/C ?\r\n\r\nThe document in https://huggingface.co/docs/datasets/dataset_script only tells me how to create datasets with subsets that is hosted on another server. How to do it if my datasets are local?",
"> The document in https://huggingface.co/docs/datasets/dataset_script only tells me how to create datasets with subsets that is hosted on another server. How to do it if my datasets are local?\r\n\r\nIt works the same - you just need to use local paths instead of URLs"
] | Currently to define several subsets/configurations of your dataset, you need to use a dataset script.
However it would be nice to have a no-code way to to this.
For example we could specify different configurations of a dataset (for example, if a dataset contains different languages) with one directory per configuration.
These structures are not supported right now, but would be nice to have:
```
my_dataset_repository/
βββ README.md
βββ en/
β βββ train.csv
β βββ test.csv
βββ fr/
βββ train.csv
βββ test.csv
```
Or with one directory per split:
```
my_dataset_repository/
βββ README.md
βββ en/
β βββ train/
β β βββ shard_0.csv
β β βββ shard_1.csv
β βββ test/
β βββ shard_0.csv
β βββ shard_1.csv
βββ fr/
βββ train/
β βββ shard_0.csv
β βββ shard_1.csv
βββ test/
βββ shard_0.csv
βββ shard_1.csv
```
cc @stevhliu @albertvillanova
This can be specified in the README as YAML with
```
configs:
- config_name: en
data_dir: en
- config_name: fr
data_dir: fr
``` | 4,578 |
https://github.com/huggingface/datasets/issues/4575 | Problem about wmt17 zh-en dataset | [
"Running into the same error with `wmt17/zh-en`, `wmt18/zh-en` and `wmt19/zh-en`.",
"@albertvillanova @lhoestq Could you take a look at this issue?",
"@winterfell2021 Hi, I wonder where the code you provided should be added. I tried to add them in the `datasets/table.py` in `array_cast` function, however, the 'zh' item is none.",
"I found some 'zh' item is none while 'c[hn]' is not.\r\nSo the code may change to:\r\n```python\r\nif 'c[hn]' in str(array.type):\r\n py_array = array.to_pylist()\r\n data_list = []\r\n for vo in py_array:\r\n tmp = {\r\n 'en': vo['en'],\r\n }\r\n if vo.get('zh'):\r\n tmp['zh'] = vo['zh']\r\n else:\r\n tmp['zh'] = vo['c[hn]']\r\n data_list.append(tmp)\r\n array = pa.array(data_list, type=pa.struct([\r\n pa.field('en', pa.string()),\r\n pa.field('zh', pa.string()),\r\n ]))\r\n```",
"I just pushed a fix, we'll do a new release of `datasets` soon to include this fix. In the meantime you can use the fixed dataset by passing `revision=\"main\"` to `load_dataset`"
] | It seems that in subset casia2015, some samples are like `{'c[hn]':'xxx', 'en': 'aa'}`.
So when using `data = load_dataset('wmt17', "zh-en")` to load the wmt17 zh-en dataset, which will raise the exception:
```
Traceback (most recent call last):
File "train.py", line 78, in <module>
data = load_dataset(args.dataset, "zh-en")
File "/usr/local/lib/python3.7/dist-packages/datasets/load.py", line 1684, in load_dataset
use_auth_token=use_auth_token,
File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 705, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 1221, in _download_and_prepare
super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 793, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 1215, in _prepare_split
num_examples, num_bytes = writer.finalize()
File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 533, in finalize
self.write_examples_on_file()
File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 410, in write_examples_on_file
self.write_batch(batch_examples=batch_examples)
File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 503, in write_batch
arrays.append(pa.array(typed_sequence))
File "pyarrow/array.pxi", line 230, in pyarrow.lib.array
File "pyarrow/array.pxi", line 110, in pyarrow.lib._handle_arrow_array_protocol
File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 198, in __arrow_array__
out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1675, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1846, in cast_array_to_feature
return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1675, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1756, in array_cast
raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{pa_type}")
TypeError: Couldn't cast array of type
struct<c[hn]: string, en: string, zh: string>
to
struct<en: string, zh: string>
```
So the solution of this problem is to change the original array manually:
```
if 'c[hn]' in str(array.type):
py_array = array.to_pylist()
data_list = []
for vo in py_array:
tmp = {
'en': vo['en'],
}
if 'zh' not in vo:
tmp['zh'] = vo['c[hn]']
else:
tmp['zh'] = vo['zh']
data_list.append(tmp)
array = pa.array(data_list, type=pa.struct([
pa.field('en', pa.string()),
pa.field('zh', pa.string()),
]))
```
Therefore, maybe a correct version of original casia2015 file need to be updated | 4,575 |
https://github.com/huggingface/datasets/issues/4572 | Dataset Viewer issue for mlsum | [
"Thanks for reporting, @lewtun.\r\n\r\nAfter investigation, it seems that the server https://gitlab.lip6.fr does not allow HTTP Range requests.\r\n\r\nWe are trying to find a workaround..."
] | ### Link
https://huggingface.co/datasets/mlsum/viewer/de/train
### Description
There's seems to be a problem with the download / streaming of this dataset:
```
Server error
Status code: 400
Exception: BadZipFile
Message: File is not a zip file
```
### Owner
No | 4,572 |
https://github.com/huggingface/datasets/issues/4571 | move under the facebook org? | [
"Related to https://github.com/huggingface/datasets/issues/4562#issuecomment-1166911751\r\n\r\nI'll assign @albertvillanova ",
"I'm just wondering why we don't have this dataset under:\r\n- the `facebook` namespace\r\n- or the canonical dataset `flores`: why does this only have 2 languages?",
"fwiw: the dataset viewer is working. Renaming the issue"
] | ### Link
https://huggingface.co/datasets/gsarti/flores_101
### Description
It seems like streaming isn't supported for this dataset:
```
Server Error
Status code: 400
Exception: NotImplementedError
Message: Extraction protocol for TAR archives like 'https://dl.fbaipublicfiles.com/flores101/dataset/flores101_dataset.tar.gz' is not implemented in streaming mode. Please use `dl_manager.iter_archive` instead.
```
### Owner
No | 4,571 |
https://github.com/huggingface/datasets/issues/4570 | Dataset sharding non-contiguous? | [
"This was silly; I was sure I'd looked for a `contiguous` argument, and was certain there wasn't one the first time I looked :smile:\r\n\r\nSorry about that.",
"Hi! You can pass `contiguous=True` to `.shard()` get contiguous shards. More info on this and the default behavior can be found in the [docs](https://huggingface.co/docs/datasets/v2.3.2/en/package_reference/main_classes#datasets.Dataset.shard).\r\n\r\nEDIT: Answered as you closed the thread π ",
"Hahaha I'm sorry; my excuse is: it's Sunday. (Which makes me all the more grateful for your response :smiley: ",
"@mariosasko Sorry for reviving this, but I was curious as to why `contiguous=False` was the default. This might be a personal bias, but I feel that a user would expect the opposite to be the default. :thinking: ",
"This project started as a fork of TFDS, and `contiguous=False` is the default behavior [there](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#shard)."
] | ## Describe the bug
I'm not sure if this is a bug; more likely normal behavior but i wanted to double check.
Is it normal that `datasets.shard` does not produce chunks that, when concatenated produce the original ordering of the sharded dataset?
This might be related to this pull request (https://github.com/huggingface/datasets/pull/4466) but I have to admit I did not properly look into the changes made.
## Steps to reproduce the bug
```python
max_shard_size = convert_file_size_to_int('300MB')
dataset_nbytes = dataset.data.nbytes
num_shards = int(dataset_nbytes / max_shard_size) + 1
num_shards = max(num_shards, 1)
print(f"{num_shards=}")
for shard_index in range(num_shards):
shard = dataset.shard(num_shards=num_shards, index=shard_index)
shard.to_parquet(f"tokenized/tokenized-{shard_index:03d}.parquet")
os.listdir('tokenized/')
```
## Expected results
I expected the shards to match the order of the data of the original dataset; i.e. `dataset[10]` being the same as `shard_1[10]` for example
## Actual results
Only the first element is the same; i.e. `dataset[0]` is the same as `shard_1[0]`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.2
- Platform: Linux-4.15.0-176-generic-x86_64-with-glibc2.31
- Python version: 3.10.4
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,570 |
https://github.com/huggingface/datasets/issues/4569 | Dataset Viewer issue for sst2 | [
"Hi @lewtun, thanks for reporting.\r\n\r\nI have checked locally and refreshed the preview and it seems working smooth now:\r\n```python\r\nIn [8]: ds\r\nOut[8]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['idx', 'sentence', 'label'],\r\n num_rows: 67349\r\n })\r\n validation: Dataset({\r\n features: ['idx', 'sentence', 'label'],\r\n num_rows: 872\r\n })\r\n test: Dataset({\r\n features: ['idx', 'sentence', 'label'],\r\n num_rows: 1821\r\n })\r\n})\r\n```\r\n\r\nCould you confirm? ",
"Thanks @albertvillanova - it is indeed working now (not sure what caused the error in the first place). Closing this :)"
] | ### Link
https://huggingface.co/datasets/sst2
### Description
Not sure what is causing this, however it seems that `load_dataset("sst2")` also hangs (even though it downloads the files without problem):
```
Status code: 400
Exception: Exception
Message: Give up after 5 attempts with ConnectionError
```
### Owner
No | 4,569 |
https://github.com/huggingface/datasets/issues/4568 | XNLI cache reload is very slow | [
"Hi,\r\nCould you tell us how you are running this code?\r\nI tested on my machine (M1 Mac). And it is running fine both on and off internet.\r\n\r\n<img width=\"1033\" alt=\"Screen Shot 2022-07-03 at 1 32 25 AM\" src=\"https://user-images.githubusercontent.com/8711912/177026364-4ad7cedb-e524-4513-97f7-7961bbb34c90.png\">\r\nTested on both stable and dev version. ",
"Sure, I was running it on a Linux machine.\r\nI found that if I turn the Internet off, it would still try to make a HTTPS call which would slow down the cache loading. If you can't reproduce then we can close the issue.",
"Hi @Muennighoff! You can set the env variable `HF_DATASETS_OFFLINE` to `1` to avoid this behavior in offline mode. More info is available [here](https://huggingface.co/docs/datasets/master/en/loading#offline)."
] | ### Reproduce
Using `2.3.3.dev0`
`from datasets import load_dataset`
`load_dataset("xnli", "en")`
Turn off Internet
`load_dataset("xnli", "en")`
I cancelled the second `load_dataset` eventually cuz it took super long. It would be great to have something to specify e.g. `only_load_from_cache` and avoid the library trying to download when there is no Internet. If I leave it running it works but takes way longer than when there is Internet. I would expect loading from cache to take the same amount of time regardless of whether there is Internet.
```
---------------------------------------------------------------------------
gaierror Traceback (most recent call last)
/opt/conda/lib/python3.7/site-packages/urllib3/connection.py in _new_conn(self)
174 conn = connection.create_connection(
--> 175 (self._dns_host, self.port), self.timeout, **extra_kw
176 )
/opt/conda/lib/python3.7/site-packages/urllib3/util/connection.py in create_connection(address, timeout, source_address, socket_options)
71
---> 72 for res in socket.getaddrinfo(host, port, family, socket.SOCK_STREAM):
73 af, socktype, proto, canonname, sa = res
/opt/conda/lib/python3.7/socket.py in getaddrinfo(host, port, family, type, proto, flags)
751 addrlist = []
--> 752 for res in _socket.getaddrinfo(host, port, family, type, proto, flags):
753 af, socktype, proto, canonname, sa = res
gaierror: [Errno -3] Temporary failure in name resolution
During handling of the above exception, another exception occurred:
KeyboardInterrupt Traceback (most recent call last)
/tmp/ipykernel_33/3594208039.py in <module>
----> 1 load_dataset("xnli", "en")
/opt/conda/lib/python3.7/site-packages/datasets/load.py 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)
1673 revision=revision,
1674 use_auth_token=use_auth_token,
-> 1675 **config_kwargs,
1676 )
1677
/opt/conda/lib/python3.7/site-packages/datasets/load.py in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, use_auth_token, **config_kwargs)
1494 download_mode=download_mode,
1495 data_dir=data_dir,
-> 1496 data_files=data_files,
1497 )
1498
/opt/conda/lib/python3.7/site-packages/datasets/load.py in dataset_module_factory(path, revision, download_config, download_mode, force_local_path, dynamic_modules_path, data_dir, data_files, **download_kwargs)
1182 download_config=download_config,
1183 download_mode=download_mode,
-> 1184 dynamic_modules_path=dynamic_modules_path,
1185 ).get_module()
1186 elif path.count("/") == 1: # community dataset on the Hub
/opt/conda/lib/python3.7/site-packages/datasets/load.py in __init__(self, name, revision, download_config, download_mode, dynamic_modules_path)
506 self.dynamic_modules_path = dynamic_modules_path
507 assert self.name.count("/") == 0
--> 508 increase_load_count(name, resource_type="dataset")
509
510 def download_loading_script(self, revision: Optional[str]) -> str:
/opt/conda/lib/python3.7/site-packages/datasets/load.py in increase_load_count(name, resource_type)
166 if not config.HF_DATASETS_OFFLINE and config.HF_UPDATE_DOWNLOAD_COUNTS:
167 try:
--> 168 head_hf_s3(name, filename=name + ".py", dataset=(resource_type == "dataset"))
169 except Exception:
170 pass
/opt/conda/lib/python3.7/site-packages/datasets/utils/file_utils.py in head_hf_s3(identifier, filename, use_cdn, dataset, max_retries)
93 return http_head(
94 hf_bucket_url(identifier=identifier, filename=filename, use_cdn=use_cdn, dataset=dataset),
---> 95 max_retries=max_retries,
96 )
97
/opt/conda/lib/python3.7/site-packages/datasets/utils/file_utils.py in http_head(url, proxies, headers, cookies, allow_redirects, timeout, max_retries)
445 allow_redirects=allow_redirects,
446 timeout=timeout,
--> 447 max_retries=max_retries,
448 )
449 return response
/opt/conda/lib/python3.7/site-packages/datasets/utils/file_utils.py in _request_with_retry(method, url, max_retries, base_wait_time, max_wait_time, timeout, **params)
366 tries += 1
367 try:
--> 368 response = requests.request(method=method.upper(), url=url, timeout=timeout, **params)
369 success = True
370 except (requests.exceptions.ConnectTimeout, requests.exceptions.ConnectionError) as err:
/opt/conda/lib/python3.7/site-packages/requests/api.py in request(method, url, **kwargs)
59 # cases, and look like a memory leak in others.
60 with sessions.Session() as session:
---> 61 return session.request(method=method, url=url, **kwargs)
62
63
/opt/conda/lib/python3.7/site-packages/requests/sessions.py in request(self, method, url, params, data, headers, cookies, files, auth, timeout, allow_redirects, proxies, hooks, stream, verify, cert, json)
527 }
528 send_kwargs.update(settings)
--> 529 resp = self.send(prep, **send_kwargs)
530
531 return resp
/opt/conda/lib/python3.7/site-packages/requests/sessions.py in send(self, request, **kwargs)
643
644 # Send the request
--> 645 r = adapter.send(request, **kwargs)
646
647 # Total elapsed time of the request (approximately)
/opt/conda/lib/python3.7/site-packages/requests/adapters.py in send(self, request, stream, timeout, verify, cert, proxies)
448 decode_content=False,
449 retries=self.max_retries,
--> 450 timeout=timeout
451 )
452
/opt/conda/lib/python3.7/site-packages/urllib3/connectionpool.py in urlopen(self, method, url, body, headers, retries, redirect, assert_same_host, timeout, pool_timeout, release_conn, chunked, body_pos, **response_kw)
708 body=body,
709 headers=headers,
--> 710 chunked=chunked,
711 )
712
/opt/conda/lib/python3.7/site-packages/urllib3/connectionpool.py in _make_request(self, conn, method, url, timeout, chunked, **httplib_request_kw)
384 # Trigger any extra validation we need to do.
385 try:
--> 386 self._validate_conn(conn)
387 except (SocketTimeout, BaseSSLError) as e:
388 # Py2 raises this as a BaseSSLError, Py3 raises it as socket timeout.
/opt/conda/lib/python3.7/site-packages/urllib3/connectionpool.py in _validate_conn(self, conn)
1038 # Force connect early to allow us to validate the connection.
1039 if not getattr(conn, "sock", None): # AppEngine might not have `.sock`
-> 1040 conn.connect()
1041
1042 if not conn.is_verified:
/opt/conda/lib/python3.7/site-packages/urllib3/connection.py in connect(self)
356 def connect(self):
357 # Add certificate verification
--> 358 self.sock = conn = self._new_conn()
359 hostname = self.host
360 tls_in_tls = False
/opt/conda/lib/python3.7/site-packages/urllib3/connection.py in _new_conn(self)
173 try:
174 conn = connection.create_connection(
--> 175 (self._dns_host, self.port), self.timeout, **extra_kw
176 )
177
KeyboardInterrupt:
``` | 4,568 |
https://github.com/huggingface/datasets/issues/4566 | Document link #load_dataset_enhancing_performance points to nowhere | [
"Hi! This is indeed the link the docstring should point to. Are you interested in submitting a PR to fix this?",
"https://github.com/huggingface/datasets/blame/master/docs/source/cache.mdx#L93\r\n\r\nThere seems already an anchor here. Somehow it doesn't work. I am not very familiar with how this online documentation works."
] | ## Describe the bug
A clear and concise description of what the bug is.
The [load_dataset_enhancing_performance](https://huggingface.co/docs/datasets/v2.3.2/en/package_reference/main_classes#load_dataset_enhancing_performance) link [here](https://huggingface.co/docs/datasets/v2.3.2/en/package_reference/main_classes#datasets.Dataset.load_from_disk.keep_in_memory) points to nowhere, I guess it should point to https://huggingface.co/docs/datasets/v2.3.2/en/cache#improve-performance?
| 4,566 |
https://github.com/huggingface/datasets/issues/4565 | Add UFSC OCPap dataset | [
"I will add this directly on the hub (same as #4486)βin https://huggingface.co/lapix"
] | ## Adding a Dataset
- **Name:** UFSC OCPap: Papanicolaou Stained Oral Cytology Dataset (v4)
- **Description:** The UFSC OCPap dataset comprises 9,797 labeled images of 1200x1600 pixels acquired from 5 slides of cancer diagnosed and 3 healthy of oral brush samples, from distinct patients.
- **Paper:** https://dx.doi.org/10.2139/ssrn.4119212
- **Data:** https://data.mendeley.com/datasets/dr7ydy9xbk/1
- **Motivation:** real data of pap stained oral cytology samples
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 4,565 |
https://github.com/huggingface/datasets/issues/4562 | Dataset Viewer issue for allocine | [
"I removed my assignment as @huggingface/datasets should be able to answer better than me\r\n",
"Let me have a look...",
"Thanks for the quick fix @albertvillanova ",
"Note that the underlying issue is that datasets containing TAR files are not streamable out of the box: they need being iterated with `dl_manager.iter_archive` to avoid performance issues because they access their file content *sequentially* (no random access).",
"> Note that the underlying issue is that datasets containing TAR files are not streamable out of the box: they need being iterated with `dl_manager.iter_archive` to avoid performance issues because they access their file content _sequentially_ (no random access).\r\n\r\nAh thanks for the clarification! I'll look out for this next time and implement the fix myself :)"
] | ### Link
https://huggingface.co/datasets/allocine
### Description
Not sure if this is a problem with `bz2` compression, but I thought these datasets could be streamed:
```
Status code: 400
Exception: AttributeError
Message: 'TarContainedFile' object has no attribute 'readable'
```
### Owner
No | 4,562 |
https://github.com/huggingface/datasets/issues/4556 | Dataset Viewer issue for conll2003 | [
"Fixed, thanks."
] | ### Link
https://huggingface.co/datasets/conll2003/viewer/conll2003/test
### Description
Seems like a cache problem with this config / split:
```
Server error
Status code: 400
Exception: FileNotFoundError
Message: [Errno 2] No such file or directory: '/cache/modules/datasets_modules/datasets/conll2003/__init__.py'
```
### Owner
No | 4,556 |
https://github.com/huggingface/datasets/issues/4555 | Dataset Viewer issue for xtreme | [
"Fixed, thanks."
] | ### Link
https://huggingface.co/datasets/xtreme/viewer/PAN-X.de/test
### Description
There seems to be a problem with the cache of this config / split:
```
Server error
Status code: 400
Exception: FileNotFoundError
Message: [Errno 2] No such file or directory: '/cache/modules/datasets_modules/datasets/xtreme/349258adc25bb45e47de193222f95e68a44f7a7ab53c4283b3f007208a11bf7e/xtreme.py'
```
### Owner
No | 4,555 |
https://github.com/huggingface/datasets/issues/4550 | imdb source error | [
"Thanks for reporting, @Muhtasham.\r\n\r\nIndeed IMDB dataset is not accessible from yesterday, because the data is hosted on the data owners servers at Stanford (http://ai.stanford.edu/) and these are down due to a power outage originated by a fire: https://twitter.com/StanfordAILab/status/1539472302399623170?s=20&t=1HU1hrtaXprtn14U61P55w\r\n\r\nAs a temporary workaroud, you can load the IMDB dataset with this tweak:\r\n```python\r\nds = load_dataset(\"imdb\", revision=\"tmp-fix-imdb\")\r\n```\r\n"
] | ## Describe the bug
imdb dataset not loading
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset("imdb")
```
## Expected results
## Actual results
```bash
06/23/2022 14:45:18 - INFO - datasets.builder - Dataset not on Hf google storage. Downloading and preparing it from source
06/23/2022 14:46:34 - INFO - datasets.utils.file_utils - HEAD request to http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz timed out, retrying... [1.0]
.....
ConnectionError: Couldn't reach http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz (ConnectTimeout(MaxRetryError("HTTPConnectionPool(host='ai.stanford.edu', port=80): Max retries exceeded with url: /~amaas/data/sentiment/aclImdb_v1.tar.gz (Caused by ConnectTimeoutError(<urllib3.connection.HTTPConnection object at 0x7f2d750cf690>, 'Connection to ai.stanford.edu timed out. (connect timeout=100)'))")))
```
## Environment info
- `datasets` version: 2.3.2
- 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
| 4,550 |
https://github.com/huggingface/datasets/issues/4549 | FileNotFoundError when passing a data_file inside a directory starting with double underscores | [
"I have consistently experienced this bug on GitHub actions when bumping to `2.3.2`",
"We're working on a fix ;)"
] | Bug experienced in the `accelerate` CI: https://github.com/huggingface/accelerate/runs/7016055148?check_suite_focus=true
This is related to https://github.com/huggingface/datasets/pull/4505 and the changes from https://github.com/huggingface/datasets/pull/4412 | 4,549 |
https://github.com/huggingface/datasets/issues/4548 | Metadata.jsonl for Imagefolder is ignored if it's in a parent directory to the splits directories/do not have "{split}_" prefix | [
"I agree it would be nice to support this. It doesn't fit really well in the current data_files.py, where files of each splits are separated in different folder though, maybe we have to modify a bit the logic here. \r\n\r\nOne idea would be to extend `get_patterns_in_dataset_repository` and `get_patterns_locally` to additionally check for `metadata.json`, but feel free to comment if you have better ideas (I feel like we're reaching the limits of what the current implementation IMO, so we could think of a different way of resolving the data files if necessary)"
] | If data contains a single `metadata.jsonl` file for several splits, it won't be included in a dataset's `data_files` and therefore ignored.
This happens when a directory is structured like as follows:
```
train/
file_1.jpg
file_2.jpg
test/
file_3.jpg
file_4.jpg
metadata.jsonl
```
or like as follows:
```
train_file_1.jpg
train_file_2.jpg
test_file_3.jpg
test_file_4.jpg
metadata.jsonl
```
The same for HF repos.
because it's ignored by the patterns [here](https://github.com/huggingface/datasets/blob/master/src/datasets/data_files.py#L29)
@lhoestq @mariosasko Do you think it's better to add this functionality in `data_files.py` or just specifically in imagefolder/audiofolder code? In `data_files.py` would me more general but I don't know if there are any other cases when that might be needed.
| 4,548 |
https://github.com/huggingface/datasets/issues/4544 | [CI] seqeval installation fails sometimes on python 3.6 | [] | The CI sometimes fails to install seqeval, which cause the `seqeval` metric tests to fail.
The installation fails because of this error:
```
Collecting seqeval
Downloading seqeval-1.2.2.tar.gz (43 kB)
|ββββββββ | 10 kB 42.1 MB/s eta 0:00:01
|βββββββββββββββ | 20 kB 53.3 MB/s eta 0:00:01
|βββββββββββββββββββββββ | 30 kB 67.2 MB/s eta 0:00:01
|ββββββββββββββββββββββββββββββ | 40 kB 76.1 MB/s eta 0:00:01
|ββββββββββββββββββββββββββββββββ| 43 kB 10.0 MB/s
Preparing metadata (setup.py) ... - error
ERROR: Command errored out with exit status 1:
command: /home/circleci/.pyenv/versions/3.6.15/bin/python3.6 -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/tmp/pip-install-1l96tbyj/seqeval_b31086f711d84743abe6905d2aa9dade/setup.py'"'"'; __file__='"'"'/tmp/pip-install-1l96tbyj/seqeval_b31086f711d84743abe6905d2aa9dade/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' egg_info --egg-base /tmp/pip-pip-egg-info-pf54_vqy
cwd: /tmp/pip-install-1l96tbyj/seqeval_b31086f711d84743abe6905d2aa9dade/
Complete output (22 lines):
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "/tmp/pip-install-1l96tbyj/seqeval_b31086f711d84743abe6905d2aa9dade/setup.py", line 56, in <module>
'Programming Language :: Python :: Implementation :: PyPy'
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/site-packages/setuptools/__init__.py", line 143, in setup
return distutils.core.setup(**attrs)
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/distutils/core.py", line 108, in setup
_setup_distribution = dist = klass(attrs)
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/site-packages/setuptools/dist.py", line 442, in __init__
k: v for k, v in attrs.items()
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/distutils/dist.py", line 281, in __init__
self.finalize_options()
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/site-packages/setuptools/dist.py", line 601, in finalize_options
ep.load()(self, ep.name, value)
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/site-packages/pkg_resources/__init__.py", line 2346, in load
return self.resolve()
File "/home/circleci/.pyenv/versions/3.6.15/lib/python3.6/site-packages/pkg_resources/__init__.py", line 2352, in resolve
module = __import__(self.module_name, fromlist=['__name__'], level=0)
File "/tmp/pip-install-1l96tbyj/seqeval_b31086f711d84743abe6905d2aa9dade/.eggs/setuptools_scm-7.0.2-py3.6.egg/setuptools_scm/__init__.py", line 5
from __future__ import annotations
^
SyntaxError: future feature annotations is not defined
----------------------------------------
WARNING: Discarding https://files.pythonhosted.org/packages/9d/2d/233c79d5b4e5ab1dbf111242299153f3caddddbb691219f363ad55ce783d/seqeval-1.2.2.tar.gz#sha256=f28e97c3ab96d6fcd32b648f6438ff2e09cfba87f05939da9b3970713ec56e6f (from https://pypi.org/simple/seqeval/). Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.
