Datasets:

---

amannagrawall002

/

github-issues

like 0

[]

[ "The dataset still can be found, if instead of using the search form you just enter the language code in the url, like https://huggingface.co/datasets?language=language:myv. \r\n\r\nBut of course having a more complete list of languages in the search form (or just a fallback to the language codes, if they are missing from the code=>language mapping) would be much more convenient!", "Hi! I've opened a PR to make these languages searchable on the Hub.", "Thanks @mariosasko!\r\nDo you think it is possible to turn this into a more scalable pipeline? Such as:\r\n1. Looping through all the datasets on the hub and collecting the set of all their language codes;\r\n2. Selecting the codes not covered yet in `Language.ts`\r\n3. Looking up their codes at https://iso639-3.sil.org/code_tables/639/data\r\n4. Adding all the newly found language codes to `Language.ts`", "@avidale This has been discussed in https://github.com/huggingface/datasets/issues/4881, so also feel free to share your opinion there." ]

[ "Hi! You can contact `[email protected]` to report this vulnerability." ]

[ "Hi ! It's been fixed by https://github.com/huggingface/datasets/pull/5598. We're doing a new release tomorrow with the fix and you'll be able to push your 100k images ;)\r\n\r\nBasically `push_to_hub` used to fail if the remote repository already exists and has a README.md without dataset_info in the YAML tags.\r\n\r\nIn the meantime you can install datasets from source", "Hi @lhoestq ,\r\n\r\nWhat version of datasets library fix this case? I am using the last `v2.10.1` and I get the same error.", "We just released 2.11 which includes a fix :)" ]

[ "Sounds like an issue with incompatible `transformers` dependencies versions.\r\n\r\nCan you try to update `transformers` ?\r\n\r\nEDIT: I checked the `transformers` dependencies and it seems like you need `tokenizers>=0.10.1,<0.11` with `transformers==4.5.1`\r\n\r\nEDIT2: this old version of `datasets` seems to import `transformers` but it's no longer the case, so you could also simply update `datasets` and `transformers` won't be imported", "Thank you for advising me to update these libraries versions.\r\n\r\nI can implement codes using `datasets==2.10.1` and `transformers==4.27.3`" ]

[ "Hi ! This sounds related to https://github.com/huggingface/datasets/issues/5406\r\n\r\nUpdating `datasets` fixes the issue ;)", "Thanks @lhoestq!\r\n\r\nI'll close this issue now." ]

[ "Do I miss something here?", "Hi! \r\n\r\nThe first dataset stores images as bytes (the \"image\" column type is `datasets.Image()`) and decodes them as `PIL.Image` objects and the second dataset stores them as variable-length lists (the \"image\" column type is `datasets.Sequence(...)`)), so I guess going from `arrow bytes -> NumPy -> decoding as PIL.Image -> PyTorch` is faster than going from `arrow list -> NumPy -> PyTorch`. \r\n\r\nTo store image bytes in the second example, you can do the following:\r\n\r\n```python\r\ndef transform(example):\r\n example[\"image2\"] = cv2.imread(example[\"image_file_path\"])\r\n return example\r\n\r\nfeatures = dataset.features.copy()\r\ndel features[\"image\"]\r\nfeatures[\"image2\"] = datasets.Image()\r\ndataset2 = dataset.map(transform, remove_columns=[\"image\"], features=features)\r\n\r\nfor x in DataLoader(dataset2.with_format(\"torch\"), batch_size=16, shuffle=True, num_workers=8):\r\n pass\r\n```", "Thanks, @mariosasko. I could not understand why a (decoded) sequence should be MUCH slower than an encoded image (that must be decoded every time). At any rate, I tried you suggestion. It made the `map` step to run extremely slow (consumes all the 16GB of memory and starts swapping)\r\n\r\nI tried also the easiest (as I see it) scenario, where images are kept as bytes, but it made things even worse: not only it was extremely slow, but also crashes\r\n\r\n```python\r\n\r\ndef transform(example):\r\n example[\"image2\"] = cv2.imread(example[\"image_file_path\"]).tobytes()\r\n return example\r\n\r\ndataset2 = dataset.map(transform, remove_columns=[\"image\"])\r\n\r\nfor x in DataLoader(dataset2.with_format(\"torch\"), batch_size=16, shuffle=True, num_workers=8):\r\n pass\r\n\r\n\r\nResource temporarily unavailable (src/thread.cpp:269)\r\nOutput exceeds the size limit. Open the full output data in a text editor\r\n---------------------------------------------------------------------------\r\nRuntimeError Traceback (most recent call last)\r\nFile ~/virtenvs/py310/lib/python3.10/site-packages/torch/utils/data/dataloader.py:1133, in _MultiProcessingDataLoaderIter._try_get_data(self, timeout)\r\n 1132 try:\r\n-> 1133 data = self._data_queue.get(timeout=timeout)\r\n 1134 return (True, data)\r\n\r\nFile ~/virtenvs/py310/lib/python3.10/multiprocessing/queues.py:113, in Queue.get(self, block, timeout)\r\n 112 timeout = deadline - time.monotonic()\r\n--> 113 if not self._poll(timeout):\r\n 114 raise Empty\r\n\r\nFile ~/virtenvs/py310/lib/python3.10/multiprocessing/connection.py:257, in _ConnectionBase.poll(self, timeout)\r\n 256 self._check_readable()\r\n--> 257 return self._poll(timeout)\r\n\r\nFile ~/virtenvs/py310/lib/python3.10/multiprocessing/connection.py:424, in Connection._poll(self, timeout)\r\n 423 def _poll(self, timeout):\r\n--> 424 r = wait([self], timeout)\r\n 425 return bool(r)\r\n\r\nFile ~/virtenvs/py310/lib/python3.10/multiprocessing/connection.py:931, in wait(object_list, timeout)\r\n 930 while True:\r\n--> 931 ready = selector.select(timeout)\r\n 932 if ready:\r\n...\r\n-> 1146 raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str)) from e\r\n 1147 if isinstance(e, queue.Empty):\r\n 1148 return (False, None)\r\n\r\nRuntimeError: DataLoader worker (pid(s) 195393) exited unexpectedly\r\nResource temporarily unavailable (src/thread.cpp:269)\r\nResource temporarily unavailable (src/thread.cpp:269)\r\nResource temporarily unavailable (src/thread.cpp:269)\r\nResource temporarily unavailable (src/thread.cpp:269)\r\nResource temporarily unavailable (src/thread.cpp:269)\r\n```\r\n", "Correction: the `beans` dataset stores the image file paths, not the bytes.\r\n\r\nFor your use case, I think it makes more sense to use `with_tranform` than `map` and lazily decode images with `cv2.imread` when indexing an example/batch:\r\n```python\r\nimport cv2\r\n\r\ndef transform(batch):\r\n batch[\"image2\"] = np.stack([cv2.imread(image_file_path) for image_file_path in batch[\"image_file_path\"]])\r\n return batch\r\n\r\ndataset = dataset.with_transform(transform)\r\n```\r\n", "This is incorrect.\n\nDid you try to run it? dataset[0] returns a tensor of numbers. dataset2[0]\nreturns the same tensor, but after a few long seconds. Looping over a\nthousand of images cannot take 15 minutes.\n\nOn Fri, 24 Mar 2023 at 19:28 Mario Šaško ***@***.***> wrote:\n\n> Correction: the beans dataset stores the image file paths, not the bytes.\n>\n> For your use case, I think it makes more sense to use with_tranform than\n> map and lazily decode images with cv2.imread when accessing an\n> example/batch:\n>\n> import cv2\n> def transform(batch):\n> batch[\"image2\"] = np.stack([cv2.imread(image_file_path) for image_file_path in batch[\"image_file_path\"]])\n> return batch\n> dataset = dataset.with_transform(transform)\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5669#issuecomment-1483084347>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AASS73SHRWXIQX6SCYCJ7ITW5XDUDANCNFSM6AAAAAAWFSHWEM>\n> .\n> You are receiving this because you authored the thread.Message ID:\n> ***@***.***>\n>\n", "I updated the transform with the NumPy -> PyTorch conversion.\r\n\r\nI'm sharing the entire code:\r\n```python\r\nimport cv2\r\nimport numpy as np\r\nimport datasets\r\nimport torch\r\nfrom datasets import load_dataset\r\nfrom torch.utils.data import DataLoader\r\n\r\ndataset = load_dataset(\"beans\", split=\"train\")\r\n\r\ndef transform(batch):\r\n # # Pillow decodes as RGB\r\n # batch[\"image\"] = torch.stack([torch.from_numpy(cv2.cvtColor(cv2.imread(image_file_path), cv2.COLOR_BGR2RGB)) for image_file_path in batch[\"image_file_path\"]])\r\n batch[\"image\"] = torch.stack([torch.from_numpy(cv2.imread(image_file_path)) for image_file_path in batch[\"image_file_path\"]])\r\n batch[\"labels\"] = torch.tensor(batch[\"labels\"])\r\n return batch\r\n\r\ndataset2 = dataset.cast_column(\"image\", datasets.Image(decode=False)).with_transform(transform)\r\n\r\nfor x in DataLoader(dataset2, batch_size=16, shuffle=True, num_workers=8):\r\n pass\r\n```\r\n\r\nThis code is ≈ 10% faster on my machine than the default decoding with Pillow and `.with_format(\"torch\")`.", "Thanks, @mariosasko \r\nMy question remain unanswered though. Why is the `map`ed dataset so slow? My understanding is that a dataset of numpy arrays should be must faster than a dataset that has to decode images into numpy arrays every time one accesses an item. " ]

[]

