Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "Hi ! This would be a nice addition indeed :) This sounds like a duplicate of https://github.com/huggingface/datasets/issues/5468\r\n\r\n> Not sure. Some of my PRs are still open and some do not have any discussions.\r\n\r\nSorry to hear that, feel free to ping me on those PRs" ]

[ "Hi! I agree it would be nice to have a method like that. Instead of `keep_columns`, we can name it `select_columns` to be more aligned with PyArrow's naming convention (`pa.Table.select`).", "Hi, I am a newbie to open source and would like to contribute. @mariosasko can I take up this issue ?", "Hey, I also want to work on this issue I am a newbie to open source. ", "This sounds related to https://github.com/huggingface/datasets/issues/5474\r\n\r\nI'm fine with `select_columns`, or we could also override `select` to also accept a list of columns maybe ?", "@lhoestq, I am planning to add a member function to the dataset class to perform the selection operation. Do you think its the right way to proceed? or there is a better option ?", "Unless @mariosasko thinks otherwise, I think it can go in `Dataset.select()` :)\r\nThough some parameters like keep_in_memory, indices_cache_file_name or writer_batch_size wouldn't when selecting columns, so we would need to update the docstring as well", "If someone wants to give it a shot, feel free to comment `#self-assign` and it will assign the issue to you.\r\n\r\nFeel free to ping us here if you have questions or if we can help :)", "I would rather have this functionality as a separate method. IMO it's always better to be explicit than to have an API where a single method can do different/uncorrelated things (somewhat reminds me of Pandas, and there is probably a good reason why PyArrow is more rigid in this aspect).", "In the end I also think it would be nice to have it as a separate method, this way we can also have it for `IterableDataset` (which can't have `select` for indices)" ]

[]

[ "Thanks for reporting, @sarahwie.\r\n\r\nPlease note this issue was already fixed in `datasets` 2.6.0 version:\r\n- #5040\r\n\r\nIf you update your `datasets` version, you will be able to load the dataset:\r\n```\r\npip install -U datasets\r\n```", "Oops, missed that I needed to upgrade. Thanks!" ]

[ "Ccing @dwofk (the author of `fast-depth`). \r\n\r\nThanks, @awsaf49 for reporting this. I believe this is because the NYU Depth V2 shipped from `fast-depth` is already preprocessed. \r\n\r\nIf you think it might be better to have the NYU Depth V2 dataset from BTS [here](https://huggingface.co/datasets/sayakpaul/nyu_depth_v2) feel free to open a PR, I am happy to provide guidance :) ", "Good catch ! Ideally it would be nice to have the datasets in the raw form, this way users can choose whatever processing they want to apply", "> Ccing @dwofk (the author of `fast-depth`).\r\n> \r\n> Thanks, @awsaf49 for reporting this. I believe this is because the NYU Depth V2 shipped from `fast-depth` is already preprocessed.\r\n> \r\n> If you think it might be better to have the NYU Depth V2 dataset from BTS [here](https://huggingface.co/datasets/sayakpaul/nyu_depth_v2) feel free to open a PR, I am happy to provide guidance :)\r\n\r\n@sayakpaul I would love to create a PR on this. As this will be my first PR here, some guidance would be helpful.\r\n\r\nNeed a bit of advice on the dataset, there are three publicly available datasets. Which one should I consider for PR?\r\n1. [BTS](https://github.com/cleinc/bts): Containst train/test: 36K/654 data, dtype = `uint16` hence more precise\r\n2. [DenseDepth](https://github.com/ialhashim/DenseDepth) It contains train/test: 50K/654 data, dtype = `uint8` hence less precise\r\n3. [Official](https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html#raw_parts): Size is big 400GB+, requires **MatLab** code for fixing **projection** and **sync**, DataType: `pgm` and `dump` hence can't be used directly.\r\n\r\ncc: @lhoestq\r\n\r\n", "I think BTS. Repositories like https://github.com/vinvino02/GLPDepth usually use BTS. Also, just for clarity, the PR will be to https://huggingface.co/datasets/sayakpaul/nyu_depth_v2. Once we have worked it out, we can update the following things:\r\n\r\n* https://github.com/huggingface/blog/pull/718\r\n* https://huggingface.co/docs/datasets/main/en/depth_estimation\r\n\r\nDon't worry about it if it seems overwhelming. We will work it out together :) \r\n\r\n@lhoestq what do you think? ", "@sayakpaul If I get this right I have to,\r\n1. Create a PR on https://huggingface.co/datasets/sayakpaul/nyu_depth_v2\r\n2. Create a PR on https://github.com/huggingface/blog\r\n3. Create a PR on https://github.com/huggingface/datasets to update https://github.com/huggingface/datasets/blob/main/docs/source/depth_estimation.mdx", "The last two are low-hanging fruits. Don't worry about them. ", "Yup opening a PR to use BTS on https://huggingface.co/datasets/sayakpaul/nyu_depth_v2 sounds good :) Thanks for the help !", "Finally, I have found the origin of the **discretized depth map**. When I first loaded the datasets from HF I noticed it was 30GB but in DenseDepth data is only 4GB with dtype=uint8. This means data from fast-depth (before loading to HF) must have high precision. So when I tried to dig deeper by directly loading depth_map from `h5py`, I found depth_map from `h5py` came with `float32`. But when the data is processed in HF with `datasets.Image()` it was directly converted to `uint8` from `float32` hence the **discretized** depth map.\r\nhttps://github.com/huggingface/datasets/blob/c78559cacbb0ca6e0bc8bfc313cc0359f8c23ead/src/datasets/features/image.py#L91-L93\r\n\r\n## Solutions:\r\n\r\n#### 1. Array2D\r\nUse `Array2D` feature with `float32` for depth_map \r\n\r\n* Code:\r\n```py\r\nFeatures({'depth_map': Array2D(shape=(480, 640), dtype='float32')})\r\n```\r\n* Pros:\r\nNo precision loss.\r\n\r\n* Cons:\r\nAs depth_map is saved as Array I think it can't be visuzlied in [hf.co/dataset](https://huggingface.co/datasets/sayakpaul/nyu_depth_v2) page like segmentation mask.\r\n\r\n#### 2. Uint16\r\nUse `uint16` as dtype for Image in `_h5_loader` for saving depth maps and accept `uint16` dtype in `datasets.Image()` feature.\r\n\r\n* Code\r\n```py\r\ndepth = np.array(h5f[\"depth\"])\r\ndepth /= 10.0 # [0, max_depth] -> [0, 1]\r\ndepth *= (2**16 -1) # transform from [0, 1] -> [0, 2^16 - 1]\r\ndepth = depth.astype('uint16')\r\n```\r\n* Pros:\r\n * We can visualize depth map in hf.co/datasets page like segmentation mask.\r\n * No need for post-processing.\r\n\r\n* Cons:\r\n * We need to make two change\r\n * Modify `_h5_loader` in https://huggingface.co/datasets/sayakpaul/nyu_depth_v2 to convert depth_map from `float32` to `uint16`.\r\n * Make sure `datasets.Image()` converts `np.ndarray` to `uint16` checking max value\r\n * Precision loss due to `float32` to `uint16`\r\n * Post-processing required for depth_map to transform from `[0, 2^16 - 1]` to `[0, max_depth]` before feeding them to model.", "Thanks so much for digging into this. \r\n\r\nSince the second solution entails changes to core datatypes in `datasets`, I think it's better to go with the first solution. \r\n\r\n@lhoestq WDYT?", "@sayakpaul Yes, Solution 1 requires minimal change and provides no precision loss. But I think support for `uint16` image would be a great addition as many datasets come with `uint16` image. For example [UW-Madison GI Tract Image Segmentation](https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation) dataset, here the image itself comes with `uint16` dtype rather than mask. So, saving `uint16` image with `uint8` will result in precision loss.\r\n\r\nPerhaps we can adapt solution 1 for this issue and Add support for `uint16` image separately?", "Using Array2D makes it not practical to use to train a model - in `transformers` we expect an image type.\r\n\r\nThere is a pull request to support more precision than uint8 in Image() here: https://github.com/huggingface/datasets/pull/5365/files\r\n\r\nwe can probably merge it today and do a release right away", "Fantastic, @lhoestq! \r\n\r\n@awsaf49 then let's wait for the PR to get merged and then take the next steps? ", "Sure", "The PR adds support for uint16 which is ok for BTS if I understand correctly, would it be ok for you ?", "If the main issue with the current version of NYU we have on the Hub is related to the precision loss stemming from `Image()`, I'd prefer if `Image()` supported float32 as well. ", "I also prefer `float32` as it offers more precision. But I'm not sure if we'll be able to visualize image with `float32` precision.", "We could have a separate loading for the float32 one using Array2D, but I feel like it's less convenient to use due to the amount of disk space and because it's not an Image() type. That's why I think uint16 is a better solution for users", "A bit confused here, If https://github.com/huggingface/datasets/pull/5365 gets merged won't this issue will be resolved automatically?", "Yes in theory :)", "actually float32 also seems to work in this PR (it just doesn't work for multi-channel)", "In that case, a new PR isn't necessary, right?", "Yep. I just tested from the PR and it works:\r\n```python\r\n>>> train_dataset = load_dataset(\"sayakpaul/nyu_depth_v2\", split=\"train\", streaming=True) \r\nDownloading readme: 100%|██████████████████| 8.71k/8.71k [00:00<00:00, 3.60MB/s]\r\n>>> next(iter(train_dataset))\r\n{'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=640x480 at 0x1382ED7F0>,\r\n 'depth_map': <PIL.TiffImagePlugin.TiffImageFile image mode=F size=640x480 at 0x1382EDF28>}\r\n>>> x = next(iter(train_dataset))\r\n>>> np.asarray(x[\"depth_map\"]) \r\narray([[0. , 0. , 0. , ..., 0. , 0. ,\r\n 0. ],\r\n [0. , 0. , 0. , ..., 0. , 0. ,\r\n 0. ],\r\n [0. , 0. , 0. , ..., 0. , 0. ,\r\n 0. ],\r\n ...,\r\n [0. , 2.2861192, 2.2861192, ..., 2.234162 , 2.234162 ,\r\n 0. ],\r\n [0. , 2.2861192, 2.2861192, ..., 2.234162 , 2.234162 ,\r\n 0. ],\r\n [0. , 2.2861192, 2.2861192, ..., 2.234162 , 2.234162 ,\r\n 0. ]], dtype=float32)\r\n```", "Great! the case is closed! This issue has been solved and I have to say, it was quite the thrill ride. I felt like Sherlock Holmes, solving a mystery and finding the bug🕵️‍♂️. But in all seriousness, it was a pleasure working on this issue and I'm glad we could get to the bottom of it.\r\n\r\nOn another note, should I consider closing the issue? I think we still need to make updates on https://github.com/huggingface/blog and https://github.com/huggingface/datasets/blob/main/docs/source/depth_estimation.mdx", "Haha thanks Mr Holmes :p\r\n\r\nmaybe let's close this issue when we're done updating the blog post and the documentation", "@awsaf49 thank you for your hard work! \r\n\r\nI am a little unsure why the other links need to be updated, though. They all rely on datasets internally. ", "I think depth_map still shows discretized version. It would be nice to have corrected one.\r\n<img src=\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/datasets/depth_est_target_viz.png\" width = 300>", "Also, I think we need to make some changes in the code to visualize depth_map as it is `float32` . `plot.imshow()` supports either [0, 1] + float32 or [0. 255] + uint8", "Oh yes! Do you want to start with the fixes? Please feel free to say no but I wanted to make sure your contributions are reflected properly in our doc and the blog :)", "Yes I think that would be nice :)", "I'll make the changes tomorrow. I hope it's okay..." ]

[ "Hi! Yes, that's correct. When `streaming` is `True`, only split names can be specified as `split`, and for slicing, you have to use `.skip`/`.take` instead.\r\n\r\nE.g. \r\n`load_dataset(\"lhoestq/demo1\",revision=None, streaming=True, split=\"train[:3]\")`\r\n\r\nrewritten with `.skip`/`.take`:\r\n`load_dataset(\"lhoestq/demo1\",revision=None, streaming=True, split=\"train\").take(3)`\r\n\r\n\r\n", "Thank you for your quick response!" ]

[ "Hi! \r\n\r\nThis issue is due to our audio/image datasets not being self-contained. This allows us to save disk space (files are written only once) but also leads to the issues like this one. We plan to make all our datasets self-contained in Datasets 3.0.\r\n\r\nIn the meantime, you can run the following map to ensure your dataset is self-contained:\r\n```python\r\nfrom datasets.table import embed_table_storage\r\n# load_dataset ...\r\ndset = dset.with_format(\"arrow\")\r\ndset.map(embed_table_storage, batched=True)\r\ndset = dset.with_format(\"python\")\r\n```\r\n", "Understood. Thank you, Mario.\r\n\r\nPerhaps the solution could be very simple - move the extracted files into the directory of the cached dataset? Which would make it self-contained already and won't require waiting for a new major release. Unless I'm missing some back-compat nuance.\r\n\r\nBut regardless if X relies on Y - it could check if Y is still there when loading X. so not checking full consistency but just the top-level directory it relies on." ]

[ "Something that'd be nice to have is \"manual update of state\". One of the learning from training LLMs is the ability to skip some batches whenever we notice huge spike might be handy.", "Your outline spec is very sound and clear, @lhoestq - thank you!\r\n\r\n@thomasw21, indeed that would be a wonderful extra feature. In Megatron-Deepspeed we manually drained the dataloader for the range we wanted. I wasn't very satisfied with the way we did it, since its behavior would change if you were to do multiple range skips. I think it should remember all the ranges it skipped and not just skip the last range - since otherwise the data is inconsistent (but we probably should discuss this in a separate issue not to derail this much bigger one).", "Hi there! I think this is a critical issue and have an urgent need for it, in my attempt to train on a super large-scale dataset using `datasets`. It is impossible to resume a time-consuming (like one month) experiment by iterating all seen data again, which could possibly cost several days.\r\n\r\n@stas00 @thomasw21 @lhoestq Any updates on this problem after 1 year passed?", "any update？", "No update so far, I wonder if someone implemented a resumable pytorch Sampler somwhere.\r\n\r\nThen regarding resuming a streaming dataset, we'd first like to have an efficient way to skip shards automatically but this is not implemented yet", "I opened a draft here for IterableDataset: https://github.com/huggingface/datasets/pull/6658\r\n\r\n\r\n\r\n```python\r\n\"\"\"Requires https://github.com/huggingface/datasets/pull/6658 (WIP)\"\"\"\r\nfrom datasets import load_dataset\r\nfrom torch.utils.data import DataLoader\r\n\r\nds = load_dataset(..., streaming=True)\r\n# ds = ds.map(tokenize)\r\n# ds = ds.shuffle(seed=42, buffer_size=1000)\r\n\r\n# Init the dataset state_dict, or load it from a checkpoint\r\ndataset_state_dict = ds.state_dict()\r\n\r\n# Resumable training loop\r\nds.load_state_dict(dataset_state_dict)\r\ndataloader = DataLoader(ds, batch_size=batch_size)\r\nfor step, batch in enumerate(dataloader):\r\n ...\r\n if step % save_steps == 0:\r\n dataset_state_dict = ds.state_dict()\r\n```", "Hi @lhoestq - can you provide more information and how to implement on saving and restoring vanilla DataLoader states with map-style datasets?\r\n\r\n", "For now the easiest is probably to use the vanilla DataLoader only for batching and multiprocessing, and implement the resuming logic using a `Dataset` (it has `.select()` to skip examples) and a `dataset_state_dict`:\r\n\r\n\r\n```python\r\nfrom datasets import load_dataset\r\nfrom torch.utils.data import DataLoader\r\n\r\nds = load_dataset(...)\r\n# ds = ds.map(tokenize)\r\n# ds = ds.shuffle(seed=42)\r\n\r\n# Init the dataset state_dict, or load it from a checkpoint\r\ndataset_state_dict = {\"step\": 0} \r\n\r\n# Resumable training loop\r\nstart_step = dataset_state_dict[\"step\"]\r\ndataloader = DataLoader(ds.select(range(start_step * batch_size, len(ds))), batch_size=batch_size)\r\nfor step, batch in enumerate(dataloader, start=start_step):\r\n ...\r\n if step % save_steps == 0:\r\n dataset_state_dict = {\"step\": step}\r\n```", "Hello, I found a similar implementation online that seems to solve your problem. https://github.com/facebookresearch/vissl/blob/main/vissl/data/data_helper.py#L93\r\nit looks like we can set_start_iter in StatefulDistributedSampler to implement the stateful resume requirement we want.\r\n\r\n", "Hi y'all, @lhoestq I wanted to flag that we currently have a StatefulDataLoader in `pytorch/data/torchdata` that has state_dict/load_state_dict methods, which will call a dataset's state_dict/load_state_dict methods but also handle multiprocessing under the hood. Any chance we can collaborate on this and try to get them to work well together? Please have a look here for some basic examples: https://github.com/pytorch/data/tree/main/torchdata/stateful_dataloader#saving-and-loading-state ", "Fantastic ! This will help pushing our IterableDataset state_dict implementation at https://github.com/huggingface/datasets/pull/6658 :) I'll check if there is anything missing to maker them work together, and add tests and some docs referring to the StatefulDataLoader :)", "Ah I just saw this disclaimer in the torchdata README and it feels like people should not rely on it. Should the StatefulDataLoader live elsewhere @andrewkho ?\r\n\r\n> ⚠️ As of July 2023, we have paused active development on TorchData and have paused new releases. We have learnt a lot from building it and hearing from users, but also believe we need to re-evaluate the technical design and approach given how much the industry has changed since we began the project. During the rest of 2023 we will be re-evaluating our plans in this space. Please reach out if you suggestions or comments (please use https://github.com/pytorch/data/issues/1196 for feedback).", "@lhoestq Good find, we are in the midst of updating this disclaimer as we're re-starting development and regular releases, though our approach will be to iterate on DL V1 (ie StatefulDataLoader) instead of continuing development on datapipes+DLV2. Let's discuss on a call at some point to figure out the best path forward! ", "As a heads up, `IterableDataset` state_dict has been added in https://github.com/huggingface/datasets/pull/6658\r\n\r\n...and it works out of the box with the `torchdata` `StatefulDataLoader` :)\r\n\r\nSee the docs at https://huggingface.co/docs/datasets/main/en/use_with_pytorch#checkpoint-and-resume", "amazing! Thank you, @lhoestq \r\n\r\ndoes it work with non-iterable dataset as well? the docs only mention iterable dataset", "It's for iterable dataset only. For regular dataset I believe the sampler should implement state_dict, but maybe @andrewkho might know best how to resume a regular dataset with torchdata", "@stas00 stateful dataloader will save and resume samplers for map style datasets. If no state_dict/load_state_dict is provided by the sampler, it will naively skip samples to fast forward. See here for more details https://github.com/pytorch/data/blob/main/torchdata/stateful_dataloader/README.md \n\nHope this helps! ", "Thank you very much for clarifying that, Andrew.\r\n\r\n" ]