```
for example in https://app.circleci.com/pipelines/github/huggingface/datasets/12665/workflows/93878eb9-a923-4b35-b2e7-c5e9b22f10ad/jobs/75300
Here is a diff of the pip install logs until the error is reached: https://www.diffchecker.com/VkQDLeQT
This could be caused by the latest updates of setuptools-scm | 4,544 |
https://github.com/huggingface/datasets/issues/4542 | [to_tf_dataset] Use Feather for better compatibility with TensorFlow ? | [
"This has so much potential to be great! Also I think you tagged some poor random dude on the internet whose name is also Joao, lol, edited that for you! ",
"cc @sayakpaul here too, since he was interested in our new approaches to converting datasets!",
"Noted and I will look into the thread in detail tomorrow once I log back in. ",
"@lhoestq I have used TFRecords with `tf.data` for both vision and text and I can say that they are quite performant. I haven't worked with Feather yet as similarly as I have with TFRecords. If you haven't started the benchmarking script yet, I can prepare a Colab notebook that loads Feather files, converts them into a `tf.data` pipeline, and does some basic preprocessing. \r\n\r\nBut in my limited understanding, Feather might be better suited for CSV files. Not yet sure if it's good for modalities like images. ",
"> Not yet sure if it's good for modalities like images.\r\n\r\nWe store images pretty much the same way as tensorflow_datasets (i.e. storing the encoded image bytes, or a path to the local image, so that the image can be decoded on-the-fly), so as long as we use something similar as TFDS for image decoding it should be ok",
"So for image datasets, we could potentially store the paths in the feather format and decode and read them on the fly? But it introduces an I/O redundancy of having to read the images every time.\r\n\r\nWith caching it could be somewhat mitigated but it's not a good solution for bigger image datasets. ",
"> So for image datasets, we could potentially store the paths in the feather format and decode and read them on the fly?\r\n\r\nhopefully yes :) \r\n\r\nI double-checked the TFDS source code and they always save the bytes actually, not the path. Anyway we'll see if we run into issues or not (as a first step we can require the bytes to be in the feather file)",
"Yes. For images, TFDS actually prepares TFRecords first for encoding and then reuses them for every subsequent call. ",
"@lhoestq @Rocketknight1 I worked on [this PoC](https://gist.github.com/sayakpaul/f7d5cc312cd01cb31098fad3fd9c6b59) that\r\n\r\n* Creates Feather files from a medium resolution dataset (`tf_flowers`).\r\n* Explores different options with TensorFlow IO to load the Feather files. \r\n\r\nI haven't benchmarked those different options yet. There's also a gotcha that I have noted in the PoC. I hope it gets us started but I'm sorry if this is redundant. ",
"Cool thanks ! If I understand correctly in your PoC you store the flattened array of pixels in the feather file. This will take a lot of disk space.\r\n\r\nMaybe we could just save the encoded bytes and let users apply a `map` to decode/transform them into the format they need for training ? Users can use tf.image to do so for example",
"@lhoestq this is what I tried:\r\n\r\n```py\r\ndef read_image(path):\r\n with open(path, \"rb\") as f:\r\n return f.read()\r\n\r\n\r\ntotal_images_written = 0\r\n\r\nfor step in tqdm.tnrange(int(math.ceil(len(image_paths) / batch_size))):\r\n batch_image_paths = image_paths[step * batch_size : (step + 1) * batch_size]\r\n batch_image_labels = all_integer_labels[step * batch_size : (step + 1) * batch_size]\r\n\r\n data = [read_image(path) for path in batch_image_paths]\r\n table = pa.Table.from_arrays([data, batch_image_labels], [\"data\", \"labels\"])\r\n write_feather(table, f\"/tmp/flowers_feather_{step}.feather\", chunksize=chunk_size)\r\n total_images_written += len(batch_image_paths)\r\n print(f\"Total images written: {total_images_written}.\")\r\n\r\n del data\r\n```\r\n\r\nI got the feather files done (no resizing required as you can see):\r\n\r\n```sh\r\nls -lh /tmp/*.feather\r\n\r\n-rw-r--r-- 1 sayakpaul wheel 64M Jun 24 09:28 /tmp/flowers_feather_0.feather\r\n-rw-r--r-- 1 sayakpaul wheel 59M Jun 24 09:28 /tmp/flowers_feather_1.feather\r\n-rw-r--r-- 1 sayakpaul wheel 51M Jun 24 09:28 /tmp/flowers_feather_2.feather\r\n-rw-r--r-- 1 sayakpaul wheel 45M Jun 24 09:28 /tmp/flowers_feather_3.feather\r\n```\r\n\r\nNow there seems to be a problem with `tfio.arrow`:\r\n\r\n```py\r\nimport tensorflow_io.arrow as arrow_io\r\n\r\n\r\ndataset = arrow_io.ArrowFeatherDataset(\r\n [\"/tmp/flowers_feather_0.feather\"],\r\n columns=(0, 1),\r\n output_types=(tf.string, tf.int64),\r\n output_shapes=([], []),\r\n batch_mode=\"auto\",\r\n)\r\n\r\nprint(dataset.element_spec) \r\n```\r\n\r\nPrints:\r\n\r\n```\r\n(TensorSpec(shape=(None,), dtype=tf.string, name=None),\r\n TensorSpec(shape=(None,), dtype=tf.int64, name=None))\r\n```\r\n\r\nBut when I do `sample = next(iter(dataset))` it goes into:\r\n\r\n```py\r\nInternalError Traceback (most recent call last)\r\nInput In [30], in <cell line: 1>()\r\n----> 1 sample = next(iter(dataset))\r\n\r\nFile ~/.local/bin/.virtualenvs/jax/lib/python3.8/site-packages/tensorflow/python/data/ops/iterator_ops.py:766, in OwnedIterator.__next__(self)\r\n 764 def __next__(self):\r\n 765 try:\r\n--> 766 return self._next_internal()\r\n 767 except errors.OutOfRangeError:\r\n 768 raise StopIteration\r\n\r\nFile ~/.local/bin/.virtualenvs/jax/lib/python3.8/site-packages/tensorflow/python/data/ops/iterator_ops.py:749, in OwnedIterator._next_internal(self)\r\n 746 # TODO(b/77291417): This runs in sync mode as iterators use an error status\r\n 747 # to communicate that there is no more data to iterate over.\r\n 748 with context.execution_mode(context.SYNC):\r\n--> 749 ret = gen_dataset_ops.iterator_get_next(\r\n 750 self._iterator_resource,\r\n 751 output_types=self._flat_output_types,\r\n 752 output_shapes=self._flat_output_shapes)\r\n 754 try:\r\n 755 # Fast path for the case `self._structure` is not a nested structure.\r\n 756 return self._element_spec._from_compatible_tensor_list(ret) # pylint: disable=protected-access\r\n\r\nFile ~/.local/bin/.virtualenvs/jax/lib/python3.8/site-packages/tensorflow/python/ops/gen_dataset_ops.py:3017, in iterator_get_next(iterator, output_types, output_shapes, name)\r\n 3015 return _result\r\n 3016 except _core._NotOkStatusException as e:\r\n-> 3017 _ops.raise_from_not_ok_status(e, name)\r\n 3018 except _core._FallbackException:\r\n 3019 pass\r\n\r\nFile ~/.local/bin/.virtualenvs/jax/lib/python3.8/site-packages/tensorflow/python/framework/ops.py:7164, in raise_from_not_ok_status(e, name)\r\n 7162 def raise_from_not_ok_status(e, name):\r\n 7163 e.message += (\" name: \" + name if name is not None else \"\")\r\n-> 7164 raise core._status_to_exception(e) from None\r\n\r\nInternalError: Invalid: INVALID_ARGUMENT: arrow data type 0x7ff9899d8038 is not supported: Type error: Arrow data type is not supported [Op:IteratorGetNext]\r\n```\r\n\r\nSome additional notes:\r\n\r\n* I can actually decode an image encoded with `read_image()` (shown earlier):\r\n\r\n ```py\r\n sample_image_path = image_paths[0]\r\n encoded_image = read_image(sample_image_path)\r\n image = tf.image.decode_png(encoded_image, 3)\r\n print(image.shape)\r\n ```\r\n\r\n* If the above `tf.data.Dataset` object would have succeeded my plan was to just map the decoder like so:\r\n\r\n ```py\r\n autotune = tf.data.AUTOTUNE\r\n dataset = dataset.map(lambda x, y: (tf.image.decode_png(x, 3), y), num_parallel_calls=autotune)\r\n ```",
"@lhoestq I think I was able to make it work in the way you were envisioning. Here's the PoC:\r\nhttps://gist.github.com/sayakpaul/f7d5cc312cd01cb31098fad3fd9c6b59#file-feather-tf-poc-bytes-ipynb\r\n\r\nSome details:\r\n\r\n* I am currently serializing the images as strings with `base64`). In comparison to the flattened arrays as before, the size of the individual feather files has reduced (144 MB -> 85 MB, largest).\r\n* When decoding, I am first decoding the base64 string and then decoding that string (with `tf.io.decode_base64`) as an image with `tf.image.decode_png()`. \r\n* The entire workflow (from generating the Feather files to loading them and preparing the batched `tf.data` pipeline) involves the following libraries: `pyarraow`, `tensorflow-io`, and `tensorflow`. \r\n\r\nCc: @Rocketknight1 @gante ",
"Cool thanks ! Too bad the Arrow binary type doesn't seem to be supported in `arrow_io.ArrowFeatherDataset` :/ We would also need it to support Arrow struct type. Indeed images in `datasets` are represented using an Arrow type\r\n```python\r\npa.struct({\"path\": pa.string(), \"bytes\": pa.binary()})\r\n```\r\nnot sure yet how hard it is to support this though.\r\n\r\nChanging the typing on our side would create concerning breaking changes, that's why it would be awesome if it could work using these types",
"If the ArrowFeatherDataset doesn't yet support it, I guess our hands are a bit tied at the moment. \r\n\r\nIIUC, in my [latest PoC notebook](https://gist.github.com/sayakpaul/f7d5cc312cd01cb31098fad3fd9c6b59#file-feather-tf-poc-bytes-ipynb), you wanted to see each entry in the feather file to be represented like so?\r\n\r\n```\r\npa.struct({\"path\": pa.string(), \"bytes\": pa.binary()})\r\n``` \r\n\r\nIn that case, `pa.binary()` isn't yet supported.",
"> IIUC, in my [latest PoC notebook](https://gist.github.com/sayakpaul/f7d5cc312cd01cb31098fad3fd9c6b59#file-feather-tf-poc-bytes-ipynb), you wanted to see each entry in the feather file to be represented like so?\r\n> \r\n> pa.struct({\"path\": pa.string(), \"bytes\": pa.binary()})\r\n\r\nYea because that's the data format we're using. If we were to use base64, then we would have to process the full dataset to convert it, which can take some time. Converting to TFRecords would be simpler than converting to base64 in Feather files.\r\n\r\nMaybe it would take too much time to be worth exploring, but according to https://github.com/tensorflow/io/issues/1361#issuecomment-819029002 it's possible to add support for binary type in ArrowFeatherDataset. What do you think ? Any other alternative in mind ?",
"> Maybe it would take too much time to be worth exploring, but according to https://github.com/tensorflow/io/issues/1361#issuecomment-819029002 it's possible to add support for binary type in ArrowFeatherDataset.\r\n\r\nShould be possible as per the comment but there hasn't been any progress and it's been more than a year. \r\n\r\n> If we were to use base64, then we would have to process the full dataset to convert it, which can take some time.\r\n\r\nI don't understand this. I would think TFRecords would also need something similar but I need the context you're coming from. \r\n\r\n> What do you think ? Any other alternative in mind ?\r\n\r\nTFRecords since the TensorFlow ecosystem has developed good support for it over the years. ",
"> I don't understand this. I would think TFRecords would also need something similar but I need the context you're coming from.\r\n\r\nUsers already have a copy of the dataset in Arrow format (we can change this to Feather). So to load the Arrow/feather files to a TF dataset we need TF IO or something like that. Otherwise the user has to convert all the files from Arrow to TFRecords to use TF data efficiently. But the conversion needs resources: CPU, disk, time. Converting the images to base64 require the same sort of resources.\r\n\r\nSo the issue we're trying to tackle is how to load the Arrow data in TF without having to convert anything ^^",
"Yeah, it looks like in its current state the tfio support for `Feather` is incomplete, so we'd end up having to write a lot of it, or do a conversion that defeats the whole point (because if we're going to convert the whole dataset we might as well convert to `TFRecord`).",
"Understood @lhoestq. Thanks for explaining!\r\n\r\nAgreed with @Rocketknight1. ",
"@lhoestq Although I think this is a dead-end for now unfortunately, because of the limitations at TF's end, we could still explore automatic conversion to TFRecord, or I could dive into refining `to_tf_dataset()` to yield unbatched samples and/or load samples with multiprocessing to improve throughput. Do you have any preferences there?",
"> @lhoestq Although I think this is a dead-end for now unfortunately, because of the limitations at TF's end, we could still explore automatic conversion to TFRecord, or I could dive into refining `to_tf_dataset()` to yield unbatched samples and/or load samples with multiprocessing to improve throughput. Do you have any preferences there?\r\n\r\nHappy to take part there @Rocketknight1.",
"If `to_tf_dataset` can be unbatched, then it should be fairly easy for users to convert the TF dataset to TFRecords right ?",
"@lhoestq why one would convert to TFRecords after unbatching? ",
"> If to_tf_dataset can be unbatched, then it should be fairly easy for users to convert the TF dataset to TFRecords right ?\r\n\r\nSort of! A `tf.data.Dataset` is more like an iterator, and does not support sample indexing. `to_tf_dataset()` creates an iterator, but to convert that to `TFRecord`, the user would have to iterate over the whole thing and manually save the stream of samples to files. ",
"Someone would like to try to dive into tfio to fix this ? Sounds like a good opportunity to learn what are the best ways to load a dataset for TF, and also the connections between Arrow and TF.\r\n\r\nIf we can at least have the Arrow `binary` type working for TF that would be awesome already (issue https://github.com/tensorflow/io/issues/1361)\r\n\r\nalso cc @nateraw in case you'd be interested ;)",
"> Sounds like a good opportunity to learn what are the best ways to load a dataset for TF\r\n\r\nThe recommended way would likely be a combination of TFRecords and `tf.data`. \r\n\r\nExploring the connection between Arrow and TensorFlow is definitely worth pursuing though. But I am not sure about the implications of storing images in a format supported by Arrow. I guess we'll know more once we have at least figured out the support for `binary` type for TFIO. I will spend some time on it and keep this thread updated. ",
"I am currently working on a fine-tuning notebook for the TFSegFormer model (Semantic Segmentation). The resolution is high for both the input images and the labels - (512, 512, 3). Here's the [Colab Notebook](https://colab.research.google.com/drive/1jAtR7Z0lYX6m6JsDI5VByh5vFaNhHIbP?usp=sharing) (it's a WIP so please bear that in mind).\r\n\r\nI think the current implementation of `to_tf_dataset()` does create a bottleneck here since the GPU utilization is quite low. ",
"Here's a notebook showing the performance difference: https://colab.research.google.com/gist/sayakpaul/d7ca67c90beb47e354942c9d8c0bd8ef/scratchpad.ipynb. \r\n\r\nNote that I acknowledge that it's not an apples-to-apples comparison in many aspects (the dataset isn't the same, data serialization format isn't the same, etc.) but this is the best I could do. ",
"Thanks ! I think the speed difference can be partly explained: you use ds.shuffle in your dataset, which is an exact shuffling (compared to TFDS which does buffer shuffling): it slows down query time by 2x to 10x since it has to play with data that are not contiguous.\r\n\r\nThe rest of the speed difference seems to be caused by image decoding (from 330Β΅s/image to 30ms/image)",
"Fair enough. Can do one without shuffling too. But it's an important one to consider I guess. "
] | To have better performance in TensorFlow, it is important to provide lists of data files in supported formats. For example sharded TFRecords datasets are extremely performant. This is because tf.data can better leverage parallelism in this case, and load one file at a time in memory.
It seems that using `tensorflow_io` we could have something similar for `to_tf_dataset` if we provide sharded Feather files: https://www.tensorflow.org/io/api_docs/python/tfio/arrow/ArrowFeatherDataset
Feather is a format almost equivalent to the Arrow IPC Stream format we're using in `datasets`: Feather V2 is equivalent to Arrow IPC File format, which is an extension of the stream format (it has an extra footer). Therefore we could store datasets as Feather instead of Arrow IPC Stream format without breaking the whole library.
Here are a few points to explore
- [ ] check the performance of ArrowFeatherDataset in tf.data
- [ ] check what would change if we were to switch to Feather if needed, in particular check that those are fine: memory mapping, typing, writing, reading to python objects, etc.
We would also need to implement sharding when loading a dataset (this will be done anyway for #546)
cc @Rocketknight1 @gante feel free to comment in case I missed anything !
I'll share some files and scripts, so that we can benchmark performance of Feather files with tf.data | 4,542 |
https://github.com/huggingface/datasets/issues/4540 | Avoid splitting by` .py` for the file. | [
"Hi @espoirMur, thanks for reporting.\r\n\r\nYou are right: that code line could be improved and made more generically valid.\r\n\r\nOn the other hand, I would suggest using `os.path.splitext` instead.\r\n\r\nAre you willing to open a PR? :)",
"I will have a look.. \r\n\r\nThis weekend .. ",
"@albertvillanova , Can you have a look at #4590. \r\n\r\nThanks ",
"#self-assign"
] | https://github.com/huggingface/datasets/blob/90b3a98065556fc66380cafd780af9b1814b9426/src/datasets/load.py#L272
Hello,
Thanks you for this library .
I was using it and I had one edge case. my home folder name ends with `.py` it is `/home/espoir.py` so anytime I am running the code to load a local module this code here it is failing because after splitting it is trying to save the code to my home directory.
Step to reproduce.
- If you have a home folder which ends with `.py`
- load a module with a local folder
`qa_dataset = load_dataset("src/data/build_qa_dataset.py")`
it is failed
A possible workaround would be to use pathlib at the mentioned line
` meta_path = Path(importable_local_file).parent.joinpath("metadata.json")` this can alivate the issue .
Let me what are your thought on this and I can try to fix it by A PR.
| 4,540 |
https://github.com/huggingface/datasets/issues/4538 | Dataset Viewer issue for Pile of Law | [
"Hi @Breakend, yes β we'll propose a solution today",
"Thanks so much, I appreciate it!",
"Thanks so much for adding the docs. I was able to successfully hide the viewer using the \r\n```\r\nviewer: false\r\n```\r\nflag in the README.md of the dataset. I'm closing the issue because this is resolved. Thanks again!",
"Awesome! Thanks for confirming. cc @severo ",
"Just for the record:\r\n\r\n- the doc\r\n \r\n<img width=\"1430\" alt=\"Capture dβeΜcran 2022-06-27 aΜ 09 29 27\" src=\"https://user-images.githubusercontent.com/1676121/175884089-bca6c0d5-6387-473e-98ca-86a910ede4bd.png\">\r\n\r\n- the dataset main page\r\n\r\n<img width=\"1134\" alt=\"Capture dβeΜcran 2022-06-27 aΜ 09 29 05\" src=\"https://user-images.githubusercontent.com/1676121/175884152-5f285bf0-3471-45de-927a-e141b00ebb33.png\">\r\n\r\n- the dataset viewer page\r\n\r\n<img width=\"567\" alt=\"Capture dβeΜcran 2022-06-27 aΜ 09 29 16\" src=\"https://user-images.githubusercontent.com/1676121/175884191-ab6a297b-1c11-417e-bbde-0b7623278a79.png\">\r\n"
] | ### Link
https://huggingface.co/datasets/pile-of-law/pile-of-law
### Description
Hi, I would like to turn off the dataset viewer for our dataset without enabling access requests. To comply with upstream dataset creator requests/licenses, we would like to make sure that the data is not indexed by search engines and so would like to turn off dataset previews. But we do not want to collect user emails because it would violate single blind review, allowing us to deduce potential reviewers' identities. Is there a way that we can turn off the dataset viewer without collecting identity information?
Thanks so much!
### Owner
Yes | 4,538 |
https://github.com/huggingface/datasets/issues/4533 | Timestamp not returned as datetime objects in streaming mode | [] | As reported in (internal) https://github.com/huggingface/datasets-server/issues/397
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset("ett", name="h2", split="test", streaming=True)
>>> d = next(iter(dataset))
>>> d['start']
Timestamp('2016-07-01 00:00:00')
```
while loading in non-streaming mode it returns `datetime.datetime(2016, 7, 1, 0, 0)` | 4,533 |
https://github.com/huggingface/datasets/issues/4531 | Dataset Viewer issue for CSV datasets | [
"this should now be fixed",
"Confirmed, it's fixed now. Thanks for reporting, and thanks @coyotte508 for fixing it\r\n\r\n<img width=\"1123\" alt=\"Capture dβeΜcran 2022-06-21 aΜ 10 28 05\" src=\"https://user-images.githubusercontent.com/1676121/174753833-1b453a5a-6a90-4717-bca1-1b5fc6b75e4a.png\">\r\n"
] | ### Link
https://huggingface.co/datasets/scikit-learn/breast-cancer-wisconsin
### Description
I'm populating CSV datasets [here](https://huggingface.co/scikit-learn) but the viewer is not enabled and it looks for a dataset loading script, the datasets aren't on queue as well.
You can replicate the problem by simply uploading any CSV dataset.
### Owner
Yes | 4,531 |
https://github.com/huggingface/datasets/issues/4529 | Ecoset | [
"Hi! Very cool dataset! I answered your questions on the forum. Also, feel free to comment `#self-assign` on this issue to self-assign it.",
"The dataset lives on the Hub [here](https://huggingface.co/datasets/kietzmannlab/ecoset), so I'm closing this issue.",
"Hey There, thanks for closing π€ \r\n\r\nForgot the issue existed, so I didn't close it after implementing the downloader :)"
] | ## Adding a Dataset
- **Name:** *Ecoset*
- **Description:** *https://www.kietzmannlab.org/ecoset/*
- **Paper:** *https://doi.org/10.1073/pnas.2011417118*
- **Data:** *https://codeocean.com/capsule/9570390/tree/v1*
- **Motivation:**
**Ecoset** was created as a clean and ecologically valid alternative to **Imagenet**.
It is a large image recognition dataset, similar to Imagenet in size and structure. However, the authors of ecoset claim several improvements over Imagenet, like:
- more ecologically valid classes (e.g. not over-focussed on distinguishing different dog breeds)
- less NSFW content
- 'pre-packed image recognition models' that come with the dataset and can be used for validation of other models.
I am working for one of the authors of the paper with the aim of bringing Ecoset to huggingface datasets. Therefore I can work on this issue personally, but could use some help from devs and experienced users if the dataset is of interest to them. I phrased some of my questions on [discuss.huggingface](https://discuss.huggingface.co/t/handling-large-image-datasets/19373).
| 4,529 |
https://github.com/huggingface/datasets/issues/4528 | Memory leak when iterating a Dataset | [
"Is someone assigned to this issue?",
"The same issue is being debugged here: https://github.com/huggingface/datasets/issues/4883\r\n",
"Here is a modified repro example that makes it easier to see the leak:\r\n\r\n```\r\n$ cat ds2.py\r\nimport gc, sys\r\nimport time\r\nfrom datasets import load_dataset\r\nimport os, psutil\r\n\r\nprocess = psutil.Process(os.getpid())\r\n\r\nprint(process.memory_info().rss/2**20)\r\n\r\ncorpus = load_dataset(\"BeIR/msmarco\", 'corpus', keep_in_memory=False, streaming=False)['corpus']\r\ncorpus = corpus.select(range(200000))\r\n\r\nprint(process.memory_info().rss/2**20)\r\n\r\nbatch = None\r\n\r\nmem_before_start = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\n\r\nstep = 20000\r\nfor i in range(0, 10*step, step):\r\n mem_before = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\n batch = corpus[i:i+step]\r\n import objgraph\r\n #objgraph.show_refs([batch])\r\n #objgraph.show_refs([corpus])\r\n #sys.exit()\r\n gc.collect()\r\n\r\n mem_after = psutil.Process(os.getpid()).memory_info().rss / 2**20\r\n print(f\"{i:6d} {mem_after - mem_before:12.4f} {mem_after - mem_before_start:12.4f}\")\r\n\r\n```\r\n\r\nLet's run:\r\n\r\n```\r\n$ python ds2.py\r\n 0 36.5391 36.5391\r\n 20000 10.4609 47.0000\r\n 40000 5.9766 52.9766\r\n 60000 7.8906 60.8672\r\n 80000 6.0586 66.9258\r\n100000 8.4453 75.3711\r\n120000 6.7422 82.1133\r\n140000 8.5664 90.6797\r\n160000 5.7344 96.4141\r\n180000 8.3398 104.7539\r\n```\r\n\r\nYou can see the last column of total RSS memory keeps on growing in MBs. The mid column is by how much it was grown during a single iteration of the repro script (20000 items)",
"@NouamaneTazi, please check my analysis here https://github.com/huggingface/datasets/issues/4883#issuecomment-1242599722 so if you agree with my research this Issue can be closed as well.\r\n\r\nI also made a suggestion at how to proceed to hunt for a real leak here https://github.com/huggingface/datasets/issues/4883#issuecomment-1242600626\r\n\r\nyou may find this one to be useful as well https://github.com/huggingface/datasets/issues/4883#issuecomment-1242597966",
"Amazing job! Thanks for taking time to debug this π€\r\n\r\nFor my side, I tried to do some more research as well, but to no avail. https://github.com/huggingface/datasets/issues/4883#issuecomment-1243415957"
] | e## Describe the bug
It seems that memory never gets freed after iterating a `Dataset` (using `.map()` or a simple `for` loop)
## Steps to reproduce the bug
```python
import gc
import logging
import time
import pyarrow
from datasets import load_dataset
from tqdm import trange
import os, psutil
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
process = psutil.Process(os.getpid())
print(process.memory_info().rss) # output: 633507840 bytes
corpus = load_dataset("BeIR/msmarco", 'corpus', keep_in_memory=False, streaming=False)['corpus'] # or "BeIR/trec-covid" for a smaller dataset
print(process.memory_info().rss) # output: 698601472 bytes
logger.info("Applying method to all examples in all splits")
for i in trange(0, len(corpus), 1000):
batch = corpus[i:i+1000]
data = pyarrow.total_allocated_bytes()
if data > 0:
logger.info(f"{i}/{len(corpus)}: {data}")
print(process.memory_info().rss) # output: 3788247040 bytes
del batch
gc.collect()
print(process.memory_info().rss) # output: 3788247040 bytes
logger.info("Done...")
time.sleep(100)
```
## Expected results
Limited memory usage, and memory to be freed after processing
## Actual results
Memory leak
You can see how the memory allocation keeps increasing until it reaches a steady state when we hit the `time.sleep(100)`, which showcases that even the garbage collector couldn't free the allocated memory
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.2
- Platform: Linux-5.4.0-90-generic-x86_64-with-glibc2.31
- Python version: 3.9.7
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,528 |
https://github.com/huggingface/datasets/issues/4527 | Dataset Viewer issue for vadis/sv-ident | [
"Fixed, thanks!\r\n![Uploading Capture dβeΜcran 2022-06-21 aΜ 18.42.40.pngβ¦]()\r\n\r\n"
] | ### Link
https://huggingface.co/datasets/vadis/sv-ident
### Description
The dataset preview does not work:
```
Server Error
Status code: 400
Exception: Status400Error
Message: The dataset does not exist.
```
However, the dataset is streamable and works locally:
```python
In [1]: from datasets import load_dataset; ds = load_dataset("sv-ident.py", split="train", streaming=True); item = next(iter(ds)); item
Using custom data configuration default
Out[1]:
{'sentence': 'Our point, however, is that so long as downward (favorable) comparisons overwhelm the potential for unfavorable comparisons, system justification should be a likely outcome amongst the disadvantaged.',
'is_variable': 1,
'variable': ['exploredata-ZA5400_VarV66', 'exploredata-ZA5400_VarV53'],
'research_data': ['ZA5400'],
'doc_id': '73106',
'uuid': 'b9fbb80f-3492-4b42-b9d5-0254cc33ac10',
'lang': 'en'}
```
CC: @e-tornike
### Owner
No | 4,527 |
https://github.com/huggingface/datasets/issues/4526 | split cache used when processing different split | [
"I was not able to reproduce this behavior (I tried without using pytorch lightning though, since I don't know what code you ran in pytorch lightning to get this).\r\n\r\nIf you can provide a MWE that would be perfect ! :)",
"Hi, I think the issue happened because I was loading datasets under an `if` ... `else` statement and the condition would change the dataset I would need to load but instead the cached one was always returned. However, I believe that is expected behaviour, if so I'll close the issue.\r\n\r\nOtherwise I will try to provide a MWE"
] | ## Describe the bug`
```
ds1 = load_dataset('squad', split='validation')
ds2 = load_dataset('squad', split='train')
ds1 = ds1.map(some_function)
ds2 = ds2.map(some_function)
assert ds1 == ds2
```
This happens when ds1 and ds2 are created in `pytorch_lightning.DataModule` through
```
class myDataModule:
def train_dataloader(self):
ds = load_dataset('squad', split='train')
ds = ds.map(some_function)
return [ds]
def val_dataloader(self):
ds = load_dataset('squad', split="validation")
ds = ds.map(some_function)
return [ds]
```
I don't know if it depends on `pytorch_lightning` or `datasets` but setting `ds.map(some_function, load_from_cache_file=False)` fixes the issue.
If this is not enough to replicate I will try and provide and MWE, I don't have time now so I thought I wuld open the issue first! | 4,526 |
https://github.com/huggingface/datasets/issues/4525 | Out of memory error on workers while running Beam+Dataflow | [
"Some naive ideas to cope with this:\r\n- enable more RAM on each worker\r\n- force the spanning of more workers\r\n- others?",
"@albertvillanova We were finally able to process the full NQ dataset on our machines using 600 gb with 5 workers. Maybe these numbers will work for you as well.",
"Thanks a lot for the hint, @seirasto.\r\n\r\nI have one question: what runner did you use? Direct, Apache Flink/Nemo/Samza/Spark, Google Dataflow...? Thank you.",
"I asked my colleague who ran the code and he said apache beam.",
"@albertvillanova Since we have already processed the NQ dataset on our machines can we upload it to datasets so the NQ PR can be merged?",
"Maybe @lhoestq can give a more accurate answer as I am not sure about the authentication requirements to upload those files to our cloud bucket.\r\n\r\nAnyway I propose to continue this discussion on the dedicated PR for Natural questions dataset:\r\n- #4368",
"> I asked my colleague who ran the code and he said apache beam.\r\n\r\nHe looked into it further and he just used DirectRunner. @albertvillanova ",
"OK, thank you @seirasto for your hint.\r\n\r\nThat explains why you did not encounter the out of memory error: this only appears when the processing is distributed (on workers memory) and DirectRunner does not distribute the processing (all is done in a single machine). ",
"@albertvillanova Doesn't DirectRunner offer distributed processing through?\r\n\r\nhttps://beam.apache.org/documentation/runners/direct/\r\n\r\n```\r\nSetting parallelism\r\n\r\nNumber of threads or subprocesses is defined by setting the direct_num_workers pipeline option. From 2.22.0, direct_num_workers = 0 is supported. When direct_num_workers is set to 0, it will set the number of threads/subprocess to the number of cores of the machine where the pipeline is running.\r\n\r\nSetting running mode\r\n\r\nIn Beam 2.19.0 and newer, you can use the direct_running_mode pipeline option to set the running mode. direct_running_mode can be one of ['in_memory', 'multi_threading', 'multi_processing'].\r\n\r\nin_memory: Runner and workersβ communication happens in memory (not through gRPC). This is a default mode.\r\n\r\nmulti_threading: Runner and workers communicate through gRPC and each worker runs in a thread.\r\n\r\nmulti_processing: Runner and workers communicate through gRPC and each worker runs in a subprocess.\r\n```",
"Unrelated to the OOM issue, but we deprecated datasets with Beam scripts in #6474. I think we can close this issue"
] | ## Describe the bug
While running the preprocessing of the natural_question dataset (see PR #4368), there is an issue for the "default" config (train+dev files).
Previously we ran the preprocessing for the "dev" config (only dev files) with success.
Train data files are larger than dev ones and apparently workers run out of memory while processing them.
Any help/hint is welcome!
Error message:
```
Data channel closed, unable to receive additional data from SDK sdk-0-0
```
Info from the Diagnostics tab:
```
Out of memory: Killed process 1882 (python) total-vm:6041764kB, anon-rss:3290928kB, file-rss:0kB, shmem-rss:0kB, UID:0 pgtables:9520kB oom_score_adj:900
The worker VM had to shut down one or more processes due to lack of memory.
```
## Additional information
### Stack trace
```
Traceback (most recent call last):
File "/home/albert_huggingface_co/natural_questions/venv/bin/datasets-cli", line 8, in <module>
sys.exit(main())
File "/home/albert_huggingface_co/natural_questions/venv/lib/python3.9/site-packages/datasets/commands/datasets_cli.py", line 39, in main
service.run()
File "/home/albert_huggingface_co/natural_questions/venv/lib/python3.9/site-packages/datasets/commands/run_beam.py", line 127, in run
builder.download_and_prepare(
File "/home/albert_huggingface_co/natural_questions/venv/lib/python3.9/site-packages/datasets/builder.py", line 704, in download_and_prepare
self._download_and_prepare(
File "/home/albert_huggingface_co/natural_questions/venv/lib/python3.9/site-packages/datasets/builder.py", line 1389, in _download_and_prepare
pipeline_results.wait_until_finish()
File "/home/albert_huggingface_co/natural_questions/venv/lib/python3.9/site-packages/apache_beam/runners/dataflow/dataflow_runner.py", line 1667, in wait_until_finish
raise DataflowRuntimeException(
apache_beam.runners.dataflow.dataflow_runner.DataflowRuntimeException: Dataflow pipeline failed. State: FAILED, Error:
Data channel closed, unable to receive additional data from SDK sdk-0-0
```
### Logs
```
Error message from worker: Data channel closed, unable to receive additional data from SDK sdk-0-0
Workflow failed. Causes: S30:train/ReadAllFromText/ReadAllFiles/Reshard/ReshufflePerKey/GroupByKey/Read+train/ReadAllFromText/ReadAllFiles/Reshard/ReshufflePerKey/GroupByKey/GroupByWindow+train/ReadAllFromText/ReadAllFiles/Reshard/ReshufflePerKey/FlatMap(restore_timestamps)+train/ReadAllFromText/ReadAllFiles/Reshard/RemoveRandomKeys+train/ReadAllFromText/ReadAllFiles/ReadRange+train/Map(_parse_example)+train/Encode+train/Count N. Examples+train/Get values/Values+train/Save to parquet/Write/WriteImpl/WindowInto(WindowIntoFn)+train/Save to parquet/Write/WriteImpl/WriteBundles+train/Save to parquet/Write/WriteImpl/Pair+train/Save to parquet/Write/WriteImpl/GroupByKey/Write failed., The job failed because a work item has failed 4 times. Look in previous log entries for the cause of each one of the 4 failures. For more information, see https://cloud.google.com/dataflow/docs/guides/common-errors. The work item was attempted on these workers: beamapp-alberthuggingface-06170554-5p23-harness-t4v9 Root cause: Data channel closed, unable to receive additional data from SDK sdk-0-0, beamapp-alberthuggingface-06170554-5p23-harness-t4v9 Root cause: The worker lost contact with the service., beamapp-alberthuggingface-06170554-5p23-harness-bwsj Root cause: The worker lost contact with the service., beamapp-alberthuggingface-06170554-5p23-harness-5052 Root cause: The worker lost contact with the service.