[ "+1", "+1" ]

[]

[]

[ "You can convert any dataset to pandas to be used with imbalanced-learn using `.to_pandas()`\r\n\r\nOtherwise if you want to keep a `Dataset` object and still use e.g. [make_imbalance](https://imbalanced-learn.org/stable/references/generated/imblearn.datasets.make_imbalance.html#imblearn.datasets.make_imbalance), you just need to pass the list of rows ids and labels:\r\n\r\n```python\r\nrow_indices = list(range(len(dataset)))\r\nresampled_row_indices, _ = make_imbalance(\r\n row_indices,\r\n dataset[\"label\"],\r\n sampling_strategy={0: 25, 1: 50, 2: 50},\r\n random_state=RANDOM_STATE,\r\n)\r\n\r\nresampled_dataset = dataset.select(resampled_row_indices)\r\n```" ]

[ "cc @polinaeterna @lhoestq ", "@sanchit-gandhi can you please also post the logs of `pip install soundfile==0.12.1`? To check what wheel is being installed or if it's being built from source (I think it's the latter case). \r\nRequired `libsndfile` binary **should** be bundeled with `soundfile` wheel but I assume it **might not** be the case for some non standard Linux distributions. \r\nThe only solution for using `soundfile` here is to build [`libsndfile`](https://github.com/libsndfile/libsndfile) from source:\r\n\r\n```bash\r\ngit clone https://github.com/libsndfile/libsndfile.git\r\ncd libsndfile/\r\nautoreconf -vif\r\n./configure --enable-werror \r\nmake\r\nmake install\r\n```\r\nfor this, some building libraries should be installed, for Debian/Ubuntu it's like:\r\n```bash\r\napt install autoconf autogen automake build-essential libasound2-dev \\\r\n libflac-dev libogg-dev libtool libvorbis-dev libopus-dev libmp3lame-dev \\\r\n libmpg123-dev pkg-config python\r\n```\r\nbut for other Linux distributions it might be different.\r\n\r\nWhen the binary is compiled, it should be put into location where `soundfile` would search for it (the directory is named `_soundfile_data`), it depends on where`libsdfile` (from the previous step) and `soundfile` were installed, might be something like this:\r\n\r\n```bash\r\ncp /usr/local/lib/libsndfile.so /usr/local/lib/python3.7/dist-packages/_soundfile_data/\r\ncp /usr/local/lib/libsndfile.la /usr/local/lib/python3.7/dist-packages/_soundfile_data/\r\n```\r\n\r\nAnother solution is to not use `soundfile` and apply custom processing function with `torchaudio` while setting `decode=False` in `Audio` feature and passing custom function to `.map`. ", "Not sure if it may help, but you could also try updating `pip` before installing soundfile", "@lhoestq @sanchit-gandhi. I encountered the same error (also on the TPU v4) when trying to run `datasets` from source.\r\n\r\nDowngrading soundfile with `pip install soundfile==0.12.0` seems to fix the issue for me.", "Maybe let's open an issue at https://github.com/bastibe/python-soundfile/issues in case they might know why you get `OSError: cannot load library 'libsndfile.so'` ?", "> @sanchit-gandhi can you please also post the logs of `pip install soundfile==0.12.1`? To check what wheel is being installed or if it's being built from source (I think it's the latter case). Required `libsndfile` binary **should** be bundeled with `soundfile` wheel but I assume it **might not** be the case for some non standard Linux distributions. The only solution for using `soundfile` here is to build [`libsndfile`](https://github.com/libsndfile/libsndfile) from source:\r\n> \r\n> ```shell\r\n> git clone https://github.com/libsndfile/libsndfile.git\r\n> cd libsndfile/\r\n> autoreconf -vif\r\n> ./configure --enable-werror \r\n> make\r\n> make install\r\n> ```\r\n\r\nThis fixed the issue for me. After installing libsndfile as described above, I had to uninstall soundfile and re-install it with this command. `pip install \"soundfile>=0.12.1\"`", "Thank you so much for the comprehensive instructions @polinaeterna! Also confirming that they worked for me 🤗 In my case, I had to run several of these commands under \"sudo\" for privileges, but otherwise this workaround gave a successful `libsndfile` install:\r\n\r\n1. Grab source code:\r\n```\r\ngit clone https://github.com/libsndfile/libsndfile.git\r\n```\r\n\r\n2. Set up a build environment:\r\n```\r\nsudo apt install autoconf autogen automake build-essential libasound2-dev \\\r\n libflac-dev libogg-dev libtool libvorbis-dev libopus-dev libmp3lame-dev \\\r\n libmpg123-dev pkg-config python\r\n```\r\n\r\n3. Build and test `libsndfile`:\r\n\r\n```\r\nautoreconf -vif\r\n./configure --enable-werror\r\nsudo make\r\nsudo make check\r\n```\r\n\r\n4. Create `_soundfile_data` submodule (if it does not exist already):\r\n```\r\nsudo mkdir /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n```\r\n\r\n5. Copy `libsndfile` files into submodule:\r\n```\r\nsudo cp /usr/local/lib/libsndfile.* /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n```", "On a different machine, I also tried separately by first upgrading pip, then installing soundfile. This worked too! Thanks @lhoestq 🙌", "> @sanchit-gandhi can you please also post the logs of `pip install soundfile==0.12.1`? To check what wheel is being installed or if it's being built from source (I think it's the latter case). Required `libsndfile` binary **should** be bundeled with `soundfile` wheel but I assume it **might not** be the case for some non standard Linux distributions. The only solution for using `soundfile` here is to build [`libsndfile`](https://github.com/libsndfile/libsndfile) from source:\r\n> \r\n> ```shell\r\n> git clone https://github.com/libsndfile/libsndfile.git\r\n> cd libsndfile/\r\n> autoreconf -vif\r\n> ./configure --enable-werror \r\n> make\r\n> make install\r\n> ```\r\n> \r\n> for this, some building libraries should be installed, for Debian/Ubuntu it's like:\r\n> \r\n> ```shell\r\n> apt install autoconf autogen automake build-essential libasound2-dev \\\r\n> libflac-dev libogg-dev libtool libvorbis-dev libopus-dev libmp3lame-dev \\\r\n> libmpg123-dev pkg-config python\r\n> ```\r\n> \r\n> but for other Linux distributions it might be different.\r\n> \r\n> When the binary is compiled, it should be put into location where `soundfile` would search for it (the directory is named `_soundfile_data`), it depends on where`libsdfile` (from the previous step) and `soundfile` were installed, might be something like this:\r\n> \r\n> ```shell\r\n> cp /usr/local/lib/libsndfile.so /usr/local/lib/python3.7/dist-packages/_soundfile_data/\r\n> cp /usr/local/lib/libsndfile.la /usr/local/lib/python3.7/dist-packages/_soundfile_data/\r\n> ```\r\n> \r\n> Another solution is to not use `soundfile` and apply custom processing function with `torchaudio` while setting `decode=False` in `Audio` feature and passing custom function to `.map`.\r\n\r\nThanks, the solution solved my problem. \r\n\r\n1. Purge uninstall libsndfile, uninstall python-soundfile.\r\n2. Build libsndfile from source code and install.\r\n3. Build python-soundfile from source code and install\r\n4. Well done.", "> Thank you so much for the comprehensive instructions @polinaeterna! Also confirming that they worked for me 🤗 In my case, I had to run several of these commands under \"sudo\" for privileges, but otherwise this workaround gave a successful `libsndfile` install:\r\n> \r\n> 1. Grab source code:\r\n> \r\n> ```\r\n> git clone https://github.com/libsndfile/libsndfile.git\r\n> ```\r\n> \r\n> 2. Set up a build environment:\r\n> \r\n> ```\r\n> sudo apt install autoconf autogen automake build-essential libasound2-dev \\\r\n> libflac-dev libogg-dev libtool libvorbis-dev libopus-dev libmp3lame-dev \\\r\n> libmpg123-dev pkg-config python\r\n> ```\r\n> \r\n> 3. Build and test `libsndfile`:\r\n> \r\n> ```\r\n> autoreconf -vif\r\n> ./configure --enable-werror\r\n> sudo make\r\n> sudo make check\r\n> ```\r\n> \r\n> 4. Create `_soundfile_data` submodule (if it does not exist already):\r\n> \r\n> ```\r\n> sudo mkdir /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n> ```\r\n> \r\n> 5. Copy `libsndfile` files into submodule:\r\n> \r\n> ```\r\n> sudo cp /usr/local/lib/libsndfile.* /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n> ```\r\n\r\nI had to run 'make install' or the `/usr/local/lib/libsndfile.*` files didn't exist.\r\n\r\nIt's working though!", "I had the same issue but it is working now! Thanks for all of your comments!", "I had the same issue on SageMaker but not on Colab;\r\nThe `soundfile` versioning was fine.\r\n\r\n my approach to solve it was to match {\"numpy\", \"numba\"} exact versions\r\n\r\n```\r\n! pip install \"numpy==1.23.5\"\r\n! pip install \"numpy==0.58.1\"\r\n\r\n```\r\nthe numbers are from Colab where successfully I could do the job.\r\n\r\n", "> Thank you so much for the comprehensive instructions @polinaeterna! Also confirming that they worked for me 🤗 In my case, I had to run several of these commands under \"sudo\" for privileges, but otherwise this workaround gave a successful `libsndfile` install:\r\n> \r\n> 1. Grab source code:\r\n> \r\n> ```\r\n> git clone https://github.com/libsndfile/libsndfile.git\r\n> ```\r\n> \r\n> 2. Set up a build environment:\r\n> \r\n> ```\r\n> sudo apt install autoconf autogen automake build-essential libasound2-dev \\\r\n> libflac-dev libogg-dev libtool libvorbis-dev libopus-dev libmp3lame-dev \\\r\n> libmpg123-dev pkg-config python\r\n> ```\r\n> \r\n> 3. Build and test `libsndfile`:\r\n> \r\n> ```\r\n> autoreconf -vif\r\n> ./configure --enable-werror\r\n> sudo make\r\n> sudo make check\r\n> ```\r\n> \r\n> 4. Create `_soundfile_data` submodule (if it does not exist already):\r\n> \r\n> ```\r\n> sudo mkdir /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n> ```\r\n> \r\n> 5. Copy `libsndfile` files into submodule:\r\n> \r\n> ```\r\n> sudo cp /usr/local/lib/libsndfile.* /usr/local/lib/python3.8/dist-packages/_soundfile_data/\r\n> ```\r\n\r\nIt works and don't forget to \"apt uninstall libsndfile1\" after installing it from source code." ]