[ "Hi ! Could you share the full stack trace ? Which dataset did you try to load ?\r\n\r\nit may be related to https://github.com/huggingface/datasets/pull/5640", "The `BadZipFile` error means the ZIP file is corrupted, so I'm closing this issue as it's not directly related to `datasets`.", "For others that find this issue following a `BadZipFile` error, I had the same problem because I had a file in a folder dataset `my-image.target` and the datasets library was incorrectly determining that the (PNG) file was a zip archive. When it tried to extract the file, this error occurred. \r\n\r\nUpdating to `datasets==2.12.0` fixed the problem for me." ]

[ "wtf", "Couldn't find what's causing this, this will need more investigation", "A possible hint: The function it seems to be spending a lot of time in (when iterating over the original dataset) is `_get_mp` in the PIL JPEG decoder: \r\n\r\n", "If \"mp\" is multiprocessing, this might suggest some kind of negative interaction between the JPEG decoder and TF's handling of processes/threads. Note that we haven't merged the parallel `to_tf_dataset` PR yet, so it's not caused by that PR!", "Update: MP isn't multiprocessing at all, it's an internal PIL method for loading metadata from JPEG files. No idea why that would be a bottleneck, but I'll see if a Python profiler can't figure out where the time is actually being spent.", "After further profiling, the slowdown is in the C methods for JPEG decoding that are included as part of PIL. Because Python profilers can't inspect inside that, I don't have any further information on which lines exactly are responsible for the slowdown or why.\r\n\r\nIn the meantime, I'm going to suggest switching from `return_tensors=\"tf\"` to `return_tensors=\"np\"` in most of our `transformers` code - this generally works better for pre-processing. Two relevant PRs are [here](https://github.com/huggingface/transformers/pull/21266) and [here](https://github.com/huggingface/notebooks/pull/308).", "Closing this issue as we've done what we can with this one! " ]

[]

[]

[ "Looks like a duplicate of https://github.com/huggingface/datasets/issues/1948. \r\n\r\nI also think these should be logged as INFO messages, but let's see what @lhoestq thinks.", "It can be considered unexpected to see a `map` function return instantaneously. The warning is here to explain this case by mentioning that the cache was used. I don't expect first time users (only seeing warnings) to guess that the cache works this way", "Oh, so it's intentional? Do all Hugging Face packages use `warning` when using cache?\r\nI guess feel free to close this issue then.", "Yes it's intentional for `map`. For `load_dataset` it's also intentional but for a different reason: it shows where in the cache the dataset is located, in case the user wants to clear the cache.", "OK I see. It's surprising to me that these are considered \"something unexpected happened\", the concept of cache is pretty common.\r\n\r\nHas a user every actually complained that they ran their code once, and it took a minute while the data downloaded, then ran their code again and it ran really fast (and completed successfully) but they were so baffled by the fact that it ran quickly, _and_ didn't set the log level to INFO, _and_ hadn't read the docs (or thought about it) to know that datasets are cached, that they logged an issue asking that this information be output as a warning every time they run their code?\r\n\r\nThat seems like a very niche scenario to cater for, given that the side effect is to flood the console with irrelevant warnings for every other user every other time they run a bit of `datasets` code. And the real world impact is that people TURN OFF warnings, which is a pretty bad habit to get into.\r\n\r\nAnyhoo, if there's no chance I'm going to change your mind, please close the issue :)", "I see your point and I'm not closed to switching to INFO, but I think those logs are important to make the library less opaque. I also just checked `transformers` scripts and they default to INFO which is nice. However for colab users the default is still WARNING iirc, and it counts as one of the main env where `datasets` is used.\r\n\r\nWe also use progress bars a lot in `datasets`, that are shown if the logger is at the WARNING level. But we offer a function to disable the progress bars if necessary.", "These kinds of messages are logged as INFO in Transformers, so we should probably be consistent with them" ]

[ "Hi! \r\n\r\nWe use [`fsspec`](https://github.com/fsspec/filesystem_spec) to integrate with storage providers. You can find more info (and the usage examples) in [our docs](https://huggingface.co/docs/datasets/v2.8.0/filesystems#download-and-prepare-a-dataset-into-a-cloud-storage).\r\n\r\n[`gdrivefs`](https://github.com/fsspec/gdrivefs) makes it possible to use Google Drive as a storage service in Datasets, but this is not the case for OneDrive, since its[ Python SDK](https://github.com/OneDrive/onedrive-sdk-python) is not integrated with `fsspec`. Can you please request the integration with `fsspec` in their repo to address this limitation?", "I'm closing this issue as implementing a fsspec-compliant OneDrive filesystem is not our responsibility." ]

[ "@polinaeterna any tentative date on when the Common Voice 12.0 dataset will be added ?", "This dataset is now hosted on the Hub here: https://huggingface.co/datasets/mozilla-foundation/common_voice_12_0" ]

[ "Hi! Can you please share the directory structure of your image folder and the `load_dataset` call? We decode images with Pillow, and Pillow supports RGBA PNGs, so this shouldn't be a problem.\r\n\r\n", "> Hi! Can you please share the directory structure of your image folder and the `load_dataset` call? We decode images with Pillow, and Pillow supports RGBA PNGs, so this shouldn't be a problem.\n> \n> \n\nI have only 1 folder that I use in the load_dataset function with the name \"IMGDATA\" and all my 9000 images are located in this folder.\n`\nfrom datasets import load_dataset\n\ndataset = load_dataset(\"IMGDATA\")\n`\nAt the same time, using another data set with images consisting of 3 RGB channels, everything works", "Okay, I figured out what was wrong. When uploading my dataset via Google Drive, the images broke and Pillow couldn't open them. As a result, I solved the problem by downloading the ZIP archive" ]

[ "Just for your information, Tensorflow confirmed this issue [here.](https://github.com/tensorflow/tensorflow/issues/59279)", "Thanks for reporting, @HaoyuYang59.\r\n\r\nPlease note that these are different \"dataset\" objects: our docs refer to Hugging Face `datasets.Dataset` and not to TensorFlow `tf.data.Dataset`.\r\n\r\nOur `datasets.Dataset.shuffle` method does not have a `reshuffle_each_iteration` argument. Therefore, I would say the statement in our docs is True because they refer to `datasets.Dataset.shuffle`, `datasets.Dataset.skip` and `datasets.Dataset.take`.\r\n\r\nI think this issue is restricted to TensorFlow dataset, and this would be addressed by them in the issue you opened in their repo: https://github.com/tensorflow/tensorflow/issues/59279", "Also note that you are referring to an outdated documentation page: datasets 1.10.2 version\r\n\r\nCurrent datasets version is 2.8.0 and the corresponding documentation page is: https://huggingface.co/docs/datasets/stream#split-dataset", "Hi @albertvillanova thanks for your reply and your explaination here. \r\n\r\nSorry for the confusion as I'm not actually a user of your repo and I just happen to find the thread by Google (and didn't read carefully).\r\n\r\nGreat to know that and you made everything very clear now.\r\n\r\nThanks for your time and sorry for the consusion.\r\n\r\nWishing you a wonderful time. \r\n\r\nRegards" ]

[ "Hi! Can you describe what the actual error is?", "working on the setfit example script\r\n\r\n from setfit import SetFitModel, SetFitTrainer, sample_dataset\r\n\r\nImportError: cannot import name 'sample_dataset' from 'setfit' (C:\\Python\\Python38\\lib\\site-packages\\setfit\\__init__.py)\r\n\r\n apart from that, I also had to hack these loads to import thses modules:\r\n from datasets.load import load_dataset \r\n from datasets.arrow_dataset import Dataset\r\n from datasets.dataset_dict import DatasetDict", "Hi! This issue is related to the [SetFit](https://github.com/huggingface/setfit) project, so can you please open it there?" ]

[ "Sorry, it was us:[^1] https://github.com/iterative/cml/pull/1317 & https://github.com/iterative/cml/issues/1319#issuecomment-1385599559; should be fixed with [v0.18.17](https://github.com/iterative/cml/releases/tag/v0.18.17).\r\n\r\n[^1]: More or less, see https://github.com/yargs/yargs/issues/873.", "Opened https://github.com/huggingface/datasets/pull/5436 unpinning again the container image.", "Hi @0x2b3bfa0, thanks a lot for the investigation, the context about the the root cause and for fixing it!!\r\n\r\nWe are reviewing your PR to unpin the container image." ]

[]

[ "Some of the shard files now have 0 number of rows.\r\n\r\nWe have opened an issue in the Apache Beam repo:\r\n- https://github.com/apache/beam/issues/25041" ]

[ "Hi! Sure, feel free to submit a PR. Maybe if we want to be consistent with the existing API, it would be cleaner to directly add support for `fsspec` paths in `Dataset.load_faiss_index`/`Dataset.save_faiss_index` in the same manner as it was done in `Dataset.load_from_disk`/`Dataset.save_to_disk`.", "That's a great idea! I'll do that instead. " ]

[ "Thanks for reporting, @ilos-vigil.\r\n\r\nWe have transferred this issue to the corresponding dataset on the Hugging Face Hub: https://huggingface.co/datasets/id_clickbait/discussions/1 " ]

[]

[ "Hi! \r\n\r\n`Dataset.sort` calls `df.sort_values` internally, and `df.sort_values` brings all the \"sort\" columns in memory, so sorting on multiple keys could be very expensive. This makes me think that maybe we can replace `df.sort_values` with `pyarrow.compute.sort_indices` - the latter can also sort on multiple keys and currently loads the data into memory; however, there is a plan to eventually implement \"memory-map\" friendly kernels for the Arrow compute ops (using the Acero execution engine). \r\n\r\nSo to address this issue, you should replace `df.sort_values` with `pyarrow.compute.sort_indices` in `Dataset.sort` and adjust the signature of this function (deprecate the `kind` parameter, etc.).\r\n\r\nPS: Feel free to ping us if you need some additional help/pointers", "@mariosasko If I understand the code right, using `pyarrow.compute.sort_indices` would also require changes to the `select` method if it is meant to sort multiple keys. That's because `select` only accepts 1D input for `indices`, not an iterable or similar which would be required for multiple keys unless you want some looping over selects. Doesn't seem that straight-forward but I might be missing something here... ", "@MichlF No, it doesn't require modifying select because sorting on multiple keys also returns a 1D array.\r\n\r\nIt's easier to understand with an example:\r\n```python\r\n>>> import pyarrow as pa\r\n>>> import pyarrow.compute as pc\r\n>>> table = pa.table({\r\n... \"name\": [\"John\", \"Eve\", \"Peter\", \"John\"],\r\n... \"surname\": [\"Johnson\", \"Smith\", \"Smith\", \"Doe\"],\r\n... \"age\": [20, 40, 30, 50],\r\n... })\r\n>>> indices = pc.sort_indices(table, sort_keys=[(\"name\", \"ascending\"), (\"surname\", \"ascending\")])\r\n>>> print(indices)\r\n[\r\n 1,\r\n 3,\r\n 0,\r\n 2\r\n]\r\n```\r\n\r\n", "Thanks for clarifying.\r\nI can prepare a PR to address this issue. This would be my first PR here so I have a few maybe silly questions but:\r\n- What is the preferred input type of `sort_keys` for the sort method? A sequence with name, order tuples like pyarrow's `sort_indices` requires?\r\n- What about backwards compatability: is it supposed to also accept the old way of calling sort() or should both `column` and `kind` be deprecated?\r\n- If `sort_keys` is provided in the same format as for pyarrow's `sort_indices` - i.e. along with order for each column -, `reverse` doesn't make much sense either and should be deprecated as well I assume.", "I think we can have the following signature:\r\n```python\r\ndef sort(\r\n self,\r\n column_names: Union[str, Sequence[str]],\r\n reverse: Union[bool, Sequence[bool]] = False,\r\n kind=\"deprecated\",\r\n null_placement: str = \"last\",\r\n keep_in_memory: bool = False,\r\n load_from_cache_file: bool = True,\r\n indices_cache_file_name: Optional[str] = None,\r\n writer_batch_size: Optional[int] = 1000,\r\n new_fingerprint: Optional[str] = None,\r\n ) -> \"Dataset\":\r\n``` \r\n\r\nSo we should:\r\n* rename`column` to `column_names`. `column` is a positional argument, so it's OK to rename it (not marked as positional-only with \"/\", but still should be fine)\r\n* deprecate `kind`\r\n* keep `reverse` instead of introducing `sort_keys`, but we should allow passing a list of booleans that defines the sort order of each column from `column_names` to it (`reverse = False` would be equal to `[False] * len(column_names)` and `reverse = True` to `[True] * len(column_names)`)", "I am pretty much done with the PR. Just one clarification: `Sequence` in `arrow_dataset.py` is a custom dataclass from `features.py` instead of the `type.hinting` class `Sequence` from Python. Do you suggest using that custom `Sequence` class somehow ? Otherwise signature currently reads instead:\r\n```Python\r\n def sort(\r\n self,\r\n column_names: Union[str, List[str]],\r\n reverse: Union[bool, List[bool]] = False,\r\n kind = \"deprecated\",\r\n null_placement: str = \"last\",\r\n keep_in_memory: bool = False,\r\n load_from_cache_file: bool = True,\r\n indices_cache_file_name: Optional[str] = None,\r\n writer_batch_size: Optional[int] = 1000,\r\n new_fingerprint: Optional[str] = None,\r\n )\r\n```\r\n\r\nAlso, to maintain backwards compatibility, I added conditionals for `null_placement`, because pyarrow's `null_placement` only accepts `at_start` and `at_end`, and not `last` and `first`.\r\nIf that is all good, I think I can open the PR.", "I meant `typing.Sequence` (`datasets.Sequence` is a feature type). \r\n\r\nRegarding `null_placement`, I think we can support both `at_start` and `at_end`, and `last` and `first` (for backward compatibility; convert internally to `at_end` and `at_start` respectively).", "> I meant typing.Sequence (datasets.Sequence is a feature type).\r\n\r\nSorry, I actually meant `typing.Sequence` and not `type.hinting`. However, the issue is still that `dataset.Sequence` is imported in `arrow_dataset.py` so I cannot import and use `typing.Sequence` for the `sort`'s signature without overwriting the `dataset.Sequence` import. The latter is used in the `align_labels_with_mapping` method so it's a necessary import for `arrow_dataset.py`. \r\nTo import `typing.Sequence` as something else than `Sequence` to avoid overwriting may only be confusing and doesn't seem good practice!? The other solution is to keep `List` type hinting as in the signature I posted in my previous post but this excludes other Sequence types and may cause problems further down the line.\r\nPlease advise,\r\nThanks for all the clarifications!", "You can avoid the name collision by renaming `typing.Sequence` to `Sequence_` when importing:\r\n```python\r\nfrom typing import Sequence as Sequence_\r\n```", "Resolved via #5502 " ]

[ "Hi! You can get a `DatasetDict` if you pass a dictionary with read instructions as follows:\r\n```python\r\ninstructions = [\r\n ReadInstruction(split_name=\"train\", from_=0, to=10, unit='%', rounding='closest'),\r\n ReadInstruction(split_name=\"dev\", from_=0, to=10, unit='%', rounding='closest'),\r\n ReadInstruction(split_name=\"test\", from_=0, to=5, unit='%', rounding='closest')\r\n]\r\n\r\ndataset = load_dataset('csv', data_dir=\"data/\", data_files={\"train\":\"train.tsv\", \"dev\":\"dev.tsv\", \"test\":\"test.tsv\"}, delimiter=\"\\t\", split={inst.split_name: inst for inst in instructions})\r\n```\r\n" ]

[ "I can confirm that the error exists!\r\nI'm trying to read 3 parquet files locally:\r\n```python\r\nfrom datasets import load_dataset, Features, Value, ClassLabel\r\n\r\nreview_dataset = load_dataset(\r\n \"parquet\",\r\n data_files={\r\n \"train\": os.path.join(sentiment_analysis_data_path, \"train.parquet\"),\r\n \"validation\": os.path.join(sentiment_analysis_data_path, \"validation.parquet\"),\r\n \"test\": os.path.join(sentiment_analysis_data_path, \"test.parquet\"),\r\n },\r\n)\r\n```\r\n\r\nBut you can fix it, by specifying `features` for `load_dataset()` function like this:\r\n```python\r\nfrom datasets import load_dataset, Features, Value, ClassLabel\r\n\r\nfeatures = Features(\r\n {\r\n \"label\": ClassLabel(\r\n num_classes=3,\r\n names=[\"negative\", \"neutral\", \"positive\"],\r\n ),\r\n \"text\": Value(dtype=\"string\"),\r\n }\r\n)\r\n\r\nreview_dataset = load_dataset(\r\n \"parquet\",\r\n data_files={\r\n \"train\": os.path.join(sentiment_analysis_data_path, \"train.parquet\"),\r\n \"validation\": os.path.join(sentiment_analysis_data_path, \"validation.parquet\"),\r\n \"test\": os.path.join(sentiment_analysis_data_path, \"test.parquet\"),\r\n },\r\n features=features,\r\n)\r\n\r\nprint(review_dataset)\r\n```", "@Extremesarova I think this is a different issue, but understand using features could be a work-around.\r\nIt seems the field `closed_at` is `null` in many cases.\r\n\r\nI've not found a way to specify only a single feature without (succesfully) specifiying the full and quite detailed set of expected features. Using this features set gives an error the column names don't match.\r\n`features = Features({'closed_at': Value(dtype='timestamp[s]', id=None)})`\r\n\r\n", "Found this when searching for the same error, looks like based on #3965 it's just an issue with the data. I found that changing `df = pd.DataFrame.from_records(all_issues)` to `df = pd.DataFrame.from_records(all_issues).dropna(axis=1, how='all').drop(['milestone'], axis=1)` from the fetch_issues function fixed the issue. \r\nThe \"milestone\" column seemed to be problematic (only ~50 non null rows) and dropped any columns that were all null as well just in case.", "I have this same issue. I saved a dataset to disk and now I can't load it.", "Ok the solution was to use load_from_disk instead of load_dataset.", "Hi @folterj , I faced same issue while creating `issues_dataset` (https://huggingface.co/learn/nlp-course/chapter5/5?fw=pt). The fix which worked for me was loading the *.jsonl file as pd.read_json and then converting it into a Dataset using datasets API.\r\n```\r\nimport pandas as pd\r\ndf=pd.read_json(\"datasets-issues.jsonl\", lines=True)\r\ndf.head()\r\n\r\nfrom datasets import Dataset\r\nissues_dataset = Dataset.from_pandas(df)\r\nissues_dataset\r\nsample = issues_dataset.shuffle(seed=666).select(range(3))\r\nsample[0]\r\n```", "I understand some work-around suggestions would be to not use load_dataset(), and instead using a different API function. Another alternative would be using the same function using streaming, as I had already suggested in my original post.\r\nHowever, the fact remains that there is an issue in this function as reported." ]