```
| 4,525 |
https://github.com/huggingface/datasets/issues/4524 | Downloading via Apache Pipeline, client cancelled (org.apache.beam.vendor.grpc.v1p43p2.io.grpc.StatusRuntimeException) | [
"Hi @dan-the-meme-man, thanks for reporting.\r\n\r\nWe are investigating a similar issue but with Beam+Dataflow (instead of Beam+Flink): \r\n- #4525\r\n\r\nIn order to go deeper into the root cause, we need as much information as possible: logs from the main process + logs from the workers are very informative.\r\n\r\nIn the case of the issue with Beam+Dataflow, the logs from the workers report an out of memory issue.",
"As I continued working on this today, I came to suspect that it is in fact an out of memory issue - I have a few more notebooks that I've left running, and if they produce the same error, I will try to get the logs. In the meantime, if there's any chance that there is a repo out there with those three languages already as .arrow files, or if you know about how much memory would be needed to actually download those sets, please let me know!"
] | ## Describe the bug
When downloading some `wikipedia` languages (in particular, I'm having a hard time with Spanish, Cebuano, and Russian) via FlinkRunner, I encounter the exception in the title. I have been playing with package versions a lot, because unfortunately, the different dependencies required by these packages seem to be incompatible in terms of versions (dill and requests, for instance). It should be noted that the following code runs for several hours without issue, executing the `load_dataset()` function, before the exception occurs.
## Steps to reproduce the bug
```python
# bash commands
!pip install datasets
!pip install apache-beam[interactive]
!pip install mwparserfromhell
!pip install dill==0.3.5.1
!pip install requests==2.23.0
# imports
import os
from datasets import load_dataset
import apache_beam as beam
import mwparserfromhell
from google.colab import drive
import dill
import requests
# mount drive
drive_dir = os.path.join(os.getcwd(), 'drive')
drive.mount(drive_dir)
# confirming the versions of these two packages are the ones that are suggested by the outputs from the bash commands
print(dill.__version__)
print(requests.__version__)
lang = 'es' # or 'ru' or 'ceb' - these are the ones causing the issue
lang_dir = os.path.join(drive_dir, 'path/to/my/folder', lang)
if not os.path.exists(lang_dir):
x = None
x = load_dataset('wikipedia', '20220301.' + lang, beam_runner='Flink',
split='train')
x.save_to_disk(lang_dir)
```
## Expected results
Although some warnings are generally produced by this code (run in Colab Notebook), most languages I've tried have been successfully downloaded. It should simply go through without issue, but for these languages, I am continually encountering this error.
## Actual results
Traceback below:
```
Exception in thread run_worker_3-1:
Traceback (most recent call last):
File "/usr/lib/python3.7/threading.py", line 926, in _bootstrap_inner
self.run()
File "/usr/lib/python3.7/threading.py", line 870, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 234, in run
for work_request in self._control_stub.Control(get_responses()):
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 426, in __next__
return self._next()
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 826, in _next
raise self
grpc._channel._MultiThreadedRendezvous: <_MultiThreadedRendezvous of RPC that terminated with:
status = StatusCode.UNAVAILABLE
details = "Socket closed"
debug_error_string = "{"created":"@1655593643.871830638","description":"Error received from peer ipv4:127.0.0.1:44441","file":"src/core/lib/surface/call.cc","file_line":952,"grpc_message":"Socket closed","grpc_status":14}"
>
Traceback (most recent call last):
File "apache_beam/runners/common.py", line 1198, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 718, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam/runners/common.py", line 782, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 426, in __getitem__
self._cache[target_window] = self._side_input_data.view_fn(raw_view)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 391, in <lambda>
lambda iterable: from_runtime_iterable(iterable, view_options))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 512, in _from_runtime_iterable
head = list(itertools.islice(it, 2))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1228, in _lazy_iterator
self._underlying.get_raw(state_key, continuation_token))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1019, in get_raw
continuation_token=continuation_token)))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1060, in _blocking_request
raise RuntimeError(response.error)
RuntimeError: Unknown process bundle instruction id '26'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 267, in _execute
response = task()
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 340, in <lambda>
lambda: self.create_worker().do_instruction(request), request)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 581, in do_instruction
getattr(request, request_type), request.instruction_id)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 618, in process_bundle
bundle_processor.process_bundle(instruction_id))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 996, in process_bundle
element.data)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 221, in process_encoded
self.output(decoded_value)
File "apache_beam/runners/worker/operations.py", line 346, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam/runners/worker/operations.py", line 348, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam/runners/worker/operations.py", line 215, in apache_beam.runners.worker.operations.SingletonConsumerSet.receive
File "apache_beam/runners/worker/operations.py", line 707, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam/runners/worker/operations.py", line 708, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam/runners/common.py", line 1200, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 1281, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam/runners/common.py", line 1198, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 718, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam/runners/common.py", line 782, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 426, in __getitem__
self._cache[target_window] = self._side_input_data.view_fn(raw_view)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 391, in <lambda>
lambda iterable: from_runtime_iterable(iterable, view_options))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 512, in _from_runtime_iterable
head = list(itertools.islice(it, 2))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1228, in _lazy_iterator
self._underlying.get_raw(state_key, continuation_token))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1019, in get_raw
continuation_token=continuation_token)))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1060, in _blocking_request
raise RuntimeError(response.error)
RuntimeError: Unknown process bundle instruction id '26' [while running 'train/Save to parquet/Write/WriteImpl/WriteBundles']
ERROR:apache_beam.runners.worker.sdk_worker:Error processing instruction 26. Original traceback is
Traceback (most recent call last):
File "apache_beam/runners/common.py", line 1198, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 718, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam/runners/common.py", line 782, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 426, in __getitem__
self._cache[target_window] = self._side_input_data.view_fn(raw_view)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 391, in <lambda>
lambda iterable: from_runtime_iterable(iterable, view_options))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 512, in _from_runtime_iterable
head = list(itertools.islice(it, 2))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1228, in _lazy_iterator
self._underlying.get_raw(state_key, continuation_token))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1019, in get_raw
continuation_token=continuation_token)))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1060, in _blocking_request
raise RuntimeError(response.error)
RuntimeError: Unknown process bundle instruction id '26'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 267, in _execute
response = task()
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 340, in <lambda>
lambda: self.create_worker().do_instruction(request), request)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 581, in do_instruction
getattr(request, request_type), request.instruction_id)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 618, in process_bundle
bundle_processor.process_bundle(instruction_id))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 996, in process_bundle
element.data)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 221, in process_encoded
self.output(decoded_value)
File "apache_beam/runners/worker/operations.py", line 346, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam/runners/worker/operations.py", line 348, in apache_beam.runners.worker.operations.Operation.output
File "apache_beam/runners/worker/operations.py", line 215, in apache_beam.runners.worker.operations.SingletonConsumerSet.receive
File "apache_beam/runners/worker/operations.py", line 707, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam/runners/worker/operations.py", line 708, in apache_beam.runners.worker.operations.DoOperation.process
File "apache_beam/runners/common.py", line 1200, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 1281, in apache_beam.runners.common.DoFnRunner._reraise_augmented
File "apache_beam/runners/common.py", line 1198, in apache_beam.runners.common.DoFnRunner.process
File "apache_beam/runners/common.py", line 718, in apache_beam.runners.common.PerWindowInvoker.invoke_process
File "apache_beam/runners/common.py", line 782, in apache_beam.runners.common.PerWindowInvoker._invoke_process_per_window
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/bundle_processor.py", line 426, in __getitem__
self._cache[target_window] = self._side_input_data.view_fn(raw_view)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 391, in <lambda>
lambda iterable: from_runtime_iterable(iterable, view_options))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/pvalue.py", line 512, in _from_runtime_iterable
head = list(itertools.islice(it, 2))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1228, in _lazy_iterator
self._underlying.get_raw(state_key, continuation_token))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1019, in get_raw
continuation_token=continuation_token)))
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/sdk_worker.py", line 1060, in _blocking_request
raise RuntimeError(response.error)
RuntimeError: Unknown process bundle instruction id '26' [while running 'train/Save to parquet/Write/WriteImpl/WriteBundles']
ERROR:root:org.apache.beam.vendor.grpc.v1p43p2.io.grpc.StatusRuntimeException: CANCELLED: client cancelled
ERROR:apache_beam.runners.worker.data_plane:Failed to read inputs in the data plane.
Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/data_plane.py", line 634, in _read_inputs
for elements in elements_iterator:
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 426, in __next__
return self._next()
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 826, in _next
raise self
grpc._channel._MultiThreadedRendezvous: <_MultiThreadedRendezvous of RPC that terminated with:
status = StatusCode.CANCELLED
details = "Multiplexer hanging up"
debug_error_string = "{"created":"@1655593654.436885887","description":"Error received from peer ipv4:127.0.0.1:43263","file":"src/core/lib/surface/call.cc","file_line":952,"grpc_message":"Multiplexer hanging up","grpc_status":1}"
>
Exception in thread read_grpc_client_inputs:
Traceback (most recent call last):
File "/usr/lib/python3.7/threading.py", line 926, in _bootstrap_inner
self.run()
File "/usr/lib/python3.7/threading.py", line 870, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/data_plane.py", line 651, in <lambda>
target=lambda: self._read_inputs(elements_iterator),
File "/usr/local/lib/python3.7/dist-packages/apache_beam/runners/worker/data_plane.py", line 634, in _read_inputs
for elements in elements_iterator:
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 426, in __next__
return self._next()
File "/usr/local/lib/python3.7/dist-packages/grpc/_channel.py", line 826, in _next
raise self
grpc._channel._MultiThreadedRendezvous: <_MultiThreadedRendezvous of RPC that terminated with:
status = StatusCode.CANCELLED
details = "Multiplexer hanging up"
debug_error_string = "{"created":"@1655593654.436885887","description":"Error received from peer ipv4:127.0.0.1:43263","file":"src/core/lib/surface/call.cc","file_line":952,"grpc_message":"Multiplexer hanging up","grpc_status":1}"
>
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
[/tmp/ipykernel_219/3869142325.py](https://localhost:8080/#) in <module>
18 x = None
19 x = load_dataset('wikipedia', '20220301.' + lang, beam_runner='Flink',
---> 20 split='train')
21 x.save_to_disk(lang_dir)
3 frames
[/usr/local/lib/python3.7/dist-packages/apache_beam/runners/portability/portable_runner.py](https://localhost:8080/#) in wait_until_finish(self, duration)
604
605 if self._runtime_exception:
--> 606 raise self._runtime_exception
607
608 return self._state
RuntimeError: Pipeline BeamApp-root-0618220708-b3b59a0e_d8efcf67-9119-4f76-b013-70de7b29b54d failed in state FAILED: org.apache.beam.vendor.grpc.v1p43p2.io.grpc.StatusRuntimeException: CANCELLED: client cancelled
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.2
- 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
| 4,524 |
https://github.com/huggingface/datasets/issues/4522 | Try to reduce the number of datasets that require manual download | [] | > Currently, 41 canonical datasets require manual download. I checked their scripts and I'm pretty sure this number can be reduced to β 30 by not relying on bash scripts to download data, hosting data directly on the Hub when the license permits, etc. Then, we will mostly be left with datasets with restricted access, which we can ignore
from https://github.com/huggingface/datasets-server/issues/12#issuecomment-1026920432 | 4,522 |
https://github.com/huggingface/datasets/issues/4521 | Datasets method `.map` not hashing | [
"Fix posted: https://github.com/huggingface/datasets/issues/4506#issuecomment-1157417219",
"Didn't realize it's a bug when I asked the question yesterday! Free free to post an answer if you are sure the cause has been addressed.\r\n\r\nhttps://stackoverflow.com/questions/72664827/can-pickle-dill-foo-but-not-lambda-x-foox",
"Thank @nalzok . That works for me:\r\n\r\n`pip install \"dill<0.3.5\"`"
] | ## Describe the bug
Datasets method `.map` not hashing, even with an empty no-op function
## Steps to reproduce the bug
```python
from datasets import load_dataset
# download 9MB dummy dataset
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean")
def prepare_dataset(batch):
return(batch)
ds = ds.map(
prepare_dataset,
num_proc=1,
desc="preprocess train dataset",
)
```
## Expected results
Hashed and cached dataset preprocessing
## Actual results
Does not hash properly:
```
Parameter 'function'=<function prepare_dataset at 0x7fccb68e9280> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.3.dev0
- Platform: Linux-5.11.0-1028-gcp-x86_64-with-glibc2.31
- Python version: 3.9.12
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
cc @lhoestq
| 4,521 |
https://github.com/huggingface/datasets/issues/4520 | Failure to hash `dataclasses` - results in functions that cannot be hashed or cached in `.map` | [
"I think this has been fixed by #4516, let me know if you encounter this again :)\r\n\r\nI re-ran your code in 3.7 and 3.9 and it works fine",
"Thank you!"
] | Dataclasses cannot be hashed. As a result, they cannot be hashed or cached if used in the `.map` method. Dataclasses are used extensively in Transformers examples scripts: (c.f. [CTC example](https://github.com/huggingface/transformers/blob/main/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py)). Since dataclasses cannot be hashed, one has to define separate variables prior to passing dataclass attributes to the `.map` method:
```python
phoneme_language = data_args.phoneme_language
```
in the example https://github.com/huggingface/transformers/blob/3c7e56fbb11f401de2528c1dcf0e282febc031cd/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L603-L630
## Steps to reproduce the bug
```python
from dataclasses import dataclass, field
from datasets.fingerprint import Hasher
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
phoneme_language: str = field(
default=None, metadata={"help": "The name of the phoneme language to use."}
)
data_args = DataTrainingArguments(phoneme_language ="foo")
Hasher.hash(data_args)
phoneme_language = data_args.phoneme_language
Hasher.hash(phoneme_language)
```
## Expected results
A hash.
## Actual results
<details>
<summary> Traceback </summary>
```
---------------------------------------------------------------------------
KeyError Traceback (most recent call last)
Input In [1], in <cell line: 16>()
10 phoneme_language: str = field(
11 default=None, metadata={"help": "The name of the phoneme language to use."}
12 )
14 data_args = DataTrainingArguments(phoneme_language ="foo")
---> 16 Hasher.hash(data_args)
18 phoneme_language = data_args. phoneme_language
20 Hasher.hash(phoneme_language)
File ~/datasets/src/datasets/fingerprint.py:237, in Hasher.hash(cls, value)
235 return cls.dispatch[type(value)](cls, value)
236 else:
--> 237 return cls.hash_default(value)
File ~/datasets/src/datasets/fingerprint.py:230, in Hasher.hash_default(cls, value)
228 @classmethod
229 def hash_default(cls, value: Any) -> str:
--> 230 return cls.hash_bytes(dumps(value))
File ~/datasets/src/datasets/utils/py_utils.py:564, in dumps(obj)
562 file = StringIO()
563 with _no_cache_fields(obj):
--> 564 dump(obj, file)
565 return file.getvalue()
File ~/datasets/src/datasets/utils/py_utils.py:539, in dump(obj, file)
537 def dump(obj, file):
538 """pickle an object to a file"""
--> 539 Pickler(file, recurse=True).dump(obj)
540 return
File ~/hf/lib/python3.8/site-packages/dill/_dill.py:620, in Pickler.dump(self, obj)
618 raise PicklingError(msg)
619 else:
--> 620 StockPickler.dump(self, obj)
621 return
File /usr/lib/python3.8/pickle.py:487, in _Pickler.dump(self, obj)
485 if self.proto >= 4:
486 self.framer.start_framing()
--> 487 self.save(obj)
488 self.write(STOP)
489 self.framer.end_framing()
File /usr/lib/python3.8/pickle.py:603, in _Pickler.save(self, obj, save_persistent_id)
599 raise PicklingError("Tuple returned by %s must have "
600 "two to six elements" % reduce)
602 # Save the reduce() output and finally memoize the object
--> 603 self.save_reduce(obj=obj, *rv)
File /usr/lib/python3.8/pickle.py:687, in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
684 raise PicklingError(
685 "args[0] from __newobj__ args has the wrong class")
686 args = args[1:]
--> 687 save(cls)
688 save(args)
689 write(NEWOBJ)
File /usr/lib/python3.8/pickle.py:560, in _Pickler.save(self, obj, save_persistent_id)
558 f = self.dispatch.get(t)
559 if f is not None:
--> 560 f(self, obj) # Call unbound method with explicit self
561 return
563 # Check private dispatch table if any, or else
564 # copyreg.dispatch_table
File ~/hf/lib/python3.8/site-packages/dill/_dill.py:1838, in save_type(pickler, obj, postproc_list)
1836 postproc_list = []
1837 postproc_list.append((setattr, (obj, '__qualname__', obj_name)))
-> 1838 _save_with_postproc(pickler, (_create_type, (
1839 type(obj), obj.__name__, obj.__bases__, _dict
1840 )), obj=obj, postproc_list=postproc_list)
1841 log.info("# %s" % _t)
1842 else:
File ~/hf/lib/python3.8/site-packages/dill/_dill.py:1140, in _save_with_postproc(pickler, reduction, is_pickler_dill, obj, postproc_list)
1137 pickler._postproc[id(obj)] = postproc_list
1139 # TODO: Use state_setter in Python 3.8 to allow for faster cPickle implementations
-> 1140 pickler.save_reduce(*reduction, obj=obj)
1142 if is_pickler_dill:
1143 # pickler.x -= 1
1144 # print(pickler.x*' ', 'pop', obj, id(obj))
1145 postproc = pickler._postproc.pop(id(obj))
File /usr/lib/python3.8/pickle.py:692, in _Pickler.save_reduce(self, func, args, state, listitems, dictitems, state_setter, obj)
690 else:
691 save(func)
--> 692 save(args)
693 write(REDUCE)
695 if obj is not None:
696 # If the object is already in the memo, this means it is
697 # recursive. In this case, throw away everything we put on the
698 # stack, and fetch the object back from the memo.
File /usr/lib/python3.8/pickle.py:560, in _Pickler.save(self, obj, save_persistent_id)
558 f = self.dispatch.get(t)
559 if f is not None:
--> 560 f(self, obj) # Call unbound method with explicit self
561 return
563 # Check private dispatch table if any, or else
564 # copyreg.dispatch_table
File /usr/lib/python3.8/pickle.py:901, in _Pickler.save_tuple(self, obj)
899 write(MARK)
900 for element in obj:
--> 901 save(element)
903 if id(obj) in memo:
904 # Subtle. d was not in memo when we entered save_tuple(), so
905 # the process of saving the tuple's elements must have saved
(...)
909 # could have been done in the "for element" loop instead, but
910 # recursive tuples are a rare thing.
911 get = self.get(memo[id(obj)][0])
File /usr/lib/python3.8/pickle.py:560, in _Pickler.save(self, obj, save_persistent_id)
558 f = self.dispatch.get(t)
559 if f is not None:
--> 560 f(self, obj) # Call unbound method with explicit self
561 return
563 # Check private dispatch table if any, or else
564 # copyreg.dispatch_table
File ~/hf/lib/python3.8/site-packages/dill/_dill.py:1251, in save_module_dict(pickler, obj)
1248 if is_dill(pickler, child=False) and pickler._session:
1249 # we only care about session the first pass thru
1250 pickler._first_pass = False
-> 1251 StockPickler.save_dict(pickler, obj)
1252 log.info("# D2")
1253 return
File /usr/lib/python3.8/pickle.py:971, in _Pickler.save_dict(self, obj)
968 self.write(MARK + DICT)
970 self.memoize(obj)
--> 971 self._batch_setitems(obj.items())
File /usr/lib/python3.8/pickle.py:997, in _Pickler._batch_setitems(self, items)
995 for k, v in tmp:
996 save(k)
--> 997 save(v)
998 write(SETITEMS)
999 elif n:
File /usr/lib/python3.8/pickle.py:560, in _Pickler.save(self, obj, save_persistent_id)
558 f = self.dispatch.get(t)
559 if f is not None:
--> 560 f(self, obj) # Call unbound method with explicit self
561 return
563 # Check private dispatch table if any, or else
564 # copyreg.dispatch_table
File ~/datasets/src/datasets/utils/py_utils.py:862, in save_function(pickler, obj)
859 if state_dict:
860 state = state, state_dict
--> 862 dill._dill._save_with_postproc(
863 pickler,
864 (
865 dill._dill._create_function,
866 (obj.__code__, globs, obj.__name__, obj.__defaults__, closure),
867 state,
868 ),
869 obj=obj,
870 postproc_list=postproc_list,
871 )
872 else:
873 closure = obj.func_closure
File ~/hf/lib/python3.8/site-packages/dill/_dill.py:1153, in _save_with_postproc(pickler, reduction, is_pickler_dill, obj, postproc_list)
1151 dest, source = reduction[1]
1152 if source:
-> 1153 pickler.write(pickler.get(pickler.memo[id(dest)][0]))
1154 pickler._batch_setitems(iter(source.items()))
1155 else:
1156 # Updating with an empty dictionary. Same as doing nothing.
KeyError: 140434581781568
```
</details>
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.3.dev0
- Platform: Linux-5.11.0-1028-gcp-x86_64-with-glibc2.29
- Python version: 3.8.10
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
cc @lhoestq | 4,520 |
https://github.com/huggingface/datasets/issues/4514 | Allow .JPEG as a file extension | [
"Hi, thanks for reporting! I've opened a PR with the fix.",
"Wow, that was quick! Thank you very much π "
] | ## Describe the bug
When loading image data, HF datasets seems to recognize `.jpg` and `.jpeg` file extensions, but not e.g. .JPEG. As the naming convention .JPEG is used in important datasets such as imagenet, I would welcome if according extensions like .JPEG or .JPG would be allowed.
## Steps to reproduce the bug
```python
# use bash to create 2 sham datasets with jpeg and JPEG ext
!mkdir dataset_a
!mkdir dataset_b
!wget https://upload.wikimedia.org/wikipedia/commons/7/71/Dsc_%28179253513%29.jpeg -O example_img.jpeg
!cp example_img.jpeg ./dataset_a/
!mv example_img.jpeg ./dataset_b/example_img.JPEG
from datasets import load_dataset
# working
df1 = load_dataset("./dataset_a", ignore_verifications=True)
#not working
df2 = load_dataset("./dataset_b", ignore_verifications=True)
# show
print(df1, df2)
```
## Expected results
```
DatasetDict({
train: Dataset({
features: ['image', 'label'],
num_rows: 1
})
}) DatasetDict({
train: Dataset({
features: ['image', 'label'],
num_rows: 1
})
})
```
## Actual results
```
FileNotFoundError: Unable to resolve any data file that matches '['**']' at /..PATH../dataset_b with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'blp', 'bmp', 'dib', 'bufr', 'cur', 'pcx', 'dcx', 'dds', 'ps', 'eps', 'fit', 'fits', 'fli', 'flc', 'ftc', 'ftu', 'gbr', 'gif', 'grib', 'h5', 'hdf', 'png', 'apng', 'jp2', 'j2k', 'jpc', 'jpf', 'jpx', 'j2c', 'icns', 'ico', 'im', 'iim', 'tif', 'tiff', 'jfif', 'jpe', 'jpg', 'jpeg', 'mpg', 'mpeg', 'msp', 'pcd', 'pxr', 'pbm', 'pgm', 'ppm', 'pnm', 'psd', 'bw', 'rgb', 'rgba', 'sgi', 'ras', 'tga', 'icb', 'vda', 'vst', 'webp', 'wmf', 'emf', 'xbm', 'xpm', 'zip']
```
I know that it can be annoying to allow seemingly arbitrary numbers of file extensions. But I think this one would be really welcome. | 4,514 |
https://github.com/huggingface/datasets/issues/4508 | cast_storage method from datasets.features | [
"Hi! We've recently added a check to the `ClassLabel` type to ensure the values are in the valid label range `-1, 0, ..., num_classes-1` (-1 is used for missing values). The error in your case happens only if the `labels` column is of type `Sequence(ClassLabel(...))` before the `map` call and can be avoided by calling `dataset = dataset.cast_column(\"labels\", Sequence(Value(\"int\")))` beforehand. The token-classification examples in Transformers introduce a new `labels` column, so their type is also `Sequence(Value(\"int\"))`, which doesn't lead to an error as this type unbounded. ",
"I'm fine with re-adding support for all negative values for unknown/missing labels @mariosasko, wdyt ?"
] | ## Describe the bug
A bug occurs when mapping a function to a dataset object. I ran the same code with the same data yesterday and it worked just fine. It works when i run locally on an old version of datasets.
## Steps to reproduce the bug
Steps are:
- load whatever datset
- write a preprocessing function such as "tokenize_and_align_labels" written in https://huggingface.co/docs/transformers/tasks/token_classification
- map the function on dataset and get "ValueError: Class label -100 less than -1" from cast_storage method from datasets.features
# Sample code to reproduce the bug
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(examples["tokens"], truncation=True, is_split_into_words=True, max_length=38,padding="max_length")
labels = []
for i, label in enumerate(examples[f"labels"]):
word_ids = tokenized_inputs.word_ids(batch_index=i) # Map tokens to their respective word.
previous_word_idx = None
label_ids = []
for word_idx in word_ids: # Set the special tokens to -100.
if word_idx is None:
label_ids.append(-100)
elif word_idx != previous_word_idx: # Only label the first token of a given word.
label_ids.append(label[word_idx])
else:
label_ids.append(-100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
dt = dataset.map(tokenize_and_align_labels, batched=True)
## Expected results
New dataset objects should load and do on older versions.
## Actual results
"ValueError: Class label -100 less than -1" from cast_storage method from datasets.features
## Environment info
everything works fine on older installations of datasets/transformers
Issue arises when installing datasets on google collab under python3.7
I can't manage to find the exact output you're requirering but version printed is datasets-2.3.2
| 4,508 |
https://github.com/huggingface/datasets/issues/4507 | How to let `load_dataset` return a `Dataset` instead of `DatasetDict` in customized loading script | [
"Hi @liyucheng09.\r\n\r\nUsers can pass the `split` parameter to `load_dataset`. For example, if your split name is \"train\",\r\n```python\r\nds = load_dataset(\"dataset_name\", split=\"train\")\r\n```\r\nwill return a `Dataset` instance.",
"@albertvillanova Thanks! I can't believe I didn't know this feature till now."
] | If the dataset does not need splits, i.e., no training and validation split, more like a table. How can I let the `load_dataset` function return a `Dataset` object directly rather than return a `DatasetDict` object with only one key-value pair.
Or I can paraphrase the question in the following way: how to skip `_split_generators` step in `DatasetBuilder` to let `as_dataset` gives a single `Dataset` rather than a list`[Dataset]`?
Many thanks for any help. | 4,507 |
https://github.com/huggingface/datasets/issues/4506 | Failure to hash (and cache) a `.map(...)` (almost always) - using this method can produce incorrect results | [
"Important info:\r\n\r\nAs hashes are generated randomly for functions, it leads to **false identifying some results as already hashed** (mapping function is not executed after a method update) when there's a `pytorch_lightning.seed_everything(123)`",
"@lhoestq\r\nseems like quite critical stuff for me, if I'm not making a mistake",
"Hi ! Thanks for reporting. This bug seems to appear in python 3.9 using dill 3.5.1\r\n\r\nAs a workaround you can use an older version of dill:\r\n```\r\npip install \"dill<0.3.5\"\r\n```",
"installing `dill<0.3.5` after installing `datasets` by pip results in dependency conflict with the version required for `multiprocess`. It can be solved by installing `pip install datasets \"dill<0.3.5\"` (simultaneously) on a clean environment",
"This has been fixed in https://github.com/huggingface/datasets/pull/4516, we will do a new release soon to include the fix :)"
] | ## Describe the bug
Sometimes I get messages about not being able to hash a method:
`Parameter 'function'=<function StupidDataModule._separate_speaker_id_from_dialogue at 0x7f1b27180d30> of the transform datasets.arrow_dataset.Dataset.
_map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.`
Whilst the function looks like this:
```python
@staticmethod
def _separate_speaker_id_from_dialogue(example: arrow_dataset.Example):
speaker_id, dialogue = tuple(zip(*(example["dialogue"])))
example["speaker_id"] = speaker_id
example["dialogue"] = dialogue
return example
```
This is the first step in my preprocessing pipeline, but sometimes the message about failure to hash is not appearing on the first step, but then appears on a later step.
This error is sometimes causing a failure to use cached data, instead of re-running all steps again.
## Steps to reproduce the bug
```python
import copy
import datasets
from datasets import arrow_dataset
def main():
dataset = datasets.load_dataset("blended_skill_talk")
res = dataset.map(method)
print(res)
def method(example: arrow_dataset.Example):
example['previous_utterance_copy'] = copy.deepcopy(example['previous_utterance'])
return example
if __name__ == '__main__':
main()
```
Run with:
```
python -m reproduce_error
```
## Expected results
Dataset is mapped and cached correctly.