[ "Upd. the above code works if we replace `25` with `1`, but the result value at key \"hr\" is not a tensor but a list of lists of lists of uint8.\r\n\r\nAdding `train_data.set_format(\"torch\")` after map fixes this, but the original issue remains\r\n\r\n", "As a workaround, one can replace\r\n`return {\"hr\": torch.stack([crop_transf(tensor) for _ in range(25)])}`\r\nwith\r\n`return {f\"hr_crop_{i}\": crop_transf(tensor) for i in range(25)}`\r\nand then choose appropriate crop randomly in further processing, but I still don't understand why the original approach doesn't work(\r\n" ]

[ "I agree this should be documented" ]

[ "`save_to_disk` should indeed expand `~`. Marking it as a \"good first issue\".", "#self-assign\r\n\r\nFile path to code: \r\n\r\nhttps://github.com/huggingface/datasets/blob/2.13.0/src/datasets/arrow_dataset.py#L1364\r\n\r\n@RmZeta2718 I created a pull request for this issue. ", "Hello, \r\nIt says `save_to_disk` is deprecated in 2.8.0, so the alternative to this will be `storage_options`? \r\n\r\nhttps://huggingface.co/docs/datasets/package_reference/main_classes#datasets.Dataset.save_to_disk", "@ashikshafi08 I think you misunderstood the warning. The method `save_to_disk` is not deprecated only the optional parameter `fs`.\r\nAlso @benjaminbrown038 as I cannot find your PR I would like to work on this if you don't mind.", "@mariosasko It's been several months and the PR is not reviewed. Could you please take a look? I assume this is not complicated and could be merged fairly soon." ]

[ "Can you post a reproducible code snippet of what you tried to do?\r\n\r\n", "> Can you post a reproducible code snippet of what you tried to do?\n> \n> \n\n```python\nfrom datasets import load_dataset\n\ndataset = load_dataset(\"my_folder_name\", split=\"train\")\n```", "hi @WiNE-iNEFF ! can you please also tell a bit more about how your data is structured (directory structure and filenames patterns)?", "> hi @WiNE-iNEFF ! can you please also tell a bit more about how your data is structured (directory structure and filenames patterns)?\n\nAll file have format .png converted in RGBA. \nIn main folder \"MyData\" contain 4 folder with images. In function load_dataset i use folder \"MyData\"", "@WiNE-iNEFF I'm sorry there is still not enough information to answer your question :( For now I can only assume that your [filenames contain split names](https://huggingface.co/docs/datasets/repository_structure#splits-and-file-names) which are somehow incorrectly parsed. \r\nWhat would be the output if you omit `split` while loading? Like just\r\n```python\r\nds = load_dataset(\"MyData\")\r\nprint(ds)\r\n```\r\n\r\n", "> @WiNE-iNEFF I'm sorry there is still not enough information to answer your question :( For now I can only assume that your [filenames contain split names](https://huggingface.co/docs/datasets/repository_structure#splits-and-file-names) which are somehow incorrectly parsed. \n> What would be the output if you omit `split` while loading? Like just\n> ```python\n> ds = load_dataset(\"MyData\")\n> print(ds)\n> ```\n> \n> \n\n```python\nDataset({\n features: ['image', 'label'],\n num_rows: 4\n})\n```", "@WiNE-iNEFF My only guess is that 4 images in your data have `\"train\"` string in their names (something like `\"train_image_0.png\"`) and others do not and the loader ignores all the files that do not contain split name in filename. If it's true, please try to remove \"train\" from filenames. Or maybe they are inside a directory named \"train\", then the directory should be renamed (unless you want to put only these 4 specific images to the train but apparently you do not).\r\n\r\nIf there is a bug I cannot investigate it unfortunately because I cannot reproduce your case without some data samples. ", "> @WiNE-iNEFF My only guess is that 4 images in your data have `\"train\"` string in their names (something like `\"train_image_0.png\"`) and others do not and the loader ignores all the files that do not contain split name in filename. If it's true, please try to remove \"train\" from filenames. Or maybe they are inside a directory named \"train\", then the directory should be renamed (unless you want to put only these 4 specific images to the train but apparently you do not).\n> \n> If there is a bug I cannot investigate it unfortunately because I cannot reproduce your case without some data samples. \n\nI checked my files and some of them do have the words train, valid and test in their names, but the number of such images is more than 500, not 4.", "@WiNE-iNEFF Probably they are named inconsistently so that the correct pattern for which files should correspond to which split cannot be inferred. You can make it clearer to the loader by removing split names from filenames and putting files in separate folder for each split (you can take a look at the [documentation for imagefolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder)):\r\n```\r\n Fuaimeanna2/\r\n├─ test\r\n│   ├─ label_0\r\n│   │   ├── filename_0.jpg\r\n│   │   └── filename_1.jpg\r\n│   │   └── ...\r\n│   ├─ label_1\r\n│   │   └── ...\r\n│   ├─ label_2\r\n│   │   └── ...\r\n│   └─ label_3\r\n│   └── ...\r\n├─ train\r\n│   ├─ label_0\r\n│   │   └── ...\r\n│   ├─ label_1\r\n│   │   └── ...\r\n│   ├─ label_2\r\n│   │   └── ...\r\n│   └─ label_3\r\n│   └── ...\r\n└── validation\r\n    ├─ label_0\r\n   │   └── ...\r\n    ├─ label_1\r\n   │   └── ...\r\n    ├─ label_2\r\n   │   └── ...\r\n └─ label_3\r\n └── ...\r\n```", "> @WiNE-iNEFF Probably they are named inconsistently so that the correct pattern for which files should correspond to which split cannot be inferred. You can make it clearer to the loader by removing split names from filenames and putting files in separate folder for each split (you can take a look at the [documentation for imagefolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder)):\n> ```\n> Fuaimeanna2/\n> ├─ test\n> │   ├─ label_0\n> │   │   ├── filename_0.jpg\n> │   │   └── filename_1.jpg\n> │   │   └── ...\n> │   ├─ label_1\n> │   │   └── ...\n> │   ├─ label_2\n> │   │   └── ...\n> │   └─ label_3\n> │   └── ...\n> ├─ train\n> │   ├─ label_0\n> │   │   └── ...\n> │   ├─ label_1\n> │   │   └── ...\n> │   ├─ label_2\n> │   │   └── ...\n> │   └─ label_3\n> │   └── ...\n> └── validation\n>    ├─ label_0\n>    │   └── ...\n>    ├─ label_1\n>    │   └── ...\n>    ├─ label_2\n>    │   └── ...\n> └─ label_3\n> └── ...\n> ```\n\nI have read this documentation more than once. It just wasn't a problem before.", "Hi,\r\n\r\nYou need to use:\r\n```\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"imagefolder\", split=\"train\", data_dir=\"path_to_your_folder\")\r\n```\r\ninstead of \r\n```\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"my_folder_name\", split=\"train\")\r\n```\r\nTo create an image dataset from your local folders.", "> Hi,\r\n> \r\n> You need to use:\r\n> \r\n> ```\r\n> from datasets import load_dataset\r\n> \r\n> dataset = load_dataset(\"imagefolder\", split=\"train\", data_dir=\"path_to_your_folder\")\r\n> ```\r\n> \r\n> instead of\r\n> \r\n> ```\r\n> from datasets import load_dataset\r\n> \r\n> dataset = load_dataset(\"my_folder_name\", split=\"train\")\r\n> ```\r\n> \r\n> To create an image dataset from your local folders.\r\n\r\nThank you, but even using the method that you wrote above absolutely nothing changes, especially without using data_dir on my other data everything works fine", "@WiNE-iNEFF have you tried the suggestion I posted above? with removing split names from filenames and structuring files in folders? \r\n\r\n\r\n> even using the method that you wrote above absolutely nothing changes\r\n\r\nfyi - nothing changed because these two approaches are basically the same. it's just that when you pass your data directory as a dataset name (`load_dataset(\"my_folder_name\"`), not as `data_dir` (`load_dataset(\"imagefolder\", data_dir=\"my_folder_name\"`), `datasets` infers what module to use (`imagefolder` in your case) automatically, by file extensions.", "Oh I didn't know that! OK but in any case, not sure why the image builder isn't working for @WiNE-iNEFF. But it's hard for us to help if we can't reproduce. I'd just check the structure of the folders, see if the splits are correctly set up, etc.", "> @WiNE-iNEFF have you tried the suggestion I posted above? with removing split names from filenames and structuring files in folders? \n> \n> \n> > even using the method that you wrote above absolutely nothing changes\n> \n> fyi - nothing changed because these two approaches are basically the same. it's just that when you pass your data directory as a dataset name (`load_dataset(\"my_folder_name\"`), not as `data_dir` (`load_dataset(\"imagefolder\", data_dir=\"my_folder_name\"`), `datasets` infers what module to use (`imagefolder` in your case) automatically, by file extensions.\n\nI'll try to try your method over the next few days, then I'll write it turned out ", "> @WiNE-iNEFF have you tried the suggestion I posted above? with removing split names from filenames and structuring files in folders? \n> \n> \n> > even using the method that you wrote above absolutely nothing changes\n> \n> fyi - nothing changed because these two approaches are basically the same. it's just that when you pass your data directory as a dataset name (`load_dataset(\"my_folder_name\"`), not as `data_dir` (`load_dataset(\"imagefolder\", data_dir=\"my_folder_name\"`), `datasets` infers what module to use (`imagefolder` in your case) automatically, by file extensions.\n\nI tried creating a `train` folder and put my image folders in it. As a result, all 18,000 images were loaded. ", "@WiNE-iNEFF great! So to explain what happened according to my assumptions:\r\n\r\nWhen you use a standard packaged loader (like `imagefolder`, `csv`, `jsonl`, and so on) and load your data like `load_dataset(\"my_folder_name\")` or `load_dataset(\"imagefolder\", data_dir=\"my_folder_name\"`, the library searches for patterns to divide files into splits. This is described a bit in [this doc](https://huggingface.co/docs/datasets/v2.10.0/en/repository_structure#splits-and-file-names). And the order to search for patterns is the following:\r\n1. first it checks for [pattern like `data/<split_name>-xxxxx-of-xxxxx`](https://huggingface.co/docs/datasets/v2.10.0/en/repository_structure#custom-split-names) (which allows to pass custom split names)\r\n2. then for directories named as splits (if you have directories named `train`, `test` etc.)\r\n3. then for [splits in filenames](https://huggingface.co/docs/datasets/v2.10.0/en/repository_structure#splits-and-file-names) (like if you have files named `train-image.jpg`, `test_0.jpg`, ...)\r\n4. then if no pattern was found, it treats all files as belonging to a single `train` split\r\n\r\nThe code is [here](https://github.com/huggingface/datasets/blob/main/src/datasets/data_files.py#L215).\r\nSo I assume that in your case, since you didn't have directories for splits (pattern 2), some files that included split keywords (pattern 3) were included and others were ignored as not matching the pattern. And when you added `train` directory, the pattern for directories (pattern 2) was triggered first and everything worked as expected. Everything worked in your previous cases probably because you didn't have split names keywords in filenames, so all the files ended up being a part of a single train split (pattern 4).\r\n\r\nAnother way to mitigate this apart from structuring your data according to the patterns is to explicitly state with files belong to which splits by passing them with `data_files` parameter:\r\n```python\r\nload_dataset(\"my_folder_name\", data_files={\"train\": \"**\"}) # to tell that all files should be included \r\n```\r\n\r\nNow I see that this order should be explained in documentation and also referenced in sections for packaged modules like `imagefolder`, thank you for pointing this out. \r\n\r\n \r\n", "@NielsRogge @polinaeterna I have a similar problem when reading my dataset. I want to use DETR for object detection, but my data is in YOLO format. With a dataset of 10k images, yolo format involves having 10k labels. As far as I read regarding [COCO format](https://auto.gluon.ai/stable/tutorials/multimodal/object_detection/data_preparation/convert_data_to_coco_format.html), there must be one JSON per split. However, as I post in the [Hugging Face forum](https://discuss.huggingface.co/t/prepare-dataset-from-yolo-format-to-coco-for-detr/34894), when it is read, the number of rows is 1, which does not make sense. \r\nThe instruction to read the train-val-test splits are: \r\n```python\r\nfrom datasets import load_dataset\r\ndata_files = {\r\n\t\"train\": './train_labels.json',\r\n\t\"validation\": './val_labels.json',\r\n\t\"test\": './test_labels.json'\r\n}\r\ndataset = load_dataset(\"json\", data_files=data_files)\r\n```\r\nAn example of the short version of the json file I read, to reproduce my error, is the following: \r\n\r\n``` json\r\n{\r\n \"info\": {},\r\n \"licenses\": [],\r\n \"images\": [\r\n {\r\n \"id\": 1,\r\n \"file_name\": \"aceca_100.mp4frame21.png\",\r\n \"width\": 1280,\r\n \"height\": 720,\r\n \"pixel_values\": null,\r\n \"pixel_mask\": null\r\n },\r\n {\r\n \"id\": 2,\r\n \"file_name\": \"aceca_100.mp4frame24.png\",\r\n \"width\": 1280,\r\n \"height\": 720,\r\n \"pixel_values\": null,\r\n \"pixel_mask\": null\r\n },\r\n {\r\n \"id\": 3,\r\n \"file_name\": \"aceca_100.mp4frame25.png\",\r\n \"width\": 1280,\r\n \"height\": 720,\r\n \"pixel_values\": null,\r\n \"pixel_mask\": null}],\r\n \"annotations\": [\r\n {\r\n \"id\": 1,\r\n \"image_id\": 1,\r\n \"category_id\": 0,\r\n \"bbox\": [0.0, 278.21896388398557, 86.94096523844935, 156.0293445072134],\r\n \"area\": 13565.341816979679,\r\n \"iscrowd\": 0\r\n },\r\n {\r\n \"id\": 2,\r\n \"image_id\": 2,\r\n \"category_id\": 0,\r\n \"bbox\": [149.28851295721816, 297.6359759754418, 34.76802347007475, 98.03908698442889],\r\n \"area\": 3408.625277259324,\r\n \"iscrowd\": 0\r\n },\r\n {\r\n \"id\": 3,\r\n \"image_id\": 3,\r\n \"category_id\": 0,\r\n \"bbox\": [153.3817197549372, 300.168969412891, 31.787555842913775, 89.69583163436312],\r\n \"area\": 2851.2112569539095,\r\n \"iscrowd\": 0\r\n }\r\n ],\r\n \"categories\": [\r\n {\r\n \"id\": 0, \"name\": \"person\"\r\n }\r\n ]\r\n }\r\n```\r\nIf full files required, my email is [email protected]", "Hi @Alberto1404, to load an object detection dataset it's recommended to make use of the metadata feature as explained [here](https://huggingface.co/docs/datasets/image_dataset#object-detection). ", "Thank you @NielsRogge! It works!!!", "You can now refer to https://huggingface.co/docs/datasets/repository_structure to learn about the `datasets`' data files inference, so I'm closing this issue." ]