[ "Closing as case-insensitivity should be only for URL redirection on the Hub. In the APIs, we will only support the canonical name (https://github.com/huggingface/moon-landing/pull/2399#issuecomment-1382085611)" ]

[ "Hi! Thanks for pointing out this inconsistency. Changing the default value at this point is probably not worth it, considering we've started discussing the state of the task API internally - we will most likely deprecate the current one and replace it with a more robust solution that relies on the `train_eval_index` field stored in the YAML section of the dataset cards.", "The task templates API has been deprecated (will be removed in version 3.0), so I'm closing this issue." ]

[ "Thanks for your proposal, @zanussbaum. Yes, I agree that would definitely be a nice feature to have!", "@albertvillanova I’m happy to make a quick PR for the feature! let me know ", "That would be awesome ! You can comment `#self-assign` to assign you to this issue and open a PR :) Will be happy to review", "Closing as this has been merged @lhoestq " ]

[]

[ "Thanks for reporting, @Nithin-Holla.\r\n\r\nThis is a known issue for multiple datasets and we are investigating it:\r\n- See e.g.: https://huggingface.co/datasets/ami/discussions/3", "Main issue:\r\n- #5415", "@albertvillanova Thanks! As a workaround for now, can I use the dataset in streaming mode?", "Yes, @Nithin-Holla, in the meantime you can use this dataset in streaming mode." ]

[ "Hi ! Thanks for reporting :)\r\n\r\nI managed to reproduce the hub using\r\n```python\r\n\r\nfrom datasets import concatenate_datasets, Dataset, load_from_disk\r\n\r\nDataset.from_dict({\"a\": range(9)}).save_to_disk(\"tmp/ds1\")\r\nds1 = load_from_disk(\"tmp/ds1\")\r\nds1 = concatenate_datasets([ds1, ds1])\r\n\r\nDataset.from_dict({\"b\": range(6)}).save_to_disk(\"tmp/ds2\")\r\nds2 = load_from_disk(\"tmp/ds2\")\r\nds2 = concatenate_datasets([ds2, ds2, ds2])\r\n\r\nconcatenate_datasets([ds1, ds2], axis=1)\r\n```\r\nand I get\r\n```python\r\nTraceback (most recent call last): \r\n File \"test.py\", line 98, in <module>\r\n dds = concatenate_datasets([ds1, ds2], axis=1)\r\n File \"/Users/.../datasets/combine.py\", line 182, in concatenate_datasets\r\n return _concatenate_map_style_datasets(dsets, info=info, split=split, axis=axis)\r\n File \"/Users/.../datasets/arrow_dataset.py\", line 5499, in _concatenate_map_style_datasets\r\n table = concat_tables([dset._data for dset in dsets], axis=axis)\r\n File \"/Users/.../datasets/table.py\", line 1778, in concat_tables\r\n return ConcatenationTable.from_tables(tables, axis=axis)\r\n File \"/Users/.../datasets/table.py\", line 1483, in from_tables\r\n blocks = _extend_blocks(blocks, table_blocks, axis=axis)\r\n File \"/Users/.../datasets/table.py\", line 1477, in _extend_blocks\r\n result[i].extend(row_blocks)\r\nIndexError: list index out of range\r\n```\r\n\r\nIt appears to happen when the two datasets have a number of shards that is not the same" ]

[ "Hi ! It fails because the dataset is already being prepared by your first run. I'd encourage you to prepare your dataset before using it for multiple trainings.\r\n\r\nYou can also specify another cache directory by passing `cache_dir=` to `load_dataset()`.", "Thank you! What do you mean by prepare it beforehand? I am unclear how to conduct dataset preparation outside of using the `load_dataset` function.", "You can have a separate script that does load_dataset + map + save_to_disk to save your prepared dataset somewhere. Then in your training script you can reload the dataset with load_from_disk", "Thank you! I believe I was running additional map steps after loading, resulting in the cache conflict. " ]

[ "Hi ! On macos I tried with\r\n- py 3.9.11\r\n- datasets 2.8.0\r\n- transformers 4.25.1\r\n- dill 0.3.4\r\n\r\nand I was able to hash `prepare_dataset` correctly:\r\n```python\r\nfrom datasets.fingerprint import Hasher\r\nHasher.hash(prepare_dataset)\r\n```\r\n\r\nWhat version of transformers do you have ? Can you try to call `Hasher.hash` on the the tokenizer and the feature extractor to see which one can't be hashed ?", "Thanks for your prompt reply.\r\n\r\nI update datasets version to 2.8.0 and the warning is gong." ]

[ "Thanks for reporting @msummerfield. We are fixing it." ]

[ "I still get this error on 2.9.0\r\n<img width=\"1925\" alt=\"image\" src=\"https://user-images.githubusercontent.com/7208470/215597359-2f253c76-c472-4612-8099-d3a74d16eb29.png\">\r\n", "Hi ! I just tested locally and or colab and it works fine for 2.9 on `sst2`.\r\n\r\nAlso the code that is shown in your stack trace is not present in the 2.9 source code - so I'm wondering how you installed `datasets` that could cause this ? (you can check by searching for `[0:{label_ids[-1] + 1}]` in the [2.9 codebase](https://github.dev/huggingface/datasets/tree/b5672a956d5de864e6f5550e493527d962d6ae55) - it doesn't find anything)\r\n\r\nAnyway you can try uninstalling `datasets` and install it again", "For what it's worth, I've also gotten this error on 2.9.0, and I've tried uninstalling an reinstalling\r\n\r\n\r\nI'm very new to this package (I was following this tutorial: https://huggingface.co/docs/transformers/training), so there's a good chance I was doing something wrong 😅 but thought I'd pass along the feedback", "@ntrpnr @mtwichel Did you install `datasets` with conda ?\r\n\r\nI suspect that `datasets` 2.9 on conda still have this issue for some reason. When I install `datasets` with `pip` I don't have this error.", "> @ntrpnr @mtwichel Did you install datasets with conda ?\r\n\r\nI did yeah, I wonder if that's the issue", "I just checked on conda at https://anaconda.org/HuggingFace/datasets/files\r\n\r\nand everything looks fine, I got\r\n```python\r\n\r\nf\"ClassLabel expected a value for all label ids [0:{int(label_ids[-1]) + 1}] but some ids are missing.\"\r\n```\r\nas expected in features.py line 1760 (notice the \"int()\") to not have the TypeError.\r\n\r\nFrom where on conda did you install `datasets` ? You should use the `HuggingFace` official channel\r\n\r\nedit: the conda-forge one [here](https://anaconda.org/conda-forge/datasets/files) seems ok as well", "Could you also try this in your notebook ? In case your python kernel doesn't match the `pip` environment in your shell\r\n```python\r\nimport datasets; datasets.__version__\r\n```\r\nand\r\n```\r\n!which python\r\n```\r\n```python\r\nimport sys; sys.executable\r\n```", "Mmmm, just a potential clue:\r\n\r\nWhere are you running your Python code? Is it the Spyder IDE?\r\n\r\nI have recently seen some users reporting conflicting Python environments while using Spyder...\r\n\r\nMaybe related:\r\n- #5487", "Other potential clue:\r\n- Had you already imported `datasets` before pip-updating it? You should first update datasets, before importing it. Otherwise, you need to restart the kernel after updating it.", "I installed `datasets` with Conda using `conda install datasets` and got this issue.\r\n\r\nThen I tried to reinstall using\r\n`\r\nconda install -c huggingface -c conda-forge datasets\r\n`\r\nThe issue is now fixed.", "I'm still getting this error on 2.13.0" ]

[ "Hi @Breakend,\r\n\r\nIndeed, the attribute `size_in_bytes` refers to the size of the entire dataset configuration, for all splits (size of downloaded files + Arrow files), not the specific split.\r\nThis is also the case for `download_size` (downloaded files) and `dataset_size` (Arrow files).\r\n\r\nThe size of the Arrow files for a specific split can be accessed: e.g. size of the \"test\" split only\r\n```python\r\nds[\"train\"].info.splits[\"test\"].num_bytes\r\n```\r\n\r\nI agree this is confusing and maybe we should improve it." ]

[ "Hi, I made my version : https://huggingface.co/datasets/tasksource/bigbench" ]

[ "`load_from_disk` must be used on datasets saved using `save_to_disk`: they correspond to fully serialized datasets including their state.\r\n\r\nOn the other hand, `download_and_prepare` just downloads the raw data and convert them to arrow (or parquet if you want). We are working on allowing you to reload a dataset saved on S3 with `download_and_prepare` using `load_dataset` in #5281 \r\n\r\nFor now I'd encourage you to keep using `save_to_disk`", "Thanks, I'll follow that issue. \r\n\r\nI was following the [cloud storage](https://huggingface.co/docs/datasets/filesystems) docs section and perhaps I'm missing some part of the flow; start with `load_dataset_builder` + `download_and_prepare`. You say I need an explicit `save_to_disk` but what object needs to be saved? the builder? is that related to the other issue?", "Right now `load_dataset_builder` + `download_and_prepare` is to be used with tools like dask or spark, but `load_dataset` will support private cloud storage soon as well so you'll be able to reload the dataset with `datasets`.\r\n\r\nRight now the only function that can load a dataset from a cloud storage is `load_from_disk`, that must be used with a dataset serialized with `save_to_disk`." ]

[]

[]

[ "I still getting the same error :\r\n\r\n`python -m spacy download fr_core_news_lg\r\n`.\r\n`import spacy`", "@MFatnassi, this issue and the corresponding fix only affect our Continuous Integration testing environment.\r\n\r\nNote that `datasets` does not depend on `spacy`." ]

[ "Hey @catswithbats! Super sorry for the late reply! This is happening because there is data with label length (504) that exceeds the model's max length (448). \r\n\r\nThere are two options here:\r\n1. Increase the model's `max_length` parameter: \r\n```python\r\nmodel.config.max_length = 512\r\n```\r\n2. Filter data with labels longer than max length: https://discuss.huggingface.co/t/open-to-the-community-whisper-fine-tuning-event/26681/21?u=sanchit-gandhi\r\n\r\nNote that the datasets repo is reserved for issues directly related to the HF datasets library. Issues related to custom fine-tuning implementations are more applicable to the HF Forum: https://discuss.huggingface.co. You're more likely to get a response by posting your issue in the most applicable place and boost the chance of someone sharing a working solution!", "@sanchit-gandhi Thank you for all your work on this topic.\r\n\r\nI'm finding that changing the `max_length` value does not make this error go away." ]

[ "Hmmm, git bisect tells me that the behavior is the same since https://github.com/huggingface/datasets/commit/67e65c90e9490810b89ee140da11fdd13c356c9c (3 Oct), i.e. https://github.com/huggingface/datasets/pull/4926", "Maybe related to the discussions in https://github.com/huggingface/datasets/pull/5196", "Maybe the current version of moonlanding in Hub CI is the issue.\r\n\r\nI relaunched tests that were working two days ago: now they are failing. https://github.com/huggingface/datasets-server/commit/746414449cae4b311733f8a76e5b3b4ca73b38a9 for example\r\n\r\ncc @huggingface/moon-landing ", "Hi! I don't think this has anything to do with `datasets`. Hub CI seems to be the culprit - the identical failure can be found in [this](https://github.com/huggingface/datasets/pull/5389) PR (with unrelated changes) opened today.", "OK! Thanks for looking at it. Closing then." ]

[ "Hi! I can't reproduce this error locally (Mac) or in Colab. What version of `datasets` are you using?", "Hi [mariosasko](https://github.com/mariosasko), the datasets version is '2.8.0'.", "@valmetisrinivas you get that error because you imported `datasets` (and thus `fsspec`) before installing `zstandard`.\r\n\r\nPlease, restart your Colab runtime and execute the install commands before importing `datasets`:\r\n```python\r\n!pip install datasets\r\n!pip install zstandard\r\n\r\nfrom datasets import load_dataset\r\n\r\nds = load_dataset(\r\n \"json\",\r\n data_files=\"https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst\",\r\n split=\"train\",\r\n streaming=True,\r\n)\r\nnext(iter(ds))\r\n```", "> @valmetisrinivas you get that error because you imported `datasets` (and thus `fsspec`) before installing `zstandard`.\r\n> \r\n> Please, restart your Colab runtime and execute the install commands before importing `datasets`:\r\n> \r\n> ```python\r\n> !pip install datasets\r\n> !pip install zstandard\r\n> \r\n> from datasets import load_dataset\r\n> \r\n> ds = load_dataset(\r\n> \"json\",\r\n> data_files=\"https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst\",\r\n> split=\"train\",\r\n> streaming=True,\r\n> )\r\n> next(iter(ds))\r\n> ```\r\n\r\nI guess that was the problem, importing datasets before the installation of zstandard. Thank you for the feedback. " ]

[ "Hi! Our documentation builder does not support links to sections, hence the bug. This is the link it should point to https://huggingface.co/docs/datasets/v2.8.0/en/cache#improve-performance." ]

[ "Hi! \r\n\r\nThis behavior stems from the fact that we don't always embed image bytes in the underlying arrow table, which can lead to bad size estimation (we use the first 1000 table rows to [estimate](https://github.com/huggingface/datasets/blob/9a7272cd4222383a5b932b0083a4cc173fda44e8/src/datasets/arrow_dataset.py#L4627) the external file size). We plan to address this in the next major release by always embedding external bytes. In the meantime, you can either shuffle the dataset with `.shuffle().flatten_indices()` to make the estimation more precise or embed the bytes in the table like so:\r\n```python\r\nfrom datasets.table import embed_table_storage\r\nformat = ds.format\r\nds = ds.with_format(\"arrow\")\r\nds = ds.map(embed_table_storage, batched=True)\r\nds = ds.with_format(**format)\r\n...\r\nds.push_to_hub(...)\r\n```", "Embedding the bytes worked like charm. Thanks @mariosasko!" ]

[ "Hi! Yes, we should deprecate the `fs` param here. Would you be interested in submitting a PR? ", "> Hi! Yes, we should deprecate the `fs` param here. Would you be interested in submitting a PR?\r\n\r\nYeah I can do that sometime next week. Should the storage_options be a new arg here? I’ll look around for anywhere else where fs is an arg.", "Closed by #5393." ]

[ "Another example is that `IterableDataset.map` does not have `fn_kwargs`, among other arguments. It makes it harder to convert code from Dataset to IterableDataset.", "Hi! `fn_kwargs` was added to `IterableDataset.map` in `datasets 2.5.0`, so please update your installation (`pip install -U datasets`) to use it.\r\n\r\nRegarding `column_names`, I agree we should add this property to `IterableDataset`. In the meantime, you can use `list(dataset.features.keys())` instead.", "Thanks! That's great news.\n\nOn Thu, Dec 22, 2022, 07:48 Mario Šaško ***@***.***> wrote:\n\n> Hi! fn_kwargs was added to IterableDataset.map in datasets 2.5.0, so\n> please update your installation (pip install -U datasets) to use it.\n>\n> Regarding column_names, I agree we should add this property to\n> IterableDataset. In the meantime, you can use\n> list(dataset.features.keys()) instead.\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5383#issuecomment-1362993633>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AAHD6N2EQUFEOUFDW3VHSILWORZ45ANCNFSM6AAAAAATGKWVGM>\n> .\n> You are receiving this because you authored the thread.Message ID:\n> ***@***.***>\n>\n", "I'm marking this issue as a \"good first issue\", as it makes sense to have `IterableDataset.column_names` in the API. Besides the case when `features` are `None` (e.g., `features` are `None` after `map`), in which we can also return `column_names` as `None`, adding this property should be straightforward,", "Hi @mariosasko, I can work on this if that's ok?", "Yes! I've assigned you the issue." ]

[ "Hi! This error can happen if there is a local file/folder with the same name as the requested dataset. And to avoid it, rename the local file/folder.\r\n\r\nSoon, it will be possible to explicitly request a Hub dataset as follows:https://github.com/huggingface/datasets/issues/5228#issuecomment-1313494020" ]