## Actual results
The code outputs this at some point:
`Parameter 'function'=<function method at 0x7faa83d2a160> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: Ubuntu 20.04.3
- Python version: 3.9.12
- PyArrow version: 8.0.0
- Datasets version: 2.3.1
| 4,506 |
https://github.com/huggingface/datasets/issues/4504 | Can you please add the Stanford dog dataset? | [
"would you like to give it a try, @dgrnd4? (maybe with the help of the dataset author?)",
"@julien-c i am sorry but I have no idea about how it works: can I add the dataset by myself, following \"instructions to add a new dataset\"?\r\nCan I add a dataset even if it's not mine? (it's public in the link that I wrote on the post)\r\n",
"Hi! The [ADD NEW DATASET](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md) instructions are indeed the best place to start. It's also perfectly fine to add a dataset if it's public, even if it's not yours. Let me know if you need some additional pointers.",
"If no one is working on this, I could take this up!",
"@khushmeeet this is the [link](https://huggingface.co/datasets/dgrnd4/stanford_dog_dataset) where I added the dataset already. If you can I would ask you to do this:\r\n1) The dataset it's all in TRAINING SET: can you please divide it in Training,Test and Validation Set? If you can for each class, take the 80% for the Training set and the 10% for Test and 10% Validation\r\n2) The images has different size, can you please resize all the images in 224,224,3? Look even at the last dimension \"3\" because some images has dimension 4!\r\n\r\nThank you!!",
"Hi @khushmeeet! Thanks for the interest. You can self-assign the issue by commenting `#self-assign` on it. \r\n\r\nAlso, I think we can skip @dgrnd4's steps as we try to avoid any custom processing on top of raw data. One can later copy the script and override `_post_process` in it to perform such processing on the generated dataset.",
"Thanks @mariosasko \r\n\r\n@dgrnd4 As dataset is there on Hub, and preprocessing is not recommended. I am not sure if there is any other task to do. However, I can't seem to find relevant `.py` files for this dataset in GitHub repo.",
"@khushmeeet @mariosasko The point is that the images must be processed and must have the same size in order to can be used for things for example \"Training\". ",
"@dgrnd4 Yes, but this can be done after loading (`map` to resize images and `train_test_split` to create extra splits)\r\n\r\n@khushmeeet The linked version is implemented as a no-code dataset and is generated directly from the ZIP archive, but our \"GitHub\" datasets (these are datasets without a user/org namespace on the Hub) need a generation script, and you can find one [here](https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/image_classification/stanford_dogs.py). `datasets` started as a fork of TFDS, so we share similar script structure, which makes it trivial to adapt it.",
"@mariosasko The point is that if I use something like this:\r\nx_train, x_test = train_test_split(dataset, test_size=0.1) \r\n\r\nto get Train 90% and Test 10%, and then to get the Validation Set (10% of the whole 100%):\r\n\r\n```\r\ntrain_ratio = 0.80\r\nvalidation_ratio = 0.10\r\ntest_ratio = 0.10\r\n\r\nx_train, x_test, y_train, y_test = train_test_split(dataX, dataY, test_size=1 - train_ratio)\r\nx_val, x_test, y_val, y_test = train_test_split(x_test, y_test, test_size=test_ratio/(test_ratio + validation_ratio)) \r\n\r\n```\r\n\r\nThe point is that the structure of the data is:\r\n```\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['image', 'label'],\r\n num_rows: 20580\r\n })\r\n})\r\n\r\n```\r\n\r\nSo how to extract images and labels?\r\n\r\nEDIT --> Split of the dataset in Train-Test-Validation:\r\n```\r\nimport datasets\r\nfrom datasets.dataset_dict import DatasetDict\r\nfrom datasets import Dataset\r\n\r\npercentage_divison_test = int(len(dataset['train'])/100 *10) # 10% --> 2058 \r\npercentage_divison_validation = int(len(dataset['train'])/100 *20) # 20% --> 4116\r\n\r\ndataset_ = datasets.DatasetDict({\"train\": Dataset.from_dict({\r\n\r\n 'image': dataset['train'][0 : len(dataset['train']) ]['image'], \r\n 'labels': dataset['train'][0 : len(dataset['train']) ]['label'] }), \r\n \r\n \"test\": Dataset.from_dict({ #20580-4116 (validation) ,20580-2058 (test)\r\n 'image': dataset['train'][len(dataset['train']) - percentage_divison_validation : len(dataset['train']) - percentage_divison_test]['image'], \r\n 'labels': dataset['train'][len(dataset['train']) - percentage_divison_validation : len(dataset['train']) - percentage_divison_test]['label'] }), \r\n \r\n \"validation\": Dataset.from_dict({ # 20580-2058 (test)\r\n 'image': dataset['train'][len(dataset['train']) - percentage_divison_test : len(dataset['train'])]['image'], \r\n 'labels': dataset['train'][len(dataset['train']) - percentage_divison_test : len(dataset['train'])]['label'] }), \r\n })\r\n```",
"@mariosasko in order to resize images I'm trying this method: \r\n```\r\nfor i in range(0,len(dataset['train'])): #len(dataset['train'])\r\n\r\n ex = dataset['train'][i] #i\r\n image = ex['image']\r\n image = image.convert(\"RGB\") # <class 'PIL.Image.Image'> <PIL.Image.Image image mode=RGB size=500x333 at 0x7F84F1948150>\r\n image_resized = image.resize(size_to_resize) # <PIL.Image.Image image mode=RGB size=224x224 at 0x7F84F17885D0>\r\n\r\n dataset['train'][i]['image'] = image_resized \r\n```\r\n\r\nBecause the DatasetDict is formed by arrows that are immutable, the changing assignment in the last line of code, doesn't work!\r\nDo you have any idea in order to get a valid result?",
"#self-assign",
"I have raised PR for adding stanford-dog dataset. I have not added any data preprocessing code. Only dataset generation script is there. Let me know any changes required, or anything to add to README.",
"Is this issue still open, i am new to open source thus want to take this one as my start.",
"@zutarich This issue should have been closed since the dataset in question is available on the Hub [here](https://huggingface.co/datasets/dgrnd4/stanford_dog_dataset)."
] | ## Adding a Dataset
- **Name:** *Stanford dog dataset*
- **Description:** *The dataset is about 120 classes for a total of 20.580 images. You can find the dataset here http://vision.stanford.edu/aditya86/ImageNetDogs/*
- **Paper:** *http://vision.stanford.edu/aditya86/ImageNetDogs/*
- **Data:** *[link to the Github repository or current dataset location](http://vision.stanford.edu/aditya86/ImageNetDogs/)*
- **Motivation:** *The dataset has been built using images and annotation from ImageNet for the task of fine-grained image categorization. It is useful for fine-grain purpose *
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 4,504 |
https://github.com/huggingface/datasets/issues/4502 | Logic bug in arrow_writer? | [
"Hi @cccntu you're right, as when `batch_examples={}` the current if-statement won't be triggered as the condition won't be satisfied, I'll prepare a PR to address it as well as add the regression tests so that this issue is handled properly.",
"Hi @alvarobartt ,\r\nThanks for answering. Do you know when and why an empty batch is passed to this function? This only happened to me when processing with multiple workers, while chunking examples, I think.",
"> Hi @alvarobartt , Thanks for answering. Do you know when and why an empty batch is passed to this function? This only happened to me when processing with multiple workers, while chunking examples, I think.\r\n\r\nSo it depends on how you're actually chunking the data as if you're not handling empty chunks `batch_examples={}` or `batch_examples=None`, you may end up running into this issue. So you could check the chunks before you actually call `ArrowWriter.write_batch`, but anyway the fix you proposed I think improves the logic of `write_batch` to avoid running into these issues.",
"Thanks, I added a if-print and I found it does return an empty examples in the chunking function that is passed to `.map()`.",
"Hi ! We consider an empty batch to look like this:\r\n```python\r\nempty_batch = {\r\n \"column_1\": [],\r\n \"column_2\": [],\r\n ...\r\n}\r\n```\r\n\r\nWhile `{}` corresponds to a batch with no columns.\r\n\r\nTherefore calling this code should fail, because the two batches don't have the same columns:\r\n```python\r\nwriter.write_batch({\"a\": [1, 2, 3]})\r\nwriter.write_batch({})\r\n```\r\n\r\nIf you want to write an empty batch, you should do this instead:\r\n```python\r\nwriter.write_batch({\"a\": [1, 2, 3]})\r\nwriter.write_batch({\"a\": []})\r\n```",
"Makes sense, then the if-statement should remain the same or is it better to handle both cases separately using `if not batch_examples or len(next(iter(batch_examples.values()))) == 0: ...`?\r\n\r\nUpdating the regressions tests with an empty batch formatted as `{\"col_1\": [], \"col_2\": []}` instead of `{}` works fine with the current if, and also with the one proposed by @cccntu.",
"> Makes sense, then the if-statement should remain the same or is it better to handle both cases separately using if not batch_examples or len(next(iter(batch_examples.values()))) == 0: ...?\r\n\r\nThere's a check later in the code that makes sure that the columns are the right ones, so I don't think we need to check for `{}` here\r\n\r\nIn particular the check `if not batch_examples or len(next(iter(batch_examples.values()))) == 0:` doesn't raise an error while it should, that why the old `if` is fine IMO\r\n\r\n> Updating the regressions tests with an empty batch formatted as {\"col_1\": [], \"col_2\": []} instead of {} works fine with the current if, and also with the one proposed by @cccntu.\r\n\r\nCool ! If you want you can update your PR to add the regression tests, to make sure that `{\"col_1\": [], \"col_2\": []}` works but not `{}`",
"Great thanks for the response! So I'll just add that regression test and remove the current if-statement.",
"Hi @lhoestq ,\r\n\r\nThanks for your explanation. Now I get it that `{}` means the columns are different. But wouldn't it be nice if the code can ignore it, like it ignores `{\"a\": []}`?\r\n\r\n\r\n--- \r\nBTW, \r\n> There's a check later in the code that makes sure that the columns are the right ones, so I don't think we need to check for {} here\r\n\r\nI remember the error happens around here:\r\nhttps://github.com/huggingface/datasets/blob/88a902d6474fae8d793542d57a4f3b0d187f3c5b/src/datasets/arrow_writer.py#L506-L507\r\nThe error says something like `arrays` and `schema` doesn't have the same length. And it's not very clear I passed a `{}`.\r\n\r\nedit: actual error message\r\n```\r\nFile \"site-packages/datasets/arrow_writer.py\", line 595, in write_batch\r\n pa_table = pa.Table.from_arrays(arrays, schema=schema)\r\n File \"pyarrow/table.pxi\", line 3557, in pyarrow.lib.Table.from_arrays\r\n File \"pyarrow/table.pxi\", line 1401, in pyarrow.lib._sanitize_arrays\r\nValueError: Schema and number of arrays unequal\r\n```",
"> But wouldn't it be nice if the code can ignore it, like it ignores {\"a\": []}?\r\n\r\nI think it would make things confusing because it doesn't follow our definition of a batch: \"the columns of a batch = the keys of the dict\". It would probably break certain behaviors as well. For example if you remove all the columns of a dataset (using `.remove_colums(...)` or `.map(..., remove_columns=...)`), the writer has to write 0 columns, and currently the only way to tell the writer to do so using `write_batch` is to pass `{}`.\r\n\r\n> The error says something like arrays and schema doesn't have the same length. And it's not very clear I passed a {}.\r\n\r\nYea the message can actually be improved indeed, it's definitely not clear. Maybe we can add a line right before the call `pa.Table.from_arrays` to make sure the keys of the batch match the field names of the schema"
] | https://github.com/huggingface/datasets/blob/88a902d6474fae8d793542d57a4f3b0d187f3c5b/src/datasets/arrow_writer.py#L475-L488
I got some error, and I found it's caused by `batch_examples` being `{}`. I wonder if the code should be as follows:
```
- if batch_examples and len(next(iter(batch_examples.values()))) == 0:
+ if not batch_examples or len(next(iter(batch_examples.values()))) == 0:
return
```
@lhoestq | 4,502 |
https://github.com/huggingface/datasets/issues/4498 | WER and CER > 1 | [
"WER can have values bigger than 1.0, this is expected when there are too many insertions\r\n\r\nFrom [wikipedia](https://en.wikipedia.org/wiki/Word_error_rate):\r\n> Note that since N is the number of words in the reference, the word error rate can be larger than 1.0"
] | ## Describe the bug
It seems that in some cases in which the `prediction` is longer than the `reference` we may have word/character error rate higher than 1 which is a bit odd.
If it's a real bug I think I can solve it with a PR changing [this](https://github.com/huggingface/datasets/blob/master/metrics/wer/wer.py#L105) line to
```python
return min(incorrect / total, 1.0)
```
## Steps to reproduce the bug
```python
from datasets import load_metric
wer = load_metric("wer")
wer_value = wer.compute(predictions=["Hi World vka"], references=["Hello"])
print(wer_value)
```
## Expected results
```
1.0
```
## Actual results
```
3.0
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.3.0
- 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 | 4,498 |
https://github.com/huggingface/datasets/issues/4494 | Patching fails for modules that are not installed or don't exist | [] | Reported in https://github.com/huggingface/huggingface_hub/runs/6894703718?check_suite_focus=true
When trying to patch `scipy.io.loadmat`:
```python
ModuleNotFoundError: No module named 'scipy'
```
Instead it shouldn't raise an error and do nothing
We use patching to extend such functions to support remote URLs and work in streaming mode | 4,494 |
https://github.com/huggingface/datasets/issues/4491 | Dataset Viewer issue for Pavithree/test | [
"This issue can be resolved according to this post https://stackoverflow.com/questions/70566660/parquet-with-null-columns-on-pyarrow. It looks like first data entry in the json file must not have any null values as pyarrow uses this first file to infer schema for entire dataset."
] | ### Link
https://huggingface.co/datasets/Pavithree/test
### Description
I have extracted the subset of original eli5 dataset found at hugging face. However, while loading the dataset It throws ArrowNotImplementedError: Unsupported cast from string to null using function cast_null error. Is there anything missing from my end? Kindly help.
### Owner
_No response_ | 4,491 |
https://github.com/huggingface/datasets/issues/4490 | Use `torch.nested_tensor` for arrays of varying length in torch formatter | [
"What's the current behavior?",
"Currently, we return a list of Torch tensors if their shapes don't match. If they do, we consolidate them into a single Torch tensor."
] | Use `torch.nested_tensor` for arrays of varying length in `TorchFormatter`.
The PyTorch API of nested tensors is in the prototype stage, so wait for it to become more mature. | 4,490 |
https://github.com/huggingface/datasets/issues/4483 | Dataset.map throws pyarrow.lib.ArrowNotImplementedError when converting from list of empty lists | [
"Hi @sanderland ! Thanks for reporting :) This is a bug, I opened a PR to fix it. We'll do a new release soon\r\n\r\nIn the meantime you can fix it by specifying in advance that the \"label\" are integers:\r\n```python\r\nimport numpy as np\r\n\r\nds = Dataset.from_dict(\r\n {\r\n \"text\": [\"the lazy dog jumps over the quick fox\", \"another sentence\"],\r\n \"label\": [[], []],\r\n }\r\n)\r\n# explicitly say that the \"label\" type is int64, even though it contains only null values\r\nds = ds.cast_column(\"label\", Sequence(Value(\"int64\")))\r\n\r\ndef mapper(features):\r\n features['label'] = [\r\n [0,0,0] for l in features['label']\r\n ]\r\n return features\r\n\r\nds_mapped = ds.map(mapper,batched=True)\r\n```"
] | ## Describe the bug
Dataset.map throws pyarrow.lib.ArrowNotImplementedError: Unsupported cast from int64 to null using function cast_null when converting from a type of 'empty lists' to 'lists with some type'.
This appears to be due to the interaction of arrow internals and some assumptions made by datasets.
The bug appeared when binarizing some labels, and then adding a dataset which had all these labels absent (to force the model to not label empty strings such with anything)
Particularly the fact that this only happens in batched mode is strange.
## Steps to reproduce the bug
```python
import numpy as np
ds = Dataset.from_dict(
{
"text": ["the lazy dog jumps over the quick fox", "another sentence"],
"label": [[], []],
}
)
def mapper(features):
features['label'] = [
[0,0,0] for l in features['label']
]
return features
ds_mapped = ds.map(mapper,batched=True)
```
## Expected results
Not crashing
## Actual results
```
../.venv/lib/python3.8/site-packages/datasets/arrow_dataset.py:2346: in map
return self._map_single(
../.venv/lib/python3.8/site-packages/datasets/arrow_dataset.py:532: in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/arrow_dataset.py:499: in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/fingerprint.py:458: in wrapper
out = func(self, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/arrow_dataset.py:2751: in _map_single
writer.write_batch(batch)
../.venv/lib/python3.8/site-packages/datasets/arrow_writer.py:503: in write_batch
arrays.append(pa.array(typed_sequence))
pyarrow/array.pxi:230: in pyarrow.lib.array
???
pyarrow/array.pxi:110: in pyarrow.lib._handle_arrow_array_protocol
???
../.venv/lib/python3.8/site-packages/datasets/arrow_writer.py:198: in __arrow_array__
out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
../.venv/lib/python3.8/site-packages/datasets/table.py:1675: in wrapper
return func(array, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/table.py:1812: in cast_array_to_feature
casted_values = _c(array.values, feature.feature)
../.venv/lib/python3.8/site-packages/datasets/table.py:1675: in wrapper
return func(array, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/table.py:1843: in cast_array_to_feature
return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
../.venv/lib/python3.8/site-packages/datasets/table.py:1675: in wrapper
return func(array, *args, **kwargs)
../.venv/lib/python3.8/site-packages/datasets/table.py:1752: in array_cast
return array.cast(pa_type)
pyarrow/array.pxi:915: in pyarrow.lib.Array.cast
???
../.venv/lib/python3.8/site-packages/pyarrow/compute.py:376: in cast
return call_function("cast", [arr], options)
pyarrow/_compute.pyx:542: in pyarrow._compute.call_function
???
pyarrow/_compute.pyx:341: in pyarrow._compute.Function.call
???
pyarrow/error.pxi:144: in pyarrow.lib.pyarrow_internal_check_status
???
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
> ???
E pyarrow.lib.ArrowNotImplementedError: Unsupported cast from int64 to null using function cast_null
pyarrow/error.pxi:121: ArrowNotImplementedError
```
## Workarounds
* Not using batched=True
* Using an np.array([],dtype=float) or similar instead of [] in the input
* Naming the output column differently from the input column
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.2
- Platform: Ubuntu
- Python version: 3.8
- PyArrow version: 8.0.0
| 4,483 |
https://github.com/huggingface/datasets/issues/4480 | Bigbench tensorflow GPU dependency | [
"Thanks for reporting ! :) cc @andersjohanandreassen can you take a look at this ?\r\n\r\nAlso @cceyda feel free to open an issue at [BIG-Bench](https://github.com/google/BIG-bench) as well regarding the `AttributeError`",
"I'm on vacation for the next week, so won't be able to do much debugging at the moment. Sorry for the inconvenience.\r\nBut I did quickly take a look:\r\n\r\n**pypi**:\r\nI managed to reproduce the above error with the pypi version begin out of date. \r\nThe version on `https://storage.googleapis.com/public_research_data/bigbench/bigbench-0.0.1.tar.gz` should be up to date, but it was my understanding that there was some issue with the pypi upload, so I don't even understand why there is a version [on pypi from April 1](https://pypi.org/project/bigbench/0.0.1/). Perhaps @ethansdyer, who's handling the pypi upload, knows the answer to that?\r\n\r\n**OOM error**:\r\nBut, I'm unable to reproduce the OOM error in a google colab with GPU enabled.\r\nThis is what I ran:\r\n```\r\n!pip install bigbench@https://storage.googleapis.com/public_research_data/bigbench/bigbench-0.0.1.tar.gz\r\n!pip install datasets\r\n\r\nfrom datasets import load_dataset\r\ndataset = load_dataset(\"bigbench\",\"swedish_to_german_proverbs\")\r\n``` \r\nThe `swedish_to_german_proverbs`task is only 72 examples, so I don't understand what could be causing the OOM error. Loading the task has no effect on the RAM for me. @cceyda Can you confirm that this does not occur in a [colab](https://colab.research.google.com/)?\r\nIf the GPU is somehow causing issues on your system, disabling the GPU from TF might be an option too\r\n```\r\nimport os\r\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"-1\"\r\n```\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n",
"Solved.\r\nYes it works on colab, and somehow magically on my machine too now. hmm not sure what was wrong before I had used a fresh venv both times with just the dataloading code, and tried multiple times. (maybe just a wrong tensorflow version got mixed up somehow) The tensorflow call seems to come from the bigbench side anyway.\r\n\r\nabout bigbench pypi version update, I opened an issue over there https://github.com/google/BIG-bench/issues/846\r\n\r\nanyway closing this now. If anyone else has the same problem can re-open."
] | ## Describe the bug
Loading bigbech
```py
from datasets import load_dataset
dataset = load_dataset("bigbench","swedish_to_german_proverbs")
```
tries to use gpu and fails with OOM with the following error
```
Downloading and preparing dataset bigbench/swedish_to_german_proverbs (download: Unknown size, generated: 68.92 KiB, post-processed: Unknown size, total: 68.92 KiB) to /home/ceyda/.cache/huggingface/datasets/bigbench/swedish_to_german_proverbs/1.0.0/7d2f6e537fa937dfaac8b1c1df782f2055071d3fd8e4f4ae93d28012a354ced0...
Generating default split: 0%| | 0/72 [00:00<?, ? examples/s]2022-06-13 14:11:04.154469: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-06-13 14:11:05.133600: F tensorflow/core/platform/statusor.cc:33] Attempting to fetch value instead of handling error INTERNAL: failed initializing StreamExecutor for CUDA device ordinal 3: INTERNAL: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_OUT_OF_MEMORY: out of memory; total memory reported: 25396838400
Aborted (core dumped)
```
I think this is because bigbench dependency (below) installs tensorflow (GPU version) and dataloading tries to use GPU as default.
`pip install bigbench@https://storage.googleapis.com/public_research_data/bigbench/bigbench-0.0.1.tar.gz`
while just doing 'pip install bigbench' results in following error
```
File "/home/ceyda/.local/lib/python3.7/site-packages/datasets/load.py", line 109, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 967, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 677, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 728, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/ceyda/.cache/huggingface/modules/datasets_modules/datasets/bigbench/7d2f6e537fa937dfaac8b1c1df782f2055071d3fd8e4f4ae93d28012a354ced0/bigbench.py", line 118, in <module>
class Bigbench(datasets.GeneratorBasedBuilder):
File "/home/ceyda/.cache/huggingface/modules/datasets_modules/datasets/bigbench/7d2f6e537fa937dfaac8b1c1df782f2055071d3fd8e4f4ae93d28012a354ced0/bigbench.py", line 127, in Bigbench
BigBenchConfig(name=name, version=datasets.Version("1.0.0")) for name in bb_utils.get_all_json_task_names()
AttributeError: module 'bigbench.api.util' has no attribute 'get_all_json_task_names'
```
## Steps to avoid the bug
Not ideal but can solve with (since I don't really use tensorflow elsewhere)
`pip uninstall tensorflow`
`pip install tensorflow-cpu`
## Environment info
- datasets @ master
- Python version: 3.7
| 4,480 |
https://github.com/huggingface/datasets/issues/4478 | Dataset slow during model training | [
"Hi ! cc @Rocketknight1 maybe you know better ?\r\n\r\nI'm not too familiar with `tf.data.experimental.save`. Note that `datasets` uses memory mapping, so depending on your hardware and the disk you are using you can expect performance differences with a dataset loaded in RAM",
"Hi @lehrig, I suspect what's happening here is that our `to_tf_dataset()` method has some performance issues when streaming samples. This is usually not a problem, but they become apparent when streaming a vision dataset into a very small vision model, which will need a lot of sample throughput to saturate the GPU.\r\n\r\nWhen you save a `tf.data.Dataset` with `tf.data.experimental.save`, all of the samples from the dataset (which are, in this case, batches of images), are saved to disk. When you load this saved dataset, you're effectively bypassing `to_tf_dataset()` entirely, which alleviates this performance bottleneck.\r\n\r\n`to_tf_dataset()` is something we're actively working on overhauling right now - particularly for image datasets, we want to make it possible to access the underlying images with `tf.data` without going through the current layer of indirection with `Arrow`, which should massively improve simplicity and performance. \r\n\r\nHowever, if you just want this to work quickly but without needing your save/load hack, my advice would be to simply load the dataset into memory if it's small enough to fit. Since all your samples have the same dimensions, you can do this simply with:\r\n\r\n```\r\ndataset = load_from_disk(prep_data_dir)\r\ndataset = dataset.with_format(\"numpy\")\r\ndata_in_memory = dataset[:]\r\n```\r\n\r\nThen you can simply do something like:\r\n\r\n```\r\nmodel.fit(data_in_memory[\"pixel_values\"], data_in_memory[\"labels\"])\r\n```",
"Thanks for the information! \r\n\r\nI have now updated the training code like so:\r\n\r\n```\r\ndataset = load_from_disk(prep_data_dir)\r\ntrain_dataset = dataset[\"train\"][:]\r\nvalidation_dataset = dataset[\"dev\"][:]\r\n\r\n...\r\n\r\nmodel.fit(\r\n train_dataset[\"pixel_values\"],\r\n train_dataset[\"label\"],\r\n epochs=epochs,\r\n validation_data=(\r\n validation_dataset[\"pixel_values\"],\r\n validation_dataset[\"label\"]\r\n ),\r\n callbacks=[earlyStopping, mcp_save, reduce_lr_loss]\r\n)\r\n```\r\n\r\n- Creating the in-memory dataset is quite quick\r\n- But: There is now a long wait (~4-5 Minutes) before the training starts (why?)\r\n- And: Training times have improved but the very first epoch leaves me wondering why it takes so long (why?)\r\n\r\n**Epoch Breakdown:**\r\n- Epoch 1/10\r\n78s 12s/step - loss: 3.1307 - accuracy: 0.0737 - val_loss: 2.2827 - val_accuracy: 0.1273 - lr: 0.0010\r\n- Epoch 2/10\r\n1s 168ms/step - loss: 2.3616 - accuracy: 0.2350 - val_loss: 2.2679 - val_accuracy: 0.2182 - lr: 0.0010\r\n- Epoch 3/10\r\n1s 189ms/step - loss: 2.0221 - accuracy: 0.3180 - val_loss: 2.2670 - val_accuracy: 0.1818 - lr: 0.0010\r\n- Epoch 4/10\r\n0s 67ms/step - loss: 1.8895 - accuracy: 0.3548 - val_loss: 2.2771 - val_accuracy: 0.1273 - lr: 0.0010\r\n- Epoch 5/10\r\n0s 67ms/step - loss: 1.7846 - accuracy: 0.3963 - val_loss: 2.2860 - val_accuracy: 0.1455 - lr: 0.0010\r\n- Epoch 6/10\r\n0s 65ms/step - loss: 1.5946 - accuracy: 0.4516 - val_loss: 2.2938 - val_accuracy: 0.1636 - lr: 0.0010\r\n- Epoch 7/10\r\n0s 63ms/step - loss: 1.4217 - accuracy: 0.5115 - val_loss: 2.2968 - val_accuracy: 0.2182 - lr: 0.0010\r\n- Epoch 8/10\r\n0s 67ms/step - loss: 1.3089 - accuracy: 0.5438 - val_loss: 2.2842 - val_accuracy: 0.2182 - lr: 0.0010\r\n- Epoch 9/10\r\n1s 184ms/step - loss: 1.2480 - accuracy: 0.5806 - val_loss: 2.2652 - val_accuracy: 0.1818 - lr: 0.0010\r\n- Epoch 10/10\r\n0s 65ms/step - loss: 1.2699 - accuracy: 0.5622 - val_loss: 2.2670 - val_accuracy: 0.2000 - lr: 0.0010\r\n\r\n",
"Regarding the new long ~5 min. wait introduced by the in-memory dataset update: this might be causing it? https://datascience.stackexchange.com/questions/33364/why-model-fit-generator-in-keras-is-taking-so-much-time-even-before-picking-the\r\n\r\nFor now, my save/load hack is still more performant, even though having more boiler-plate code :/ ",
"That 5 minute wait is quite surprising! I don't have a good explanation for why it's happening, but it can't be an issue with `datasets` or `tf.data` because you're just fitting directly on Numpy arrays at this point. All I can suggest is seeing if you can isolate the issue - for example, does fitting on a smaller dataset containing only 10% of the original data reduce the wait? This might indicate the delay is caused by your data being copied or converted somehow. Alternatively, you could try removing things like callbacks and seeing if you could isolate the issue there."
] | ## Describe the bug
While migrating towards π€ Datasets, I encountered an odd performance degradation: training suddenly slows down dramatically. I train with an image dataset using Keras and execute a `to_tf_dataset` just before training.
First, I have optimized my dataset following https://discuss.huggingface.co/t/solved-image-dataset-seems-slow-for-larger-image-size/10960/6, which actually improved the situation from what I had before but did not completely solve it.
Second, I saved and loaded my dataset using `tf.data.experimental.save` and `tf.data.experimental.load` before training (for which I would have expected no performance change). However, I ended up with the performance I had before tinkering with π€ Datasets.
Any idea what's the reason for this and how to speed-up training with π€ Datasets?
## Steps to reproduce the bug
```python
# Sample code to reproduce the bug
from datasets import load_dataset
import os
dataset_dir = "./dataset"
prep_dataset_dir = "./prepdataset"
model_dir = "./model"
# Load Data
dataset = load_dataset("Lehrig/Monkey-Species-Collection", "downsized")
def read_image_file(example):
with open(example["image"].filename, "rb") as f:
example["image"] = {"bytes": f.read()}
return example
dataset = dataset.map(read_image_file)
dataset.save_to_disk(dataset_dir)
# Preprocess
from datasets import (
Array3D,
DatasetDict,
Features,
load_from_disk,
Sequence,
Value
)
import numpy as np
from transformers import ImageFeatureExtractionMixin
dataset = load_from_disk(dataset_dir)
num_classes = dataset["train"].features["label"].num_classes
one_hot_matrix = np.eye(num_classes)
feature_extractor = ImageFeatureExtractionMixin()
def to_pixels(image):
image = feature_extractor.resize(image, size=size)
image = feature_extractor.to_numpy_array(image, channel_first=False)
image = image / 255.0
return image
def process(examples):
examples["pixel_values"] = [
to_pixels(image) for image in examples["image"]
]
examples["label"] = [
one_hot_matrix[label] for label in examples["label"]
]
return examples
features = Features({
"pixel_values": Array3D(dtype="float32", shape=(size, size, 3)),
"label": Sequence(feature=Value(dtype="int32"), length=num_classes)
})
prep_dataset = dataset.map(
process,
remove_columns=["image"],
batched=True,
batch_size=batch_size,
num_proc=2,
features=features,
)
prep_dataset = prep_dataset.with_format("numpy")
# Split
train_dev_dataset = prep_dataset['test'].train_test_split(
test_size=test_size,
shuffle=True,
seed=seed
)
train_dev_test_dataset = DatasetDict({
'train': train_dev_dataset['train'],
'dev': train_dev_dataset['test'],
'test': prep_dataset['test'],
})
train_dev_test_dataset.save_to_disk(prep_dataset_dir)
# Train Model
import datetime
import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.applications import InceptionV3
from tensorflow.keras.layers import Dense, Dropout, GlobalAveragePooling2D, BatchNormalization
from tensorflow.keras.callbacks import ReduceLROnPlateau, ModelCheckpoint, EarlyStopping
from transformers import DefaultDataCollator
dataset = load_from_disk(prep_data_dir)
data_collator = DefaultDataCollator(return_tensors="tf")
train_dataset = dataset["train"].to_tf_dataset(
columns=['pixel_values'],
label_cols=['label'],
shuffle=True,
batch_size=batch_size,
collate_fn=data_collator
)
validation_dataset = dataset["dev"].to_tf_dataset(
columns=['pixel_values'],
label_cols=['label'],
shuffle=False,
batch_size=batch_size,
collate_fn=data_collator
)
print(f'{datetime.datetime.now()} - Saving Data')
tf.data.experimental.save(train_dataset, model_dir+"/train")
tf.data.experimental.save(validation_dataset, model_dir+"/val")
print(f'{datetime.datetime.now()} - Loading Data')
train_dataset = tf.data.experimental.load(model_dir+"/train")
validation_dataset = tf.data.experimental.load(model_dir+"/val")
shape = np.shape(dataset["train"][0]["pixel_values"])
backbone = InceptionV3(
include_top=False,
weights='imagenet',
input_shape=shape
)
for layer in backbone.layers:
layer.trainable = False
model = Sequential()
model.add(backbone)
model.add(GlobalAveragePooling2D())
model.add(Dense(128, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.3))
model.add(Dense(64, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.3))
model.add(Dense(10, activation='softmax'))
model.compile(
optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy']
)
print(model.summary())
earlyStopping = EarlyStopping(
monitor='val_loss',
patience=10,
verbose=0,
mode='min'
)
mcp_save = ModelCheckpoint(
f'{model_dir}/best_model.hdf5',
save_best_only=True,
monitor='val_loss',
mode='min'
)
reduce_lr_loss = ReduceLROnPlateau(
monitor='val_loss',
factor=0.1,
patience=7,
verbose=1,
min_delta=0.0001,
mode='min'
)
hist = model.fit(
train_dataset,
epochs=epochs,
validation_data=validation_dataset,
callbacks=[earlyStopping, mcp_save, reduce_lr_loss]
)
```
## Expected results
Same performance when training without my "save/load hack" or a good explanation/recommendation about the issue.
## Actual results
Performance slower without my "save/load hack".