[ "Thanks for reporting, @lsb. \r\n\r\nWe are investigating it.\r\n\r\nOn the other hand, please note that in the next `datasets` release, the index will not be created by default (see #5583). If you would like to have it, you will need to explicitly pass `index=True`. ", "I think this is low enough priority for me to close this as Won't Fix. If I need any primary keys I can generate them beforehand. Feel free to reopen." ]

[ "Thanks for reporting! This can be fixed by setting the format to `arrow` in `flatten_indices` and restoring the original format after the flattening. I'm working on a PR that reduces the number of the `flatten_indices` calls in our codebase and makes `flatten_indices` a no-op when a dataset does not have an indices mapping, so I'll incorporate the fix in that PR." ]

[ "related to #5444 ", "We now log these messages instead of printing them (addressed in #6019), so I'm closing this issue." ]

[ "It looks like an error that we observed once in https://github.com/huggingface/datasets/pull/5166\r\n\r\nCan you try to update `datasets` ?\r\n\r\n```\r\npip install -U datasets\r\n```\r\n\r\nif it doesn't work, can you make sure you don't have packages installed that may modify `dill`'s behavior, such as `apache-beam` ?", "@lhoestq That fixed the problem, Thanks :)" ]

[]

[]

[ " I think https://github.com/fsspec/universal_pathlib is the project you are looking for.\r\n\r\n`xPath` has the methods often used in dataset scripts, and `mkdir` is not one of them (`dl_manager`'s role is to \"interact\" with the file system, so using `mkdir` is discouraged).", "Right is there a difference between UPath and xPath? Typically is xPath less well implemented compared to Upath, ie missing some implementations of some methods? Or are there methods in xPath that are not implemented with UPath?", "`xPath` is an internal component (it doesn't have a leading underscore in the name, but it should) not meant to be used outside of `datasets`, and it's only tested on HTTP URLs, not S3.\r\n\r\n", "Okay I understand that xPath won't support my usecase. What I was perhaps getting to is why not use UPath in `datasets` instead of `xPath` if UPath seems to have strictly more robust implementations.", "It seems like `universal_pathlib` does not support `fsspec` URL chaining (`::` is the chaining symbol) and \"compression\" filesystems (e.g., `zip`), but this is what we need to access and stream files from within an archive (e.g., we want to stream URLs such as this one: `zip://data.parquet::https://www.dummyurl.com/archive.zip`)" ]

[ "Hi! Yes, only `torch` is currently supported. Unlike `Dataset`, `IterableDataset` is not PyArrow-backed, so we cannot simply call `to_numpy` on the underlying subtables to format them numerically. Instead, we must manually convert examples to (numeric) arrays while preserving consistency with `Dataset`, which is not trivial, so this is still a to-do.", "Any plans to support it in the future? Or would streaming dataset be left without support for jax and tensorflow?" ]