[ "Hi! I agree `skip` can be inefficient to use in the current state.\r\n\r\nTo make it fast, we could use \"statistics\" stored in Parquet metadata and read only the chunks needed to form a dataset. \r\n\r\nAnd thanks to the \"datasets-server\" project, which aims to store the Parquet versions of the Hub datasets (only the smaller datasets are covered currently), this solution can also be applied to datasets stored in formats other than Parquet. (cc @severo)", "@mariosasko do the current parquet files created by the datasets-server already have the required \"statistics\"? If not, please open an issue on https://github.com/huggingface/datasets-server with some details to make sure we implement it.", "Yes, nothing has to be changed on the datasets-server side. What I mean by \"statistics\" is that we can use the \"row_group\" metadata embedded in a Parquet file (by default) to fetch the requested rows more efficiently.", "Glad to see the feature could be of interest. \r\n\r\nI'm sure there are many possible ways to implement this feature. I don't know enough about the datasets-server, but I guess that it is not instantaneous, in the sense that user-owned private datasets might need hours or days until they are ported to the datasets-server (if at all), which could be cumbersome. Having optionally that information in the `dataset_infos.json` file would make it easier for users to control the skip process a bit.", "re: statistics:\r\n\r\n- https://arrow.apache.org/docs/python/generated/pyarrow.parquet.FileMetaData.html\r\n- https://arrow.apache.org/docs/python/generated/pyarrow.parquet.RowGroupMetaData.html\r\n\r\n```python\r\n>>> import pyarrow.parquet as pq\r\n>>> import hffs\r\n>>> fs = hffs.HfFileSystem(\"glue\", repo_type=\"dataset\", revision=\"refs/convert/parquet\")\r\n>>> metadata = pq.read_metadata(\"ax/glue-test.parquet\", filesystem=fs)\r\n>>> metadata\r\n<pyarrow._parquet.FileMetaData object at 0x7f4537cec400>\r\n created_by: parquet-cpp-arrow version 7.0.0\r\n num_columns: 4\r\n num_rows: 1104\r\n num_row_groups: 2\r\n format_version: 1.0\r\n serialized_size: 2902\r\n>>> metadata.row_group(0)\r\n<pyarrow._parquet.RowGroupMetaData object at 0x7f45564bcbd0>\r\n num_columns: 4\r\n num_rows: 1000\r\n total_byte_size: 164474\r\n>>> metadata.row_group(1)\r\n<pyarrow._parquet.RowGroupMetaData object at 0x7f455005c400>\r\n num_columns: 4\r\n num_rows: 104\r\n total_byte_size: 13064\r\n```", "> user-owned private datasets might need hours or days until they are ported to the datasets-server (if at all)\r\n\r\nprivate datasets are not supported yet (https://github.com/huggingface/datasets-server/issues/39)", "@versae `Dataset.push_to_hub` writes shards in Parquet, so this solution would also work for such datasets (immediately after the push). ", "@mariosasko that is right. However, there are still a good amount of datasets for which the shards are created manually. In our very specific case, we create medium-sized datasets (rarely over 100-200GB) of both text and audio, we prepare the shards by hand and then upload then. It would be great to have immediate access to this download skipping feature for them too.", "From looking at Arrow's source, it seems Parquet stores metadata at the end, which means one needs to iterate over a Parquet file's data before accessing its metadata. We could mimic Dask to address this \"limitation\" and write metadata in a `_metadata`/`_common_metadata` file in `to_parquet`/`push_to_hub`, which we could then use to optimize reads (if present). Plus, it's handy that PyArrow can also parse these metadata files.", "So if Parquet metadata needs to be in its own file anyway, why not implement this skipping feature by storing the example counts per shard in `dataset_infos.json`? That would allow:\r\n- Support both private and public datasets\r\n- Immediate access to the feature upon uploading of shards\r\n- Use any dataset, not only those uploaded using `.push_to_hub()`\r\n\r\nA proper Parquet metadata file could still be created and \"overwrite\" the `dataset_infos.json` info in the datasets-server." ]

[ "Thanks for reporting @shaoyuta. We are investigating it.\r\n\r\nWe are transferring the issue to \"the_pile\" Community tab on the Hub: https://huggingface.co/datasets/the_pile/discussions/4" ]

[ "The data files are hosted on HF at https://huggingface.co/datasets/allenai/c4/tree/main\r\n\r\nYou have 200 runs streaming the same files in parallel. So this is probably a Hub limitation. Maybe rate limiting ? cc @julien-c \r\n\r\nMaybe you can also try to reduce the number of HTTP requests by increasing the block size of each request. This can be done by increasing `DEFAULT_BLOCK_SIZE` in `fsspec.implementations.http`. Default is `5 * 2**20` (5MiB)\r\n\r\nAnyway maybe it's just better to save the dataset locally in that case ?", "you don't get an HTTP error code or something in your stack trace? Kinda hard to debug with this info", "You could try to re-run using this `datasets` branch: [raise-err-when-disconnect](https://github.com/huggingface/datasets/compare/raise-err-when-disconnect?expand=1)\r\nIt should raise the fsspec error", "The weird thing is that I already have it saved locally & it seems to indeed be using the cached one 🧐 ; I'm also using offline mode, so I don't think it has something to do with the Hub.\r\n```\r\nWARNING:datasets.load:Using the latest cached version of the module from /users/muennighoff/.cache/huggingface/modules/datasets_modules/datasets/c4/df532b158939272d032cc63ef19cd5b83e9b4d00c922b833e4cb18b2e9869b01 (last modified on Mon Dec 12 10:45:02 2022) since it couldn't be found locally at c4.\r\n```\r\n\r\n", "No, you passed `streaming=True` so it streams the data from the Hub.\r\nThis message just shows that you use the cached version of the `c4` **module**, aka the python script that is run to generate the examples from the raw data files.\r\n\r\nMaybe the offline mode should also disable `fsspec`/`aiohttp` HTTP calls in `datasets` and not just the `requests` ones.", "> This message just shows that you use the cached version of the c4 module\r\n\r\nAh my bad you're right about the module, but it's also using the downloaded & cached c4 dataset. There's no internet during the runs so it wouldn't work otherwise", "You don't have internet, therefore you get an error while trying to stream ;)" ]

[ "Ccing @nazneenrajani @lvwerra @osanseviero " ]

[]

[ "Thanks for reporting, @shaoyuta.\r\n\r\nWe have checked and yes, apparently there is an issue with the server hosting the data of the \"enron_emails\" subset of \"the_pile\" dataset: http://eaidata.bmk.sh/data/enron_emails.jsonl.zst\r\nIt seems to be down: The connection has timed out.\r\n\r\nPlease note that at the Hugging Face Hub, we are not hosting their data for this dataset, but only a script that downloads the data from their servers. We are updating the data URL to one in another server.\r\n\r\nIn the meantime, please note that you can train your model in the entire \"the_pile\" dataset, by passing the \"all\" config (instead of the \"enron_emails\" one).", "We have transferred this issue to the corresponding dataset Community tab: https://huggingface.co/datasets/the_pile/discussions/2\r\n\r\nPlease, follow the updates there." ]

[ "You can try something like this ?\r\n```python\r\ndef mapper_function(batch):\r\n return {\"concatenated_audio\": [np.concatenate([audio[\"array\"] for audio in batch[\"audio\"]])]}\r\n\r\ndataset = dataset.map(\r\n mapper_function,\r\n batched=True,\r\n batch_size=3,\r\n remove_columns=list(dataset.features),\r\n)\r\n```", "Thanks for the snippet!\r\n\r\nOne more question. I wonder why those two mappers are working so different that one taking 4 sec while other taking over 1 min :\r\n\r\n```python\r\n%%time\r\ndef mapper_function1(batch):\r\n # list_audio\r\n return {\r\n \"audio\": [\r\n {\r\n \"array\": np.concatenate([audio[\"array\"] for audio in batch[\"audio\"]]),\r\n \"sampling_rate\": 16_000,\r\n }\r\n ]\r\n }\r\n\r\ndataset.map(\r\n mapper_function1,\r\n batched=True,\r\n batch_size=3,\r\n remove_columns=list(dataset.features),\r\n)\r\n\r\n# 100%\r\n# 135/135 [01:13<00:00, 1.93ba/s]\r\n# CPU times: user 1min 10s, sys: 3.21 s, total: 1min 13s\r\n# Wall time: 1min 13s\r\n# Dataset({\r\n# features: ['audio'],\r\n# num_rows: 135\r\n# })\r\n\r\n# --------------------------------\r\n%%time\r\ndef mapper_function2(batch):\r\n # list_audio\r\n return {\"audio\": [np.concatenate([audio[\"array\"] for audio in batch[\"audio\"]])]}\r\n\r\ndataset.map(\r\n mapper_function2,\r\n batched=True,\r\n batch_size=3,\r\n remove_columns=list(dataset.features),\r\n)\r\n\r\n# 100%\r\n# 135/135 [00:03<00:00, 40.69ba/s]\r\n# CPU times: user 1.88 s, sys: 1.48 s, total: 3.36 s\r\n# Wall time: 4.8 s\r\n# Dataset({\r\n# features: ['audio'],\r\n# num_rows: 135\r\n# })\r\n```\r\n", "In the first one you get a dataset with an Audio type, and in the second one you get a dataset with a sequence of floats type.\r\n\r\nThe Audio type encodes the data as WAV to save disk space, so it takes more time to create.\r\nThe Audio type is automatically inferred because you modify the column \"audio\" which was already an Audio type. If you name it to something else, type inference will use a type struct with array and sampling rate fields." ]

[ "If you use huggingface trainer, you will find the trainer has wrapped a `IterableDatasetShard` to avoid duplication.\r\nSee:\r\nhttps://github.com/huggingface/transformers/blob/dfd818420dcbad68e05a502495cf666d338b2bfb/src/transformers/trainer.py#L835\r\n", "If you want to support it by datasets natively, maybe we also need to change the code in `transformers` ?", "Opened https://github.com/huggingface/transformers/issues/20770 to discuss this :)", "Maybe something like this then ?\r\n```python\r\nfrom datasets.distributed import split_dataset_by_node\r\nds = split_dataset_by_node(ds, rank=rank, world_size=world_size)\r\n```\r\n\r\nFor map-style datasets the implementation is trivial (it can simply use `.shard()`).\r\n\r\nFor iterable datasets we would need to implement a new ExamplesIterable that would only iterate on a subset of the (possibly shuffled and re-shuffled after each epoch) list of shards, based on the rank and world size.", "My plan is to skip examples by default to not end up with duplicates.\r\n\r\nAnd if a dataset has a number of shards that is a factor of the world size, then I'd make it more optimized by distributing the shards evenly across nodes instead.", "Opened a PR here: https://github.com/huggingface/datasets/pull/5369\r\n\r\nfeel free to play with it and share your feedbacks :)", "@lhoestq I add shuffle after split_dataset_by_node, duplicated data still exist. \r\nFor example, we have a directory named `mock_pretraining_data`, which has three files, `part-00000`, `part-00002`,`part-00002`. \r\nText in `part-00000` is like this: \r\n{\"id\": 0}\r\n{\"id\": 1}\r\n{\"id\": 2}\r\n{\"id\": 3}\r\n{\"id\": 4}\r\n{\"id\": 5}\r\n{\"id\": 6}\r\n{\"id\": 7}\r\n{\"id\": 8}\r\n{\"id\": 9}\r\n\r\nand `part-00001`\r\n{\"id\": 10}\r\n{\"id\": 11}\r\n{\"id\": 12}\r\n{\"id\": 13}\r\n{\"id\": 14}\r\n{\"id\": 15}\r\n{\"id\": 16}\r\n{\"id\": 17}\r\n{\"id\": 18}\r\n{\"id\": 19}\r\n\r\nand `part-00002`\r\n{\"id\": 20}\r\n{\"id\": 21}\r\n{\"id\": 22}\r\n{\"id\": 23}\r\n{\"id\": 24}\r\n{\"id\": 25}\r\n{\"id\": 26}\r\n{\"id\": 27}\r\n{\"id\": 28}\r\n{\"id\": 29}\r\n\r\nAnd code in `test_dist.py` like this,\r\n```python\r\nimport torch\r\nfrom torch.utils.data import Dataset, DataLoader\r\nfrom datasets import load_dataset\r\nimport os\r\nfrom transformers import AutoTokenizer, NezhaForPreTraining\r\nfrom transformers import AdamW, get_linear_schedule_with_warmup\r\nimport torch.nn.functional as F\r\nimport torch.nn as nn\r\nimport torch.distributed as dist\r\nfrom datasets.distributed import split_dataset_by_node\r\nfrom torch.nn.parallel import DistributedDataParallel as DDP\r\n\r\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = '5,6,7'\r\n\r\ndist.init_process_group(\"nccl\")\r\nlocal_rank = int(os.environ['LOCAL_RANK'])\r\nworld_size = torch.distributed.get_world_size()\r\ndevice = torch.device('cuda', local_rank)\r\ndata_dir = './'\r\n\r\ndef load_trainset(train_path):\r\n dataset = load_dataset('json', data_dir=os.path.join(data_dir, train_path), split='train', streaming=True)\r\n return dataset\r\n\r\ndef collate_fn(examples):\r\n input_ids = []\r\n for example in examples:\r\n input_ids.append(example['id'])\r\n return torch.LongTensor(input_ids).to(device)\r\n\r\n\r\ndataset = load_trainset('mock_pretraining_data')\r\ndataset = split_dataset_by_node(dataset, rank=local_rank, world_size=world_size).shuffle(buffer_size=512)\r\n# train_sampler = torch.utils.data.distributed.DistributedSampler(dataset)\r\nbatch_size = 3\r\nprint('batch_size: {}'.format(batch_size))\r\ntrain_dataloader = DataLoader(dataset, batch_size=batch_size, collate_fn=collate_fn)\r\n\r\nfor x in train_dataloader:\r\n print({'rank': local_rank, 'id': x})\r\n```\r\nrun `python -m torch.distributed.launch --nproc_per_node=3 test_dist.py`\r\nThe output is\r\n```\r\n{'rank': 1, 'id': tensor([12, 15, 14], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([16, 10, 18], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([17, 13, 19], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([11], device='cuda:1')}\r\n{'rank': 0, 'id': tensor([0, 2, 9], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([4, 8, 1], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([5, 3, 6], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([7], device='cuda:0')}\r\n{'rank': 2, 'id': tensor([13, 15, 14], device='cuda:2')}\r\n{'rank': 2, 'id': tensor([19, 17, 18], device='cuda:2')}\r\n{'rank': 2, 'id': tensor([12, 16, 11], device='cuda:2')}\r\n{'rank': 2, 'id': tensor([10], device='cuda:2')}\r\n```\r\n`part-00001` is loaded twice, `part-00002` isn't loaded.\r\n\r\nIf I run `python -m torch.distributed.launch --nproc_per_node=2 test_dist.py`\r\nThe output is weirder，many numbers appear twice\r\n```\r\n{'rank': 1, 'id': tensor([26, 8, 13], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([22, 19, 20], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([12, 28, 11], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([24, 2, 14], device='cuda:1')}\r\n{'rank': 1, 'id': tensor([ 6, 27, 3], device='cuda:1')}\r\n{'rank': 0, 'id': tensor([ 8, 25, 1], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([20, 4, 12], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([14, 29, 5], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([ 7, 18, 23], device='cuda:0')}\r\n{'rank': 0, 'id': tensor([19, 17, 11], device='cuda:0')}\r\n``` ", "Hi ! Thanks for reporting, you need to pass `seed=` to `shuffle()` or the processes won't use the same seed to shuffle the shards order before assigning each shard to a node.\r\n\r\nThe issue is that the workers are not using the same seed to shuffle the shards before splitting the shards list by node.", "Opened https://github.com/huggingface/datasets/issues/5696", "I have the same issue\r\n```\r\nds['train'] = load_dataset(streaming=True)\r\nds['train'] = split_dataset_by_node(ds['train'], rank=int(os.environ[\"RANK\"]), world_size=int(os.environ[\"WORLD_SIZE\"]))\r\nvectorized_datasets = ds.map(\r\n prepare_dataset,\r\n remove_columns=raw_datasets_features,\r\n).with_format(\"torch\")\r\n\r\nvectorized_datasets[\"train\"] = vectorized_datasets[\"train\"].shuffle(\r\n buffer_size=500,\r\n seed=42,\r\n)\r\n\r\ndef prepare_dataset(batch):\r\n ....\r\n print(f\"sentence: {batch['sentence']}, target_text: {batch['target_text']}\")\r\n return batch\r\n```\r\nWhen using split_dataset_by_node(), the data being read is indeed different for each GPU ID.\r\n\r\n```\r\ntrainer = Trainer(\r\n model=model,\r\n data_collator=data_collator,\r\n args=training_args,\r\n compute_metrics=compute_metrics,\r\n train_dataset=vectorized_datasets[\"train\"] if training_args.do_train else None,\r\n eval_dataset=vectorized_datasets[\"eval\"] if training_args.do_eval else None,\r\n tokenizer=processor,\r\n callbacks=[ShuffleCallback()],\r\n )\r\n...\r\ntrain_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n```\r\nHowever, when I execute trainer.train(), the data being read is different from what I expected.\r\nBecause I print the batch value in prepare_dataset() , I observe that the data is the same for each GPU ID.\r\n\r\nHow should I handle this issue?\r\n\r\n\r\n", "There are two ways an iterable dataset can be split by node:\r\n1. if the number of shards is a factor of number of GPUs: in that case the shards are evenly distributed per GPU\r\n2. otherwise, each GPU iterate on the data and at the end keeps 1 sample out of n(GPUs) - skipping the others.\r\n\r\nIn case 2. it's therefore possible to have the same examples passed to `prepare_dataset` for each GPU.\r\n\r\nThis doesn't sound optimized though, because it runs the preprocessing on samples that won't be used in the end.\r\n\r\nCould you open a new issue so that we can discuss about this and find a solution ?" ]