**Epoch Breakdown (without my "save/load hack"):**
- Epoch 1/10
41s 2s/step - loss: 1.6302 - accuracy: 0.5048 - val_loss: 1.4713 - val_accuracy: 0.3273 - lr: 0.0010
- Epoch 2/10
32s 2s/step - loss: 0.5357 - accuracy: 0.8510 - val_loss: 1.0447 - val_accuracy: 0.5818 - lr: 0.0010
- Epoch 3/10
36s 3s/step - loss: 0.3547 - accuracy: 0.9231 - val_loss: 0.6245 - val_accuracy: 0.7091 - lr: 0.0010
- Epoch 4/10
36s 3s/step - loss: 0.2721 - accuracy: 0.9231 - val_loss: 0.3395 - val_accuracy: 0.9091 - lr: 0.0010
- Epoch 5/10
32s 2s/step - loss: 0.1676 - accuracy: 0.9856 - val_loss: 0.2187 - val_accuracy: 0.9636 - lr: 0.0010
- Epoch 6/10
42s 3s/step - loss: 0.2066 - accuracy: 0.9615 - val_loss: 0.1635 - val_accuracy: 0.9636 - lr: 0.0010
- Epoch 7/10
32s 2s/step - loss: 0.1814 - accuracy: 0.9423 - val_loss: 0.1418 - val_accuracy: 0.9636 - lr: 0.0010
- Epoch 8/10
32s 2s/step - loss: 0.1301 - accuracy: 0.9856 - val_loss: 0.1388 - val_accuracy: 0.9818 - lr: 0.0010
- Epoch 9/10
loss: 0.1102 - accuracy: 0.9856 - val_loss: 0.1185 - val_accuracy: 0.9818 - lr: 0.0010
- Epoch 10/10
32s 2s/step - loss: 0.1013 - accuracy: 0.9808 - val_loss: 0.0978 - val_accuracy: 0.9818 - lr: 0.0010
**Epoch Breakdown (with my "save/load hack"):**
- Epoch 1/10
13s 625ms/step - loss: 3.0478 - accuracy: 0.1146 - val_loss: 2.3061 - val_accuracy: 0.0727 - lr: 0.0010
- Epoch 2/10
0s 80ms/step - loss: 2.3105 - accuracy: 0.2656 - val_loss: 2.3085 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 3/10
0s 77ms/step - loss: 1.8608 - accuracy: 0.3542 - val_loss: 2.3130 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 4/10
1s 98ms/step - loss: 1.8677 - accuracy: 0.3750 - val_loss: 2.3157 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 5/10
1s 204ms/step - loss: 1.5561 - accuracy: 0.4583 - val_loss: 2.3049 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 6/10
1s 210ms/step - loss: 1.4657 - accuracy: 0.4896 - val_loss: 2.2944 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 7/10
1s 205ms/step - loss: 1.4018 - accuracy: 0.5312 - val_loss: 2.2917 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 8/10
1s 207ms/step - loss: 1.2370 - accuracy: 0.5729 - val_loss: 2.2814 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 9/10
1s 214ms/step - loss: 1.1190 - accuracy: 0.6250 - val_loss: 2.2733 - val_accuracy: 0.0909 - lr: 0.0010
- Epoch 10/10
1s 207ms/step - loss: 1.1484 - accuracy: 0.6302 - val_loss: 2.2624 - val_accuracy: 0.0909 - lr: 0.0010
## Environment info
- `datasets` version: 2.2.2
- Platform: Linux-4.18.0-305.45.1.el8_4.ppc64le-ppc64le-with-glibc2.17
- Python version: 3.8.13
- PyArrow version: 7.0.0
- Pandas version: 1.4.2
- TensorFlow: 2.8.0
- GPU (used during training): Tesla V100-SXM2-32GB
| 4,478 |
https://github.com/huggingface/datasets/issues/4477 | Dataset Viewer issue for fgrezes/WIESP2022-NER | [
"https://huggingface.co/datasets/fgrezes/WIESP2022-NER\r\n\r\nThe error:\r\n\r\n```\r\nMessage: Couldn't find a dataset script at /src/services/worker/fgrezes/WIESP2022-NER/WIESP2022-NER.py or any data file in the same directory. Couldn't find 'fgrezes/WIESP2022-NER' on the Hugging Face Hub either: FileNotFoundError: Unable to resolve any data file that matches ['**test*', '**eval*'] in dataset repository fgrezes/WIESP2022-NER with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'blp', 'bmp', 'dib', 'bufr', 'cur', 'pcx', 'dcx', 'dds', 'ps', 'eps', 'fit', 'fits', 'fli', 'flc', 'ftc', 'ftu', 'gbr', 'gif', 'grib', 'h5', 'hdf', 'png', 'apng', 'jp2', 'j2k', 'jpc', 'jpf', 'jpx', 'j2c', 'icns', 'ico', 'im', 'iim', 'tif', 'tiff', 'jfif', 'jpe', 'jpg', 'jpeg', 'mpg', 'mpeg', 'msp', 'pcd', 'pxr', 'pbm', 'pgm', 'ppm', 'pnm', 'psd', 'bw', 'rgb', 'rgba', 'sgi', 'ras', 'tga', 'icb', 'vda', 'vst', 'webp', 'wmf', 'emf', 'xbm', 'xpm', 'zip']\r\n```\r\n\r\nI understand the issue is not related to the dataset viewer in itself, but with the autodetection of the data files without a loading script in the datasets library. cc @lhoestq @albertvillanova @mariosasko ",
"Apparently it finds `scoring-scripts/compute_seqeval.py` which matches `**eval*`, a regex that detects a test split. We should probably improve the regex because it's not supposed to catch this kind of files. It must also only check for files with supported extensions: txt, csv, png etc."
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,477 |
https://github.com/huggingface/datasets/issues/4476 | `to_pandas` doesn't take into account format. | [
"Thanks for opening a discussion :)\r\n\r\nNote that you can use `.remove_columns(...)` to keep only the ones you're interested in before calling `.to_pandas()`",
"Yes I can do that thank you!\r\n\r\nDo you think that conceptually my example should work? If not, I'm happy to close this issue. \r\n\r\nIf yes, I can start working on it.",
"Hi! Instead of `with_format(columns=['a', 'b']).to_pandas()`, use `with_format(\"pandas\", columns=[\"a\", \"b\"])` for easy conversion of the parts of the dataset to pandas via indexing/slicing.\r\n\r\nThe full code:\r\n```python\r\nfrom datasets import Dataset\r\n\r\nds = Dataset.from_dict({'a': [1,2,3], 'b': [5,6,7], 'c': [8,9,10]})\r\npandas_df = ds.with_format(\"pandas\", columns=['a', 'b'])[:]\r\n```",
"Ahhhh Thank you!\r\n\r\nclosing then :)"
] | **Is your feature request related to a problem? Please describe.**
I have a large dataset that I need to convert part of to pandas to do some further analysis. Calling `to_pandas` directly on it is expensive. So I thought I could simply select the columns that I want and then call `to_pandas`.
**Describe the solution you'd like**
```python
from datasets import Dataset
ds = Dataset.from_dict({'a': [1,2,3], 'b': [5,6,7], 'c': [8,9,10]})
pandas_df = ds.with_format(columns=['a', 'b']).to_pandas()
# I would expect `pandas_df` to only include a,b as column.
```
**Describe alternatives you've considered**
I could remove all columns that I don't want? But I don't know all of them in advance.
**Additional context**
I can probably make a PR with some pointers.
| 4,476 |
https://github.com/huggingface/datasets/issues/4471 | CI error with repo lhoestq/_dummy | [
"fixed by https://github.com/huggingface/datasets/pull/4472"
] | ## Describe the bug
CI is failing because of repo "lhoestq/_dummy". See: https://app.circleci.com/pipelines/github/huggingface/datasets/12461/workflows/1b040b45-9578-4ab9-8c44-c643c4eb8691/jobs/74269
```
requests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/api/datasets/lhoestq/_dummy?full=true
```
The repo seems to no longer exist: https://huggingface.co/api/datasets/lhoestq/_dummy
```
error: "Repository not found"
```
CC: @lhoestq | 4,471 |
https://github.com/huggingface/datasets/issues/4467 | Transcript string 'null' converted to [None] by load_dataset() | [
"Hi @mbarnig, thanks for reporting.\r\n\r\nPlease note that is an expected behavior by `pandas` (we use the `pandas` library to parse CSV files): https://pandas.pydata.org/docs/reference/api/pandas.read_csv.html\r\n```\r\nBy default the following values are interpreted as NaN: \r\nββ, β#N/Aβ, β#N/A N/Aβ, β#NAβ, β-1.#INDβ, β-1.#QNANβ, β-NaNβ, β-nanβ, β1.#INDβ, β1.#QNANβ, β<NA>β, βN/Aβ, βNAβ, βNULLβ, βNaNβ, βn/aβ, βnanβ, βnullβ.\r\n```\r\n(see \"null\" in the last position in the above list).\r\n\r\nIn order to prevent `pandas` from performing that automatic conversion from the string \"null\" to a NaN value, you should pass the `pandas` parameter `keep_default_na=False`:\r\n```python\r\nIn [2]: dataset = load_dataset('csv', data_files={'train': 'null-test.csv'}, keep_default_na=False)\r\nIn [3]: dataset[\"train\"][0][\"transcript\"]\r\nOut[3]: 'null'\r\n```",
"Thanks for the quick answer.",
"@albertvillanova I also ran into this issue, it had me scratching my head for a while! In my case it was tripped by a literal \"NA\" comment collected from a user-facing form (e.g., this question does not apply to me). Thankfully this answer was here, but I feel it is such a common trap that it deserves to be noted in the official docs, maybe [here](https://huggingface.co/docs/datasets/loading#csv)? \r\n\r\nI'm happy to submit a PR if you agree!"
] | ## Issue
I am training a luxembourgish speech-recognition model in Colab with a custom dataset, including a dictionary of luxembourgish words, for example the speaken numbers 0 to 9. When preparing the dataset with the script
`ds_train1 = mydataset.map(prepare_dataset)`
the following error was issued:
```
ValueError Traceback (most recent call last)
<ipython-input-69-1e8f2b37f5bc> in <module>()
----> 1 ds_train = mydataset_train.map(prepare_dataset)
11 frames
/usr/local/lib/python3.7/dist-packages/transformers/tokenization_utils_base.py in __call__(self, text, text_pair, add_special_tokens, padding, truncation, max_length, stride, is_split_into_words, pad_to_multiple_of, return_tensors, return_token_type_ids, return_attention_mask, return_overflowing_tokens, return_special_tokens_mask, return_offsets_mapping, return_length, verbose, **kwargs)
2450 if not _is_valid_text_input(text):
2451 raise ValueError(
-> 2452 "text input must of type str (single example), List[str] (batch or single pretokenized example) "
2453 "or List[List[str]] (batch of pretokenized examples)."
2454 )
ValueError: text input must of type str (single example), List[str] (batch or single pretokenized example) or List[List[str]] (batch of pretokenized examples).
```
Debugging this problem was not easy, all transcriptions in the dataset are correct strings. Finally I discovered that the transcription string 'null' is interpreted as [None] by the `load_dataset()` script. By deleting this row in the dataset the training worked fine.
## Expected result:
transcription 'null' interpreted as 'str' instead of 'None'.
## Reproduction
Here is the code to reproduce the error with a one-row-dataset.
```
with open("null-test.csv") as f:
reader = csv.reader(f)
for row in reader:
print(row)
```
['wav_filename', 'wav_filesize', 'transcript']
['wavs/female/NULL1.wav', '17530', 'null']
```
dataset = load_dataset('csv', data_files={'train': 'null-test.csv'})
```
Using custom data configuration default-81ac0c0e27af3514
Downloading and preparing dataset csv/default to /root/.cache/huggingface/datasets/csv/default-81ac0c0e27af3514/0.0.0/433e0ccc46f9880962cc2b12065189766fbb2bee57a221866138fb9203c83519...
Downloading data files: 100%
1/1 [00:00<00:00, 29.55it/s]
Extracting data files: 100%
1/1 [00:00<00:00, 23.66it/s]
Dataset csv downloaded and prepared to /root/.cache/huggingface/datasets/csv/default-81ac0c0e27af3514/0.0.0/433e0ccc46f9880962cc2b12065189766fbb2bee57a221866138fb9203c83519. Subsequent calls will reuse this data.
100%
1/1 [00:00<00:00, 25.84it/s]
```
print(dataset['train']['transcript'])
```
[None]
## Environment info
```
!pip install datasets==2.2.2
!pip install transformers==4.19.2
``` | 4,467 |
https://github.com/huggingface/datasets/issues/4462 | BigBench: NonMatchingSplitsSizesError when passing a dataset configuration parameter | [
"Why not adding `max_examples` as part of the config name?",
"Yup it can also work, and maybe it's simpler this way. Opening a PR to fix bigbench instead of https://github.com/huggingface/datasets/pull/4463",
"Hi @lhoestq,\r\n\r\nThank you for taking a look at this issue, and proposing a solution. \r\nUnfortunately, after trying the fix in #4465 I still see the same issue.\r\n\r\nI think there is some subtlety where the config name gets overwritten somewhere when `BUILDER_CONFIGS`[(link)](https://github.com/huggingface/datasets/blob/master/datasets/bigbench/bigbench.py#L126) is defined. \r\n\r\nIf I print out the `self.config.name` in the current version (with the fix in #4465), I see just the task name, but if I comment out `BUILDER_CONFIGS`, the `num_shots` and `max_examples` gets appended as was meant by #4465.\r\n\r\nI haven't managed to track down where this happens, but I thought you might know? \r\n\r\n(Another comment on your fix: the `name` variable is used to fetch the task from the bigbench API, so modifying it causes an error if it's actually called. This can easily be fixed by having `config_name` variable in addition to the `task_name`)\r\n\r\n\r\n"
] | As noticed in https://github.com/huggingface/datasets/pull/4125 when a dataset config class has a parameter that reduces the number of examples (e.g. named `max_examples`), then loading the dataset and passing `max_examples` raises `NonMatchingSplitsSizesError`.
This is because it will check for expected the number of examples of the config with the same name without taking into account the `max_examples` parameter. This can be fixed by checking the expected number of examples using the **config id** instead of name. Indeed the config id corresponds to the config name + an optional suffix that depends on the config parameters | 4,462 |
https://github.com/huggingface/datasets/issues/4461 | AttributeError: module 'datasets' has no attribute 'load_dataset' | [
"I'm having the same issueοΌCan you tell me how to solve itοΌ",
"I have the same issue, can you tell me how to solve it? Thanks",
"I had a folder named 'datasets' so this is why it can't find the import, it's looking in the wrong place",
"@briandw your comment saved my day π "
] | ## Describe the bug
I have piped install datasets, but this package doesn't have these attributes: load_dataset, load_metric.
## Environment info
- `datasets` version: 1.9.0
- Platform: Linux-5.13.0-44-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.6.13
- PyArrow version: 6.0.1
| 4,461 |
https://github.com/huggingface/datasets/issues/4456 | Workflow for Tabular data | [
"I use below to load a dataset:\r\n```\r\ndataset = datasets.load_dataset(\"scikit-learn/auto-mpg\")\r\ndf = pd.DataFrame(dataset[\"train\"])\r\n```\r\nTBH as said, tabular folk split their own dataset, they sometimes have two splits, sometimes three. Maybe somehow avoiding it for tabular datasets might be good for later. (it's just UX improvement) ",
"is very slow batch access of a dataset (tabular, csv) with many columns to be expected?",
"Define \"many\" ? x)",
"~20k! I was surprised batch loading with as few as 32 samples was really slow. I was speculating the columnar format was the cause -- or do you see good performance with this approx size of tabular data?",
"20k can be a lot for a columnar format but maybe we can optimize a few things.\r\n\r\nIt would be cool to profile the code to see if there's an unoptimized part of the code that slows everything down.\r\n\r\n(it's also possible to kill the job when it accesses the batch, it often gives you the traceback at the location where the code was running)",
"FWIW I've worked with tabular data with 540k columns.",
"thats awesome, whats your secret? would love to see an example!",
"@wconnell I'm not sure what you mean by my secret, I load them into a numpy array π \r\n\r\nAn example dataset is [here](https://portal.gdc.cancer.gov/repository?facetTab=files&filters=%7B%22content%22%3A%5B%7B%22content%22%3A%7B%22field%22%3A%22cases.project.project_id%22%2C%22value%22%3A%5B%22TCGA-CESC%22%5D%7D%2C%22op%22%3A%22in%22%7D%2C%7B%22content%22%3A%7B%22field%22%3A%22files.data_category%22%2C%22value%22%3A%5B%22DNA%20Methylation%22%5D%7D%2C%22op%22%3A%22in%22%7D%5D%2C%22op%22%3A%22and%22%7D&searchTableTab=files) which is a dataset of DNA methylation reads. This dataset is about 950 rows and 450k columns. "
] | Tabular data are treated very differently than data for NLP, audio, vision, etc. and therefore the worflow for tabular data in `datasets` is not ideal.
For example for tabular data, it is common to use pandas/spark/dask to process the data, and then load the data into X and y (X is an array of features and y an array of labels), then train_test_split and finally feed the data to a machine learning model.
In `datasets` the workflow is different: we use load_dataset, then map, then train_test_split (if we only have a train split) and we end up with columnar dataset splits, not formatted as X and y.
Right now, it is already possible to convert a dataset from and to pandas, but there are still many things that could improve the workflow for tabular data:
- be able to load the data into X and y
- be able to load a dataset from the output of spark or dask (as far as I know it's usually csv or parquet files on S3/GCS/HDFS etc.)
- support "unsplit" datasets explicitly, instead of putting everything in "train" by default
cc @adrinjalali @merveenoyan feel free to complete/correct this :)
Feel free to also share ideas of APIs that would be super intuitive in your opinion ! | 4,456 |
https://github.com/huggingface/datasets/issues/4454 | Dataset Viewer issue for Yaxin/SemEval2015 | [
"Closing since it's a duplicate of https://github.com/huggingface/datasets/issues/4453"
] | ### Link
_No response_
### Description
the link could not visit
### Owner
_No response_ | 4,454 |
https://github.com/huggingface/datasets/issues/4453 | Dataset Viewer issue for Yaxin/SemEval2015 | [
"I understand that the issue is that a remote file (URL) is being loaded as a local file. Right @albertvillanova @lhoestq?\r\n\r\n```\r\nMessage: [Errno 2] No such file or directory: 'https://raw.githubusercontent.com/YaxinCui/ABSADataset/main/SemEval2015Task12Corrected/train/restaurants_train.xml'\r\n```",
"`xml.dom.minidom.parse` is not supported in streaming mode. I opened a PR here to fix it:\r\nhttps://huggingface.co/datasets/Yaxin/SemEval2015/discussions/1\r\n\r\nPlease review the PR @WithYouTo and let me know if it works !",
"Additionally, I'm also patching our library, so that we support streaming datasets that use `xml.dom.minidom.parse`."
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,453 |
https://github.com/huggingface/datasets/issues/4452 | Trying to load FEVER dataset results in NonMatchingChecksumError | [
"Thanks for reporting @santhnm2. We are fixing it.\r\n\r\nData owners updated their URLs recently. We have to align with them, otherwise you do not download anything (that is why ignore_verifications does not work).",
"Hello! Is there any update on this? I am having the same issue 6 months later."
] | ## Describe the bug
Trying to load the `fever` dataset fails with `datasets.utils.info_utils.NonMatchingChecksumError`.
I tried with `download_mode="force_redownload"` but that did not fix the error. I also tried with `ignore_verification=True` but then that raised a `json.decoder.JSONDecodeError`.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('fever', 'v1.0') # Fails with NonMatchingChecksumError
dataset = load_dataset('fever', 'v1.0', download_mode="force_redownload") # Fails with NonMatchingChecksumError
dataset = load_dataset('fever', 'v1.0', ignore_verification=True)` # Fails with JSONDecodeError
```
## Expected results
I expect this call to return with no error raised.
## Actual results
With `ignore_verification=False`:
```
*** datasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://s3-eu-west-1.amazonaws.com/fever.public/train.jsonl', 'https://s3-eu-west-1.amazonaws.com/fever.public/shared_task_dev.jsonl', 'https://s3-eu-west-1.amazonaws.com/fever.public/shared_task_dev_public.jsonl', 'https://s3-eu-west-1.amazonaws.com/fever.public/shared_task_test.jsonl', 'https://s3-eu-west-1.amazonaws.com/fever.public/paper_dev.jsonl', 'https://s3-eu-west-1.amazonaws.com/fever.public/paper_test.jsonl']
```
With `ignore_verification=True`:
```
*** json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.3.dev0
- Platform: Linux-4.15.0-50-generic-x86_64-with-glibc2.10
- Python version: 3.8.13
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,452 |
https://github.com/huggingface/datasets/issues/4449 | Rj | [] | import android.content.DialogInterface;
import android.database.Cursor;
import android.os.Bundle;
import android.view.View;
import android.widget.ArrayAdapter;
import android.widget.Button;
import android.widget.EditText;
import android.widget.Toast;
import androidx.appcompat.app.AlertDialog;
import androidx.appcompat.app.AppCompatActivity;
public class MainActivity extends AppCompatActivity {
private EditText editTextID;
private EditText editTextName;
private EditText editTextNum;
private String name;
private int number;
private String ID;
private dbHelper db;
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
setContentView(R.layout.activity_main);
db = new dbHelper(this);
editTextID = findViewById(R.id.editText1);
editTextName = findViewById(R.id.editText2);
editTextNum = findViewById(R.id.editText3);
Button buttonSave = findViewById(R.id.button);
Button buttonRead = findViewById(R.id.button2);
Button buttonUpdate = findViewById(R.id.button3);
Button buttonDelete = findViewById(R.id.button4);
Button buttonSearch = findViewById(R.id.button5);
Button buttonDeleteAll = findViewById(R.id.button6);
buttonSave.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
name = editTextName.getText().toString();
String num = editTextNum.getText().toString();
if (name.isEmpty() || num.isEmpty()) {
Toast.makeText(MainActivity.this, "Cannot Submit Empty Fields", Toast.LENGTH_SHORT).show();
} else {
number = Integer.parseInt(num);
try {
// Insert Data
db.insertData(name, number);
// Clear the fields
editTextID.getText().clear();
editTextName.getText().clear();
editTextNum.getText().clear();
} catch (Exception e) {
e.printStackTrace();
}
}
}
});
buttonRead.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
final ArrayAdapter<String> adapter = new ArrayAdapter<>(MainActivity.this, android.R.layout.simple_list_item_1);
String name;
String num;
String id;
try {
Cursor cursor = db.readData();
if (cursor != null && cursor.getCount() > 0) {
while (cursor.moveToNext()) {
id = cursor.getString(0); // get data in column index 0
name = cursor.getString(1); // get data in column index 1
num = cursor.getString(2); // get data in column index 2
// Add SQLite data to listView
adapter.add("ID :- " + id + "\n" +
"Name :- " + name + "\n" +
"Number :- " + num + "\n\n");
}
} else {
adapter.add("No Data");
}
cursor.close();
} catch (Exception e) {
e.printStackTrace();
}
// show the saved data in alertDialog
AlertDialog.Builder builder = new AlertDialog.Builder(MainActivity.this);
builder.setTitle("SQLite saved data");
builder.setIcon(R.mipmap.app_icon_foreground);
builder.setAdapter(adapter, new DialogInterface.OnClickListener() {
@Override
public void onClick(DialogInterface dialog, int which) {
}
});
builder.setPositiveButton("OK", new DialogInterface.OnClickListener() {
@Override
public void onClick(DialogInterface dialog, int which) {
dialog.cancel();
}
});
AlertDialog dialog = builder.create();
dialog.show();
}
});
buttonUpdate.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
name = editTextName.getText().toString();
String num = editTextNum.getText().toString();
ID = editTextID.getText().toString();
if (name.isEmpty() || num.isEmpty() || ID.isEmpty()) {
Toast.makeText(MainActivity.this, "Cannot Submit Empty Fields", Toast.LENGTH_SHORT).show();
} else {
number = Integer.parseInt(num);
try {
// Update Data
db.updateData(ID, name, number);
// Clear the fields
editTextID.getText().clear();
editTextName.getText().clear();
editTextNum.getText().clear();
} catch (Exception e) {
e.printStackTrace();
}
}
}
});
buttonDelete.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
ID = editTextID.getText().toString();
if (ID.isEmpty()) {
Toast.makeText(MainActivity.this, "Please enter the ID", Toast.LENGTH_SHORT).show();
} else {
try {
// Delete Data
db.deleteData(ID);
// Clear the fields
editTextID.getText().clear();
editTextName.getText().clear();
editTextNum.getText().clear();
} catch (Exception e) {
e.printStackTrace();
}
}
}
});
buttonDeleteAll.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
// Delete all data
// You can simply delete all the data by calling this method --> db.deleteAllData();
// You can try this also
AlertDialog.Builder builder = new AlertDialog.Builder(MainActivity.this);
builder.setIcon(R.mipmap.app_icon_foreground);
builder.setTitle("Delete All Data");
builder.setCancelable(false);
builder.setMessage("Do you really need to delete your all data ?");
builder.setPositiveButton("Yes", new DialogInterface.OnClickListener() {
@Override
public void onClick(DialogInterface dialog, int which) {
// User confirmed , now you can delete the data
db.deleteAllData();
// Clear the fields
editTextID.getText().clear();
editTextName.getText().clear();
editTextNum.getText().clear();
}
});
builder.setNegativeButton("No", new DialogInterface.OnClickListener() {
@Override
public void onClick(DialogInterface dialog, int which) {
// user not confirmed
dialog.cancel();
}
});
AlertDialog dialog = builder.create();
dialog.show();
}
});
buttonSearch.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
ID = editTextID.getText().toString();
if (ID.isEmpty()) {
Toast.makeText(MainActivity.this, "Please enter the ID", Toast.LENGTH_SHORT).show();
} else {
try {
// Search data
Cursor cursor = db.searchData(ID);
if (cursor.moveToFirst()) {
editTextName.setText(cursor.getString(1));
editTextNum.setText(cursor.getString(2));
Toast.makeText(MainActivity.this, "Data successfully searched", Toast.LENGTH_SHORT).show();
} else {
Toast.makeText(MainActivity.this, "ID not found", Toast.LENGTH_SHORT).show();
editTextNum.setText("ID Not found");
editTextName.setText("ID not found");
}
cursor.close();
} catch (Exception e) {
e.printStackTrace();
}
}
}
});
}
} | 4,449 |
https://github.com/huggingface/datasets/issues/4448 | New Preprocessing Feature - Deduplication [Request] | [
"Hi! The [datasets_sql](https://github.com/mariosasko/datasets_sql) package lets you easily find distinct rows in a dataset (an example with `SELECT DISTINCT` is in the readme). Deduplication is (still) not part of the official API because it's hard to implement for datasets bigger than RAM while only using the native PyArrow ops.\r\n\r\n(Btw, this is a duplicate of https://github.com/huggingface/datasets/issues/2514)",
"Here is an example using the [datasets_sql](https://github.com/mariosasko/datasets_sql) mentioned \r\n\r\n```python \r\nfrom datasets_sql import query\r\n\r\ndataset = load_dataset(\"imdb\", split=\"train\")\r\n\r\n# If you dont have an id column just add one by enumerating\r\ndataset=dataset.add_column(\"id\", range(len(dataset)))\r\n\r\nid_column='id'\r\nunique_column='text'\r\n\r\n# always selects min id\r\nunique_dataset = query(f\"SELECT dataset.* FROM dataset JOIN (SELECT MIN({id_column}) as unique_id FROM dataset group by {unique_column}) ON unique_id=dataset.{id_column}\")\r\n```\r\nNot ideal for large datasets but good enough for basic cases.\r\nSure would be nice to have in the library π€ "
] | **Is your feature request related to a problem? Please describe.**
Many large datasets are full of duplications and it has been shown that deduplicating datasets can lead to better performance while training, and more truthful evaluation at test-time.
A feature that allows one to easily deduplicate a dataset can be cool!
**Describe the solution you'd like**
We can define a function and keep only the first/last data-point that yields the value according to this function.
**Describe alternatives you've considered**
The clear alternative is to repeat a clear boilerplate every time someone want to deduplicate a dataset.
| 4,448 |
https://github.com/huggingface/datasets/issues/4443 | Dataset Viewer issue for openclimatefix/nimrod-uk-1km | [
"If I understand correctly, this is due to the key `split` missing in the line https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/blob/main/nimrod-uk-1km.py#L41 of the script.\r\nMaybe @albertvillanova could confirm.",
"I'm having a look.",
"Indeed there are several issues in this dataset loading script.\r\n\r\nThe one pointed out by @severo: for the default configuration \"crops\": https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/blob/main/nimrod-uk-1km.py#L244\r\n- The download manager downloads `_URL`\r\n- But `_URL` is not defined: https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/blob/main/nimrod-uk-1km.py#L41\r\n ```python\r\n _URL = {'train': []}\r\n ```\r\n- Afterwards, for each split, a different key in `_ULR` is used, but it only contains one key: \"train\"\r\n - \"valid\" key: https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/blob/main/nimrod-uk-1km.py#L260\r\n - \"test key: https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/blob/main/nimrod-uk-1km.py#L269\r\n \r\nThese keys do not exist inside `_URL`, thus the error message reported in the viewer: \r\n```\r\nException: KeyError\r\nMessage: 'valid'\r\n```",
"Would anyone want to submit a Hub PR (or open a Discussion for the authors to be aware) to this dataset? https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km",
"Hi, I'm the main author for that dataset, so I'll work on updating it! I was working on debugging some stuff awhile ago, which is what broke it. ",
"I've opened a Discussion page, so that we can ask/answer and propose fixes until the script works properly: https://huggingface.co/datasets/openclimatefix/nimrod-uk-1km/discussions/1\r\n\r\nCC: @julien-c @jacobbieker ",
"can we close this issue and followup in the discussion?"
] | ### Link
_No response_
### Description
_No response_
### Owner
_No response_ | 4,443 |
https://github.com/huggingface/datasets/issues/4442 | Dataset Viewer issue for amazon_polarity | [
"Thanks, looking at it",
"Not sure what happened π¬, but it's fixed"
] | ### Link
https://huggingface.co/datasets/amazon_polarity/viewer/amazon_polarity/test
### Description
For some reason the train split is OK but the test split is not for this dataset:
```
Server error
Status code: 400
Exception: FileNotFoundError
Message: [Errno 2] No such file or directory: '/cache/modules/datasets_modules/datasets/amazon_polarity/__init__.py'
```
### Owner
No | 4,442 |
https://github.com/huggingface/datasets/issues/4441 | Dataset Viewer issue for aeslc | [
"Not sure what happened π¬, but it's fixed"
] | ### Link
https://huggingface.co/datasets/aeslc
### Description
The dataset viewer can't find `dataset_infos.json` in it's cache:
```
Server error
Status code: 400
Exception: FileNotFoundError
Message: [Errno 2] No such file or directory: '/cache/modules/datasets_modules/datasets/aeslc/eb8e30234cf984a58ebe9f205674597ac1db2ec91e7321cd7f36864f7e3671b8/dataset_infos.json'
```
### Owner
No | 4,441 |
https://github.com/huggingface/datasets/issues/4439 | TIMIT won't load after manual download: Errors about files that don't exist | [
"To have some context, please see:\r\n- #4145\r\n\r\nPlease, also note that we have recently made some fixes to the script, which are in our GitHub master branch but not yet released:\r\n- #4422\r\n- #4425 \r\n- #4436",
"Thanks Albert! I'll try pulling `datasets` from the git repo instead of PyPI, and/or just wait for the next release.\r\n",
"I'm closing this issue then. Please, feel free to reopen it again if the problem persists."
] | ## Describe the bug
I get the message from HuggingFace that it must be downloaded manually. From the URL provided in the message, I got to UPenn page for manual download. (UPenn apparently want $250? for the dataset??) ...So, ok, I obtained a copy from a friend and also a smaller version from Kaggle. But in both cases the HF dataloader fails; it is looking for files that don't exist anywhere in the dataset: it is looking for files with lower-case letters like "**test*" (all the filenames in both my copies are uppercase) and certain file extensions that exclude the .DOC which is provided in TIMIT:
## Steps to reproduce the bug
```python
data = load_dataset('timit_asr', 'clean')['train']
```
## Expected results
The dataset should load with no errors.