[ "Hi! \r\n\r\nBy default, tqdm has `leave=True` to \"keep all traces of the progress bar upon the termination of iteration\". However, we use `leave=False` in some places (as of recently), which removes the bar once the iteration is over.\r\n\r\nI feel like our TQDM bars are noisy, so I think we should always set `leave=False` and also use the `delay` parameter to display progress bars only for tasks that take time (e.g., more than 3s). What do you think about this? Do you find these bars useful (after the dataset generation is over)?\r\n", "Hi sorry for the late update. I think the problem still exists despite the `leave` flag\r\n\r\n<img width=\"1105\" alt=\"image\" src=\"https://user-images.githubusercontent.com/110427462/226501615-5b02fb02-fd5f-4eda-b1f7-a7ed6570892d.png\">\r\n\r\n\r\n```\r\nPackage Version\r\n------------------------ ---------\r\naiofiles 22.1.0\r\naiohttp 3.8.4\r\naiosignal 1.3.1\r\naiosqlite 0.18.0\r\nanyio 3.6.2\r\nappnope 0.1.3\r\nargon2-cffi 21.3.0\r\nargon2-cffi-bindings 21.2.0\r\narrow 1.2.3\r\nasttokens 2.2.1\r\nasync-generator 1.10\r\nasync-timeout 4.0.2\r\nattrs 22.2.0\r\nBabel 2.12.1\r\nbackcall 0.2.0\r\nbeautifulsoup4 4.11.2\r\nbleach 6.0.0\r\nbrotlipy 0.7.0\r\ncertifi 2022.12.7\r\ncffi 1.15.1\r\ncfgv 3.3.1\r\ncharset-normalizer 2.1.1\r\ncomm 0.1.2\r\nconda 22.9.0\r\nconda-package-handling 2.0.2\r\nconda_package_streaming 0.7.0\r\ncoverage 7.2.1\r\ncryptography 38.0.4\r\ndatasets 2.8.0\r\ndebugpy 1.6.6\r\ndecorator 5.1.1\r\ndefusedxml 0.7.1\r\ndill 0.3.6\r\ndistlib 0.3.6\r\ndistro 1.4.0\r\nentrypoints 0.4\r\nexceptiongroup 1.1.0\r\nexecuting 1.2.0\r\nfastjsonschema 2.16.3\r\nfilelock 3.9.0\r\nflaky 3.7.0\r\nfqdn 1.5.1\r\nfrozenlist 1.3.3\r\nfsspec 2023.3.0\r\nhuggingface-hub 0.10.1\r\nidentify 2.5.18\r\nidna 3.4\r\niniconfig 2.0.0\r\nipykernel 6.12.1\r\nipyparallel 8.4.1\r\nipython 7.32.0\r\nipython-genutils 0.2.0\r\nipywidgets 8.0.4\r\nisoduration 20.11.0\r\njedi 0.18.2\r\nJinja2 3.1.2\r\njson5 0.9.11\r\njsonpointer 2.3\r\njsonschema 4.17.3\r\njupyter_client 8.0.3\r\njupyter_core 5.2.0\r\njupyter-events 0.6.3\r\njupyter_server 2.4.0\r\njupyter_server_fileid 0.8.0\r\njupyter_server_terminals 0.4.4\r\njupyter_server_ydoc 0.6.1\r\njupyter-ydoc 0.2.2\r\njupyterlab 3.6.1\r\njupyterlab-pygments 0.2.2\r\njupyterlab_server 2.20.0\r\njupyterlab-widgets 3.0.5\r\nlibmambapy 1.1.0\r\nmamba 1.1.0\r\nMarkupSafe 2.1.2\r\nmatplotlib-inline 0.1.6\r\nmistune 2.0.5\r\nmultidict 6.0.4\r\nmultiprocess 0.70.14\r\nnbclassic 0.5.3\r\nnbclient 0.7.2\r\nnbconvert 7.2.9\r\nnbformat 5.7.3\r\nnest-asyncio 1.5.6\r\nnodeenv 1.7.0\r\nnotebook 6.5.3\r\nnotebook_shim 0.2.2\r\nnumpy 1.24.2\r\noutcome 1.2.0\r\npackaging 23.0\r\npandas 1.5.3\r\npandocfilters 1.5.0\r\nparso 0.8.3\r\npexpect 4.8.0\r\npickleshare 0.7.5\r\npip 22.3.1\r\nplatformdirs 3.0.0\r\nplotly 5.13.1\r\npluggy 1.0.0\r\npre-commit 3.1.0\r\nprometheus-client 0.16.0\r\nprompt-toolkit 3.0.38\r\npsutil 5.9.4\r\nptyprocess 0.7.0\r\npure-eval 0.2.2\r\npyarrow 11.0.0\r\npycosat 0.6.4\r\npycparser 2.21\r\nPygments 2.14.0\r\npyOpenSSL 22.1.0\r\npyrsistent 0.19.3\r\nPySocks 1.7.1\r\npytest 7.2.1\r\npytest-asyncio 0.20.3\r\npytest-cov 4.0.0\r\npytest-timeout 2.1.0\r\npython-dateutil 2.8.2\r\npython-json-logger 2.0.7\r\npytz 2022.7.1\r\nPyYAML 6.0\r\npyzmq 25.0.0\r\nrequests 2.28.1\r\nresponses 0.18.0\r\nrfc3339-validator 0.1.4\r\nrfc3986-validator 0.1.1\r\nruamel-yaml-conda 0.15.80\r\nSend2Trash 1.8.0\r\nsetuptools 65.6.3\r\nsimplegeneric 0.8.1\r\nsix 1.16.0\r\nsniffio 1.3.0\r\nsortedcontainers 2.4.0\r\nsoupsieve 2.4\r\nstack-data 0.6.2\r\ntenacity 8.2.2\r\nterminado 0.17.1\r\ntinycss2 1.2.1\r\ntomli 2.0.1\r\ntoolz 0.12.0\r\ntornado 6.2\r\ntqdm 4.65.0\r\ntraitlets 5.8.1\r\ntrio 0.22.0\r\ntyping_extensions 4.5.0\r\nuri-template 1.2.0\r\nurllib3 1.26.13\r\nvirtualenv 20.19.0\r\nwcwidth 0.2.6\r\nwebcolors 1.12\r\nwebencodings 0.5.1\r\nwebsocket-client 1.5.1\r\nwheel 0.38.4\r\nwidgetsnbextension 4.0.5\r\nxxhash 3.2.0\r\ny-py 0.5.9\r\nyarl 1.8.2\r\nypy-websocket 0.8.2\r\nzstandard 0.19.0\r\n```\r\n\r\nAny idea why this is happening? I debugged this to know the tqdm.pbar value is not being updated properly and its not the kernel not sending the comm messages to the IProgress bar" ]

[ "Okay, the issue was likely caused by mixing `conda` and `pip` usage - I forgot that I have already used `pip` in this environment previously and that it was 'spoiled' because of it. Creating another environment and installing `datasets` by pip with other packages from the `requirements.txt` file solved the problem." ]

[ "Answered on the forum:\r\n\r\n> To fix the overflow error, we need to merge [support LargeListArray in pyarrow by xwwwwww · Pull Request #4800 · huggingface/datasets · GitHub](https://github.com/huggingface/datasets/pull/4800), which adds support for the large lists. However, before merging it, we need to come up with a cleaner API for large lists. I hope to find some time to address this before Datasets 3.0." ]

[ "Hi!\r\n\r\nYou can also use names other than \"train\", \"validation\" and \"test\". As an example, check the [script](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/blob/e095840f23f3dffc1056c078c2f9320dad9ca74d/common_voice_11_0.py#L139) of the Common Voice 11 dataset. " ]

[ "Hi! You seem to be using an outdated version of `datasets` that downloads the older script version. To avoid the error, you can either pass `revision=\"main\"` to `load_dataset` (this can fail if a script uses newer features of the lib) or update your installation with `pip install -U datasets` (better solution)." ]

[ "The AutoTrain format is not supported right now. I think it would require a dedicated dataset builder", "Okay, good to know. Thanks for the reply. For now I will just have to\nmanage the split manually before training, because I can’t find any way of\npulling out file indices or file names from the autogenerated split. The\nfile names field of the image dataset (loaded directly from arrow file) is\nmissing, just fyi (for anyone else this might be relevant too).\n\nOn Fri, Mar 10, 2023 at 7:02 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> The AutoTrain format is not supported right now. I think it would require\n> a dedicated dataset builder\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5627#issuecomment-1464734308>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ACBJ4F5A353MCZ76OGRJ6CTW3PFI7ANCNFSM6AAAAAAVWXNUTE>\n> .\n> You are receiving this because you authored the thread.Message ID:\n> ***@***.***>\n>\n" ]

[ "You can use \"json\" instead. It doesn't work by extension names, but rather by dataset builder names, e.g. \"text\", \"imagefolder\", etc. I don't think the example in `transformers` is correct because of that", "Yes, I understand the reasoning but this issue is to propose that the example in transformers (while incorrect) \"makes sense\" in terms of user expectation. So the question is whether it would be possible to add \"aliases\" for common types (like \"json\" and \"text\") based on common extensions (like jsonl and txt)?" ]

[ "Hi @lithafnium, thanks for reporting.\r\n\r\nPlease note that you can use the \"Community\" tab in the corresponding dataset page to start any discussion: https://huggingface.co/datasets/glue/discussions\r\n\r\nIndeed this issue was already raised there (https://huggingface.co/datasets/glue/discussions/5) and answered: https://huggingface.co/datasets/glue/discussions/5#63907885937867f0cb3cde31\r\n> The test labels are not public.\r\n>\r\n> Note this dataset belongs to a benchmark: people send their predictions for the test split to GLUE (https://gluebenchmark.com/) and then they get a score in their leaderboard...\r\n" ]

[]

[]

[ "Hi! You can use `concatenate_datasets([ids1, ids2], axis=1)` to do this." ]