[ "Hi! Linking a comment to provide more info on the issue: https://stackoverflow.com/a/39458225. This means we should replace all (most of) the occurrences of `List` with `Sequence` in function signatures.\r\n\r\n@tranhd95 Would you be interested in submitting a PR?", "Hi all! I tried to reproduce this issue and didn't work for me. Also in your example i noticed that the variables have different names: `list_of_filenames` and `list_of_files`, could this be related to that?\r\n```python\r\n#I found random data in parquet format:\r\n!wget \"https://github.com/Teradata/kylo/raw/master/samples/sample-data/parquet/userdata1.parquet\"\r\n!wget \"https://github.com/Teradata/kylo/raw/master/samples/sample-data/parquet/userdata2.parquet\"\r\n\r\n#Then i try reproduce\r\nlist_of_files = [\"userdata1.parquet\", \"userdata2.parquet\"]\r\nds = Dataset.from_parquet(list_of_files)\r\n```\r\n**My output:**\r\n```python\r\nWARNING:datasets.builder:Using custom data configuration default-e287d097dc54e046\r\nDownloading and preparing dataset parquet/default to /root/.cache/huggingface/datasets/parquet/default-e287d097dc54e046/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec...\r\nDownloading data files: 100%\r\n1/1 [00:00<00:00, 40.38it/s]\r\nExtracting data files: 100%\r\n1/1 [00:00<00:00, 23.43it/s]\r\nDataset parquet downloaded and prepared to /root/.cache/huggingface/datasets/parquet/default-e287d097dc54e046/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec. Subsequent calls will reuse this data.\r\n```\r\nP.S. This is my first experience with open source. So do not judge strictly if I do not understand something)", "@dantema There is indeed a typo in variable names. Nevertheless, I'm sorry if I was not clear but the output is from `mypy` type checker. You can run the code snippet without issues. The problem is with the type checking.", "However, I found out that the type annotation is actually misleading. The [`from_parquet`](https://github.com/huggingface/datasets/blob/5ef1ab1cc06c2b7a574bf2df454cd9fcb071ccb2/src/datasets/arrow_dataset.py#L1039) method should also accept list of [`PathLike`](https://github.com/huggingface/datasets/blob/main/src/datasets/utils/typing.py#L8) objects which includes [`os.PathLike`](https://docs.python.org/3/library/os.html#os.PathLike). But if I would ran the code snippet below, an exception is thrown.\r\n\r\n**Code**\r\n```py\r\nfrom pathlib import Path\r\n\r\nlist_of_filenames = [Path(\"foo.parquet\"), Path(\"bar.parquet\")]\r\nds = Dataset.from_parquet(list_of_filenames)\r\n```\r\n**Output**\r\n```py\r\n[/usr/local/lib/python3.8/dist-packages/datasets/arrow_dataset.py](https://localhost:8080/#) in from_parquet(path_or_paths, split, features, cache_dir, keep_in_memory, columns, **kwargs)\r\n 1071 from .io.parquet import ParquetDatasetReader\r\n 1072 \r\n-> 1073 return ParquetDatasetReader(\r\n 1074 path_or_paths,\r\n 1075 split=split,\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/io/parquet.py](https://localhost:8080/#) in __init__(self, path_or_paths, split, features, cache_dir, keep_in_memory, streaming, **kwargs)\r\n 35 path_or_paths = path_or_paths if isinstance(path_or_paths, dict) else {self.split: path_or_paths}\r\n 36 hash = _PACKAGED_DATASETS_MODULES[\"parquet\"][1]\r\n---> 37 self.builder = Parquet(\r\n 38 cache_dir=cache_dir,\r\n 39 data_files=path_or_paths,\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/builder.py](https://localhost:8080/#) in __init__(self, cache_dir, config_name, hash, base_path, info, features, use_auth_token, repo_id, data_files, data_dir, name, **config_kwargs)\r\n 298 \r\n 299 if data_files is not None and not isinstance(data_files, DataFilesDict):\r\n--> 300 data_files = DataFilesDict.from_local_or_remote(\r\n 301 sanitize_patterns(data_files), base_path=base_path, use_auth_token=use_auth_token\r\n 302 )\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/data_files.py](https://localhost:8080/#) in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)\r\n 794 for key, patterns_for_key in patterns.items():\r\n 795 out[key] = (\r\n--> 796 DataFilesList.from_local_or_remote(\r\n 797 patterns_for_key,\r\n 798 base_path=base_path,\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/data_files.py](https://localhost:8080/#) in from_local_or_remote(cls, patterns, base_path, allowed_extensions, use_auth_token)\r\n 762 ) -> \"DataFilesList\":\r\n 763 base_path = base_path if base_path is not None else str(Path().resolve())\r\n--> 764 data_files = resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)\r\n 765 origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)\r\n 766 return cls(data_files, origin_metadata)\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/data_files.py](https://localhost:8080/#) in resolve_patterns_locally_or_by_urls(base_path, patterns, allowed_extensions)\r\n 357 data_files = []\r\n 358 for pattern in patterns:\r\n--> 359 if is_remote_url(pattern):\r\n 360 data_files.append(Url(pattern))\r\n 361 else:\r\n\r\n[/usr/local/lib/python3.8/dist-packages/datasets/utils/file_utils.py](https://localhost:8080/#) in is_remote_url(url_or_filename)\r\n 62 \r\n 63 def is_remote_url(url_or_filename: str) -> bool:\r\n---> 64 parsed = urlparse(url_or_filename)\r\n 65 return parsed.scheme in (\"http\", \"https\", \"s3\", \"gs\", \"hdfs\", \"ftp\")\r\n 66 \r\n\r\n[/usr/lib/python3.8/urllib/parse.py](https://localhost:8080/#) in urlparse(url, scheme, allow_fragments)\r\n 373 Note that we don't break the components up in smaller bits\r\n 374 (e.g. netloc is a single string) and we don't expand % escapes.\"\"\"\r\n--> 375 url, scheme, _coerce_result = _coerce_args(url, scheme)\r\n 376 splitresult = urlsplit(url, scheme, allow_fragments)\r\n 377 scheme, netloc, url, query, fragment = splitresult\r\n\r\n[/usr/lib/python3.8/urllib/parse.py](https://localhost:8080/#) in _coerce_args(*args)\r\n 125 if str_input:\r\n 126 return args + (_noop,)\r\n--> 127 return _decode_args(args) + (_encode_result,)\r\n 128 \r\n 129 # Result objects are more helpful than simple tuples\r\n\r\n[/usr/lib/python3.8/urllib/parse.py](https://localhost:8080/#) in _decode_args(args, encoding, errors)\r\n 109 def _decode_args(args, encoding=_implicit_encoding,\r\n 110 errors=_implicit_errors):\r\n--> 111 return tuple(x.decode(encoding, errors) if x else '' for x in args)\r\n 112 \r\n 113 def _coerce_args(*args):\r\n\r\n[/usr/lib/python3.8/urllib/parse.py](https://localhost:8080/#) in <genexpr>(.0)\r\n 109 def _decode_args(args, encoding=_implicit_encoding,\r\n 110 errors=_implicit_errors):\r\n--> 111 return tuple(x.decode(encoding, errors) if x else '' for x in args)\r\n 112 \r\n 113 def _coerce_args(*args):\r\n\r\nAttributeError: 'PosixPath' object has no attribute 'decode'\r\n```\r\n\r\n@mariosasko Should I create a new issue? ", "@mariosasko I would like to take this issue up. ", "@avinashsai Hi, I've assigned you the issue.\r\n\r\n@tranhd95 Yes, feel free to report this in a new issue.", "@avinashsai Are you still working on this? If not I would like to give it a try.", "@mariosasko I would like to take this issue up!", "Hi @tranhd95 @mariosasko ,I hope you all are doing well.\r\n\r\nI am interested in this issue, is this still open and unresolved ?\r\n\r\nThanks and Regards" ]

[ "Just seen https://github.com/huggingface/datasets/issues/5281" ]

[ "Hi @J-shel, thanks for reporting.\r\n\r\nI think the issue comes from your call to `load_dataset`. As first argument, you should pass:\r\n- either the name of your dataset (\"mrf\") if this is already published on the Hub\r\n- or the path to the loading script of your dataset (\"path/to/your/local/mrf.py\").", "Hi, following your suggestion, I changed my call to load_dataset. Below is the latest:\r\nreader = load_dataset('data/mrf.py',\"default\", input_size=1024, split=split, streaming=True, keep_in_memory=None)\r\nHowever, I still got the same error.\r\nI have one question that is if I only define input_size=2048 in BUILDER_CONFIGS, may I specify input_size=1024 when loading the dataset? Cause I found that I could only specify name=\"default\" since I only define name=\"default\" in BUILDER_CONFIGS." ]

[ "Hi! `DatasetBuilder` is a parent class for concrete builders: `GeneratorBasedBuilder`, `ArrowBasedBuilder` and `BeamBasedBuilder`. When writing a builder script, these classes are the ones you should inherit from. And since all of them implement `_prepare_split`, you only have to implement the three methods mentioned above.", "Thanks so much @mariosasko for the fast response! I've been referencing [this page in the docs](https://huggingface.co/docs/datasets/v2.4.0/en/about_dataset_load) because it it pretty comprehensive in terms of what we have to do and I figured since we subclass the `BuilderConfig` the same pattern would hold, but I've also seen the page with those sub-classed builders as well, so that fills in a knowledge gap for me.", "cc @stevhliu who may have some ideas on how to improve this part of the docs.", "one more question for my understanding @mariosasko. the requirement of a loading script has always seemed counterintuitive to me. if i have to provide a script with every dataset, what is the point of using `datasets` if we're doing all the work of loading it, I can just do that in my code and skip the datasets integration (this of course discounts other potential benefits around metadata management, etc., my example is just simplest use case though for the sake of discussion).\r\n\r\nso i figured I would implement my own `BuilderConfig` and `DatasetBuilder` to handle that portion of it and not have to make a script. i _thought_ this would result in `datasets` (via `download_and_prepare`) then making me something that I could load using `load_dataset` moving forward.\r\n\r\nConcretely, i envisioned this pattern being possible:\r\n\r\n ```\r\nclass MyBuilderConfig(BuilderConfig):\r\n def __init__(self, name=\"my_named_dataset\", ...):\r\n super().__init__(name, ...)\r\n\r\nclass MyDatasetBuilder(GeneratorBasedBuilder):\r\n BUILDER_CONFIG_CLASS = MyBuilderConfig\r\n ....\r\n\r\nmy_builder = MyDatasetBuilder(...)\r\n\r\n# this doesn't exactly work like I thought; I don't get a dataset back, but NoneType instead\r\n# though I can see it loading the files and it generates the cache, etc.\r\nmy_dataset = my_builder.download_and_prepare()\r\n\r\n# load the dataset in the future by referencing it by name and loading from the cached arrow version\r\nnew_instance_of_my_dataset = load_dataset(\"my_named_dataset\")\r\n```\r\n\r\nI've seen references to the `save_to_disk` method which might be the next step I need in order to load it by name, in which case, that makes sense, then i just need to debug why `download_and_prepare` isn't returning me a dataset, but I feel like I still have a larger conceptual knowledge gap on how to use the library correctly.\r\n\r\nThanks again in advance!", "> the requirement of a loading script has always seemed counterintuitive to me\r\n\r\nThis is a requirement only for datasets not stored in standard formats such as CSV, JSON, SQL, Parquet, ImageFolder, etc. \r\n\r\n> if i have to provide a script with every dataset, what is the point of using datasets if we're doing all the work of loading it, I can just do that in my code and skip the datasets integration (this of course discounts other potential benefits around metadata management, etc., my example is just simplest use case though for the sake of discussion)\r\n\r\nOur README/documentation lists the main features... \r\n\r\nOne of the main ones is that our library makes it easy to work with datasets larger than RAM (thanks to Arrow and the caching mechanism), and this is not trivial to implement.\r\n\r\nRegarding the step-by-step builder, this is the pattern:\r\n```python\r\nfrom datasets import load_dataset_builder\r\nbuilder = load_dataset_builder(\"path/to/script\") # or direct instantiation with MyDatasetBuilder(...)\r\nbuilder.download_and_prepare()\r\ndset = builder.as_dataset()\r\n```", "ok, that makes sense. thank you @mariosasko. I realized i'd never looked on the hub at any of the files associated with any datasets. just did that now and it appears that i'll need to have a script regardless _but_ that will just contain my custom config and builder classes, so without realizing it I was already making my script, I just need to wrap that in a file that sits alongside my data (I looked at Glue and realized I was already doing what I thought didn't make sense to have to do, lol).\r\n\r\n`download_and_prepare` isn't returning me a dataset though, but I'll look into that and open another issue if I can't figure it out.", "`download_and_prepare` downloads and prepares the arrow files. You need to call `as_dataset` on the builder to get the dataset.", "ok, I think I was assigning the output of `builder.download_and_prepare` but it's an inplace op, so that explains the `NoneType` i was getting back. Now I'm getting:\r\n\r\n```\r\nArrowInvalid Traceback (most recent call last)\r\n<ipython-input-7-3ed50fb87c70> in <module>\r\n----> 1 ds = dataset_builder.as_dataset()\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/builder.py in as_dataset(self, split, run_post_process, ignore_verifications, in_memory)\r\n 1020 \r\n 1021 # Create a dataset for each of the given splits\r\n-> 1022 datasets = map_nested(\r\n 1023 partial(\r\n 1024 self._build_single_dataset,\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/utils/py_utils.py in map_nested(function, data_struct, dict_only, map_list, map_tuple, map_numpy, num_proc, parallel_min_length, types, disable_tqdm, desc)\r\n 442 num_proc = 1\r\n 443 if num_proc <= 1 or len(iterable) < parallel_min_length:\r\n--> 444 mapped = [\r\n 445 _single_map_nested((function, obj, types, None, True, None))\r\n 446 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/utils/py_utils.py in <listcomp>(.0)\r\n 443 if num_proc <= 1 or len(iterable) < parallel_min_length:\r\n 444 mapped = [\r\n--> 445 _single_map_nested((function, obj, types, None, True, None))\r\n 446 for obj in logging.tqdm(iterable, disable=disable_tqdm, desc=desc)\r\n 447 ]\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/utils/py_utils.py in _single_map_nested(args)\r\n 344 # Singleton first to spare some computation\r\n 345 if not isinstance(data_struct, dict) and not isinstance(data_struct, types):\r\n--> 346 return function(data_struct)\r\n 347 \r\n 348 # Reduce logging to keep things readable in multiprocessing with tqdm\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/builder.py in _build_single_dataset(self, split, run_post_process, ignore_verifications, in_memory)\r\n 1051 \r\n 1052 # Build base dataset\r\n-> 1053 ds = self._as_dataset(\r\n 1054 split=split,\r\n 1055 in_memory=in_memory,\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/builder.py in _as_dataset(self, split, in_memory)\r\n 1120 \"\"\"\r\n 1121 cache_dir = self._fs._strip_protocol(self._output_dir)\r\n-> 1122 dataset_kwargs = ArrowReader(cache_dir, self.info).read(\r\n 1123 name=self.name,\r\n 1124 instructions=split,\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/arrow_reader.py in read(self, name, instructions, split_infos, in_memory)\r\n 236 msg = f'Instruction \"{instructions}\" corresponds to no data!'\r\n 237 raise ValueError(msg)\r\n--> 238 return self.read_files(files=files, original_instructions=instructions, in_memory=in_memory)\r\n 239 \r\n 240 def read_files(\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/arrow_reader.py in read_files(self, files, original_instructions, in_memory)\r\n 257 \"\"\"\r\n 258 # Prepend path to filename\r\n--> 259 pa_table = self._read_files(files, in_memory=in_memory)\r\n 260 # If original_instructions is not None, convert it to a human-readable NamedSplit\r\n 261 if original_instructions is not None:\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/arrow_reader.py in _read_files(self, files, in_memory)\r\n 192 f[\"filename\"] = os.path.join(self._path, f[\"filename\"])\r\n 193 for f_dict in files:\r\n--> 194 pa_table: Table = self._get_table_from_filename(f_dict, in_memory=in_memory)\r\n 195 pa_tables.append(pa_table)\r\n 196 pa_tables = [t for t in pa_tables if len(t) > 0]\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/arrow_reader.py in _get_table_from_filename(self, filename_skip_take, in_memory)\r\n 327 filename_skip_take[\"take\"] if \"take\" in filename_skip_take else None,\r\n 328 )\r\n--> 329 table = ArrowReader.read_table(filename, in_memory=in_memory)\r\n 330 if take == -1:\r\n 331 take = len(table) - skip\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/arrow_reader.py in read_table(filename, in_memory)\r\n 348 \"\"\"\r\n 349 table_cls = InMemoryTable if in_memory else MemoryMappedTable\r\n--> 350 return table_cls.from_file(filename)\r\n 351 \r\n 352 \r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/table.py in from_file(cls, filename, replays)\r\n 1034 @classmethod\r\n 1035 def from_file(cls, filename: str, replays=None):\r\n-> 1036 table = _memory_mapped_arrow_table_from_file(filename)\r\n 1037 table = cls._apply_replays(table, replays)\r\n 1038 return cls(table, filename, replays)\r\n\r\n/databricks/python/lib/python3.8/site-packages/datasets/table.py in _memory_mapped_arrow_table_from_file(filename)\r\n 48 def _memory_mapped_arrow_table_from_file(filename: str) -> pa.Table:\r\n 49 memory_mapped_stream = pa.memory_map(filename)\r\n---> 50 opened_stream = pa.ipc.open_stream(memory_mapped_stream)\r\n 51 pa_table = opened_stream.read_all()\r\n 52 return pa_table\r\n\r\n/databricks/python/lib/python3.8/site-packages/pyarrow/ipc.py in open_stream(source)\r\n 152 reader : RecordBatchStreamReader\r\n 153 \"\"\"\r\n--> 154 return RecordBatchStreamReader(source)\r\n 155 \r\n 156 \r\n\r\n/databricks/python/lib/python3.8/site-packages/pyarrow/ipc.py in __init__(self, source)\r\n 43 \r\n 44 def __init__(self, source):\r\n---> 45 self._open(source)\r\n 46 \r\n 47 \r\n\r\n/databricks/python/lib/python3.8/site-packages/pyarrow/ipc.pxi in pyarrow.lib._RecordBatchStreamReader._open()\r\n\r\n/databricks/python/lib/python3.8/site-packages/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()\r\n\r\n/databricks/python/lib/python3.8/site-packages/pyarrow/error.pxi in pyarrow.lib.check_status()\r\n\r\nArrowInvalid: Tried reading schema message, was null or length 0\r\n```\r\n\r\n", "looks like my arrow files are all empty @mariosasko \r\n\r\n\r\n\r\n\r\ni also see the `incomplete_info.lock` file a level up too. seems like the data isn't being persisted to disk when I call `download_and_prepare`. is there something else i need to do before then, perhaps?", "quick update @mariosasko. i got it working! i had to downgrade to `datasets==2.4.0`. testing other versions now and will let you know the results.", "I've tested with every version of `datasets>2.4.0` and i get the same error with all of them." ]

[ "note, that `datasets` already did some of that umask fixing in the past and also at the hub - the recent work on the hub about the same: https://github.com/huggingface/huggingface_hub/pull/1220\r\n\r\nAlso I noticed that each file has a .json counterpart and the latter always has the correct perms:\r\n\r\n```\r\n-rw------- 1 uue59kq cnw 173M Dec 9 01:37 537596e64721e2ae3d98785b91d30fda0360c196a8224e29658ad629e7303a4d\r\n-rw-rw---- 1 uue59kq cnw 101 Dec 9 01:37 537596e64721e2ae3d98785b91d30fda0360c196a8224e29658ad629e7303a4d.json\r\n```\r\n\r\nso perhaps cheating is possible and syncing the perms between the 2 will do the trick." ]

[ "As the survey is finished, can we close this issue, @LysandreJik ?", "Yes! I'll post a public summary on the forums shortly.", "Is the summary available? I would be interested in reading your findings." ]