## Actual results
This error message:
```
File "/home/ubuntu/envs/data2vec/lib/python3.9/site-packages/datasets/data_files.py", line 201, in resolve_patterns_locally_or_by_urls
raise FileNotFoundError(error_msg)
FileNotFoundError: Unable to resolve any data file that matches '['**test*', '**eval*']' at /home/ubuntu/datasets/timit with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'blp', 'bmp', 'dib', 'bufr', 'cur', 'pcx', 'dcx', 'dds', 'ps', 'eps', 'fit', 'fits', 'fli', 'flc', 'ftc', 'ftu', 'gbr', 'gif', 'grib', 'h5', 'hdf', 'png', 'apng', 'jp2', 'j2k', 'jpc', 'jpf', 'jpx', 'j2c', 'icns', 'ico', 'im', 'iim', 'tif', 'tiff', 'jfif', 'jpe', 'jpg', 'jpeg', 'mpg', 'mpeg', 'msp', 'pcd', 'pxr', 'pbm', 'pgm', 'ppm', 'pnm', 'psd', 'bw', 'rgb', 'rgba', 'sgi', 'ras', 'tga', 'icb', 'vda', 'vst', 'webp', 'wmf', 'emf', 'xbm', 'xpm', 'zip']
```
But this is a strange sort of error: why is it looking for lower-case file names when all the TIMIT dataset filenames are uppercase? Why does it exclude .DOC files when the only parts of the TIMIT data set with "TEST" in them have ".DOC" extensions? ...I wonder, how was anyone able to get this to work in the first place?
The files in the dataset look like the following:
```
Β³ PHONCODE.DOC
Β³ PROMPTS.TXT
Β³ SPKRINFO.TXT
Β³ SPKRSENT.TXT
Β³ TESTSET.DOC
```
...so why are these being excluded by the dataset loader?
## Environment info
- `datasets` version: 2.2.2
- Platform: Linux-5.4.0-1060-aws-x86_64-with-glibc2.27
- Python version: 3.9.9
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,439 |
https://github.com/huggingface/datasets/issues/4435 | Load a local cached dataset that has been modified | [
"Hi! `datasets` caches every modification/loading, so you can either rerun the pipeline up to the `map` call or use `Dataset.from_file(modified_dataset)` to load the dataset directly from the cache file.",
"Awesome, hvala Mario! This works. "
] | ## Describe the bug
I have loaded a dataset as follows:
```
d = load_dataset("emotion", split="validation")
```
Afterwards I make some modifications to the dataset via a `map` call:
```
d.map(some_update_func, cache_file_name=modified_dataset)
```
This generates a cached version of the dataset on my local system in the same directory as the original download of the data (/path/to/cache). Running an `ls` returns:
```
modified_dataset
dataset_info.json
emotion-test.arrow
emotion-train.arrow
emotion-validation.arrow
```
as expected. However, when I try to load up the modified cached dataset via a call to
```
modified = load_dataset("emotion", split="validation", data_files="/path/to/cache/modified_dataset")
```
it simply redownloads a new version of the dataset and dumps to a new cache rather than loading up the original modified dataset:
```
Using custom data configuration validation-cdbf51685638421b
Downloading and preparing dataset emotion/validation to ...
```
How am I supposed to load the original modified local cache copy of the dataset?
## Environment info
- `datasets` version: 2.2.2
- Platform: Linux-5.4.0-113-generic-x86_64-with-glibc2.17
- Python version: 3.8.13
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,435 |
https://github.com/huggingface/datasets/issues/4430 | Add ability to load newer, cleaner version of Multi-News | [
"Hi! Our versioning is based on Git revisions (the `revision` param in `load_dataset`), so you can just replace the old URL with the new one and open a PR :). I can also give you some pointers if needed.",
"@mariosasko Awesome thanks! I will do that. Looks like this new version of the data is not available as a zip but as three files (train/dev/test). How is this usually handled in HF Datasets, should `_URL` be a dict with keys `train`, `val`, `test` perhaps?",
"Yes! Let me help you with more detailed instructions.\r\n\r\nIn the first step, we need to update the URLs. One of the possible dictionary structures is as follows:\r\n```python\r\n_URLs = {\r\n \"train\": {\"src\": \"https://drive.google.com/uc?export=download&id=1wHAWDOwOoQWSj7HYpyJ3Aeud8WhhaJ7P\", \"tgt\": \"https://drive.google.com/uc?export=download&id=1QVgswwhVTkd3VLCzajK6eVkcrSWEK6kq\"}\r\n \"val\": ...\r\n \"test\": ...\r\n}\r\n```\r\n\r\n(You can use this page to generate direct download links: https://sites.google.com/site/gdocs2direct/)\r\n\r\nThen we move to the `split_generators` method:\r\n```python\r\ndef _split_generators(self, dl_manager):\r\n \"\"\"Returns SplitGenerators.\"\"\"\r\n files = dl_manager.download(_URLs)\r\n return [\r\n datasets.SplitGenerator(\r\n name=datasets.Split.TRAIN,\r\n gen_kwargs={\"src_file\": files[\"train\"][\"src\"], \"tgt_file\": files[\"train\"][\"tgt\"]},\r\n ),\r\n ... # same for val and test\r\n ]\r\n```\r\nFinally, we adjust the signature of `_generate_examples`:\r\n```python\r\ndef _generate_examples(self, src_file, tgt_file):\r\n \"\"\"Yields examples.\"\"\"\r\n with open(src_file, encoding=\"utf-8\") as src_f, open(\r\n tgt_file, encoding=\"utf-8\"\r\n ) as tgt_f:\r\n ... # the rest is the same\r\n```\r\n\r\nAnd that's it!\r\n\r\nPS: Let me know if you need help updating the dummy data and regenerating the metadata file.",
"Awesome! Thanks for the detailed help, that was straightforward with your instruction. However, I think I am being blocked by this issue: https://github.com/huggingface/datasets/issues/4428",
"Feel free to open a PR, and I can fix this manually.",
"Awsome, done in #4451!"
] | **Is your feature request related to a problem? Please describe.**
The [Multi-News dataloader points to the original version of the Multi-News dataset](https://github.com/huggingface/datasets/blob/12540dd75015678ec6019f258d811ee107439a73/datasets/multi_news/multi_news.py#L47), but this has [known errors in it](https://github.com/Alex-Fabbri/Multi-News/issues/11). There exists a [newer version which fixes some of these issues](https://drive.google.com/open?id=1jwBzXBVv8sfnFrlzPnSUBHEEAbpIUnFq).
Unfortunately I don't think you can just replace this old URL with the new one, otherwise this could lead to issues with reproducibility.
**Describe the solution you'd like**
Add a new version to the Multi-News dataloader that points to the updated dataset which has fixes for some known issues.
**Describe alternatives you've considered**
Replace the current URL to the original version to the dataset with the URL to the version with fixes.
**Additional context**
Would be happy to make a PR for this, could someone maybe point me to another dataloader that has multiple versions so I can see how this is handled in `datasets`?
| 4,430 |
https://github.com/huggingface/datasets/issues/4428 | Errors when building dummy data if you use nested _URLS | [] | ## Describe the bug
When making dummy data with the `datasets-cli dummy_data` tool,
an error will be raised if you use a nested _URLS in your dataset script.
Traceback (most recent call last):
File "/home/name/LCCC/datasets/src/datasets/commands/datasets_cli.py", line 43, in <module>
main()
File "/home/name/LCCC/datasets/src/datasets/commands/datasets_cli.py", line 39, in main
service.run()
File "/home/name/LCCC/datasets/src/datasets/commands/dummy_data.py", line 311, in run
self._autogenerate_dummy_data(
File "/home/name/LCCC/datasets/src/datasets/commands/dummy_data.py", line 337, in _autogenerate_dummy_data
dataset_builder._split_generators(dl_manager)
File "/home/name/.cache/huggingface/modules/datasets_modules/datasets/personal_dialog/559332bced5eeafa7f7efc2a7c10ce02cee2a8116bbab4611c35a50ba2715b77/personal_dialog.py", line 108, in _split_generators
data_dir = dl_manager.download_and_extract(urls)
File "/home/name/LCCC/datasets/src/datasets/commands/dummy_data.py", line 56, in download_and_extract
dummy_output = self.mock_download_manager.download(url_or_urls)
File "/home/name/LCCC/datasets/src/datasets/download/mock_download_manager.py", line 130, in download
return self.download_and_extract(data_url)
File "/home/name/LCCC/datasets/src/datasets/download/mock_download_manager.py", line 122, in download_and_extract
return self.create_dummy_data_dict(dummy_file, data_url)
File "/home/name/LCCC/datasets/src/datasets/download/mock_download_manager.py", line 165, in create_dummy_data_dict
if isinstance(first_value, str) and len(set(dummy_data_dict.values())) < len(dummy_data_dict.values()):
TypeError: unhashable type: 'list'
## Steps to reproduce the bug
You can use my dataset script implemented here:
https://github.com/silverriver/datasets/blob/2ecd36760c40b8e29b1137cd19b5bad0e19c76fd/datasets/personal_dialog/personal_dialog.py
```python
datasets_cli dummy_data datasets/personal_dialog --auto_generate
```
You can change https://github.com/silverriver/datasets/blob/2ecd36760c40b8e29b1137cd19b5bad0e19c76fd/datasets/personal_dialog/personal_dialog.py#L54
to
```
"train": "https://huggingface.co/datasets/silver/personal_dialog/resolve/main/dev_random.jsonl.gz"
```
before runing the above script to avoid downloading a large training data.
## Expected results
The dummy data should be generated
## Actual results
An error is raised.
It seems that in https://github.com/huggingface/datasets/blob/12540dd75015678ec6019f258d811ee107439a73/src/datasets/download/mock_download_manager.py#L165
We only check if the first item of dummy_data_dict.values() is str.
However, dummy_data_dict.values() may have the type of [str, list, list].
A simple fix would be changing https://github.com/huggingface/datasets/blob/12540dd75015678ec6019f258d811ee107439a73/src/datasets/download/mock_download_manager.py#L165 to
```python
if all([isinstance(value, str) for value in dummy_data_dict.values()]) and len(set(dummy_data_dict.values())) < len(dummy_data_dict.values()):
```
But I don't know if this kinds of change may bring any side effect since I am not sure about the detail logic here.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- Platform: Linux
- Python version: Python 3.9.10
- PyArrow version: 7.0.0
| 4,428 |
https://github.com/huggingface/datasets/issues/4426 | Add loading variable number of columns for different splits | [
"Hi! Indeed the column is missing, but you shouldn't get an error? Have you made some modifications (locally) to the loading script? I've opened a PR to add the missing columns to the script. "
] | **Is your feature request related to a problem? Please describe.**
The original dataset `blended_skill_talk` consists of different sets of columns for the different splits: (test/valid) splits have additional data column `label_candidates` that the (train) doesn't have.
When loading such data, an exception occurs at table.py:cast_table_to_schema, because of mismatched columns. | 4,426 |
https://github.com/huggingface/datasets/issues/4422 | Cannot load timit_asr data set | [
"Thanks for reporting, @bhaddow.\r\n\r\nI'm fixing it.",
"Thanks for the quick fix!",
"@bhaddow we have also made a fix so that you don't have to convert to uppercase the file extensions of the LDC data.\r\n\r\nWould you mind checking if it works OK now for you and reporting if there are any issues? Thanks. ",
"Hi @albertvillanova -It loads fine on a copy of the data from deepai - although I have to remove the copies of the .WAV files (with extension .WAV,wav). On a copy of the data that was obtained from the LDC, the glob still fails to find the files. The LDC copy looks like it was copied from CD, in 2004, so the structure may be different to a current download.",
"Ah, if I change the train/ and test/ directories to TRAIN/ and TEST/ then it works!",
"Thanks for your investigation and report, @bhaddow. I'm adding another fix for the TRAIN/train and TEST/test directory names."
] | ## Describe the bug
I am trying to load the timit_asr data set. I have tried with a copy from the LDC, and a copy from deepai. In both cases they fail with a "duplicate key" error. With the LDC version I have to convert the file extensions all to upper-case before I can load it at all.
## Steps to reproduce the bug
```python
timit = datasets.load_dataset("timit_asr", data_dir = "/path/to/dataset")
# Sample code to reproduce the bug
```
## Expected results
The data set should load without error. It worked for me before the LDC url change.
## Actual results
```
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: SA1
Keys should be unique and deterministic in nature
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version:
- `datasets` version: 2.2.2
- Platform: Linux-5.4.0-90-generic-x86_64-with-glibc2.17
- Python version: 3.8.12
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,422 |
https://github.com/huggingface/datasets/issues/4420 | Metric evaluation problems in multi-node, shared file system | [
"If you call `metric.compute` in a distributed setup like yours, then `metric.compute` is called in each process. `metric.compute` first calls `metric.add_batch`, and it looks like your error appears at that stage.\r\n\r\nTo make sure that all the processes have started writing their predictions/references at the same time, each process waits for process 0 to lock `slurm-{world_size}-0.arrow.lock`. Process 0 locks this file when `metric.add_batch` is called, so here when `metric.compute` is called.\r\n\r\nTherefore your error can happen when process 0 takes too much time to call `metric.compute` compared to process 3 (>100 seconds by default). I haven't tried running your code but could it be the case ?\r\n\r\nI guess it could also happen if you run multiple times the same distributed job at the same time with the same `experiment_id` because they would collide.\r\n",
"We've finally been able to isolate the problem, it wasn't a timing problem, but rather a file locking one. \r\nThe locks produced by calling `flock` where not visible between nodes (so the master node couldn't check other node's locks nor the other way around). \r\n\r\nWe are now having issues with the pre-processing in our runner script, but are not related with the rendezvous process during the evaluation phase. We will let you know about it once we address it. \r\n\r\nOur solution to the rendezvous is as follows:\r\n- We solved the problem by calling `lockf` instead of `flock`.\r\n- We had to change slightly the `_check_all_processes_locks` method so that the main process (i.e. process 0) didn't check it's own lock (because `lockf` permits recursive locks and thus checking it only replaced the current lock with a new one). \r\n\r\nWe use a shared file system between nodes using GPFS in our cluster setup. Maybe the difference between the behavior we see with respect to your usage in multi-node executions comes from that fact. Which file system scheme do you use for the multi-node executions? \r\n\r\n`lockf` seems to work in more settings than `flock`, so maybe we could write a PR so you could test it in your environment. ",
"Cool, I'm glad you managed to make evaluation work :)\r\n\r\nI'm not completely aware of the differences between lockf and flock, but I've read somewhere that flock is preferable over lockf in multithreading and multiprocessing situations. Here we definitely are in such a situation so unless it is super important I don't think we will switch to lockf",
"> * We had to change slightly the `_check_all_processes_locks` method so that the main process (i.e. process 0) didn't check it's own lock (because `lockf` permits recursive locks and thus checking it only replaced the current lock with a new one).\r\n\r\nHi @panserbjorn , Can you share your `_check_all_processes_locks` function? thanks!",
"```\r\ndef _check_all_processes_locks(self):\r\n expected_lock_file_names = [\r\n os.path.join(self.data_dir, f\"{self.experiment_id}-{self.num_process}-{process_id}.arrow.lock\")\r\n for process_id in range(self.num_process)\r\n ]\r\n #for expected_lock_file_name in expected_lock_file_names: # OUR CHANGE process 0 shouldn't check its own lock\r\n for expected_lock_file_name in expected_lock_file_names[1:]:\r\n nofilelock = FileFreeLock(expected_lock_file_name)\r\n try:\r\n nofilelock.acquire(timeout=self.timeout)\r\n except Timeout:\r\n raise ValueError(\r\n f\"Expected to find locked file {expected_lock_file_name} from process {self.process_id} but it doesn't exist.\"\r\n )\r\n else:\r\n nofilelock.release()\r\n```\r\n\r\n### Changed files:\r\n- metric.py file in the datasets library \r\n- filelock.py file in the datasets/utils library. \r\n\r\n\r\nChanges we made:\r\n\r\n1. We changed the flock for lockf \r\n flock and lockf both perform a lock over a file (like the lock for writing). \r\n The difference is that flock only works in local file systems, but if you have a shared file system (like what we have in the clusters) the flock fails to βseeβ the lock of another node. The only disadvantage we had was that a single process couldnβt detect itβs own lock so we did the second change.\r\n2. We prevented the process 0 (which is the one that coordinates the rendezvous) from checking its own lock on its arrow because it didn't work with lockf (as stated in the previous change). \r\n3. We made a second rendezvous so that all the process had the results of the metrics (other than the loss) and not only the process 0.\r\n What happened was that only process 0 computed the metric and that didnβt present any problem if you are using the loss. However, if you are using another metric, the only process which had the information to choose the best checkpoint at evaluation time was the process 0. But since the evaluation was performed over all processes, every process except the process 0 chose a bad check point (bad meaning it wasnβt the best one) because they didnβt have the information of the metric of the best checkpoint. \r\n The consequence was that the evaluation was different from what would result if using only the best checkpoint, because each process chose a different checkpoint to run the evaluation and thus the numbers were often worse than the numbers that would be obtained if all processes choose the best checkpoint (correct one) to perform the evaluation of their samples. \r\n We performed a second rendezvous so that all processes had the same best_metric and best_model as process 0 after the evaluation cycle. \r\n",
"Metrics are deprecated in `datasets` and `evaluate` should be used instead: https://github.com/huggingface/evaluate"
] | ## Describe the bug
Metric evaluation fails in multi-node within a shared file system, because the master process cannot find the lock files from other nodes. (This issue was originally mentioned in the transformers repo https://github.com/huggingface/transformers/issues/17412)
## Steps to reproduce the bug
1. clone [this huggingface model](https://huggingface.co/PereLluis13/wav2vec2-xls-r-300m-ca-lm) and replace the `run_speech_recognition_ctc.py` script with the version in the gist [here](https://gist.github.com/gullabi/3f66094caa8db1c1e615dd35bd67ec71#file-run_speech_recognition_ctc-py).
2. Setup the `venv` according to the requirements of the model file plus `datasets==2.0.0`, `transformers==4.18.0` and `torch==1.9.0`
3. Launch the runner in a distributed environment which has a shared file system for two nodes, preferably with SLURM. Example [here](https://gist.github.com/gullabi/3f66094caa8db1c1e615dd35bd67ec71)
Specifically for the datasets, for the distributed setup the `load_metric` is called as:
```
process_id=int(os.environ["RANK"])
num_process=int(os.environ["WORLD_SIZE"])
eval_metrics = {metric: load_metric(metric,
process_id=process_id,
num_process=num_process,
experiment_id="slurm")
for metric in data_args.eval_metrics}
```
## Expected results
The training should not fail, due to the failure of the `Metric.compute()` step.
## Actual results
For the test I am executing the world size is 4, with 2 GPUs in 2 nodes. However the process is not finding the necessary lock files
```
File "/gpfs/projects/bsc88/speech/asr/wav2vec2-xls-r-300m-ca-lm/run_speech_recognition_ctc.py", line 841, in <module>
main()
File "/gpfs/projects/bsc88/speech/asr/wav2vec2-xls-r-300m-ca-lm/run_speech_recognition_ctc.py", line 792, in main
train_result = trainer.train(resume_from_checkpoint=checkpoint)
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/transformers/trainer.py", line 1497, in train
self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/transformers/trainer.py", line 1624, in _maybe_log_save_evaluate
metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/transformers/trainer.py", line 2291, in evaluate
metric_key_prefix=metric_key_prefix,
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/transformers/trainer.py", line 2535, in evaluation_loop
metrics = self.compute_metrics(EvalPrediction(predictions=all_preds, label_ids=all_labels))
File "/gpfs/projects/bsc88/speech/asr/wav2vec2-xls-r-300m-ca-lm/run_speech_recognition_ctc.py", line 742, in compute_metrics
metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
File "/gpfs/projects/bsc88/speech/asr/wav2vec2-xls-r-300m-ca-lm/run_speech_recognition_ctc.py", line 742, in <dictcomp>
metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/datasets/metric.py", line 419, in compute
self.add_batch(**inputs)
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/datasets/metric.py", line 465, in add_batch
self._init_writer()
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/datasets/metric.py", line 552, in _init_writer
self._check_rendez_vous() # wait for master to be ready and to let everyone go
File "/gpfs/projects/bsc88/projects/speech-tech-resources/venv_amd_speech/lib/python3.7/site-packages/datasets/metric.py", line 342, in _check_rendez_vous
) from None
ValueError: Expected to find locked file /home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-0.arrow.lock from process 3 but it doesn't exist.
```
When I look at the cache directory, I can see all the lock files in principle:
```
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-0.arrow
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-0.arrow.lock
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-1.arrow
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-1.arrow.lock
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-2.arrow
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-2.arrow.lock
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-3.arrow
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-3.arrow.lock
/home/bsc88/bsc88474/.cache/huggingface/metrics/wer/default/slurm-4-rdv.lock
```
I see that there was another related issue here https://github.com/huggingface/datasets/issues/1942, but it seems to have resolved via https://github.com/huggingface/datasets/pull/1966. Let me know if there is problem with how I am calling the `load_metric` or whether I need to make changes to the `.compute()` steps.
## Environment info
- `datasets` version: 2.0.0
- Platform: Linux-4.18.0-147.8.1.el8_1.x86_64-x86_64-with-centos-8.1.1911-Core
- Python version: 3.7.4
- PyArrow version: 7.0.0
- Pandas version: 1.3.0
| 4,420 |
https://github.com/huggingface/datasets/issues/4419 | Update `unittest` assertions over tuples from `assertEqual` to `assertTupleEqual` | [
"Hi! If the only goal is to improve readability, it's better to use `assertTupleEqual` than `assertSequenceEqual` for Python tuples. Also, note that this function is called internally by `assertEqual`, but I guess we can accept a PR to be more verbose.",
"Hi @mariosasko, right! I'll update the issue title/desc with `assertTupleEqual` even though as you said it seems to be internally using `assertEqual` so I'm not sure whether it's worth it or not...\r\n\r\nhttps://docs.python.org/3/library/unittest.html#unittest.TestCase.assertTupleEqual",
"I thought we were supposed to move gradually from `unittest` to `pytest`..."
] | **Is your feature request related to a problem? Please describe.**
So this is more a readability improvement rather than a proposal, wouldn't it be better to use `assertTupleEqual` over the tuples rather than `assertEqual`? As `unittest` added that function in `v3.1`, as detailed at https://docs.python.org/3/library/unittest.html#unittest.TestCase.assertTupleEqual, so maybe it's worth updating.
Find an example of an `assertEqual` over a tuple in π€ `datasets` unit tests over an `ArrowDataset` at https://github.com/huggingface/datasets/blob/0bb47271910c8a0b628dba157988372307fca1d2/tests/test_arrow_dataset.py#L570
**Describe the solution you'd like**
Start slowly replacing all the `assertEqual` statements with `assertTupleEqual` if the assertion is done over a Python tuple, as we're doing with the Python lists using `assertListEqual` rather than `assertEqual`.