[ "Sorry, I just found https://github.com/apache/beam/issues/24458. It seems this issue is being worked on. ", "Reopening, since I think the docs should inform the user of this problem. For example, [this page](https://huggingface.co/docs/datasets/installation) says \r\n> Datasets is tested on Python 3.7+.\r\n\r\nbut it should probably say that Beam Datasets do not work with Python 3.10 (or link to a known issues page). ", "Same problem on Colab using a vanilla setup running :\r\nPython 3.10.11 \r\napache-beam 2.47.0\r\ndatasets 2.12.0", "Same problem, \r\npy 3.10.11\r\napache-beam==2.47.0\r\ndatasets==2.12.0", "I have made a workaround by forcing an install of the version of `multiprocess` version `0.70.15` (after installing `datasets` and `apache-beam`). I can confirm that (on Python 3.10 in [this colab notebook](https://colab.research.google.com/drive/1PTeGlshamFcJZix_GiS3vMXX_YzAhGv0?usp=sharing)) `datasets` can download pre-processed Wikipedia dumps and can download non-pre-processed dumps using `beam_runner=\"DirectRunner\"`. I don't know if/how other `beam_runner`s can be made compatible.", "Same problem.\r\n\r\n```\r\npython = \"^3.10\"\r\napache-beam = { extras = [\"gcp\"], version = \"2.54.0\" }\r\ndatasets = \"^2.18.0\"\r\n```" ]

[ "I'm attaching a minimal reproducible example:\r\n```python\r\nfrom datasets import load_dataset\r\nimport pyarrow as pa\r\nimport pyarrow.parquet as pq\r\n\r\ntable_with_map = pa.Table.from_pydict(\r\n {\"a\": [1, 2], \"b\": [[(\"a\", 2)], [(\"b\", 4)]]},\r\n schema=pa.schema({\"a\": pa.int32(), \"b\": pa.map_(pa.string(), pa.int32())})\r\n)\r\npq.write_table(table_with_map, \"parquet_with_map.parquet\")\r\ndset = load_dataset(\"parquet\", data_files=\"parquet_with_map.parquet\", split=\"train\") # error unless streaming=True\r\n``` \r\n\r\nFor a dataset generated with the packaged loaders (CSV, JSON, Parquet), `streaming=True` sets the dataset's features to `None` (unless explicitly provided in `load_dataset`), hence no error will be thrown as long as the features stay \"unresolved\" (resolving the features with `_resolve_features` will lead to an error).", "I've also been wondering about datasets support for Arrow Map datatypes. I had a situation where I had a pandas series of dict[str, float] with hundreds of different possible key values (ie. not bounded), and this got converted to a sequence of structs where every single struct had the entire set of keys.\r\n\r\nI worked around it, by explicitly creating a sequence of [str, float], but given that pyarrow has an explicit Map datatype, it would be good to be able to explicitly cast/force this data type combination.", "(feel free to ignore) polars will not support this type: https://github.com/pola-rs/polars/issues/3942#issuecomment-1202331210\r\n\r\n> Polars will not add the map dtype. It's benefit do not outweigh the extra complexity. Maybe we can investigate conversion of maps to struct. But I will have to explore that.", "Looks like they chose to convert every instance with https://github.com/pola-rs/polars/pull/4226" ]

[ "Same problem, \r\ntransformers 4.28.1\r\ndatasets 2.12.0\r\n\r\nleak around 100Mb per 10 seconds when use dataloader_num_werker > 0 in training argumennts for transformer train, possile bug in transformers repo, but still not found solution :(\r\n", "found an article described a problem, may be helpful for somebody:\r\nhttps://ppwwyyxx.com/blog/2022/Demystify-RAM-Usage-in-Multiprocess-DataLoader/\r\nI confirm, it`s not memory leak, after some time memory growing has stopped", "\"After some time\" - from your description, it sounds like memory growth can happen for 12 hours+, even days, before it stops? That seems very scary." ]

[ "Hi! We've recently made some improvements to `save_to_disk`/`list_to_disk` (100x faster in some scenarios), so it would help if you could install `datasets` directly from `main` (`pip install git+https://github.com/huggingface/datasets.git`) and re-run the \"benchmark\".", "Great to hear! I'll give it a try when I've got a moment.", "@mariosasko is that fix released to pip in the meantime? Asking cause im facing still the same issue (regarding loading images from local paths):\r\n```\r\ndataset = load_dataset(\"csv\", cache_dir=\"cache\", data_files=[\"/STORAGE/DATA/mijam/vit/code/list_filtered.csv\"], num_proc=16, split=\"train\").cast_column(\"image\", Image())\r\ndataset = dataset.class_encode_column(\"label\")\r\n```\r\nquite fast. \r\n\r\nThen I do `save_to_disk()` and some time later:\r\n```\r\ndataset = load_from_disk('/STORAGE/DATA/mijam/accel/saved_arrow_big')\r\n```\r\nreally slow. In theory it should be quicked since it only loads arrow files, no conversions and so on.\r\n", "@mjamroz I assume your CSV file stores image file paths. This means `save_to_disk` needs to embed the image bytes resulting in a much bigger Arrow file (than the initial one). Maybe specifying `num_shards` to make the Arrow files smaller can help (large Arrow files on some systems take a long time to load)." ]

[ "Hi!\r\n\r\n> naming convention of mp3 files\r\n\r\nYes, this could be the problem. MP3 files should end with `.mp3`/`.MP3` to be recognized as audio files.\r\n\r\nIf the file names are not the culprit, can you paste the audio folder's directory structure to help us reproduce the error (e.g., by running the `tree \"x\"` command)?", "Hi! I'm sorry, I don't want to reveal my entire dataset, but here's a snippet (all of the mp3 files below are some of the ones not being recognized by audiofolder. Also, for another dataset, audiofolder loaded zero mp3 files because \"train\" was in the name of one of the mp3 files. \r\nmy_dataset\r\n├── data\r\n│   ├── VHA_Innovation_Stories_-_Day_2-123.mp3\r\n│   ├── VHA_Innovation_Stories_-_Day_2-124.mp3\r\n│   ├── ASSOCIATION_OF_GENERAL_PRACTITIONERS_OF_JAMAICA_NEPHROLOGY_CONFERENCE_-_JULY_3,_2022-93.mp3\r\n│   ├── ASSOCIATION_OF_GENERAL_PRACTITIONERS_OF_JAMAICA_NEPHROLOGY_CONFERENCE_-_JULY_3,_2022-94.mp3\r\n│   ├── ASSOCIATION_OF_GENERAL_PRACTITIONERS_OF_JAMAICA_NEPHROLOGY_CONFERENCE_-_JULY_3,_2022-95.mp3\r\n│   ├── Your_Impact\\357\\274\\232_Neurosurgery_equipment-5.mp3\r\n│   └── Your_Impact\\357\\274\\232_Neurosurgery_equipment-6.mp3\r\n└── metadata.csv\r\n\r\nHere's a few of the 13 files recognized by the dataset:\r\nBritish_Heart_Foundation_-_Your_guide_to_a_Coronary_Angiogram,_a_test_for_heart_disease-1.mp3\r\nBritish_Heart_Foundation_-_Your_guide_to_a_Coronary_Angiogram,_a_test_for_heart_disease-2.mp3\r\nBritish_Heart_Foundation_-_Your_guide_to_a_Coronary_Angiogram,_a_test_for_heart_disease-3.mp3\r\nIVP_⧸_IVU_test_Procedure_for_Kidneys_intravenous_pyelogram_-_medical_radiology_X-ray_ivp-1.mp3\r\nIVP_⧸_IVU_test_Procedure_for_Kidneys_intravenous_pyelogram_-_medical_radiology_X-ray_ivp-2.mp3" ]

[ "Hello, I have an interest in this issue.\r\nIs the `Dataset.to_dict` you are describing correct in the code here?\r\n\r\nhttps://github.com/huggingface/datasets/blob/35b789e8f6826b6b5a6b48fcc2416c890a1f326a/src/datasets/arrow_dataset.py#L4633-L4667", "Yes, this is where `Dataset.to_dict` is defined.", "#self-assign" ]