[ "After some more investigation, this is due to [this line of code](https://github.com/huggingface/datasets/blob/main/src/datasets/features/audio.py#L279). The function `sf.read(file)` should be updated to `sf.read(file, dtype=\"float32\")`\r\n\r\nIndeed, the default value in soundfile is `float64` ([see here](https://pysoundfile.readthedocs.io/en/latest/#soundfile.read)). \r\n", "@qmeeus I agree, decoding of different audio formats should return the same dtypes indeed!\r\n\r\nBut note that here you are concatenating datasets with different sampling rates: 48000 for CommonVoice and 16000 for Voxpopuli. So you should cast them to the same sampling rate value before interleaving, for example:\r\n```\r\ncv = cv.cast_column(\"audio\", Audio(sampling_rate=16000))\r\n```\r\notherwise you would get the same error because features of the same column (\"audio\") are not the same.\r\n\r\nAlso, the error you get is unexpected. Could you please confirm that you use the latest main version of the `datasets`? We had an issue that could lead to an error like this after using `rename_column` method, but it was fixed in https://github.com/huggingface/datasets/pull/5287 ", "Hi Polina,\r\nSorry for the late answer\r\nIt is possible that the issue was due to a bug that is now fixed. I installed an editable version of datasets from github, but I don't recall whether I had updated it at the time of the issue. My research led me to other directions so I did not follow through on the interleave datasets.\r\n" ]

[]

[ "Hi @cbarond there's already an open issue at https://github.com/dair-ai/emotion_dataset/issues/5, as the data seems to be missing now, so check that issue instead 👍🏻 ", "Thanks @cbarond for reporting and @alvarobartt for pointing to the issue we opened in the author's repo.\r\n\r\nIndeed, this issue was first raised in the \"emotion\" dataset Community tab: https://huggingface.co/datasets/emotion/discussions/3\r\n\r\nI'm closing this issue and leave the issue above for the subsequent updates.\r\n\r\nDuplicate of: https://huggingface.co/datasets/emotion/discussions/3", "try using \"SetFit/emotion\" instead", "> try using \"SetFit/emotion\" instead\r\n\r\nI' replaced \"emotion\" with \"SetFit/Emotion\", but the code is getting stuck at\r\n\r\n`emotions = load_dataset(\"SetFit/emotion\")`\r\n\r\nI pause execution using the debugger, and it takes me to filelock.py:226\r\n\r\n`with self._thread_lock:`\r\n\r\nDo you know a way to get past this issue?", "thanks @honeyimholm - worked for me", "> try using \"SetFit/emotion\" instead\r\n\r\nIt really helps a lot, thank you!", "The dataset loading script has been fixed: https://huggingface.co/datasets/emotion/discussions/4" ]

[ "Hi !\r\n\r\n> It starts using all the cores (I am not sure why because I did not pass num_proc)\r\n\r\nThe tokenizer uses Rust code that is multithreaded. And maybe the `feature_extractor` might run some things in parallel as well - but I'm not super familiar with its internals.\r\n\r\n> then progress bar stops at every 1k steps. (starts using a single core)\r\n\r\nEvery 1000 examples we flush the processed examples to disk. It is this way because Arrow is a columnar format: you must write data chunk by chunk. The processing in on hold while writing right now - maybe this can be improved in the future.", "Hi @lhoestq \r\nThanks for the explanation! it was so helpful! Let me check why `feature_extractor` is running on multiple cpus." ]

[ "I like the idea of having `webdataset` as an optional dependency to ensure our loader generates web datasets the same way as the main project.", "Webdataset is the one of the most popular dataset formats for large scale computer vision tasks. Upvote for this issue. ", "Any updates on this?", "We haven't had the bandwidth to implement it so far, but if someone wants to give it a shot please don't hesitate ^^", "Done in #6391 " ]

[ "Should `datasets` allow `ClassLabel` input parameter to be an `np.array` even though internally we need to cast it to a Python list? @lhoestq @mariosasko ", "Hi! No, I don't think so. The `names` parameter is [annotated](https://github.com/huggingface/datasets/blob/582236640b9109988e5f7a16a8353696ffa09a16/src/datasets/features/features.py#L892) as `List[str]` (**NumPy arrays are not lists**), and considering that type checking is not a common practice in Python, I think we can leave the code as-is.", "I appreciate it is the wrong type, and that type checking is not common, but I think there's a few circumstances that make it a good idea from a usability perspective.\r\n\r\nIt's quite a difficult error to debug because it comes from a utility function (so it's not immediately obvious which parameter caused it). What makes it even more difficult is the exception happens when the features instance is used to instantiate the dataset, **not** when when the wrong type is actually passed when the features is instantiated. When I was debugging the error, I didn't really consider it could be an issue with the features instance because it had instantiated fine. It's also not one of the more common exceptions caused by trying to use a non-list as a list.\r\n\r\nIt's also relatively easy to accidentally get a numpy array of class types (e.g. calling `unique()` on a pandas dataframe column). Additionally, passing in a `set` instead of the list (again, relatively easy because people may run `set(classes)` to generate uniques) causes an error when the features instance is used, albeit a slightly more obvious one.\r\n\r\nThe names list is already being processed and validated in the `__post_init__` method anyway, so it would not really be adding any complexity to check it is actually a list here too. I'm happy to contribute this change if you change your mind about whether it's worthwhile.", "I agree that it's not easy to debug this issue, so perhaps we could add some basic type checking (e.g. `not isinstance(names, list)` -> error) to make debugging easier. Feel free to submit a PR.\r\n\r\n> Additionally, passing in a set instead of the list (again, relatively easy because people may run set(classes) to generate uniques) causes an error when the features instance is used, albeit a slightly more obvious one.\r\n\r\n`set` is an unordered structure (it's ordered in Python 3.6+, but this is CPython's implementation detail), and the order of ClassLabel `names` matters, so this doesn't require a fix.", "What about checking for `Sequence` instead? I think users can pass a list or a tuple as well." ]

[]

[ "Hi! I agree it makes sense for `IterableDataset.map` to support the `batch_size=None` case. This should be super easy to fix.", "@mariosasko as this is something simple maybe I can include it as part of https://github.com/huggingface/datasets/pull/5311? Let me know :+1:", "#self-assign", "Feel free to close this @lhoestq as part of https://github.com/huggingface/datasets/pull/5336 :hugs:", "Thanks again :)\r\n\r\n> For practical usages, an alternative to this would be to convert from an iterable dataset to a map-style dataset, but it is not obvious how to do this.\r\n\r\nThis is interesting as well, if anyone wants to explore" ]

[ "I agree we have a mess with docstrings...", "Ok, I believe we've cleaned up most of the old syntax we were using for the user-facing docs! There are still a couple of `:obj:`'s and `:class:` floating around in the docstrings we don't expose that I'll track down :)", "Hi @polinaeterna @albertvillanova @stevhliu, I hope you all are doing well.\r\n\r\nIs this issue still unresolved as I am interested in it?", "It should be mostly fixed for the user-facing APIs, but there may be some Sphinx syntax still lurking around in the non-public APIs. Feel free to open a PR to fix those if you catch any! 🤗 ", "Thanks for your reply @stevhliu :)\r\nSure, I will try to find out the remaining and fix that.\r\n\r\n" ]

[ "Thanks for reporting, @liaeh.\r\n\r\nWe are having a look at it. ", "I have transferred the discussion to the Community tab of the dataset: https://huggingface.co/datasets/taskmaster2/discussions/1" ]

[ "Hi ! ImageFolder is made for small scale datasets indeed. For large scale image datasets you better group your images in TAR archives or Arrow/Parquet files. This is true not just for ImageFolder loading performance, but also because having millions of files is not ideal for your filesystem or when moving the data around.\r\n\r\nOption 1. use TAR archives\r\n\r\nI'd suggest you to take a look at how we load [Imagenet](https://huggingface.co/datasets/imagenet-1k/tree/main) for example. The dataset is sharded in multiple TAR archives and there is a [script](https://huggingface.co/datasets/imagenet-1k/blob/main/imagenet-1k.py) that iterates over the archives to load the images.\r\n\r\nOption 2. use Arrow/Parquet\r\n\r\nYou can load your images as an Arrow Dataset with\r\n```python\r\nfrom datasets import Dataset, Image, load_from_disk, load_dataset\r\n\r\nds = Dataset.from_dict({\"image\": list(glob.glob(\"path/to/dir/**/*.jpg\"))})\r\n\r\ndef add_metadata(example):\r\n ...\r\n\r\nds = ds.map(add_metadata, num_proc=16) # num_proc for multiprocessing\r\nds = ds.cast_column(\"image\", Image())\r\n\r\n# save as Arrow locally\r\nds.save_to_disk(\"output_dir\")\r\nreloaded = load_from_disk(\"output_dir\")\r\n\r\n# OR save as Parquet on the HF Hub\r\nds.push_to_hub(\"username/dataset_name\")\r\nreloaded = load_dataset(\"username/dataset_name\")\r\n# reloaded = load_dataset(\"username/dataset_name\", num_proc=16) # to use multiprocessing\r\n```\r\n\r\nPS: maybe we can actually have something similar to ImageFolder but for image archives at one point ?", "@lhoestq Thanks!\r\n\r\nPerhaps it'd be worth adding a note on the documentation that `ImageFolder` is not intended for large datasets? This limitation is not intuitively obvious to someone who has not used it before, I think.", "Thanks for the feedback @salieri! I opened #5329 to make it clear `ImageFolder` is not intended for large datasets. Please feel free to comment if you have any other feedback! 🙂 " ]

[ "Thanks for reporting, @adampauls.\r\n\r\nWe are having a look at it. " ]

[ "EDIT:\r\nI think in this case, the metadata files (either README or JSON) should not be read (i.e. `self.info.splits` should be None).\r\n\r\nOne idea: \r\n- I think ideally we should set this behavior when we pass `--save_info` to the CLI `test`\r\n- However, currently, the builder is unaware of this: `save_info` arg is not passed to it", "> I think in this case\r\n\r\n@albertvillanova You mean in cases when the script was changed? \r\n\r\nI suggest that we:\r\n* add a check on the slice (like 'split_name[n%]) kind of format here: https://github.com/huggingface/datasets/blob/main/src/datasets/splits.py#L523 to catch things like this. \r\n* Error here happens before splits verification, but in `_prepare_split`, and `_prepare_split` doesn't perform any verification and don't know about it. so we can pass this parameter and take splits from `split_generator`, not from `split.info` in case when `verify_infos` is False\r\n* we can check if split **names** from split_generators and self.info.splits are the same **before** preparing splits (if `verify_info=True`) so that we don't spend time on generating unwanted data. \r\n* provide some user-friendly warnings about `ignore_verifications` parameter so that users know that if something is not matching they can ignore it\r\n\r\nI started it here: https://github.com/huggingface/datasets/pull/5327/files\r\n\r\nWhat do you think @albertvillanova ?", "I edited my previous comment:\r\n- First I proposed setting `self.info.splits` to None when `ignore_verifications=True`\r\n - I thought it was the easiest implementation because `ignore_verifications` is passed to `DatasetBuilder.download_and_prepare`\r\n - However, afterwards, I realized this might not be a good idea for this use case:\r\n - A user wants to optimize the loading of the dataset, and passes `ignore_verifications=False` to avoid all the verifications\r\n - In this case, we want `self.info.splits` to be read from metadata file\r\n- Then, I thought that it might be better to set `self.info.splits` to None when we pass `--save_info` to the CLI test: if we are going to save the info to the metadata file, it makes no sense to read the info from the metadata file\r\n - This implementation is not so easy because the Builder knows nothing about `--save_info`\r\n\r\nI agree with you there are 2 things to be addressed here:\r\n- One is what I have just commented: `self.info.splits` should be None in this case\r\n- The other, a validation should be implemented when calling `make_file_instructions` and/or `SplitDict.__getitem__`, so that when passing \"training\" to it, we get a more descriptive error other than `TypeError: expected str, bytes or os.PathLike object, not NoneType` " ]

[ "This seems similar to https://github.com/huggingface/datasets/issues/2512 Can you try to update seqeval ? ", "@JonathanAlis also note that the metrics are deprecated in our `datasets` library.\r\n\r\nPlease, use the new library 🤗 Evaluate instead: https://huggingface.co/docs/evaluate" ]

[ "Forgot to actually convert the feature dict to a Feature object. Closing." ]

[ "Thanks for reporting @JoelNiklaus.\r\n\r\nPlease note that since we moved all dataset loading scripts to the Hub, the issues and pull requests relative to specific datasets are directly handled on the Hub, in their Community tab. I'm transferring this issue there: https://huggingface.co/datasets/joelito/mc4_legal/discussions\r\n\r\nI am also having a look at the bug in your script.", "Issue transferred to: https://huggingface.co/datasets/joelito/mc4_legal/discussions/1" ]

[ "The [timit_asr.py](https://huggingface.co/datasets/timit_asr/blob/main/timit_asr.py) script iterates over the WAV files per split directory using this:\r\n```python\r\nwav_paths = sorted(Path(data_dir).glob(f\"**/{split}/**/*.wav\"))\r\nwav_paths = wav_paths if wav_paths else sorted(Path(data_dir).glob(f\"**/{split.upper()}/**/*.WAV\"))\r\n```\r\n\r\nCan you check that there is a directory named \"test\" somewhere in your timit data directory ?" ]

[]

[]

[ "Hi ! Thanks for reporting. Indeed the lock file should be placed in a directory with write permission (e.g. in the directory where the archive is extracted).", "I opened https://github.com/huggingface/datasets/pull/5320 to fix this - it places the lock file in the cache directory instead of trying to put in next to the ZIP where it's read-only" ]

[]

[ "- Build docs for 2.6.2:\r\n - Commit: a6a5a1cf4cdf1e0be65168aed5a327f543001fe8\r\n - Build docs GH Action: https://github.com/huggingface/datasets/actions/runs/3539470622/jobs/5941404044\r\n- Build docs for 2.7.1:\r\n - Commit: 5ef1ab1cc06c2b7a574bf2df454cd9fcb071ccb2\r\n - Build docs GH Action: https://github.com/huggingface/datasets/actions/runs/3539574442/jobs/5941636792" ]

[ "Hi ! JSON is a lossy format indeed. If you want to keep the feature types or other metadata I'd encourage you to store them as well. For example you can use `dataset.info.write_to_directory` and `DatasetInfo.from_directory` to store the feature types, split info, description, license etc." ]

[ "I found a solution \r\n\r\nIf you specifically install datasets==1.18 and then run\r\n\r\nimport datasets\r\nwiki = datasets.load_dataset('wikipedia', '20200501.en')\r\nthen this should work (it worked for me.)", "I have the same problem here but installing datasets==1.18 wont work for me\r\n" ]

[ "Related to https://github.com/huggingface/datasets/issues/5245", "#self-assign", "Thanks @lhoestq and @alvarobartt!\r\n\r\nThis would be extremely helpful to have working for the Whisper fine-tuning event - we're **only** training using streaming mode, so it'll be quite important to have this feature working to make training as easy as possible!\r\n\r\n_c.f._ https://twitter.com/sanchitgandhi99/status/1592188332171493377", "> Thanks @lhoestq and @alvarobartt!\n> \n> \n> \n> This would be extremely helpful to have working for the Whisper fine-tuning event - we're **only** training using streaming mode, so it'll be quite important to have this feature working to make training as easy as possible!\n> \n> \n> \n> _c.f._ https://twitter.com/sanchitgandhi99/status/1592188332171493377\n\nI'm almost done with at least a temporary fix to `rename_column`, `rename_columns`, and `remove_columns`, just trying to figure out how to extend it to the `map` function itself!\n\nI'll probably open the PR for review either tomorrow or Sunday hopefully! Glad I can help you and HuggingFace 🤗 ", "Awesome - thank you so much for this PR @alvarobartt! Is much appreciated!", "@sanchit-gandhi PR is ready and open for review at #5287, but there's still one issue I may need @lhoestq's input :hugs:", "Let us know @sanchit-gandhi if you need a new release of `datasets` soon with this fix included :)", "Thanks for the fix guys! We can direct people to install `datasets` from main if that's easier!", "Hey guys, any update around this? I'm facing the same issue with a streamable dataset. ", "Hi @asennoussi so this was already fixed and released as part of https://github.com/huggingface/datasets/releases/tag/2.8.0, so you should be able to install it as `pip install datasets==2.8.0` or just to use `pip install datasets --upgrade` to get the latest version, as of now, the https://github.com/huggingface/datasets/releases/tag/2.9.0 released last week! 🤗", "Still facing the same issue though: \r\n```\r\nfrom datasets import IterableDatasetDict, load_dataset\r\n\r\nraw_datasets = vectorized_datasets = IterableDatasetDict()\r\n\r\n\r\nraw_datasets[\"train\"] = load_dataset(\"asennoussi/private\", split=\"train\", use_auth_token=True, streaming=True)\r\nraw_datasets[\"test\"] = load_dataset(\"asennoussi/private\", split=\"test\", use_auth_token=True, streaming=True)\r\n\r\nprint(\"Original features: \", raw_datasets['train'].features.keys())\r\n\r\n...\r\n\r\ndef prepare_dataset(batch):\r\n\r\n # load and (possibly) resample audio datato 16kHz\r\n audio = batch[\"audio\"]\r\n\r\n # compute log-Mel input features from input audio array \r\n batch[\"input_features\"] = processor.feature_extractor(audio[\"array\"], sampling_rate=audio[\"sampling_rate\"]).input_features[0]\r\n # compute input length of audio sample in seconds\r\n batch[\"input_length\"] = len(audio[\"array\"]) / audio[\"sampling_rate\"]\r\n \r\n # optional pre-processing steps\r\n transcription = batch[\"sentence\"]\r\n \r\n # encode target text to label ids\r\n batch[\"labels\"] = processor.tokenizer(transcription).input_ids\r\n batch[\"labels_length\"] = len(batch[\"labels\"])\r\n return batch\r\n...\r\nvectorized_datasets = vectorized_datasets.remove_columns(['input_length', 'labels_length']+list(next(iter(raw_datasets.values())).features))\r\nprint(\"Processed features: \", vectorized_datasets['train'].features)\r\nprint(\"First sample:\", next(iter(vectorized_datasets['train'])))\r\n\r\n```\r\n\r\nOutput: \r\n```\r\nOriginal features: dict_keys(['path', 'audio', 'sentence'])\r\nProcessed features: None\r\n```", "Hmm weird, could you try to print\r\n\r\n```python\r\nprint(\"Processed features: \", vectorized_datasets['train'].features)\r\n```\r\n\r\nagain after iterating over the `vectorized_datasets`? In the code above, should be last line :)", "Didn't seem to fix it: \r\n```\r\nOriginal features: dict_keys(['path', 'audio', 'sentence'])\r\nProcessed features: None\r\nProcessed features: None\r\n```", "Actually the culprit looks to be this one: \r\n`vectorized_datasets = raw_datasets.map(prepare_dataset).with_format(\"torch\")`\r\nWhen I remove this line: `vectorized_datasets = vectorized_datasets.remove_columns(['input_length', 'labels_length']+list(next(iter(raw_datasets.values())).features))`\r\n\r\nI still get \r\n```\r\nProcessed features: None\r\n```", "The culprit is definitely `.map` \r\nJust validated it. \r\nAny idea please? ", "> The culprit is definitely `.map` Just validated it. Any idea please?\r\n\r\nYes, indeed `.map` losses the features, because AFAIK pre-fetching the data to infer the features is expensive and not ideal, that's part of this issue https://github.com/huggingface/datasets/issues/3888\r\n\r\nAnyway, now you can pass the `features` as a param to `.map` as follows:\r\n\r\n```python\r\nfrom datasets import Features\r\nvectorized_datasets = raw_datasets.map(\r\n prepare_dataset,\r\n features=Features(\r\n {\"path\": raw_datasets[\"train\"].info.features[\"path\"], \"audio\": raw_datasets[\"train\"].info.features[\"audio\"], \"sentence\": raw_datasets[\"train\"].info.features[\"sentence\"]}\r\n ),\r\n).with_format(\"torch\")\r\n```\r\n\r\nAlso, to let you know, when calling `.remove_columns` over an `IterableDataset`, the `features` are not lost, as well as `.rename_column` and `rename_columns` :)\r\n\r\nMore information about the latter at https://github.com/huggingface/datasets/pull/5287", "@asennoussi alternatively you can just call `._resolve_features()` from your `IterableDataset` and it will pre-fetch the data to resolve the features, but note that feature-inference is not as accurate as if you manually specify which features and feature-types the `IterableDataset` has, as mentioned in the comment above, the alternative is to provide `features` param to `.map` :hugs:", "Got it thanks a lot! " ]