**Additional context**
If so, please let me know and I'll try to go over the tests and create a PR if applicable, otherwise, if you consider this should stay as `assertEqual` rather than `assertSequenceEqual` feel free to close this issue! Thanks π€
| 4,419 |
https://github.com/huggingface/datasets/issues/4417 | how to convert a dict generator into a huggingface dataset. | [
"@albertvillanova @lhoestq , could you please help me on this issue. ",
"Hi ! As mentioned on the [forum](https://discuss.huggingface.co/t/how-to-wrap-a-generator-with-hf-dataset/18464), the simplest for now would be to define a [dataset script](https://huggingface.co/docs/datasets/dataset_script) which can contain your generator. But we can also explore adding something like `ds = Dataset.from_iterable(seqio_dataset)`",
"@lhoestq , hey i did as you instructed, but sadly i cannot get pass through the download_manager, as i dont have anything to download. i was skipping the ` def _split_generators(self, dl_manager):` function. but i cannot get around it. I get a `NotImplementedError: `\r\n\r\nthe following is my code for the same: \r\n\r\n\r\n\r\n```\r\nimport datasets \r\nimport functools\r\nimport glob \r\nfrom datasets import load_from_disk\r\nimport seqio\r\nimport tensorflow as tf\r\nimport t5.data\r\nfrom datasets import load_dataset\r\nfrom t5.data import postprocessors\r\nfrom t5.data import preprocessors\r\nfrom t5.evaluation import metrics\r\nfrom seqio import FunctionDataSource, utils\r\n\r\nTaskRegistry = seqio.TaskRegistry\r\n\r\ndata_path = glob.glob(\"/home/stephen/Desktop/MEGA_CORPUS/COMBINED_CORPUS/*\", recursive=False)\r\n\r\n\r\ndef gen_dataset(split, shuffle=False, seed=None, column=\"text\", dataset_path=None):\r\n dataset = load_from_disk(dataset_path)\r\n if shuffle:\r\n if seed:\r\n dataset = dataset.shuffle(seed=seed)\r\n else:\r\n dataset = dataset.shuffle()\r\n while True:\r\n for item in dataset[str(split)]:\r\n yield item[column]\r\n\r\n\r\ndef dataset_fn(split, shuffle_files, seed=None, dataset_path=None):\r\n return tf.data.Dataset.from_generator(\r\n functools.partial(gen_dataset, split, shuffle_files, seed, dataset_path=dataset_path),\r\n output_signature=tf.TensorSpec(shape=(), dtype=tf.string, name=dataset_path)\r\n )\r\n\r\[email protected]_over_dataset\r\ndef target_to_key(x, key_map, target_key):\r\n \"\"\"Assign the value from the dataset to target_key in key_map\"\"\"\r\n return {**key_map, target_key: x}\r\n\r\n\r\n_CITATION = \"Not ready yet\"\r\n_DESCRIPTION = \"a custom seqio based mixed samples on a given temperature value, that again returns a dataset in HF dataset format well samples on the Mixture temperature\"\r\n_HOMEPAGE = \"ldcil.org\"\r\n\r\nclass CustomSeqio(datasets.GeneratorBasedBuilder):\r\n\r\n def _info(self):\r\n return datasets.DatasetInfo(\r\n description=_DESCRIPTION,\r\n features=datasets.Features(\r\n {\r\n \"text\": datasets.Value(\"string\"),\r\n }\r\n ),\r\n homepage=\"https://ldcil.org\",\r\n citation=_CITATION,)\r\n\r\ndef generate_examples(self):\r\n seqio_train_list = []\r\n for lang in data_path:\r\n dataset_name = lang.split(\"/\")[-1]\r\n dataset_shapes = None \r\n\r\n TaskRegistry.add(\r\n str(dataset_name),\r\n source=seqio.FunctionDataSource(\r\n dataset_fn=functools.partial(dataset_fn, dataset_path=lang),\r\n splits=(\"train\", \"test\"),\r\n caching_permitted=False,\r\n num_input_examples=dataset_shapes,\r\n ),\r\n preprocessors=[\r\n functools.partial(\r\n target_to_key, key_map={\r\n \"targets\": None,\r\n }, target_key=\"targets\")],\r\n output_features={\"targets\": seqio.Feature(vocabulary=seqio.PassThroughVocabulary, add_eos=False, dtype=tf.string, rank=0)},\r\n metric_fns=[]\r\n )\r\n\r\n seqio_train_dataset = seqio.get_mixture_or_task(dataset_name).get_dataset(\r\n sequence_length=None,\r\n split=\"train\",\r\n shuffle=True,\r\n num_epochs=1,\r\n shard_info=seqio.ShardInfo(index=0, num_shards=10),\r\n use_cached=False,\r\n seed=42)\r\n seqio_train_list.append(seqio_train_dataset)\r\n \r\n lang_name_list = []\r\n for lang in data_path:\r\n lang_name = lang.split(\"/\")[-1]\r\n lang_name_list.append(lang_name)\r\n\r\n seqio_mixture = seqio.MixtureRegistry.add(\r\n \"seqio_mixture\",\r\n lang_name_list,\r\n default_rate=0.7)\r\n \r\n seqio_mixture_dataset = seqio.get_mixture_or_task(\"seqio_mixture\").get_dataset(\r\n sequence_length=None,\r\n split=\"train\",\r\n shuffle=True,\r\n num_epochs=1,\r\n shard_info=seqio.ShardInfo(index=0, num_shards=10),\r\n use_cached=False,\r\n seed=42)\r\n\r\n for id, ex in enumerate(seqio_mixture_dataset):\r\n yield id, {\"text\": ex[\"targets\"].numpy().decode()}\r\n```\r\n\r\nand i load it by:\r\n\r\n`seqio_mixture = load_dataset(\"seqio_loader\")`",
"@lhoestq , just to make things clear ... \r\n\r\nthe following is my original code, thats not in the HF dataset loading script: \r\n\r\n```\r\nimport functools\r\nimport seqio\r\nimport tensorflow as tf\r\nimport t5.data\r\nfrom datasets import load_from_disk\r\nfrom t5.data import postprocessors\r\nfrom t5.data import preprocessors\r\nfrom t5.evaluation import metrics\r\nfrom seqio import FunctionDataSource, utils\r\nimport glob \r\n\r\nTaskRegistry = seqio.TaskRegistry\r\n\r\n\r\n\r\ndef gen_dataset(split, shuffle=False, seed=None, column=\"text\", dataset_path=None):\r\n dataset = load_from_disk(dataset_path)\r\n if shuffle:\r\n if seed:\r\n dataset = dataset.shuffle(seed=seed)\r\n else:\r\n dataset = dataset.shuffle()\r\n while True:\r\n for item in dataset[str(split)]:\r\n yield item[column]\r\n\r\n\r\ndef dataset_fn(split, shuffle_files, seed=None, dataset_path=None):\r\n return tf.data.Dataset.from_generator(\r\n functools.partial(gen_dataset, split, shuffle_files, seed, dataset_path=dataset_path),\r\n output_signature=tf.TensorSpec(shape=(), dtype=tf.string, name=dataset_path)\r\n )\r\n\r\n\r\[email protected]_over_dataset\r\ndef target_to_key(x, key_map, target_key):\r\n \"\"\"Assign the value from the dataset to target_key in key_map\"\"\"\r\n return {**key_map, target_key: x}\r\n\r\ndata_path = glob.glob(\"/home/stephen/Desktop/MEGA_CORPUS/COMBINED_CORPUS/*\", recursive=False)\r\n\r\nseqio_train_list = []\r\n\r\nfor lang in data_path:\r\n dataset_name = lang.split(\"/\")[-1]\r\n dataset_shapes = None \r\n\r\n TaskRegistry.add(\r\n str(dataset_name),\r\n source=seqio.FunctionDataSource(\r\n dataset_fn=functools.partial(dataset_fn, dataset_path=lang),\r\n splits=(\"train\", \"test\"),\r\n caching_permitted=False,\r\n num_input_examples=dataset_shapes,\r\n ),\r\n preprocessors=[\r\n functools.partial(\r\n target_to_key, key_map={\r\n \"targets\": None,\r\n }, target_key=\"targets\")],\r\n output_features={\"targets\": seqio.Feature(vocabulary=seqio.PassThroughVocabulary, add_eos=False, dtype=tf.string, rank=0)},\r\n metric_fns=[]\r\n )\r\n\r\n seqio_train_dataset = seqio.get_mixture_or_task(dataset_name).get_dataset(\r\n sequence_length=None,\r\n split=\"train\",\r\n shuffle=True,\r\n num_epochs=1,\r\n shard_info=seqio.ShardInfo(index=0, num_shards=10),\r\n use_cached=False,\r\n seed=42)\r\n seqio_train_list.append(seqio_train_dataset)\r\n\r\nlang_name_list = []\r\nfor lang in data_path:\r\n lang_name = lang.split(\"/\")[-1]\r\n lang_name_list.append(lang_name)\r\n\r\nseqio_mixture = seqio.MixtureRegistry.add(\r\n \"seqio_mixture\",\r\n lang_name_list,\r\n default_rate=0.7\r\n)\r\n\r\nseqio_mixture_dataset = seqio.get_mixture_or_task(\"seqio_mixture\").get_dataset(\r\n sequence_length=None,\r\n split=\"train\",\r\n shuffle=True,\r\n num_epochs=1,\r\n shard_info=seqio.ShardInfo(index=0, num_shards=10),\r\n use_cached=False,\r\n seed=42)\r\n\r\nfor _, ex in zip(range(15), seqio_mixture_dataset):\r\n print(ex[\"targets\"].numpy().decode())\r\n```\r\n\r\nwhere the seqio_mixture_dataset is the generator that i wanted to be wrapped in HF dataset. \r\n\r\nalso additionally, could you please tell me how do i set the `default_rate=0.7` args where `seqio_mixture` is defined to be made as a custom option in the HF load_dataset() method,\r\n\r\nmaybe like this: \r\n`seqio_mixture_dataset = datasets.load_dataset(\"seqio_loader\",temperature=0.5)`",
"I like the idea of having `Dataset.from_iterable(iterable)` in the API. The only problem is that we also want to make this part cachable, which is tricky if `iterable` is a generator. \r\n\r\nSome resources on this issue:\r\n* https://github.com/uqfoundation/dill/issues/311\r\n* https://stackoverflow.com/questions/7180212/why-cant-generators-be-pickled\r\n* https://github.com/tonyroberts/generator_tools - python package for pickling generators; pickles bytecode, so it creates version-specific dumps",
"For the caching maybe we can have `Dataset.from_generator` as TF and pickle+hash the generator function (not the generator object itself) ?\r\n\r\nAnd then keep `Dataset.from_iterable` fo pickable objects like lists",
"@lhoestq, @mariosasko do you too have any examples where the dataset is a generator and needs to be wrapped into hf dataset ? ",
"@lhoestq, following to my previous question ... what possibly could be done in this [link1](https://github.com/huggingface/datasets/issues/4417#issuecomment-1146627404) [link2](https://github.com/huggingface/datasets/issues/4417#issuecomment-1146627593) case? do you have any ideas? ",
"@lhoestq +1 for the `Dataset.from_generator` idea.\r\n\r\nHaving thought about it, let's avoid adding `Dataset.from_iterable` to the API since dictionaries are technically iteralbles (\"iterable\" is a broad term in Python), and we already provide `Dataset.from_dict`. And for lists maybe we can add `Dataset.from_list` similar to `pa.Table.from_pylist`. WDYT?\r\n",
"Hi @StephennFernandes!\r\n\r\nTo fix the issues in the copied code, rename `generate_examples` to` _generate_examples` and add one level of indentation as this is a method of `GeneratorBasedBuilder` and define `_split_generators` as follows (again as a method of `GeneratorBasedBuilder):\r\n```python\r\n def _split_generators(self, dl_manager):\r\n return [\r\n datasets.SplitGenerator(\r\n name=datasets.Split.TRAIN,\r\n gen_kwargs={},\r\n ),\r\n ]\r\n```\r\n\r\nAnd if you are feeling extra adventurous, you can try to use ArrowWriter to directly create a cache file:\r\n```python\r\nfrom datasets import Dataset\r\nfrom datasets.arrow_writer import ArrowWriter\r\n\r\nwriter = ArrowWriter(path=\"path/to/cache_file.arrow\", writer_batch_size=1000)\r\n\r\nwith writer:\r\n for ex in generator:\r\n writer.write(ex) \r\n writer.finalize()\r\n\r\ndset = Dataset.from_file(\"path/to/cache_file.arrow\")\r\n```\r\n\r\n",
"I have a problem which I think is very similar: I would like to \"stream\" data to a HF Array (memory-mapped) Dataset, where the final size of the dataset is unknown, but could be much larger than what fits into memory.\r\nWhat I want to end up with is an Array Dataset which I can open using `Dataset.load_from_disk(dataset_path=\"somename\")` and use e.g. as the training set. \r\n\r\nFor this I would have thought there should be an API which allows me to open/create the dataset (and define the features etc), then write examples to the dataset, but I could not find a way to do this. \r\n\r\nI tried doing this and it looks like it works, but it feels very hacky and I am not sure if this might fail to update some of the fields in the json files which may turn out to be important:\r\n```\r\nfrom datasets import Dataset, Features, ClassLabel, Sequence, Value\r\nfrom datasets.arrow_writer import ArrowWriter \r\n# 1) define the features\r\nfeatures = Features(dict(\r\n id=Value(dtype=\"string\"),\r\n tokens=Sequence(feature=Value(dtype=\"string\")),\r\n ner_tags=Sequence(feature=ClassLabel(names=['O', 'B-corporation', 'I-corporation', 'B-creative-work', 'I-creative-work', 'B-group', 'I-group', 'B-location', 'I-location', 'B-person', 'I-person', 'B-product', 'I-product'])),\r\n))\r\n# 2) create empty dataset for examples with these features and store to disk\r\nempty = dict(\r\n id = [],\r\n tokens = [],\r\n ner_tags = [],\r\n)\r\nds = Dataset.from_dict(empty, features=features)\r\nds.save_to_disk(dataset_path=\"debug_ds1\")\r\n\r\n# 3) directly write all the examples to the arrow dataset \r\nwith ArrowWriter(path=\"debug_ds1/dataset.arrow\") as writer: \r\n writer.write(dict(id=0, tokens=[\"a\", \"b\"], ner_tags=[0, 0])) \r\n writer.write(dict(id=1, tokens=[\"x\", \"y\"], ner_tags=[1, 0])) \r\n writer.finalize() \r\n \r\nds2 = Dataset.load_from_disk(dataset_path=\"debug_ds1\")\r\nlen(ds2)\r\n```\r\nIs there a cleaner/proper way to do this?\r\n\r\nI like the sound of `Dataset.from_iterable` or `Dataset.from_generator` (should not from iterable be able to handle from generator too as all generators are iterables?) but how would I define the features for me examples there? ",
"Hi @johann-petrak! You can pass the features directly to ArrowWriter's initializer like so `ArrowWriter(..., features=features)`.\r\n\r\nAnd the reason why I prefer `Dataset.from_generator` over `Dataset.from_iterable` is mentioned in one of my previous comments.",
"@mariosasko so at the moment we still have to create a fake `Dataset` first and then use `ArrowWriter` to write an actual dataset? I'm using the latest version of `datasets` on pypi but my final file is always empty. Is there anything wrong with the code below?\r\n\r\n```python\r\n total = 0\r\n with ArrowWriter(path=str(final_data_path), features=features) as writer:\r\n for batch in loader:\r\n for traj in batch:\r\n for generator in question_generators:\r\n for xi in generator(traj):\r\n # print(f\"Question: {xi.question}, answer: {xi.answer}\")\r\n total += 1\r\n writer.write(\r\n {\r\n \"id\": f\"qa_{total}\",\r\n \"question\": xi.question,\r\n \"answer\": xi.answer,\r\n }\r\n )\r\n writer.finalize()\r\n print(f\"Total #questions = {total}\") # this prints 402\r\n```",
"This works for me if I then (actually I also close the writer: `writer.close()`) open the Arrow file as a dataset using `ds=Dataset.from_file(final_data_path)` then `ds.save_to_disk(somedir)`. The Dataset created that way contains the expected examples.",
"Oh thanks. That did the trick I believe. Shouldn't ArrowWriter have a context manager that does these operations?",
"You can just use `Dataset.from_file` to get your dataset, no need to do an extra `save_to_disk` somewhere else ;)",
"I was thinking that `save_to_disk` is necessary when one wants to re-use that dataset as a proper HF dataset later, no?\r\nAt least what I wanted to achieve is create a dataset that can be opened like any other local or remote dataset. ",
"`save_to_disk`/`load_from_disk` is indeed more general, e.g. it supports datasets that consist in several files, and saves some extra info in a dataset_info.json file (description, citation, split sizes, etc.)\r\n\r\nIf you have one single file it's fine to simply do `.from_file()`"
] | ### Link
_No response_
### Description
Hey there, I have used seqio to get a well distributed mixture of samples from multiple dataset. However the resultant output from seqio is a python generator dict, which I cannot produce back into huggingface dataset.
The generator contains all the samples needed for training the model but I cannot convert it into a huggingface dataset.
The code looks like this:
```
for ex in seqio_data:
print(ex[βtextβ])
```
I need to convert the seqio_data (generator) into huggingface dataset.
the complete seqio code goes here:
```
import functools
import seqio
import tensorflow as tf
import t5.data
from datasets import load_dataset
from t5.data import postprocessors
from t5.data import preprocessors
from t5.evaluation import metrics
from seqio import FunctionDataSource, utils
TaskRegistry = seqio.TaskRegistry
def gen_dataset(split, shuffle=False, seed=None, column="text", dataset_params=None):
dataset = load_dataset(**dataset_params)
if shuffle:
if seed:
dataset = dataset.shuffle(seed=seed)
else:
dataset = dataset.shuffle()
while True:
for item in dataset[str(split)]:
yield item[column]
def dataset_fn(split, shuffle_files, seed=None, dataset_params=None):
return tf.data.Dataset.from_generator(
functools.partial(gen_dataset, split, shuffle_files, seed, dataset_params=dataset_params),
output_signature=tf.TensorSpec(shape=(), dtype=tf.string, name=dataset_name)
)
@utils.map_over_dataset
def target_to_key(x, key_map, target_key):
"""Assign the value from the dataset to target_key in key_map"""
return {**key_map, target_key: x}
dataset_name = 'oscar-corpus/OSCAR-2109'
subset= 'mr'
dataset_params = {"path": dataset_name, "language":subset, "use_auth_token":True}
dataset_shapes = None
TaskRegistry.add(
"oscar_marathi_corpus",
source=seqio.FunctionDataSource(
dataset_fn=functools.partial(dataset_fn, dataset_params=dataset_params),
splits=("train", "validation"),
caching_permitted=False,
num_input_examples=dataset_shapes,
),
preprocessors=[
functools.partial(
target_to_key, key_map={
"targets": None,
}, target_key="targets")],
output_features={"targets": seqio.Feature(vocabulary=seqio.PassThroughVocabulary, add_eos=False, dtype=tf.string, rank=0)},
metric_fns=[]
)
dataset = seqio.get_mixture_or_task("oscar_marathi_corpus").get_dataset(
sequence_length=None,
split="train",
shuffle=True,
num_epochs=1,
shard_info=seqio.ShardInfo(index=0, num_shards=10),
use_cached=False,
seed=42
)
for _, ex in zip(range(5), dataset):
print(ex['targets'].numpy().decode())
```
### Owner
_No response_ | 4,417 |
https://github.com/huggingface/datasets/issues/4413 | Dataset Viewer issue for ett | [
"Thanks for reporting @dgcnz.\r\n\r\nI have checked that the dataset works fine in streaming mode.\r\n\r\nAdditionally, other datasets containing timestamps are properly rendered by the viewer: https://huggingface.co/datasets/blbooks\r\n\r\nI have tried to force the refresh of the preview, but the endpoint is not responsive: Connection timed out\r\n\r\nCC: @severo ",
"I've just resent the refresh of the preview to the new endpoint, without success.\r\n\r\nCC: @severo ",
"Fixed!\r\n\r\nhttps://huggingface.co/datasets/ett/viewer/h1/test\r\n\r\n<img width=\"982\" alt=\"Capture dβeΜcran 2022-06-15 aΜ 09 30 22\" src=\"https://user-images.githubusercontent.com/1676121/173769035-a075d753-ecfc-4a43-b54b-973105d464d3.png\">\r\n"
] | ### Link
https://huggingface.co/datasets/ett
### Description
Timestamp is not JSON serializable.
```
Status code: 500
Exception: Status500Error
Message: Type is not JSON serializable: Timestamp
```
### Owner
No | 4,413 |
https://github.com/huggingface/datasets/issues/4407 | Dataset Viewer issue for conll2012_ontonotesv5 | [
"Thanks for reporting, @jiangwy99.\r\n\r\nI guess this could be addressed only once we fix our issue with irresponsive backend endpoint.\r\n\r\nCC: @severo ",
"I've just sent the forcing of the refresh of the preview to the new endpoint.",
"Fixed, thanks for the patience. The issue was the amount of RAM allowed to extract the first rows of the dataset was not sufficient."
] | ### Link
https://huggingface.co/datasets/conll2012_ontonotesv5
### Description
Dataset viewer outage.
### Owner
No | 4,407 |
https://github.com/huggingface/datasets/issues/4405 | [TypeError: Couldn't cast array of type] Cannot process dataset in v2.2.2 | [
"And if the problem is that the way I am to construct the {Entity Type: list of spans} makes entity types without any spans hard to handle, is there a better way to meet the demand? Although I have verified that to make entity types without any spans to behave like `entity_chunk[label] = [[\"\"]]` can perform normally, I still wonder if there is a more elegant way?"
] | ## Describe the bug
I am trying to process the [conll2012_ontonotesv5](https://huggingface.co/datasets/conll2012_ontonotesv5) dataset in `datasets` v2.2.2 and am running into a type error when casting the features.
## Steps to reproduce the bug
```python
import os
from typing import (
List,
Dict,
)
from collections import (
defaultdict,
)
from dataclasses import (
dataclass,
)
from datasets import (
load_dataset,
)
@dataclass
class ConllConverter:
path: str
name: str
cache_dir: str
def __post_init__(
self,
):
self.dataset = load_dataset(
path=self.path,
name=self.name,
cache_dir=self.cache_dir,
)
def convert(
self,
):
class_label = self.dataset["train"].features["sentences"][0]["named_entities"].feature
# label_set = list(set([
# label.split("-")[1] if label != "O" else label for label in class_label.names
# ]))
def prepare_chunk(token, entity):
assert len(token) == len(entity)
# Sequence length
length = len(token)
# Variable used
entity_chunk = defaultdict(list)
idx = flag = 0
# While loop
while idx < length:
if entity[idx] == "O":
flag += 1
idx += 1
else:
iob_tp, lab_tp = entity[idx].split("-")
assert iob_tp == "B"
idx += 1
while idx < length and entity[idx].startswith("I-"):
idx += 1
entity_chunk[lab_tp].append(token[flag: idx])
flag = idx
entity_chunk = dict(entity_chunk)
# for label in label_set:
# if label != "O" and label not in entity_chunk.keys():
# entity_chunk[label] = None
return entity_chunk
def prepare_features(
batch: Dict[str, List],
) -> Dict[str, List]:
sentence = [
sent for doc_sent in batch["sentences"] for sent in doc_sent
]
feature = {
"sentence": list(),
}
for sent in sentence:
token = sent["words"]
entity = class_label.int2str(sent["named_entities"])
entity_chunk = prepare_chunk(token, entity)
sent_feat = {
"token": token,
"entity": entity,
"entity_chunk": entity_chunk,
}
feature["sentence"].append(sent_feat)
return feature
column_names = self.dataset.column_names["train"]
dataset = self.dataset.map(
function=prepare_features,
with_indices=False,
batched=True,
batch_size=3,
remove_columns=column_names,
num_proc=1,
)
dataset.save_to_disk(
dataset_dict_path=os.path.join("data", self.path, self.name)
)
if __name__ == "__main__":
converter = ConllConverter(
path="conll2012_ontonotesv5",
name="english_v4",
cache_dir="cache",
)
converter.convert()
```
## Expected results
I want to use the dataset to perform NER task and to change the label list into a {Entity Type: list of spans} format.
## Actual results
<details>
<summary>Traceback</summary>
```python
Traceback (most recent call last): | 0/81 [00:00<?, ?ba/s]
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/multiprocess/pool.py", line 125, in worker
result = (True, func(*args, **kwds))
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 532, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 499, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/fingerprint.py", line 458, in wrapper
out = func(self, *args, **kwargs)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 2751, in _map_single
writer.write_batch(batch)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_writer.py", line 503, in write_batch
arrays.append(pa.array(typed_sequence))
File "pyarrow/array.pxi", line 230, in pyarrow.lib.array
File "pyarrow/array.pxi", line 110, in pyarrow.lib._handle_arrow_array_protocol
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_writer.py", line 198, in __arrow_array__
out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/table.py", line 1675, in wrapper
return func(array, *args, **kwargs)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/table.py", line 1793, in cast_array_to_feature
arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/table.py", line 1793, in <listcomp>
arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/table.py", line 1675, in wrapper
return func(array, *args, **kwargs)
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/table.py", line 1844, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{feature}")
TypeError: Couldn't cast array of type
struct<CARDINAL: list<item: list<item: string>>, DATE: list<item: list<item: string>>, EVENT: list<item: list<item: string>>, FAC: list<item: list<item: string>>, GPE: list<item: list<item: string>>, LANGUAGE: list<item: list<item: string>>, LAW: list<item: list<item: string>>, LOC: list<item: list<item: string>>, MONEY: list<item: list<item: string>>, NORP: list<item: list<item: string>>, ORDINAL: list<item: list<item: string>>, ORG: list<item: list<item: string>>, PERCENT: list<item: list<item: string>>, PERSON: list<item: list<item: string>>, QUANTITY: list<item: list<item: string>>, TIME: list<item: list<item: string>>, WORK_OF_ART: list<item: list<item: string>>>
to
{'CARDINAL': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'DATE': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'EVENT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'FAC': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'GPE': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'LAW': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'LOC': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'MONEY': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'NORP': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'ORDINAL': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'ORG': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PERCENT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PERSON': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PRODUCT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'QUANTITY': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'TIME': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'WORK_OF_ART': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None)}
"""
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home2/jiangwangyi/workspace/work/Entity/dataconverter.py", line 110, in <module>
converter.convert()
File "/home2/jiangwangyi/workspace/work/Entity/dataconverter.py", line 91, in convert
dataset = self.dataset.map(
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/dataset_dict.py", line 770, in map
{
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/dataset_dict.py", line 771, in <dictcomp>
k: dataset.map(
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 2459, in map
transformed_shards[index] = async_result.get()
File "/home2/jiangwangyi/miniconda3/lib/python3.9/site-packages/multiprocess/pool.py", line 771, in get
raise self._value
TypeError: Couldn't cast array of type
struct<CARDINAL: list<item: list<item: string>>, DATE: list<item: list<item: string>>, EVENT: list<item: list<item: string>>, FAC: list<item: list<item: string>>, GPE: list<item: list<item: string>>, LANGUAGE: list<item: list<item: string>>, LAW: list<item: list<item: string>>, LOC: list<item: list<item: string>>, MONEY: list<item: list<item: string>>, NORP: list<item: list<item: string>>, ORDINAL: list<item: list<item: string>>, ORG: list<item: list<item: string>>, PERCENT: list<item: list<item: string>>, PERSON: list<item: list<item: string>>, QUANTITY: list<item: list<item: string>>, TIME: list<item: list<item: string>>, WORK_OF_ART: list<item: list<item: string>>>
to
{'CARDINAL': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'DATE': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'EVENT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'FAC': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'GPE': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'LAW': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'LOC': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'MONEY': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'NORP': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'ORDINAL': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'ORG': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PERCENT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PERSON': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'PRODUCT': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'QUANTITY': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'TIME': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None), 'WORK_OF_ART': Sequence(feature=Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), length=-1, id=None)}
```
</details>
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.2
- Platform: Ubuntu 18.04
- Python version: 3.9.7
- PyArrow version: 7.0.0
| 4,405 |
https://github.com/huggingface/datasets/issues/4404 | Dataset should have a `.name` field | [
"Hi! You can already use `dset.builder_name` and `dset.config_name` for that purpose. And when it comes to versioning, it's better to use `dset._fingerprint` than the `version` attribute as the former represents a deterministic hash that encodes all the mutable ops executed on a dataset, and the latter stays the same unless it's manually updated after each op.",
"@mariosasko Can we make ._fingerprint not private? seems a critical component for tracking how a model was generated to ensure reproducibility."
] | **Is your feature request related to a problem? Please describe.**
If building pipelines that can evaluate on more than one dataset, it would be nice to be able to log results of things like `Evaluating on {dataset.name}` or `results for {dataset.name} are: {results}`
Without some way of concisely identifying a dataset from the dataset object, tools which might run on more than one dataset must be passed the dataset object _and_ the name/id of the dataset being used.
**Describe the solution you'd like**
The DatasetInfo class should have a `name` field which is the name of a dataset. then for a given dataset if it evolves in time the `version` can be updated but its different versions of the same dataset with a unique `name`. The name could then all be accessed by `dataset.name`
**Describe alternatives you've considered**
For my own purposes I am considering making `NamedDataset[Dataset]` where the subclass just has a .name field.
**Additional context**
My guess is that most usecases are not working with more than one dataset in a given pipeline so a name is not really needed. This has surprised me though as one of the advantages of a standard dataset interface is to be able to build pipelines which can be passed in a dataset and separate responsibilities of the dataset loading from the train or eval pipeline.
| 4,404 |
https://github.com/huggingface/datasets/issues/4401 | "NonMatchingChecksumError" when importing 'spider' dataset | [
"Thanks for reporting, @OmarAlaaeldein.\r\n\r\nDatasets hosted at Google Drive give problems quite often due to a change in their service:\r\n- #3786 \r\n\r\nRelated to:\r\n- #3906\r\n\r\nI'm having a look.",
"We have made a Pull Request to replace the Google Drive URL. This fix will be accessible in our next `datasets` library release.\r\n\r\nIn the meantime, once the PR merged into master, you can get this fix by installing our library from the GitHub master branch:\r\n```shell\r\npip install git+https://github.com/huggingface/datasets#egg=datasets\r\n```\r\nThen, if you had previously tried to load the data and got the checksum error, you should force the redownload of the data (before the fix, you just downloaded and cached the virus scan warning page, instead of the data file):\r\n```shell\r\nload_dataset(\"...\", download_mode=\"force_redownload\")\r\n```"
] | ## Describe the bug
When importing 'spider' dataset [https://huggingface.co/datasets/spider] an error occurs
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('spider')
```
## Expected results
Dataset object
## Actual results
NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://drive.google.com/uc?export=download&id=1_AckYkinAnhqmRQtGsQgUKAnTHxxX5J0']
## Environment info
- `datasets` version: 2.2.2
- Platform: Windows-10-10.0.19041-SP0
- Python version: 3.7.11
- PyArrow version: 6.0.1
- Pandas version: 1.3.5
| 4,401 |
https://github.com/huggingface/datasets/issues/4400 | load dataset wikitext-2-raw-v1 failed. Could not reach wikitext-2-raw-v1.py. | [
"I tried in this way.\r\n\r\n```python\r\nfrom datasets import load_dataset\r\ndataset = load_dataset(path=\"wikitext\", name=\"wikitext-103-v1\", split=\"train\")\r\n```"
] | ## Describe the bug
Could not reach wikitext-2-raw-v1.py
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("wikitext-2-raw-v1")
```
## Expected results
Download `wikitext-2-raw-v1` dataset successfully.
## Actual results
```
File "load_datasets.py", line 13, in <module>
load_dataset("wikitext-2-raw-v1")
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 1715, in load_dataset
**config_kwargs,
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 1536, in load_dataset_builder
data_files=data_files,
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 1282, in dataset_module_factory
raise e1 from None
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 1224, in dataset_module_factory
dynamic_modules_path=dynamic_modules_path,
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 559, in get_module
local_path = self.download_loading_script(revision)
File "/root/miniconda3/lib/python3.6/site-packages/datasets/load.py", line 539, in download_loading_script
return cached_path(file_path, download_config=download_config)
File "/root/miniconda3/lib/python3.6/site-packages/datasets/utils/file_utils.py", line 246, in cached_path
download_desc=download_config.download_desc,
File "/root/miniconda3/lib/python3.6/site-packages/datasets/utils/file_utils.py", line 582, in get_from_cache
raise ConnectionError(f"Couldn't reach {url} ({repr(head_error)})")
ConnectionError: Couldn't reach https://raw.githubusercontent.com/huggingface/datasets/2.2.2/datasets/wikitext-2-raw-v1/wikitext-2-raw-v1.py (ReadTimeout(ReadTimeoutError("HTTPSConnectionPool(host='raw.githubusercontent.com', port=443): Read timed out. (read timeout=100)",),))
```
I tried to download wikitext-2-raw-v1.py by chrome and got:
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.2
- Platform: CentOS 7
- Python version: 3.6
- PyArrow version: 3.0.0
| 4,400 |
https://github.com/huggingface/datasets/issues/4399 | LocalDatasetModuleFactoryWithoutScript extracts invalid builder name | [
"Ok, so\r\n```\r\nos.path.basename(\"/home/user/\")\r\n```\r\ngives `''` while \r\n```\r\nos.path.basename(\"/home/user\")\r\n```\r\ngives `user`. \r\nThe code should check if the last char is a slash.\r\n",
"The fix is:\r\n```\r\n\"name\": os.path.basename(self.path[:-1] if self.path[-1] == \"/\" else self.path)\r\n```",
"I came through the same issue , just removing the last slash in the dataset path fixed it for me, may be this repo moderators could accept this as an accepted answer atleast if this could not be integrated\r\n\r\n> The fix is:\r\n> \r\n> ```\r\n> \"name\": os.path.basename(self.path[:-1] if self.path[-1] == \"/\" else self.path)\r\n> ```\r\n\r\n@apohllo consider making a pull request on this \r\n\r\nThanks for the amazing contributions from huggingface people !!\r\n",
"@apohllo Would you be interested in submitting a PR with the fix?",
"@mariosasko here we go:\r\n\r\nhttps://github.com/huggingface/datasets/pull/4967\r\n\r\nTBH I haven't tested it yet, but should work, since this is a basic change."
] | ## Describe the bug
Trying to load a local dataset raises an error indicating that the config builder has to have a name.
No error should be reported, since the call is completly valid.
## Steps to reproduce the bug
```python
load_dataset("./data/some-dataset/", name="some-name")
```
## Expected results
The dataset should be loaded.
## Actual results
```
Traceback (most recent call last):
File "train_lquad.py", line 19, in <module>
load(tokenize_target_function, tokenize_target_function, {}, tokenizer)
File "train_lquad.py", line 14, in load
dataset = load_dataset("./data/lquad/", name="lquad")
File "/net/pr2/scratch/people/plgapohl/python-3.8.6/lib/python3.8/site-packages/datasets/load.py", line 1708, in load_dataset
builder_instance = load_dataset_builder(
File "/net/pr2/scratch/people/plgapohl/python-3.8.6/lib/python3.8/site-packages/datasets/load.py", line 1560, in load_dataset_builder
builder_instance: DatasetBuilder = builder_cls(
File "/net/pr2/scratch/people/plgapohl/python-3.8.6/lib/python3.8/site-packages/datasets/builder.py", line 269, in __init__
self.config, self.config_id = self._create_builder_config(
File "/net/pr2/scratch/people/plgapohl/python-3.8.6/lib/python3.8/site-packages/datasets/builder.py", line 403, in _create_builder_config
raise ValueError(f"BuilderConfig must have a name, got {builder_config.name}")
ValueError: BuilderConfig must have a name, got
```
## Environment info
- `datasets` version: 2.2.2
- Platform: Linux-4.18.0-348.20.1.el8_5.x86_64-x86_64-with-glibc2.2.5
- Python version: 3.8.6
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
The error is probably in line 795 in load.py:
```
builder_kwargs = {
"hash": hash,
"data_files": data_files,
"name": os.path.basename(self.path),
"base_path": self.path,
**builder_kwargs,
}
```
`os.path.basename` for a directory returns an empty string, rather than the name of the directory.
| 4,399 |
https://github.com/huggingface/datasets/issues/4398 | Calling `cast_column`/`remove_columns` and a sequence of `map` operations ends up making `faiss` fail with `ValueError` | [
"It works if we either remove the `ds = ds.cast_column(\"id\", Value(\"int32\"))` line from the code above, or if instead calling `ds.remove_columns()` we remove the columns inside each mapping as `ds.map(..., remove_columns=[...])` instead of right after the mapping.\r\n\r\nBoth of those solutions seem to fix the issue, so the root cause of it may be around that. Sorry I cannot provide you more insights, in case I get to fix it I'll submit a PR, in the meanwhile the code that I'm using as a workaround is the following:\r\n\r\n```python\r\nfrom transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\nimport torch\r\n\r\ntorch.set_grad_enabled(False)\r\nctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\nctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n\r\nfrom datasets import load_dataset, Value\r\n\r\nds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\nds = ds.cast_column(\"id\", Value(\"int32\"))\r\nds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])}, remove_columns=[\"title\", \"summary\"])\r\n\r\ndef generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n\r\nds = ds.map(generate_embeddings, remove_columns=[\"inputs\"])\r\nds.add_faiss_index(column=\"embeddings\")\r\n```",
"FYI the main reason I want to use `dataset.remove_columns` rather than the function inside `dataset.map` is because according to the π€ Datasets documentation, it's faster.\r\n\r\n\"π€ Datasets also has a [Dataset.remove_columns()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.Dataset.remove_columns) method that is functionally identical, but faster, because it doesnβt copy the data of the remaining columns.\"\r\n\r\nMore information at https://huggingface.co/docs/datasets/process#map",
"Here I'm presenting all the scenarios so that you can further investigate the issue:\r\n\r\n- β
`cast_column` -> `map` with `remove_columns` -> `map` with `remove_columns` -> `add_faiss_index`\r\n\r\n ```python\r\n from transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\n import torch\r\n \r\n torch.set_grad_enabled(False)\r\n ctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n \r\n from datasets import load_dataset, Value\r\n \r\n ds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\n ds = ds.cast_column(\"id\", Value(\"int32\"))\r\n ds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])}, remove_columns=[\"title\", \"summary\"])\r\n \r\n def generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n \r\n ds = ds.map(generate_embeddings, remove_columns=[\"inputs\"])\r\n ds.add_faiss_index(column=\"embeddings\")\r\n ```\r\n\r\n- β `cast_column` -> `map` -> `remove_columns` -> `map` -> `remove_columns` -> `add_faiss_index`\r\n\r\n ```python\r\n from transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\n import torch\r\n \r\n torch.set_grad_enabled(False)\r\n ctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n \r\n from datasets import load_dataset, Value\r\n \r\n ds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\n ds = ds.cast_column(\"id\", Value(\"int32\"))\r\n ds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])})\r\n ds = ds.remove_columns([\"title\", \"summary\"])\r\n \r\n def generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n \r\n ds = ds.map(generate_embeddings)\r\n ds = ds.remove_columns([\"inputs\"])\r\n ds.add_faiss_index(column=\"embeddings\")\r\n ```\r\n\r\n- β `cast_column` -> `map` with `remove_columns` -> `map` -> `remove_columns` -> `add_faiss_index`\r\n\r\n\r\n ```python\r\n from transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\n import torch\r\n \r\n torch.set_grad_enabled(False)\r\n ctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n \r\n from datasets import load_dataset, Value\r\n \r\n ds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\n ds = ds.cast_column(\"id\", Value(\"int32\"))\r\n ds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])}, remove_columns=[\"title\", \"summary\"])\r\n \r\n def generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n \r\n ds = ds.map(generate_embeddings)\r\n ds = ds.remove_columns([\"inputs\"])\r\n ds.add_faiss_index(column=\"embeddings\")\r\n ```\r\n\r\n- β
`cast_column` -> `map` -> `remove_columns` -> `map` with `remove_columns` -> `add_faiss_index`\r\n\r\n\r\n ```python\r\n from transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\n import torch\r\n \r\n torch.set_grad_enabled(False)\r\n ctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n \r\n from datasets import load_dataset, Value\r\n \r\n ds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\n ds = ds.cast_column(\"id\", Value(\"int32\"))\r\n ds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])})\r\n ds = ds.remove_columns([\"title\", \"summary\"])\r\n \r\n def generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n \r\n ds = ds.map(generate_embeddings, remove_columns=[\"inputs\"])\r\n ds.add_faiss_index(column=\"embeddings\")\r\n ```\r\n\r\n- β
`map` -> `remove_columns` -> `map` -> `remove_columns` -> `add_faiss_index`\r\n\r\n\r\n ```python\r\n from transformers import DPRContextEncoder, DPRContextEncoderTokenizer\r\n import torch\r\n \r\n torch.set_grad_enabled(False)\r\n ctx_encoder = DPRContextEncoder.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n \r\n from datasets import load_dataset, Value\r\n \r\n ds = load_dataset(\"csv\", data_files=[\"sample.csv\"], split=\"train\")\r\n ds = ds.map(lambda x: {\"inputs\": f\"{ctx_tokenizer.sep_token}\".join([\"title\", \"summary\"])})\r\n ds = ds.remove_columns([\"title\", \"summary\"])\r\n \r\n def generate_embeddings(x):\r\n return {\"embeddings\": ctx_encoder(**ctx_tokenizer(x[\"inputs\"], return_tensors=\"pt\"))[0][0].numpy()}\r\n \r\n ds = ds.map(generate_embeddings)\r\n ds = ds.remove_columns([\"inputs\"])\r\n ds.add_faiss_index(column=\"embeddings\")\r\n ```",
"So on, I've created #4411 so as to fix the bug with `remove_columns` under certain conditions before `add_faiss_index`, which means that the scenarios not working above are already working fine."