[ "Hi! \r\n\r\n\r\nYou can specify `download_config=DownloadConfig(resume_download=True))` in `load_dataset` to resume the download when re-running the code after the timeout error:\r\n```python\r\nfrom datasets import load_dataset, DownloadConfig\r\ndataset = load_dataset('the_pile', split='train', cache_dir='F:\\datasets', download_config=DownloadConfig(resume_download=True))\r\n```\r\n\r\n", "@mariosasko , I used your suggestion but its not saving anything , just stops and runs from the same point .\r\nbelow is the script to download and save on disk .\r\n\r\n```\r\nfrom datasets import load_dataset, DownloadConfig\r\n\r\n\r\n#load the Pile dataset from Hugging Face Datasets\r\n#dataset = load_dataset('the_pile')\r\ndataset = load_dataset('the_pile', split='train', cache_dir='datasets', download_config=DownloadConfig(resume_download=True))\r\n\r\n\r\n# save each file in the dataset to disk\r\nfor i, example in enumerate(dataset['train']):\r\n filename = f'pile_file_{i}.json'\r\n with open(filename, 'w') as f:\r\n f.write(str(example))\r\n\r\nprint(\"Finished saving Pile dataset files to disk.\")\r\n```\r\n", "@mariosasko , it shows nothing in dataset folder\r\n\r\n```\r\n du -sh /mnt/nlp/hugging_face/*\r\n20K /mnt/nlp/hugging_face/datasets\r\n4.0K /mnt/nlp/hugging_face/download_pile.py\r\n```\r\n", "@mariosasko \r\n\r\n```\r\nroot@d20f0ab8f4f8:/mnt/hugging_face# python3 download_pile.py\r\nNo config specified, defaulting to: the_pile/all\r\nDownloading and preparing dataset the_pile/all to /mnt/hugging_face/datasets/the_pile/all/0.0.0/6fadc480ecb32470826cbf5900a9558b791ce55d5e9a0fdc8ad653e7b64bb349...\r\nDownloading data files: 0%| | 0/3 [00:00<?, ?it/s]\r\n\r\n\r\n\r\n\r\n\r\nDownloading data: 70%|████████████████████████████████████████████████████████████████████▊ | 10.7G/15.2G [12:09<11:53, 6.36MB/s]\r\nDownloading data: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 15.2G/15.2G [22:15<00:00, 7.25MB/s]\r\nDownloading data: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 15.2G/15.2G [46:17<00:00, 5.48MB/s]\r\nDownloading data: 40%|██████████████████████████████████████▏ | 6.07G/15.3G [50:49<1:17:02, 1.99MB/s]\r\nTraceback (most recent call last):██████████████████████████▊ | 6.07G/15.3G [50:49<25:35:23, 99.9kB/s]\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 444, in _error_catcher\r\n yield\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 567, in read\r\n data = self._fp_read(amt) if not fp_closed else b\"\"\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 525, in _fp_read\r\n data = self._fp.read(chunk_amt)\r\n File \"/usr/lib/python3.8/http/client.py\", line 459, in read\r\n n = self.readinto(b)\r\n File \"/usr/lib/python3.8/http/client.py\", line 503, in readinto\r\n n = self.fp.readinto(b)\r\n File \"/usr/lib/python3.8/socket.py\", line 669, in readinto\r\n return self._sock.recv_into(b)\r\n File \"/usr/lib/python3.8/ssl.py\", line 1241, in recv_into\r\n return self.read(nbytes, buffer)\r\n File \"/usr/lib/python3.8/ssl.py\", line 1099, in read\r\n return self._sslobj.read(len, buffer)\r\nConnectionResetError: [Errno 104] Connection reset by peer\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.8/dist-packages/requests/models.py\", line 816, in generate\r\n yield from self.raw.stream(chunk_size, decode_content=True)\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 628, in stream\r\n data = self.read(amt=amt, decode_content=decode_content)\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 593, in read\r\n raise IncompleteRead(self._fp_bytes_read, self.length_remaining)\r\n File \"/usr/lib/python3.8/contextlib.py\", line 131, in __exit__\r\n self.gen.throw(type, value, traceback)\r\n File \"/usr/local/lib/python3.8/dist-packages/urllib3/response.py\", line 461, in _error_catcher\r\n raise ProtocolError(\"Connection broken: %r\" % e, e)\r\nurllib3.exceptions.ProtocolError: (\"Connection broken: ConnectionResetError(104, 'Connection reset by peer')\", ConnectionResetError(104, 'Connection reset by peer'))\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"download_pile.py\", line 6, in <module>\r\n dataset = load_dataset('the_pile', split='train', cache_dir='datasets', download_config=DownloadConfig(resume_download=True))\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/load.py\", line 1782, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/builder.py\", line 872, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/builder.py\", line 1649, in _download_and_prepare\r\n super()._download_and_prepare(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/builder.py\", line 945, in _download_and_prepare\r\n split_generators = self._split_generators(dl_manager, **split_generators_kwargs)\r\n File \"/root/.cache/huggingface/modules/datasets_modules/datasets/the_pile/6fadc480ecb32470826cbf5900a9558b791ce55d5e9a0fdc8ad653e7b64bb349/the_pile.py\", line 192, in _split_generators\r\n data_dir = dl_manager.download(_DATA_URLS[self.config.name])\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/download/download_manager.py\", line 427, in download\r\n downloaded_path_or_paths = map_nested(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/py_utils.py\", line 443, in map_nested\r\n mapped = [\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/py_utils.py\", line 444, in <listcomp>\r\n _single_map_nested((function, obj, types, None, True, None))\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/py_utils.py\", line 363, in _single_map_nested\r\n mapped = [_single_map_nested((function, v, types, None, True, None)) for v in pbar]\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/py_utils.py\", line 363, in <listcomp>\r\n mapped = [_single_map_nested((function, v, types, None, True, None)) for v in pbar]\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/py_utils.py\", line 346, in _single_map_nested\r\n return function(data_struct)\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/download/download_manager.py\", line 453, in _download\r\n return cached_path(url_or_filename, download_config=download_config)\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/file_utils.py\", line 182, in cached_path\r\n output_path = get_from_cache(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/file_utils.py\", line 575, in get_from_cache\r\n http_get(\r\n File \"/usr/local/lib/python3.8/dist-packages/datasets/utils/file_utils.py\", line 379, in http_get\r\n for chunk in response.iter_content(chunk_size=1024):\r\n File \"/usr/local/lib/python3.8/dist-packages/requests/models.py\", line 818, in generate\r\n raise ChunkedEncodingError(e)\r\nrequests.exceptions.ChunkedEncodingError: (\"Connection broken: ConnectionResetError(104, 'Connection reset by peer')\", ConnectionResetError(104, 'Connection reset by peer'))\r\n```\r\n", "Users with slow internet speed are doomed (4MB/s). The dataset downloads fine at minimum speed 10MB/s.\n\nAlso, when the train splits were generated and then I removed the downloads folder to save up disk space, it started redownloading the whole dataset. Is there any way to use the already generated splits instead?", "@sentialx @mariosasko , anytime on my above script , am I downloading and saving dataset correctly . Please suggest :)", "@sentialx probably worth noting that `resume_download=True` doesn't directly save the dataset to disk, but instead just helps in resuming the dataset resume on interruption as @mariosasko mentions. resolving resumptions after a crash is [an open issue](https://github.com/huggingface/datasets/issues/5380) at the moment." ]

[ "Hi! \r\n\r\nIt's better to report this kind of issue in the `huggingface_hub` repo, so if you still haven't resolved it, I suggest you open an issue there.", "Yeah, I solved it. Problem was in osxkeychain. When I do `hugginface-cli login` it's add token with default account (username)`hg_user` but my repo contain other username. When I changed username in keychain - it works now." ]

[ "Hi! \r\n\r\n> (see example of DreamboothDatasets)\r\n\r\n\r\nCould you please provide a link to it? If you are referring to the example in the `diffusers` repo, your issue is unrelated to `datasets` as that example uses `Dataset` from PyTorch to load data." ]

[ "We won't support in-place modifications since `datasets` is based on the Apache Arrow format which doesn't support in-place modifications.\r\n\r\nIn your case the old dataset is garbage collected pretty quickly so you won't have memory issues.\r\n\r\nNote that datasets loaded from disk (memory mapped) are not loaded in memory, and therefore the new dataset actually use the same buffers as the old one.", "Thank you for your detailed reply.\r\n\r\n> In your case the old dataset is garbage collected pretty quickly so you won't have memory issues.\r\n\r\nI understand this, but it still copies the old dataset to create the new one, is this correct? So maybe it is not memory-consuming, but time-consuming?", "Indeed, and because of that it is more efficient to add multiple rows at once instead of one by one, using `concatenate_datasets` for example." ]

[ "Apparently some JSON objects have a `\"labels\"` field. Since this field is not present in every object, you must specify all the fields types in the README.md\r\n\r\nEDIT: actually specifying the feature types doesn’t solve the issue, it raises an error because “labels” is missing in the data", "We've updated the dataset to remove the extra `labels` field from some files, closing this issue. Thanks!", "A similar error occurs in the Pile dataset (EleutherAI/the_pile)\r\n\r\nLoading the dataset produces the following error.\r\n\r\n```\r\nTypeError: Couldn't cast array of type\r\nstruct<file: string, id: string>\r\nto\r\n{'id': Value(dtype='string', id=None)}\r\n```\r\n", "I think this was fixed in https://huggingface.co/datasets/EleutherAI/the_pile/discussions/11", "i have the same problem ，how to solve :\r\n raise TypeError(f\"Couldn't cast array of type\\n{array.type}\\nto\\n{feature}\")\r\nTypeError: Couldn't cast array of type\r\nlist<item: string>\r\nto\r\n{'content': Value(dtype='string', id=None), 'role': Value(dtype='string', id=None)}" ]