[ "Or for example an archive on GitHub releases! Before I added support for JXL (locally only, PR still pending) I was considering hosting my files on GitHub instead...", "+1 to this. I would like to use 'audiofolder' with a data_dir that's on S3, for example. I don't want to upload my dataset to the Hub, but I would find all the fingerprinting/caching features useful.", "Adding to the conversation, Dask also uses `fsspec` for this feature.\r\n\r\n[Dask: How to connect to remote data](https://docs.dask.org/en/stable/how-to/connect-to-remote-data.html)\r\n\r\nHappy to help on this feature :D ", "+1 to this feature request since I think it also tackles my use-case. I am collaborating with a team, working with a loading script which takes some time to generate the dataset artifacts. It would be very handy to use this as a cloud cache to avoid duplicating the effort. \r\n\r\nCurrently we could use `builder.download_and_prepare(path_to_cloud_storage, storage_options, ...)` to cache the artifacts to cloud storage, but then `builder.as_dataset()` yields `NotImplementedError: Loading a dataset cached in SomeCloudFileSystem is not supported`", "Makes sense ! If you want to load locally a dataset that you download_and_prepared on a cloud storage, you would use `load_dataset(path_to_cloud_storage)` indeed. It would download the data from the cloud storage, cache them locally, and return a `Dataset`.", "It seems currently the `cached_path` function handles all URLs by `get_from_cache` that only supports `ftp` and `http(s)` here:\r\nhttps://github.com/huggingface/datasets/blob/b5672a956d5de864e6f5550e493527d962d6ae55/src/datasets/utils/file_utils.py#L181\r\n\r\nI guess one can add another condition that handles `s3://` or `gs://` URLs via `fsspec` here.", "I could use this functionality, so I put together a PR using @kyamagu's suggestion to use `fsspec` in `datasets.utils.file_utils`\r\n\r\nhttps://github.com/huggingface/datasets/pull/5580", "Thanks @dwyatte for adding support for fsspec urls\r\n\r\nLet me just reopen this since the original issue is not resolved", "I'm not yet understanding how to use https://github.com/huggingface/datasets/pull/5580 in order to use `load_dataset(data_files=\"s3://...\")`. Any help/example would be much appreciated :) thanks! ", "It's still not officially supported x) But you can try to update `request_etag` in `file_utils.py` to use `fsspec_head` instead of `http_head`. It is responsible of getting the ETags of the remote files for caching. This change may do the trick for S3 urls", "Thank you for your guys help on this and merging in #5580. I manually pulled the changes to my local datasets package (datasets.utils.file_utils.py) since it only seemed to be this file that was changed in the PR and I'm getting the error: \r\nInvalidSchema: No connection adapters were found for 's3://bucket/folder/'. I'm calling load_dataset using the S3 URI. When I use the S3 URL I get HTTPError: 403 Client Error. \r\nAm I not supposed to use the S3 URI? How do I pull in the changes from this merge? I'm running datasets 2.10.1. ", "The current implementation depends on gcsfs/s3fs being able to authenticate through some other means e.g., environmental variables. For AWS, it looks like you can set `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, and `AWS_SESSION_TOKEN`\r\n\r\nNote that while testing this just now, I did note a discrepancy between gcsfs and s3fs that we might want to address where gcsfs passes the timeout from `storage_options` [here](https://github.com/huggingface/datasets/blob/3e6269979fc80ae8939294d26298897f0db5b84d/src/datasets/utils/file_utils.py#L333) down into the `aiohttp.ClientSession.request`, but s3fs does not handle this (tries to pass to the `aiobotocore.session.AioSession` constructor raising `TypeError: __init__() got an unexpected keyword argument 'requests_timeout'`).\r\n\r\nIt seems like some work trying to unify kwargs across different fsspec implementations, so if the plan is to pass down `storage_options`, I wonder if we should just let users control the timeout (and other kwargs) using that and if not specified, use the default?", "> Note that while testing this just now, I did note a discrepancy between gcsfs and s3fs that we might want to address where gcsfs passes the timeout from storage_options [here](https://github.com/huggingface/datasets/blob/3e6269979fc80ae8939294d26298897f0db5b84d/src/datasets/utils/file_utils.py#L333) down into the aiohttp.ClientSession.request, but s3fs does not handle this (tries to pass to the aiobotocore.session.AioSession constructor raising TypeError: __init__() got an unexpected keyword argument 'requests_timeout').\r\n\r\n> It seems like some work trying to unify kwargs across different fsspec implementations, so if the plan is to pass down storage_options, I wonder if we should just let users control the timeout (and other kwargs) and if not specified, use the default?\r\n\r\n@lhoestq here's a small PR for this: https://github.com/huggingface/datasets/pull/5673\r\n\r\n", "@lhoestq sorry for being a little dense here but I am very keen to use fsspec / adlfs for for a larger image dataset I have for object detection. I have to keep it on Azure storage and would also like to avoid a full download or zipping (so use `load_dataset(..., streaming=True)`. So this development is godsend :) only... I am unable to make it work.\r\n\r\nWould you expect the setup to work for:\r\n- azure blob storage\r\n- image files (not the standard formats `json`, `parquet`....\r\n\r\n? I appreciate that you mostly focus on s3 but it seems that, similar to the [remaining cloud storage functionality](https://huggingface.co/docs/datasets/filesystems#cloud-storage), it should also work for Azure blob storage.\r\n\r\n\r\nI would imagine that something like (Streaming true or false):\r\n```python\r\nd = load_dataset(\"new_dataset.py\", storage_options=storage_options, split=\"train\")\r\n```\r\nwould work with \r\n```python\r\n# new_dataset.py\r\n....\r\n_URL=\"abfs://container/image_folder``` \r\n\r\narchive_path = dl_manager.download(_URL)\r\nsplit_metadata_paths = dl_manager.download(_METADATA_URLS)\r\nreturn [\r\n datasets.SplitGenerator(\r\n name=datasets.Split.TRAIN,\r\n gen_kwargs={\r\n \"annotation_file_path\": split_metadata_paths[\"train\"],\r\n \"files\": dl_manager.iter_files(archive_path)\r\n},\r\n ),\r\n...\r\n```\r\nbut I get \r\n```\r\nTraceback (most recent call last):\r\n... ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/load.py\", line 1797, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/builder.py\", line 890, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/builder.py\", line 1649, in _download_and_prepare\r\n super()._download_and_prepare(\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/builder.py\", line 963, in _download_and_prepare\r\n split_generators = self._split_generators(dl_manager, **split_generators_kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"~/.cache/huggingface/modules/datasets_modules/datasets/new_dataset/dd26a081eab90074f41fa2c821b458424fde393cc73d3d8241aca956d1fb3aa0/new_dataset_script.py\", line 56, in _split_generators\r\n archive_path = dl_manager.download(_URL)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/download/download_manager.py\", line 427, in download\r\n downloaded_path_or_paths = map_nested(\r\n ^^^^^^^^^^^\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/utils/py_utils.py\", line 435, in map_nested\r\n return function(data_struct)\r\n ^^^^^^^^^^^^^^^^^^^^^\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/download/download_manager.py\", line 453, in _download\r\n return cached_path(url_or_filename, download_config=download_config)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"~/miniconda3/envs/hf/lib/python3.11/site-packages/datasets/utils/file_utils.py\", line 206, in cached_path\r\n raise ValueError(f\"unable to parse {url_or_filename} as a URL or as a local path\")\r\nValueError: unable to parse abfs://container/image_folder as a URL or as a local path\r\n\r\n```\r\n", "What version of `datasets` are you using ?", "@lhoestq \r\nhello, i still have problem with loading json from S3:\r\n\r\n\r\nstorage_options = {\r\n \"key\": xxxx,\r\n \"secret\": xxx,\r\n \"endpoint_url\": xxxx\r\n}\r\npath = 's3://xxx/xxxxxxx.json'\r\ndataset = load_dataset(\"json\", data_files=path, storage_options=storage_options)\r\n\r\nand it throws an error:\r\nTypeError: AioSession.__init__() got an unexpected keyword argument 'hf'\r\nand I use the lastest 2.14.4_dev0 version", "Hi @lhoestq, thanks for getting back to me :) you have been busy over the summer I see... I was on `2.12.0`. I have updated to `2.14.4`. \r\n\r\nNow `d = load_dataset(\"new_dataset.py\", storage_options=storage_options, split=\"train\", streaming=True)` works for Azure blob storage (with a local data loader script) when I explicitly list all blobs (I am struggling to make `fs.ls(<path>)` work in the script to make the list available to the download manager).\r\n\r\n Any chance that it could work out-of-the-box by supplying just the image folder, not the full list of image filenames? It seems that `dl_manager.download(_URL)` always wants one or more (possibly archived) files. In my situation, where I don't want to archive or download, it would be great to just supply the folder (seems reasonably doable with fsspec).\r\n\r\nLet me know if there is anything I can do to help.\r\n\r\nThanks,", "> Any chance that it could work out-of-the-box by supplying just the image folder, not the full list of image filenames? It seems that dl_manager.download(_URL) always wants one or more (possibly archived) files. In my situation, where I don't want to archive or download, it would be great to just supply the folder (seems reasonably doable with fsspec).\r\n\r\n@mayorblock This is not supported right now, you have to use archives or implement a way to get the list by yourself\r\n\r\n> TypeError: AioSession.init() got an unexpected keyword argument 'hf'\r\n\r\n@hjq133 Can you update `fsspec` and try again ?\r\n\r\n```python\r\npip install -U fsspec\r\n```", "thanks for your suggestion，it works now ! ", "I'm seeing same problem as @hjq133 with following versions:\r\n\r\n```\r\ndatasets==2.15.0\r\n(venv) ➜ finetuning-llama2 git:(main) ✗ pip freeze | grep s3fs \r\ns3fs==2023.10.0\r\n(venv) ➜ finetuning-llama2 git:(main) ✗ pip freeze | grep fsspec\r\nfsspec==2023.10.0\r\n```", "> @lhoestq hello, i still have problem with loading json from S3:\r\n> \r\n> storage_options = { \"key\": xxxx, \"secret\": xxx, \"endpoint_url\": xxxx } path = 's3://xxx/xxxxxxx.json' dataset = load_dataset(\"json\", data_files=path, storage_options=storage_options)\r\n> \r\n> and it throws an error: TypeError: AioSession.**init**() got an unexpected keyword argument 'hf' and I use the lastest 2.14.4_dev0 version\r\n\r\nI am trying to do the same thing, but the loading is just hanging, without any error. \r\n@lhoestq is there any documentation how to load from private s3 buckets?\r\n", "Hi ! S3 support is still experimental. It seems like there is an extra `hf` field passed to the `s3fs` storage_options that causes this error. I just check the source code of `_prepare_single_hop_path_and_storage_options` and I think you can try passing explicitly your own `storage_options={\"s3\": {...}}`. Also note that it's generally better to load datasets from HF (we run extensive tests and benchmarks for speed and robustness)", "That worked! Thanks\r\nIt seems thought that `data_dir=...` doesn't work on s3, only `data_files`.\r\n", "@lhoestq Would this work either with an Azure Blob Storage Container or its respective Azure Machine Learning Datastore? If yes, what would that look like in code? I've tried a couple of combinations but no success so far, on the latest version of `datasets`. I need to migrate a dataset to the Azure cloud, `load_dataset(\"path_to_data\")` worked perfectly while the files were local only. Thank you!\r\n\r\n@mayorblock would you mind sharing how you got it to work? What did you pass as `storage_options`? Would it maybe work without a custom data loader script?", "This ticket would be of so much help.", "@lhoestq I've been using this feature for the last year on GCS without problem, but I think we need to fix an issue with S3 and then document the supported calling patterns to reduce confusion\r\n\r\nIt looks like `datasets` uses a default `DownloadConfig` which is where some potentially unintended storage options are getting passed to fsspec\r\n\r\n```\r\nDownloadConfig(\r\n\tcache_dir=None, \r\n\tforce_download=False, \r\n\tresume_download=False, \r\n\tlocal_files_only=False, \r\n\tproxies=None, \r\n\tuser_agent=None, \r\n\textract_compressed_file=False, \r\n\tforce_extract=False, \r\n\tdelete_extracted=False, \r\n\tuse_etag=True, \r\n\tnum_proc=None, \r\n\tmax_retries=1, \r\n\ttoken=None, \r\n\tignore_url_params=False, \r\n\tstorage_options={'hf': {'token': None, 'endpoint': 'https://huggingface.co'}}, \r\n\tdownload_desc=None\r\n)\r\n```\r\n(specifically the `storage_options={'hf': {'token': None, 'endpoint': 'https://huggingface.co'}}` part)\r\n\r\n`gcsfs` is robust to the extra key in storage options for whatever reason, but `s3fs` is not (haven't dug into why). I'm unable to test `adlfs` but it looks like people here got it working\r\n\r\nIs this an issue that needs fixed with s3fs? Or can we avoid passing these default storage options in some cases?\r\n\r\nUpdate: I think probably https://github.com/huggingface/datasets/pull/6127 is where these default storage options were introduced", "Hmm not sure, maybe it has to do with `_prepare_single_hop_path_and_storage_options` returning the \"hf\" storage options when it shouldn't", "Also running into this issue downloading a parquet dataset from S3 (Upload worked fine using [current main branch](https://github.com/huggingface/datasets/commit/90b896193a0a315789022580eb4c80305d168d4d)).\r\n`dataset = Dataset.from_parquet('s3://path-to-file')`\r\nraises \r\n`TypeError: AioSession.__init__() got an unexpected keyword argument 'hf'`\r\nFound that the issue is introduced in [#6028](https://github.com/huggingface/datasets/commit/14f6edd9222e577dccb962ed5338b79b73502fa5#diff-8a868b8c1127c5cbe7870bbeec30bd2cbec6696a378e65b66782a6935e3f412e)\r\n\r\nWhen commenting out the __post_init__ part to set 'hf', I am able to download the dataset." ]

[ "Hi ! Can you \r\n```python\r\nimport platform\r\nprint(platform.python_version())\r\n```\r\nto see that it returns ?", "Hi,\n\n3.8.13\n\nGet Outlook for Android<https://aka.ms/AAb9ysg>\n________________________________\nFrom: Quentin Lhoest ***@***.***>\nSent: Tuesday, November 22, 2022 2:37:02 PM\nTo: huggingface/datasets ***@***.***>\nCc: feketedavid1012 ***@***.***>; Author ***@***.***>\nSubject: Re: [huggingface/datasets] Import error (Issue #5280)\n\n\nHi ! Can you\n\nimport platform\nprint(platform.python_version())\n\nto see that it returns ?\n\n—\nReply to this email directly, view it on GitHub<https://github.com/huggingface/datasets/issues/5280#issuecomment-1323691385>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AJW7F5YGG32W6WABYC25NJTWJTD75ANCNFSM6AAAAAASHZJ2AU>.\nYou are receiving this because you authored the thread.Message ID: ***@***.***>\n", "Then it should work as expected if you use the same python when using `datasets`\r\n\r\nPlease make sure you're running your code in the right environment", "It's the right environment. But in if statement I have\n\"3.8.13\" < 3.7\nAnd in the error message is Python>=3.7 which is true in my case (3.8.13 is greater then 3.7), so I don't understand my python should be below the 3.7 which case the if statement is right, but the message is wrong, or above 3.7 which case if statement is wrong, error message is right.\n\nGet Outlook for Android<https://aka.ms/AAb9ysg>\n________________________________\nFrom: Quentin Lhoest ***@***.***>\nSent: Tuesday, November 22, 2022 2:41:43 PM\nTo: huggingface/datasets ***@***.***>\nCc: feketedavid1012 ***@***.***>; Author ***@***.***>\nSubject: Re: [huggingface/datasets] Import error (Issue #5280)\n\n\nThen it should work as expected if you use the same python when using datasets\n\nPlease make sure you're running your code in the right environment\n\n—\nReply to this email directly, view it on GitHub<https://github.com/huggingface/datasets/issues/5280#issuecomment-1323697094>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AJW7F54JURTAJJWWDO2QGI3WJTERPANCNFSM6AAAAAASHZJ2AU>.\nYou are receiving this because you authored the thread.Message ID: ***@***.***>\n", "If you're having an error then you're not running your code in the right environment." ]

[ "Can you try to remove \"drop_labels=false\" ? It may force the loader to infer the labels instead of reading the metadata", "Hi, thanks for responding. I tried that, but it does not change anything.", "Can you try updating `datasets` ? Metadata support was added in `datasets` 2.4", "Probably the issue, will report back asap!", "Okay, now it seems to actually load the metadata and create the train_split, but it still says only returns \"image\" and \"label\", which is always 0 since all images are from same folder", "> Can you try updating `datasets` ? Metadata support was added in `datasets` 2.4\r\n\r\nUpdate: This was the issue." ]