] | First of all, sorry in advance for the unclear title, but this bug is weird to explain (at least for me), so I tried my best to summarize all the information in this issue.
## Describe the bug
Calling a certain combination of operations over a π€ `Dataset` and then trying to calculate the `faiss` index with `.add_faiss_index` ends up throwing an exception while trying to set the format back of a previously removed column. But this just happens over certain conditions... I'll present some scenarios below!
## Steps to reproduce the bug
Assuming the following dataset named `sample.csv` with some IMDb data:
```csv
id,title,summary
1877830,"The Batman","When a sadistic serial killer begins murdering key political figures in Gotham, Batman is forced to investigate the city's hidden corruption and question his family's involvement."
9419884,"Doctor Strange in the Multiverse of Madness","Doctor Strange teams up with a mysterious teenage girl from his dreams who can travel across multiverses, to battle multiple threats, including other-universe versions of himself, which threaten to wipe out millions across the multiverse. They seek help from Wanda the Scarlet Witch, Wong and others."
11138512,"The Northman","From visionary director Robert Eggers comes The Northman, an action-filled epic that follows a young Viking prince on his quest to avenge his father's murder."
1745960,"Top Gun: Maverick","After more than thirty years of service as one of the Navy's top aviators, Pete Mitchell is where he belongs, pushing the envelope as a courageous test pilot and dodging the advancement in rank that would ground him."
```
We'll be able to reproduce the bug using the following piece of code:
```python
# Sample code to reproduce the bug
from transformers import DPRContextEncoder, DPRContextEncoderTokenizer
import torch
torch.set_grad_enabled(False)
ctx_encoder = DPRContextEncoder.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base")
ctx_tokenizer = DPRContextEncoderTokenizer.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base")
from datasets import load_dataset, Value
ds = load_dataset("csv", data_files=["sample.csv"], split="train")
ds = ds.cast_column("id", Value("int32")) # from `int64` to `int32`
ds = ds.map(lambda x: {"inputs": f"{ctx_tokenizer.sep_token}".join(["title", "summary"])})
ds = ds.remove_columns(["title", "summary"])
def generate_embeddings(x):
return {"embeddings": ctx_encoder(**ctx_tokenizer(x["inputs"], return_tensors="pt"))[0][0].numpy()}
ds = ds.map(generate_embeddings)
ds = ds.remove_columns("inputs")
ds.add_faiss_index(column="embeddings") # It fails here!
```
The code above is an adaptation of https://huggingface.co/docs/datasets/faiss_es, for the sake of presenting the bug with a simple example.
## Expected results
Ideally, the `faiss` index should be calculated over the π€ `Dataset` and no exception should be triggered.
## Actual results
But what happens instead is that a `ValueError: Columns ['inputs'] not in the dataset. Current columns in the dataset: ['id', 'embeddings']`, which makes no sense as that column has been previously dropped.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.2
- Platform: Linux-5.4.0-1074-azure-x86_64-with-glibc2.31
- Python version: 3.9.5
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,398 |
https://github.com/huggingface/datasets/issues/4394 | trainer became extremely slow after reload dataset by `load_from_disk` | [
"I tried to make the dataset much more smaller (100000 rows) , then the speed became `33.88it/s` from`8.62s/it`. It's nearly 200 times... Do you have any idea? Thank you!",
"Similar issue: https://github.com/huggingface/transformers/issues/8818\r\n\r\nI changed `RandomSampler` to `SequentialSampler` in the `trainer.py`, but the speed didn't become faster.",
"I changed\r\n```\r\ntokenized_datasets = load_from_disk(\r\n \"/pathto/dataset\"\r\n )\r\n```\r\nto\r\n```\r\ntokenized_datasets = load_from_disk(\r\n \"/pathto/dataset\", keep_in_memory=True\r\n )\r\n```\r\nand obtained normal speed. It's seems that the problem is on the os's speed limit.",
"Hi ! Currently `save_to_disk` saves one big Arrow file, which causes some slow downs. This has been discussed in #3735 and we'll implement sharding pretty soon to solve this\r\n\r\nFor now you can try splitting and saving your dataset in several Arrow files. Then you can load them one by one and use `concatenate_datasets` to have your big dataset again and hopefully with a better speed",
"Any update on fixing this? The issue still seems to be present."
] | ## Describe the bug
Due to memory problem, I need to save my tokenized datasets locally by CPU and reload it by multi GPU for running training script. However, after I reload it by `load_from_disk` and start training, the speed is extremely slow. It says I need about 1500 hours with 8 A100 cards. Before this, I can run the whole script in one day with a single A100 card.
Since I am try to pre-train a BERT, **my dataset is very large(29058165 rows)**
## Steps to reproduce the bug
```python
tokenized_datasets.save_to_disk(
"/pathto/dataset"
)
tokenized_datasets = load_from_disk(
"/pathto/dataset"
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval
else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
train_result = trainer.train(resume_from_checkpoint=checkpoint)
```
## Expected results
Without the save and reload process, I only need about one day to run the whole script with one A100 card.
## Actual results
```
[INFO|trainer.py:1290] 2022-05-23 22:49:46,266 >> ***** Running training *****
[INFO|trainer.py:1291] 2022-05-23 22:49:46,266 >> Num examples = 29058165
[INFO|trainer.py:1292] 2022-05-23 22:49:46,266 >> Num Epochs = 5
[INFO|trainer.py:1293] 2022-05-23 22:49:46,266 >> Instantaneous batch size per device = 16
[INFO|trainer.py:1294] 2022-05-23 22:49:46,266 >> Total train batch size (w. parallel, distributed & accumulation) = 256
[INFO|trainer.py:1295] 2022-05-23 22:49:46,266 >> Gradient Accumulation steps = 2
[INFO|trainer.py:1296] 2022-05-23 22:49:46,266 >> Total optimization steps = 567540
0%| | 1/567540 [00:09<1544:49:04, 9.80s/it]
0%| | 2/567540 [00:17<1320:00:17, 8.37s/it]
0%| | 3/567540 [00:26<1393:10:17, 8.84s/it]
0%| | 4/567540 [00:34<1344:56:33, 8.53s/it]
0%| | 5/567540 [00:43<1359:36:12, 8.62s/it]
```
## Environment info
```
torch 1.11.0+cu113
torchaudio 0.11.0+cu113
torchvision 0.12.0+cu113
transformers 4.18.0
datasets 2.2.2
``` | 4,394 |
https://github.com/huggingface/datasets/issues/4387 | device/google/accessory/adk2012 - Git at Google | [] | "git clone https://android.googlesource.com/device/google/accessory/adk2012"
https://android.googlesource.com/device/google/accessory/adk2012/#:~:text=git%20clone%20https%3A//android.googlesource.com/device/google/accessory/adk2012 | 4,387 |
https://github.com/huggingface/datasets/issues/4386 | Bug for wiki_auto_asset_turk from GEM | [
"Thanks for reporting, @StevenTang1998.\r\n\r\nI'm looking into it. ",
"Hi @StevenTang1998,\r\n\r\nWe have fixed the issue:\r\n- #4389\r\n\r\nThe fix will be available in our next `datasets` library release. In the meantime, you can incorporate that fix by installing `datasets` from our GitHub repo:\r\n```\r\npip install git+https://github.com/huggingface/datasets#egg=datasets\r\n```",
"Thanks for your reply!!\r\nAnd the totto dataset has the same problem. The url should be change to [https://storage.googleapis.com/totto-public/totto_data.zip](https://storage.googleapis.com/totto-public/totto_data.zip).",
"Hi again @StevenTang1998,\r\n\r\nI don't see any problem when loading `totto` dataset:\r\n```python\r\nIn [4]: import datasets\r\n ...: ds = datasets.load_dataset(\"totto\")\r\nDownloading builder script: 5.58kB [00:00, 5.33MB/s] \r\nDownloading metadata: 2.78kB [00:00, 2.96MB/s] \r\nUsing custom data configuration default\r\nDownloading and preparing dataset totto/default (download: 179.03 MiB, generated: 706.59 MiB, post-processed: Unknown size, total: 885.62 MiB) to .../.cache/huggingface/datasets/totto/default/1.0.0/263c85871e5451bc892c65ca0306c0629eb7beb161e0eb998f56231562335dd2...\r\nDownloading data: 100%|βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 188M/188M [00:32<00:00, 5.77MB/s]\r\nDataset totto downloaded and prepared to .../.cache/huggingface/datasets/totto/default/1.0.0/263c85871e5451bc892c65ca0306c0629eb7beb161e0eb998f56231562335dd2. Subsequent calls will reuse this data.\r\n100%|βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 3/3 [00:00<00:00, 147.95it/s]\r\n\r\nIn [5]: ds\r\nOut[5]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['id', 'table_page_title', 'table_webpage_url', 'table_section_title', 'table_section_text', 'table', 'highlighted_cells', 'example_id', 'sentence_annotations', 'overlap_subset'],\r\n num_rows: 120761\r\n })\r\n validation: Dataset({\r\n features: ['id', 'table_page_title', 'table_webpage_url', 'table_section_title', 'table_section_text', 'table', 'highlighted_cells', 'example_id', 'sentence_annotations', 'overlap_subset'],\r\n num_rows: 7700\r\n })\r\n test: Dataset({\r\n features: ['id', 'table_page_title', 'table_webpage_url', 'table_section_title', 'table_section_text', 'table', 'highlighted_cells', 'example_id', 'sentence_annotations', 'overlap_subset'],\r\n num_rows: 7700\r\n })\r\n})\r\n```",
"Sorry, I didn't express it clearly. It's the totto dataset from gem.\r\ndatasets.load_dataset('gem', 'totto')\r\n",
"@StevenTang1998 fixed in:\r\n- #4396",
"Thanks!!"
] | ## Describe the bug
The script of wiki_auto_asset_turk for GEM may be out of date.
## Steps to reproduce the bug
```python
import datasets
datasets.load_dataset('gem', 'wiki_auto_asset_turk')
```
## Actual results
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/load.py", line 1731, in load_dataset
builder_instance.download_and_prepare(
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/builder.py", line 640, in download_and_prepare
self._download_and_prepare(
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/builder.py", line 1158, in _download_and_prepare
super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/builder.py", line 707, in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
File "/home/tangtianyi/.cache/huggingface/modules/datasets_modules/datasets/gem/982a54473b12c6a6e40d4356e025fb7172a5bb2065e655e2c1af51f2b3cf4ca1/gem.py", line 538, in _split_generators
dl_dir = dl_manager.download_and_extract(_URLs[self.config.name])
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/download_manager.py", line 416, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/download_manager.py", line 294, in download
downloaded_path_or_paths = map_nested(
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/py_utils.py", line 351, in map_nested
mapped = [
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/py_utils.py", line 352, in <listcomp>
_single_map_nested((function, obj, types, None, True, None))
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/py_utils.py", line 288, in _single_map_nested
return function(data_struct)
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/download_manager.py", line 320, in _download
return cached_path(url_or_filename, download_config=download_config)
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/file_utils.py", line 234, in cached_path
output_path = get_from_cache(
File "/home/tangtianyi/miniconda3/lib/python3.8/site-packages/datasets/utils/file_utils.py", line 579, in get_from_cache
raise FileNotFoundError(f"Couldn't find file at {url}")
FileNotFoundError: Couldn't find file at https://github.com/facebookresearch/asset/raw/master/dataset/asset.test.orig
``` | 4,386 |
https://github.com/huggingface/datasets/issues/4383 | L | [] | ## Describe the L
L
## Expected L
A clear and concise lmll
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:
- Python version:
- PyArrow version: | 4,383 |
https://github.com/huggingface/datasets/issues/4382 | First time trying | [] | ## Adding a Dataset
- **Name:** *name of the dataset*
- **Description:** *short description of the dataset (or link to social media or blog post)*
- **Paper:** *link to the dataset paper if available*
- **Data:** *link to the Github repository or current dataset location*
- **Motivation:** *what are some good reasons to have this dataset*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 4,382 |
https://github.com/huggingface/datasets/issues/4381 | Bug in caching 2 datasets both with the same builder class name | [
"Hi @NouamaneTazi, thanks for reporting.\r\n\r\nPlease note that both datasets are cached in the same directory because their loading builder classes have the same name: `class MTOP(datasets.GeneratorBasedBuilder)`.\r\n\r\nYou should name their builder classes differently, e.g.:\r\n- `MtopDomain`\r\n- `MtopIntent`",
"Hi @NouamaneTazi, please note that after our fix:\r\n- #4388\r\n\r\nwe do not consider the class name anymore, but the name of the file where the loading builder class is implemented. "
] | ## Describe the bug
The two datasets `mteb/mtop_intent` and `mteb/mtop_domain `use both the same cache folder `.cache/huggingface/datasets/mteb___mtop`. So if you first load `mteb/mtop_intent` then datasets will not load `mteb/mtop_domain`.
If you delete this cache folder and flip the order how you load the two datasets , you will get the opposite datasets loaded (difference is here in terms of the label and label_text).
## Steps to reproduce the bug
```python
import datasets
dataset = datasets.load_dataset("mteb/mtop_intent", "en")
print(dataset['train'][0])
dataset = datasets.load_dataset("mteb/mtop_domain", "en")
print(dataset['train'][0])
```
## Expected results
```
Reusing dataset mtop (/home/nouamane/.cache/huggingface/datasets/mteb___mtop_intent/en/0.0.0/f930e32a294fed424f70263d8802390e350fff17862266e5fc156175c07d9c35)
100%|βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 3/3 [00:00<00:00, 920.14it/s]
{'id': 3232343436343136, 'text': 'Has Angelika Kratzer video messaged me?', 'label': 1, 'label_text': 'GET_MESSAGE'}
Reusing dataset mtop (/home/nouamane/.cache/huggingface/datasets/mteb___mtop_domain/en/0.0.0/f930e32a294fed424f70263d8802390e350fff17862266e5fc156175c07d9c35)
100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 3/3 [00:00<00:00, 1307.59it/s]
{'id': 3232343436343136, 'text': 'Has Angelika Kratzer video messaged me?', 'label': 0, 'label_text': 'messaging'}
```
## Actual results
```
Reusing dataset mtop (/home/nouamane/.cache/huggingface/datasets/mteb___mtop/en/0.0.0/f930e32a294fed424f70263d8802390e350fff17862266e5fc156175c07d9c35)
100%|βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 3/3 [00:00<00:00, 920.14it/s]
{'id': 3232343436343136, 'text': 'Has Angelika Kratzer video messaged me?', 'label': 1, 'label_text': 'GET_MESSAGE'}
Reusing dataset mtop (/home/nouamane/.cache/huggingface/datasets/mteb___mtop/en/0.0.0/f930e32a294fed424f70263d8802390e350fff17862266e5fc156175c07d9c35)
100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 3/3 [00:00<00:00, 1307.59it/s]
{'id': 3232343436343136, 'text': 'Has Angelika Kratzer video messaged me?', 'label': 1, 'label_text': 'GET_MESSAGE'}
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.1
- Platform: macOS-12.1-arm64-arm-64bit
- Python version: 3.9.12
- PyArrow version: 8.0.0
- Pandas version: 1.4.2
| 4,381 |
https://github.com/huggingface/datasets/issues/4379 | Latest dill release raises exception | [
"Fixed by:\r\n- #4380 ",
"Just an additional insight, the latest dill (either 0.3.5 or 0.3.5.1) also broke the hashing/fingerprinting of any mapping function.\r\n\r\nFor example:\r\n```\r\nfrom datasets import load_dataset\r\n\r\nd = load_dataset(\"rotten_tomatoes\")\r\nd.map(lambda x: x)\r\n```\r\n\r\nReturns the standard non-dillable error:\r\n```\r\nParameter 'function'=<function <lambda> at 0x7fe7d18c9560> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly....\r\n```",
"@albertvillanova ExamplesTests.test_run_speech_recognition_seq2seq is in which file?",
"Thanks a lot @gugarosa for the insight: we will incorporate it in our CI as regression testing for future dill releases.",
"Hi @anivegesana, that test is in `transformers` library:\r\n- https://github.com/huggingface/transformers/blob/main/examples/pytorch/test_pytorch_examples.py#L449\r\n- https://github.com/huggingface/transformers/blob/main/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py ",
"@albertvillanova\n\nI did a deep dive into @gugarosa's problem and found the issue and it might be related to the one @sgugger discovered. In dill 0.3.5(.1), I created a new `save_function` that fixes a bug in dill that prevented the pickling of recursive inner functions. It was a more complete solution to the problem that `dill._dill.stack` tried to solve in the internal API of dill. Since `dill._dill.stack` was no longer needed, I removed it. Since datasets copies the `save_function` directly from the dill API, it stops working with the new dill version since `dill._dill.stack` is no longer present and the `save_function` has been updated with new code.\r\n\r\nhttps://github.com/huggingface/datasets/blob/95193ae61e92aa537d0c65d37a1fd9d2393aae89/src/datasets/utils/py_utils.py#L607-L678\r\n\r\n~If the dill version is below 0.3.5, you should keep this function. If it is after, you would need to update your copy of `save_function` to use the code I introduced, or manually add a `stack` variable to `dill._dill` if it doesn't exist. Fortunately, in any version of Python 3.7+, dictionaries are always in insertion order and dill no longer supports Python 3.6 or older. So, any globals dictionary saved by dill 0.3.5+ will be deterministic given that the version of dill is held constant and this save_function is unnecessary for newer versions of dill.~\r\n\r\nAh. I see what is happening. I guess a different copy of the function code is needed that sorts the global variables by name.\r\n\r\n```py\r\nif dill.__version__.split('.') < ['0', '3', '5']:\r\n # current save_function code inside here\r\nelse:\r\n # new save_function code inside here with the following line inserted after creating the globals\r\n globs = {k: globs[k] for k in sorted(globs.keys())} \r\n```\r\n\r\nWill look into the test case @sgugger pointed out after that and verify if this is causing the problem.\r\n\r\nI am actually looking into rewriting the global variables code in uqfoundation/dill#466 and will keep this in mind and will try to create an easy way to modify the global variables in dill 0.3.6 (for example, sort them by key like datasets does).",
"Thanks a lot for your investigation @anivegesana.\r\n\r\nYes, we copied-pasted the old `save_function` function from `dill`, just adding a line to make deterministic the order of global variables `globs`. \r\n\r\nHowever, this function has changed a lot from version 0.3.5, after your PR (thank you for the fix in recursiveness, indeed):\r\n- uqfoundation/dill#443\r\n\r\nWe have to address this change.\r\n\r\nIf finally your PR to sort global variables is merged into dill 0.3.6, that will make our life easier, as the tweak will no longer be necessary. ;)\r\n\r\nI have included a regression test so that we are sure future releases of dill do not break `datasets`:\r\n- #4385 ",
"I should note that because Python 3.6 and older are now deprecated and Python 3.7 has insertion order dictionaries, the globals in dill will have a deterministic order, just not sorted. I would still keep it sorted like you have it to help with stability (for example, if someone reorders variables in a file, then sorting the globals would not invalidate the cache.)\n\nIt seems that the order is not quite deterministic in IPython. Huggingface datasets seems to do well in Jupyter regardless, so it is not a good idea to remove the sorting. uqfoundation/dill#19"
] | ## Describe the bug
As reported by @sgugger, latest dill release is breaking things with Datasets.
```
______________ ExamplesTests.test_run_speech_recognition_seq2seq _______________
self = <multiprocess.pool.ApplyResult object at 0x7fa5981a1cd0>, timeout = None
def get(self, timeout=None):
self.wait(timeout)
if not self.ready():
raise TimeoutError
if self._success:
return self._value
else:
> raise self._value
E TypeError: '>' not supported between instances of 'NoneType' and 'float'
```
| 4,379 |
https://github.com/huggingface/datasets/issues/4376 | irc_disentagle viewer error | [
"DUPLICATED comment from https://github.com/huggingface/datasets/issues/3807:\r\n\r\nmy code:\r\n```\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"irc_disentangle\", download_mode=\"force_redownload\")\r\n```\r\nhowever, it produces the same error\r\n```\r\n[38](file:///Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/datasets/utils/info_utils.py?line=37) if len(bad_urls) > 0:\r\n [39](file:///Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/datasets/utils/info_utils.py?line=38) error_msg = \"Checksums didn't match\" + for_verification_name + \":\\n\"\r\n---> [40](file:///Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/datasets/utils/info_utils.py?line=39) raise NonMatchingChecksumError(error_msg + str(bad_urls))\r\n [41](file:///Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/site-packages/datasets/utils/info_utils.py?line=40) logger.info(\"All the checksums matched successfully\" + for_verification_name)\r\n\r\nNonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['https://github.com/jkkummerfeld/irc-disentanglement/tarball/master']\r\n```\r\nI attempted to use the `ignore_verifications' as such:\r\n\r\n```\r\nds = datasets.load_dataset('irc_disentangle', download_mode=\"force_redownload\", ignore_verifications=True)\r\n\r\nDownloading builder script: 12.0kB [00:00, 5.92MB/s] \r\nDownloading metadata: 7.58kB [00:00, 3.48MB/s] \r\nNo config specified, defaulting to: irc_disentangle/ubuntu\r\nDownloading and preparing dataset irc_disentangle/ubuntu (download: 112.98 MiB, generated: 60.05 MiB, post-processed: Unknown size, total: 173.03 MiB) to /Users/laylabouzoubaa/.cache/huggingface/datasets/irc_disentangle/ubuntu/1.0.0/0f24ab262a21d8c1d989fa53ed20caa928f5880be26c162bfbc02445dbade7e5...\r\nDownloading data: 118MB [00:09, 12.1MB/s] \r\n \r\nDataset irc_disentangle downloaded and prepared to /Users/laylabouzoubaa/.cache/huggingface/datasets/irc_disentangle/ubuntu/1.0.0/0f24ab262a21d8c1d989fa53ed20caa928f5880be26c162bfbc02445dbade7e5. Subsequent calls will reuse this data.\r\n100%|ββββββββββ| 3/3 [00:00<00:00, 675.38it/s]\r\n```\r\nbut, this returns an empty set?\r\n```\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['id', 'raw', 'ascii', 'tokenized', 'date', 'connections'],\r\n num_rows: 0\r\n })\r\n test: Dataset({\r\n features: ['id', 'raw', 'ascii', 'tokenized', 'date', 'connections'],\r\n num_rows: 0\r\n })\r\n validation: Dataset({\r\n features: ['id', 'raw', 'ascii', 'tokenized', 'date', 'connections'],\r\n num_rows: 0\r\n })\r\n})\r\n```\r\nnot sure what else to try at this point?\r\nThanks in advancedπ€",
"Thanks for reporting, @labouz. I'm addressing it. ",
"The issue with checksum and empty dataset has been fixed by:\r\n- #4377\r\n\r\nTo load the dataset, you should force the re-generation of the dataset from the downloaded file by passing `download_mode=\"reuse_cache_if_exists\"` to `load_dataset`.\r\n\r\nIn relation with the issue with the dataset viewer, first the dataset should be refactored to support streaming.",
"parfait!\r\nit works now, thank you π ",
"Hi there, \r\nI see this issue is closed, but I am wondering if there is any chance the source files have been moved since this fix? I am stumbling into the same NonMatchingChecksumError noted by lebouz's second post once 118MB of data has been downloaded, and have tried the solutions noted in the various fix checksum posts linked here and in other posts regarding passing in \"reuse_cache_if_exists\" to download_mode. Any suggestions? Thank you!\r\n\r\n"
] | the dataviewer shows this message for "ubuntu" - "train", "test", and "validation" splits:
```
Server error
Status code: 400
Exception: ValueError
Message: Cannot seek streaming HTTP file
```
it appears to give the same message for the "channel_two" data as well.
I get a Checksums error when using `load_data()` with this dataset. Even with the `download_mode` and `ignore_verifications` options set. i referenced the issue here: https://github.com/huggingface/datasets/issues/3807 | 4,376 |
https://github.com/huggingface/datasets/issues/4374 | extremely slow processing when using a custom dataset | [
"Hi !\r\n\r\nMy guess is that some examples in your dataset are bigger than your RAM, and therefore loading them in RAM to pass them to `remove_non_indic_sentences` takes forever because it might use SWAP memory.\r\n\r\nMaybe several examples in your dataset are grouped together, can you check `len(lang_dataset[\"train\"])` and `lang_dataset[\"train\"].data.nbytes` of both datasets please ? It can also be helpful to check the distribution of lengths of each examples in your dataset.",
"Closing due to inactivity"
] | ## processing a custom dataset loaded as .txt file is extremely slow, compared to a dataset of similar volume from the hub
I have a large .txt file of 22 GB which i load into HF dataset
`lang_dataset = datasets.load_dataset("text", data_files="hi.txt")`
further i use a pre-processing function to clean the dataset
`lang_dataset["train"] = lang_dataset["train"].map(
remove_non_indic_sentences, num_proc=12, batched=True, remove_columns=lang_dataset['train'].column_names), batch_size=64)`
the following processing takes astronomical time to process, while hoging all the ram.
similar dataset of same size that's available in the huggingface hub works completely fine. which runs the same processing function and has the same amount of data.
`lang_dataset = datasets.load_dataset("oscar-corpus/OSCAR-2109", "hi", use_auth_token=True)`
the hours predicted to preprocess are as follows:
huggingface hub dataset: 6.5 hrs
custom loaded dataset: 7000 hrs
note: both the datasets are almost actually same, just provided by different sources with has +/- some samples, only one is hosted on the HF hub and the other is downloaded in a text format.
## Steps to reproduce the bug
```
import datasets
import psutil
import sys
import glob
from fastcore.utils import listify
import re
import gc
def remove_non_indic_sentences(example):
tmp_ls = []
eng_regex = r'[. a-zA-Z0-9ΓΓΓ
ΓΆΓ€Γ₯ _.,!"\'\/$]*'
for e in listify(example['text']):
matches = re.findall(eng_regex, e)
for match in (str(match).strip() for match in matches if match not in [""," ", " ", ",", " ,", ", ", " , "]):
if len(list(match.split(" "))) > 2:
e = re.sub(match," ",e,count=1)
tmp_ls.append(e)
gc.collect()
example['clean_text'] = tmp_ls
return example
lang_dataset = datasets.load_dataset("text", data_files="hi.txt")
lang_dataset["train"] = lang_dataset["train"].map(
remove_non_indic_sentences, num_proc=12, batched=True, remove_columns=lang_dataset['train'].column_names), batch_size=64)
## same thing work much faster when loading similar dataset from hub
lang_dataset = datasets.load_dataset("oscar-corpus/OSCAR-2109", "hi", split="train", use_auth_token=True)
lang_dataset["train"] = lang_dataset["train"].map(
remove_non_indic_sentences, num_proc=12, batched=True, remove_columns=lang_dataset['train'].column_names), batch_size=64)
```
## Actual results
similar dataset of same size that's available in the huggingface hub works completely fine. which runs the same processing function and has the same amount of data.
`lang_dataset = datasets.load_dataset("oscar-corpus/OSCAR-2109", "hi", use_auth_token=True)
**the hours predicted to preprocess are as follows:**
huggingface hub dataset: 6.5 hrs
custom loaded dataset: 7000 hrs
**i even tried the following:**
- sharding the large 22gb text files into smaller files and loading
- saving the file to disk and then loading
- using lesser num_proc
- using smaller batch size
- processing without batches ie : without `batched=True`
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 2.2.2.dev0
- Platform: Ubuntu 20.04 LTS
- Python version: 3.9.7
- PyArrow version:8.0.0
| 4,374 |
https://github.com/huggingface/datasets/issues/4366 | TypeError: __init__() missing 1 required positional argument: 'scheme' | [
"Duplicate of:\r\n- #3956\r\n\r\nI think you should report that issue to `elasticsearch` library: https://github.com/elastic/elasticsearch-py"
] | "name" : "node-1",
"cluster_name" : "elasticsearch",
"cluster_uuid" : "",
"version" : {
"number" : "7.5.0",
"build_flavor" : "default",
"build_type" : "tar",
"build_hash" : "",
"build_date" : "2019-11-26T01:06:52.518245Z",
"build_snapshot" : false,
"lucene_version" : "8.3.0",
"minimum_wire_compatibility_version" : "6.8.0",
"minimum_index_compatibility_version" : "6.0.0-beta1"
when I run the order:
nohup python3 custom_service.pyc > service.log 2>&1&
the log:
nohup: εΏ½η₯θΎε
₯
Traceback (most recent call last):
File "/home/xfz/p3_custom_test/custom_service.py", line 55, in <module>
File "/home/xfz/p3_custom_test/custom_service.py", line 48, in doInitialize
File "custom_impl.py", line 286, in custom_setup
File "custom_impl.py", line 127, in create_es_index
File "/usr/local/lib/python3.7/site-packages/elasticsearch/_sync/client/__init__.py", line 345, in __init__
ssl_show_warn=ssl_show_warn,
File "/usr/local/lib/python3.7/site-packages/elasticsearch/_sync/client/utils.py", line 105, in client_node_configs
node_configs = hosts_to_node_configs(hosts)
File "/usr/local/lib/python3.7/site-packages/elasticsearch/_sync/client/utils.py", line 154, in hosts_to_node_configs
node_configs.append(host_mapping_to_node_config(host))
File "/usr/local/lib/python3.7/site-packages/elasticsearch/_sync/client/utils.py", line 221, in host_mapping_to_node_config
return NodeConfig(**options) # type: ignore
TypeError: __init__() missing 1 required positional argument: 'scheme'
[1]+ ιεΊ 1 nohup python3 custom_service.pyc > service.log 2>&1
custom_service_pyc can't running
| 4,366 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.