[ "Sounds like one of the file is not a valid one, can you make sure you uploaded valid mp3 files ?", "Well I just sharded the original commonVoice dataset and pushed a small chunk of it in a private rep\n\nWhat did go wrong?\n\nHolen Sie sich Outlook für iOS<https://aka.ms/o0ukef>\n________________________________\nVon: Quentin Lhoest ***@***.***>\nGesendet: Tuesday, November 22, 2022 3:03:40 PM\nAn: huggingface/datasets ***@***.***>\nCc: capsabogdan ***@***.***>; Author ***@***.***>\nBetreff: Re: [huggingface/datasets] Bug in downloading common_voice data and snall chunk of it to one's own hub (Issue #5276)\n\n\nSounds like one of the file is not a valid one, can you make sure you uploaded valid mp3 files ?\n\n—\nReply to this email directly, view it on GitHub<https://github.com/huggingface/datasets/issues/5276#issuecomment-1323727434>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ALSIFOAPAL2V4TBJTSPMAULWJTHDZANCNFSM6AAAAAASHQJ63U>.\nYou are receiving this because you authored the thread.Message ID: ***@***.***>\n", "It should be all good then !\r\nCould you share a link to your repository for me to investigate what went wrong ?", "https://huggingface.co/datasets/DTU54DL/common-voice-test16k\n\nAm Di., 22. Nov. 2022 um 16:43 Uhr schrieb Quentin Lhoest <\n***@***.***>:\n\n> It should be all good then !\n> Could you share a link to your repository for me to investigate what went\n> wrong ?\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5276#issuecomment-1323876682>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ALSIFOEUJRZWXAM7DYA5VJDWJTS3NANCNFSM6AAAAAASHQJ63U>\n> .\n> You are receiving this because you authored the thread.Message ID:\n> ***@***.***>\n>\n", "I see ! This is a bug with MP3 files.\r\n\r\nWhen we store audio data in parquet, we store the bytes and the file name. From the file name extension we know if it's a WAV, an MP3 or else. But here it looks like the paths are all None.\r\n\r\nIt looks like it comes from here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/7feeb5648a63b6135a8259dedc3b1e19185ee4c7/src/datasets/features/audio.py#L212\r\n\r\nCc @polinaeterna maybe we should simply put the file name instead of None values ?", "@lhoestq I remember we wanted to avoid storing redundant data but maybe it's not that crucial indeed to store one more string value. \r\nOr we can store paths only for mp3s, considering that for other formats we don't have such a problem with reading from bytes without format specified. ", "It doesn't cost much to always store the file name IMO", "thanks for the help!\n\ncan I do anything on my side? we are doing a DL project and we need the\ndata really quick.\n\nthanks\nbogdan\n\n> Message ID: ***@***.***>\n>\n", "I opened a pull requests here: https://github.com/huggingface/datasets/pull/5285, we'll do a new release soon with this fix.\r\n\r\nOtherwise if you're really in a hurry you can install `datasets` from this PR", "[image: image.png]\n\n> Message ID: ***@***.***>\n>\n", "any idea on what's going wrong here?\n\nthanks\n\nAm So., 27. Nov. 2022 um 13:53 Uhr schrieb Bogdan Capsa <\n***@***.***>:\n\n> [image: image.png]\n>\n>> Message ID: ***@***.***>\n>>\n>\n", "hi @capsabogdan! \r\ncould you please share more specifically what problem do you have now?", "I have attached this screenshot above . can u pls help? So can not pip from pull request\r\n\r\n\r\n", "The pull request has been merged on `main`.\r\nYou can install `datasets` from `main` using\r\n```\r\npip install git+https://github.com/huggingface/datasets.git\r\n```", "I've tried to load this dataset DTU54DL/common-voice-test16k, but am\ngetting the same error.\n\nSo the bug fix will fix only if I upload a new dataset, or also loading\npreviously uploaded datasets?\n\nthanks\n\nAm Mo., 28. Nov. 2022 um 19:51 Uhr schrieb Quentin Lhoest <\n***@***.***>:\n\n> The pull request has been merged on main.\n> You can install datasets from main using\n>\n> pip install git+https://github.com/huggingface/datasets.git\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5276#issuecomment-1329587334>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ALSIFOCNYYIGHM2EX3ZIO6DWKT5MXANCNFSM6AAAAAASHQJ63U>\n> .\n> You are receiving this because you were mentioned.Message ID:\n> ***@***.***>\n>\n", "> So the bug fix will fix only if I upload a new dataset, or also loading\r\npreviously uploaded datasets?\r\n\r\nYou have to reupload the dataset, sorry for the inconvenience", "thank you so much for the help! works like a charm!\n\nAm Di., 29. Nov. 2022 um 12:15 Uhr schrieb Quentin Lhoest <\n***@***.***>:\n\n> So the bug fix will fix only if I upload a new dataset, or also loading\n> previously uploaded datasets?\n>\n> You have to reupload the dataset, sorry for the inconvenience\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5276#issuecomment-1330468393>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ALSIFOBKEFZO57BAKY4IGW3WKXQUZANCNFSM6AAAAAASHQJ63U>\n> .\n> You are receiving this because you were mentioned.Message ID:\n> ***@***.***>\n>\n" ]

[ "@huggingface/datasets if you agree, I can make the bulk edit on the Hub to fix integer keys into strings.", "Ok for me, and we can merge (internal) https://github.com/huggingface/moon-landing/pull/4609", "FYI there are still 2k+ weekly users on `datasets` 2.6.1 which doesn't support the string label format for class labels. And among those, some are using datasets with class labels like imdb (60 users), conllpp (40), msra_ner (40), peoples_daily_enr (40), weibo_ner (30), conll2003 (20), etc. And renaming to string would break these users code.", "but isn't `datasets 2.6.1` downloading files from the Hub with the corresponding tag? I thought we had something like this before", "We're using `main` as models do. Some datasets need to be updated from time to time, e.g. when a link to download the data is dead.\r\n\r\nBut yea a year ago we had those tags, we just ended up not using them", "I opened https://github.com/huggingface/datasets/issues/5406 to communicate on this. Let me know what you think, and if it sounds good to you I can pin this issue", "So, is it OK to make the bulk edit on the Hub now or should we wait longer? If the latter, how long?", "I think we can do it. If you want to be extra cautious you can do it for all datasets except imdb and conllpp for now which are actively used by 2.6.1 users. For those two we can keep the YAML like this for some more time, or alternatively use the old dataset_infos.json file", "The bulk edit of canonical datasets (except imdb and conllpp) is running. \r\n\r\nSee e.g.: https://huggingface.co/datasets/acronym_identification/discussions/3\r\n\r\nEDITED: \r\nDone, except for \"universal_morphologies\", where I get\r\n```\r\nHTTPError: 413 Client Error: Payload Too Large for url: https://huggingface.co/api/validate-yaml\r\n```\r\n\r\nAlso not done for the datasets missing matadata \"dataset_info\":\r\n- mc4: https://huggingface.co/datasets/mc4/discussions/3\r\n- the_pile: https://huggingface.co/datasets/the_pile/discussions/6\r\n- timit_asr: https://huggingface.co/datasets/timit_asr/discussions/1", "Thank you !", "@lhoestq, there are 6 community datasets with YAML integer keys in their `dataset_info` `class_label`:\r\n- indonlp/indonlu\r\n- rcds/swiss_judgment_prediction\r\n- Jean-Baptiste/wikiner_fr\r\n- Bingsu/Cat_and_Dog\r\n- taskydata/tasky_or_not\r\n- RCC-MSU/collection3\r\n\r\nMaybe we could open a PR on them as well?", "Let's do this then:\r\n\r\n- [x] [indonlp/indonlu](https://huggingface.co/datasets/indonlp/indonlu/discussions/3)\r\n- [x] rcds/swiss_judgment_prediction\r\n- [x] Jean-Baptiste/wikiner_fr\r\n- [x] Bingsu/Cat_and_Dog -> merged\r\n- [x] taskydata/tasky_or_not (was already using quotes)\r\n- [x] RCC-MSU/collection3\r\n\r\nEDIT: all done :)", "@lhoestq I was not asking you to do it, but asking if you agree me to do it... :man_facepalming: \r\nAs I self-assigned this issue... :sweat_smile: " ]

[ "@BradleyHsu", "Btw, thanks very much for finding the hub rollback temporary fix and bringing the issue to our attention @KhoomeiK!", "I see the same issue when calling `load_dataset('poloclub/diffusiondb', 'large_random_1k')` with `datasets==2.7.1` and `huggingface-hub=0.11.0`. No issue with `datasets=2.6.1` and `huggingface_hub==0.10.1`.\r\n\r\nhttps://github.com/poloclub/diffusiondb/issues/7", "I fixed my issue by specifying `repo_type` in `hf_hub_url()`. https://github.com/poloclub/diffusiondb/commit/9eb91c79aaca98b0515a0ce45778b8af65b84652\r\n\r\nI opened a PR on the Winoground's repo: https://huggingface.co/datasets/facebook/winoground/discussions/2", "This is a bug in the script, indeed. The most robust fix is to use a relative path instead of `hf_hub_url`, which does not depend on `huggingface_hub`'s version 🙂. I've opened a PR here: https://huggingface.co/datasets/facebook/winoground/discussions/3.", "Awesome, big thanks to both @xiaohk and @mariosasko!", "so, if i reproduce the bug, what should i do ? with huggingface_hub0.13.3 dataset2.6.1", "huggingface_hub.utils._validators.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name':\r\n\r\n tokenizer = AutoTokenizer.from_pretrained(ARGS.model_path, trust_remote_code=True)\r\n\r\nPlease handle automatically for local path and repo name inside, otherwise users always get confused about this", "I think I'm also hitting this error, trying to load a model from a local path." ]

[]

[ "Hi ! We're using the Arrow format for the datasets, and PyArrow tensors are not part of the Arrow format AFAIK:\r\n\r\n> There is no direct support in the arrow columnar format to store Tensors as column values.\r\n\r\nsource: https://github.com/apache/arrow/issues/4802#issuecomment-508494694", "@wesm @rok its been around three years. any updates, regarding dataset arrow tensor support? 🙏 I know you must be very busy, would appreciate to learn what is the state of art. I saw the PR is still open [#8510](https://github.com/apache/arrow/pull/8510)", "Hey @franz101 & @lhoestq!\r\nThere is a plan and a PR to create an [ExtensionArray of Tensors](https://github.com/apache/arrow/pull/8510) of equal sizes as well as a plan to do the same for Tensors of different sizes [ARROW-8714](https://issues.apache.org/jira/browse/ARROW-8714).", "The work stalled a little because it was not clear where TensorArray would live. However Arrow community recently agreed to make a [well-known-extension-type document](https://lists.apache.org/thread/sxd5fhc42hb6svs79t3fd79gkqj83pfh) and I would like https://github.com/apache/arrow/pull/8510 to land there and add an implementation to C++/Python + another language. Is that something you would find beneficial to you?", "that is a great update, thank you.\r\nit looks like this feature would benefit datasets implementation of [ArrayExtensionArray](https://github.com/huggingface/datasets/blob/9f2ff14673cac1f1ad56d80221a793f5938b68c7/src/datasets/features/features.py#L585-L641). Is that correct @eladsegal @lhoestq?\r\n\r\n", "TensorArray sounds great ! Looking forward to it :)\r\n\r\nWe've had our own ExtensionArray for fixed shape tensors for a while now, hoping to see something more standardized by the arrow community.\r\n\r\nAlso super interested in the extension array for tensors of different sizes cc @mariosasko ", "[FixedShapeTensor ExtensionType](https://github.com/apache/arrow/pull/8510) was merged and will be in Arrow 12.0.0 (release is planned mid April).\r\n", "@rok Thanks for keeping us updated! I think it's best to introduce a new feature type that would use this extension type under the hood. I'll create an issue to discuss the design with the community in the coming days.\r\n\r\nAlso, is there a tentative time frame for the variable-shape Tensor extension type?", "@mariosasko please tag me in the discussion, perhaps I can contribute.\r\n\r\nAs for the [variable shape tensor array](https://github.com/apache/arrow/issues/24868) - I'd be interested in working on it but didn't see much interest in community yet. Are you saying `huggingface/datasets` could use it?", "pyarrow 12 is out 🎉, will have a look if I can work on it for the ExtensionArray", "I think these two issues need to be fixed first on the Arrow side before adding the tensor feature type here: https://github.com/apache/arrow/issues/35573 and https://github.com/apache/arrow/issues/35599.\r\n\r\n@rok We've had a couple of requests for supporting variable-shape tensors on the forum/GH, but I did not manage to find the concrete issues using the search. TF/TFDS (and PyTorch with the `nested_tensor` API) support them, so it makes sense for us to do the same eventually (the Ray project has an [extension](https://github.com/ray-project/ray/blob/42a8d1489b37243f203120899a23d919dc85bf2a/python/ray/air/util/tensor_extensions/arrow.py#L634) type to support this case)", "> @rok We've had a couple of requests for supporting variable-shape tensors on the forum/GH, but I did not manage to find the concrete issues using the search. TF/TFDS (and PyTorch with the `nested_tensor` API) support them, so it makes sense for us to do the same eventually (the Ray project has an [extension](https://github.com/ray-project/ray/blob/42a8d1489b37243f203120899a23d919dc85bf2a/python/ray/air/util/tensor_extensions/arrow.py#L634) type to support this case)\r\n\r\nThat does make sense indeed. We should probably also be careful about memory layout to enable zero-copy interface to TF/PyTorch.", "So there is no way we can use [pyarrow.Tensor](https://arrow.apache.org/docs/python/generated/pyarrow.Tensor.html#pyarrow.Tensor) ?", "Not with with the Arrow format, and therefore not in `datasets`. But they released a new [FixedShapeTensorArray](https://arrow.apache.org/docs/python/extending_types.html#fixed-size-tensor) to store tensors in Arrow format. We plan to support this in `datasets` at one point !", "There is also an open issue to enable the conversion of `pyarrow.Tensor` to `pyarrow.FixedShapeTensorType`: https://github.com/apache/arrow/issues/35068. This way one could indirectly use `pyarrow.Tensor` in Arrow format.", "We started a [mailing list discussion](https://lists.apache.org/thread/qc9qho0fg5ph1dns4hjq56hp4tj7rk1k) about potential `VariableShapeTensor` extension array, please check it out and give feedback. For more details here's also a PR https://github.com/apache/arrow/pull/37166." ]

[ "It can fix the bug temporarily.\r\n```python\r\nfrom datasets import DownloadConfig\r\nconfig = DownloadConfig(num_proc=8)\r\nIn [5]: dataset = load_dataset('Freed-Wu/kodak', split='test', download_config=config)\r\nDownloading and preparing dataset kodak/default to /home/wzy/.cache/huggingface/datasets/Freed-Wu___kodak/default/0.0.1/6cf51f2b3d686d24a33fe86945f9e16802def212325f9345cf3cbb1b9f5f4a57...\r\nDownloading data files #4: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.39obj/s]\r\nDownloading data files #2: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.38obj/s]\r\nDownloading data files #3: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.13obj/s]\r\nDownloading data files #7: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.09obj/s]\r\nDownloading data files #5: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.08obj/s]\r\nDownloading data files #0: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:02<00:00, 1.08obj/s]\r\nDownloading data files #1: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:10<00:00, 3.36s/obj]\r\nDownloading data: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 492k/492k [00:01<00:00, 253kB/s]\r\nDownloading data files #6: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:13<00:00, 4.63s/obj]\r\nExtracting data files #0: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1407.17obj/s]\r\nExtracting data files #1: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1325.91obj/s]\r\nExtracting data files #3: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1524.46obj/s]\r\nExtracting data files #2: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1404.66obj/s]\r\nExtracting data files #4: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1538.63obj/s]\r\nExtracting data files #6: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1711.73obj/s]\r\nExtracting data files #7: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 2144.33obj/s]\r\nExtracting data files #5: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 1964.85obj/s]\r\nDataset kodak downloaded and prepared to /home/wzy/.cache/huggingface/datasets/Freed-Wu___kodak/default/0.0.1/6cf51f2b3d686d24a33fe86945f9e16802def212325f9345cf3cbb1b9f5f4a57. Subsequent calls will reuse this data.\r\n```", "Thanks for reporting ! This sounds like an issue with python multiprocessing. If we switch to multithreading for the downloads it should be much more robust - let me know if this is something you'd like to contribute, I'd be happy to help and give you some pointers", "> an issue with python multiprocessing\r\n\r\nIf it is an issue with multiprocessing, should we report it to upstream?", "Debugging this would require quite some work in my opinion, and I've often failed to make reproducible examples, since it's pretty correlated to one's environment + hardware. So I wouldn't spend too much time on this unless we manage to reproduce this on another machine consistently.\r\n\r\nInstead I'd encourage a more pragmatic fix that is: not create tons of processes (on regular machines it may slow things down anyway), and instead use multithreading by default.", "I am not expert of python. I hear about python has GIL, which result in multi processing is worse than multi threading. So I am not sure if this change makes sense?\r\n\r\nAnd if this is a bug of multi processing, why not report to upstream and let them fix? And even if change it to multi threading, how can we make sure it can truly fix this problem?", "Just my 2c. No offense.", "> Just my 2c. No offense.\r\n\r\nsure np ^^\r\n\r\n> I hear about python has GIL, which result in multi processing is worse than multi threading. So I am not sure if this change makes sense?\r\n\r\nHere the bottleneck speed is the bandwidth used to download the files. When downloading, the GIL is released, so multithreading gives the same speed as multiprocessing.\r\n\r\n> And if this is a bug of multi processing, why not report to upstream and let them fix?\r\n\r\nUsually to fix a bug it's important to be able to reproduce it. This way you can share it, experiment with it, and then make sure it's fixed. Here I'm afraid it's not easy to reproduce. Though I think that spawning too many processes for your machine can lead to this kind of issues.\r\n\r\n> And even if change it to multi threading, how can we make sure it can truly fix this problem?\r\n\r\nMultithreading is more robust in python because IIRC there are less locks involved which are often the cause of code hanging for no reason." ]

[ "I don't think we need completion on the datasets-cli, since we're mainly developing huggingface-cli", "I see." ]