[ "Hi! I cannot reproduce this error on my machine.\r\n\r\nThe raised error could mean that one of the downloaded files is corrupted. To verify this is not the case, you can run `load_dataset` as follows:\r\n```python\r\ntrain_dataset = load_dataset('xtreme', 'udpos.English', split=\"train\", cache_dir=args.cache_dir, download_mode=\"force_redownload\", verification_mode=\"all_checks\")\r\n```", "Hi! Apologies for the delayed response! I tried the above and it doesn't solve the issue. Actually, the dataset gets downloaded most times, but sometimes this error occurs (at random afaik). Is it possible that there is a server issue for this particular dataset? I am able to download other datasets using the same code on the same machine with no issues :( I get this error now : \r\n```\r\nDownloading data: 16%|███████████████▌ | 55.9M/355M [04:45<25:25, 196kB/s]\r\nTraceback (most recent call last):\r\n File \"/home/skhanuja/Optimal-Resource-Allocation-for-Multilingual-Finetuning/src/train_al.py\", line 1107, in <module>\r\n main()\r\n File \"/home/skhanuja/Optimal-Resource-Allocation-for-Multilingual-Finetuning/src/train_al.py\", line 439, in main\r\n en_dataset = load_dataset(\"xtreme\", \"udpos.English\", split=\"train\", download_mode=\"force_redownload\", verification_mode=\"all_checks\")\r\n File \"/home/skhanuja/miniconda3/envs/multilingual_ft/lib/python3.10/site-packages/datasets/load.py\", line 1782, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/home/skhanuja/miniconda3/envs/multilingual_ft/lib/python3.10/site-packages/datasets/builder.py\", line 872, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/home/skhanuja/miniconda3/envs/multilingual_ft/lib/python3.10/site-packages/datasets/builder.py\", line 1649, in _download_and_prepare\r\n super()._download_and_prepare(\r\n File \"/home/skhanuja/miniconda3/envs/multilingual_ft/lib/python3.10/site-packages/datasets/builder.py\", line 949, in _download_and_prepare\r\n verify_checksums(\r\n File \"/home/skhanuja/miniconda3/envs/multilingual_ft/lib/python3.10/site-packages/datasets/utils/info_utils.py\", line 62, in verify_checksums\r\n raise NonMatchingChecksumError(\r\ndatasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['https://lindat.mff.cuni.cz/repository/xmlui/bitstream/handle/11234/1-3105/ud-treebanks-v2.5.tgz']\r\nSet `verification_mode='no_checks'` to skip checksums verification and ignore this error\r\n```", "If this happens randomly, then this means the data file from the error message is not always downloaded correctly. \r\n\r\nThe only solution in this scenario is to download the dataset again by passing `download_mode=\"force_redownload\"` to the `load_dataset` call.", "Wow. I effectively have to redownload a dataset of 1TB because of this now?\r\nBecause 3% of its parts are broken?\r\n\r\nWhy is this downloader library so sh*t and badly documented also? I found almost nothing on the net, at least finally this issue about the problem here.\r\nNo words to express how disappointed I am by that dataset tool provided by Huggingface here, which I sadly have to use because HF is the only place where the Dataset I plan to work with is hosted....\r\n\r\nI mean... checksum check after download... or hitting timeout of a part... and redownload if not matching... that's content of every junior developer training session.\r\n\r\nI added `verification_mode=\"all_checks\"`. And it really calculated checksums for 4096 parts of ~350 MB... But then did nothing and tried to extract still, hitting the error again. \r\n\r\nEDIT: Apparently it is able to fix it by getting a little help: Just delete the broken parts and associated files from `~/.cache/huggingface/datasets/downloads`", "I'm getting it too, although just retrying fixed it. Nevertheless, the dataset is too large to have re-downloaded the whole thing, for it's probably just one file with an issue. It would be good to know if there's a way people could manually examine the files (first for sizes, then possibly checksums)... going to the web or elsewhere to compare and correct it by hand, if ever needed.", "Okay, no, it got further but it is repeatedly giving me:\r\n```/home/jaggz/.cache/huggingface/modules/datasets_modules/datasets/mozilla-foundation--common_voice_11_0/3f27acf10f303eac5b6fbbbe02495aeddb46ecffdb0a2fe3507fcfbf89094631/common_voice_11_0.py\", line 195, in _generate_examples\r\nresult[\"audio\"] = {\"path\": path, \"bytes\": file.read()}\r\n^^^^^^^^^^^\r\nFile \"/usr/lib/python3.11/tarfile.py\", line 687, in read\r\nraise ReadError(\"unexpected end of data\")\r\ntarfile.ReadError: unexpected end of data\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\nFile \"/home/jaggz/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 625, in <module>\r\nmain()\r\nFile \"/home/jaggz/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 360, in main\r\nraw_datasets[\"train\"] = load_dataset(\r\n^^^^^^^^^^^^^\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/load.py\", line 2153, in load_dataset\r\nbuilder_instance.download_and_prepare(\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/builder.py\", line 954, in download_and_prepare\r\nself._download_and_prepare(\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/builder.py\", line 1717, in _download_and_prepare\r\nsuper()._download_and_prepare(\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/builder.py\", line 1049, in _download_and_prepare\r\nself._prepare_split(split_generator, **prepare_split_kwargs)\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/builder.py\", line 1555, in _prepare_split\r\nfor job_id, done, content in self._prepare_split_single(\r\nFile \"/home/jaggz/venvs/pynow/lib/python3.11/site-packages/datasets/builder.py\", line 1712, in _prepare_split_single\r\nraise DatasetGenerationError(\"An error occurred while generating the dataset\") from e\r\ndatasets.builder.DatasetGenerationError: An error occurred while generating the datase\r\n", "@RuntimeRacer \r\n> EDIT: Apparently it is able to fix it by getting a little help: Just delete the broken parts and associated files from `~/.cache/huggingface/datasets/downloads`\r\n\r\nHow do you know the broken parts?\r\nMine's consistently erroring and.. yeah, really this thing should be able to check the files (but where's that even done)...\r\n\r\n2023-11-02 00:14:09.846055: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\r\nTo enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\n/home/j/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py:299: FutureWarning: The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.\r\n warnings.warn(\r\n11/02/2023 00:14:37 - WARNING - __main__ - Process rank: 0, device: cuda:0, n_gpu: 1, distributed training: False, 16-bits training: True\r\n11/02/2023 00:14:37 - INFO - __main__ - Training/evaluation parameters Seq2SeqTrainingArguments(\r\n_n_gpu=1,\r\nadafactor=False,\r\nadam_beta1=0.9,\r\nadam_beta2=0.999,\r\n...\r\nlogging_dir=./whisper-tiny-en/runs/Nov02_00-14-28_jsys,\r\n...\r\nrun_name=./whisper-tiny-en,\r\n...\r\nweight_decay=0.0,\r\n)\r\n11/02/2023 00:14:37 - INFO - __main__ - Training/evaluation parameters Seq2SeqTrainingArguments(\r\n_n_gpu=1,\r\nadafactor=False,\r\n...\r\nlogging_dir=./whisper-tiny-en/runs/Nov02_00-14-28_jsys,\r\n...\r\nweight_decay=0.0,\r\n)\r\n\r\nDownloading data files: 0%| | 0/5 [00:00<?, ?it/s]\r\nDownloading data files: 100%|██████████| 5/5 [00:00<00:00, 2426.42it/s]\r\n\r\nExtracting data files: 0%| | 0/5 [00:00<?, ?it/s]\r\nExtracting data files: 100%|██████████| 5/5 [00:00<00:00, 421.16it/s]\r\n\r\nDownloading data files: 0%| | 0/5 [00:00<?, ?it/s]\r\nDownloading data files: 100%|██████████| 5/5 [00:00<00:00, 18707.87it/s]\r\n\r\nExtracting data files: 0%| | 0/5 [00:00<?, ?it/s]\r\nExtracting data files: 100%|██████████| 5/5 [00:00<00:00, 3754.97it/s]\r\n\r\nGenerating train split: 0 examples [00:00, ? examples/s]\r\n\r\nReading metadata...: 0it [00:00, ?it/s]\u001b[A\r\n...\r\nReading metadata...: 948736it [00:23, 40632.92it/s] \r\n\r\nGenerating train split: 1 examples [00:23, 23.37s/ examples]\r\n...\r\nGenerating train split: 948736 examples [08:28, 1866.15 examples/s]\r\n\r\nGenerating validation split: 0 examples [00:00, ? examples/s]\r\n\r\nReading metadata...: 0it [00:00, ?it/s]\u001b[A\r\n\r\nReading metadata...: 16089it [00:00, 157411.88it/s]\u001b[A\r\nReading metadata...: 16354it [00:00, 158233.27it/s]\r\n\r\nGenerating validation split: 1 examples [00:00, 7.60 examples/s]\r\nGenerating validation split: 16354 examples [00:14, 1154.77 examples/s]\r\n\r\nGenerating test split: 0 examples [00:00, ? examples/s]\r\n\r\nReading metadata...: 0it [00:00, ?it/s]\u001b[A\r\nReading metadata...: 16354it [00:00, 194855.03it/s]\r\n\r\nGenerating test split: 1 examples [00:00, 4.53 examples/s]\r\nGenerating test split: 16354 examples [00:07, 2105.43 examples/s]\r\n\r\nGenerating other split: 0 examples [00:00, ? examples/s]\r\n\r\nReading metadata...: 0it [00:00, ?it/s]\u001b[A\r\nReading metadata...: 290846it [00:01, 235823.90it/s]\r\n\r\nGenerating other split: 1 examples [00:01, 1.27s/ examples]\r\n...\r\nGenerating other split: 290846 examples [02:12, 2196.96 examples/s]\r\nGenerating invalidated split: 0 examples [00:00, ? examples/s]\r\nReading metadata...: 252599it [00:01, 241965.85it/s]\r\n\r\nGenerating invalidated split: 1 examples [00:01, 1.08s/ examples]\r\n...\r\nGenerating invalidated split: 60130 examples [00:34, 1764.14 examples/s]\r\nTraceback (most recent call last):\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 1676, in _prepare_split_single\r\n for key, record in generator:\r\n File \"/home/j/.cache/huggingface/modules/datasets_modules/datasets/mozilla-foundation--common_voice_11_0/3f27acf10f303eac5b6fbbbe02495aeddb46ecffdb0a2fe3507fcfbf89094631/common_voice_11_0.py\", line 195, in _generate_examples\r\n result[\"audio\"] = {\"path\": path, \"bytes\": file.read()}\r\n ^^^^^^^^^^^\r\n File \"/usr/lib/python3.11/tarfile.py\", line 687, in read\r\n raise ReadError(\"unexpected end of data\")\r\ntarfile.ReadError: unexpected end of data\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"/home/j/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 625, in <module>\r\n main()\r\n File \"/home/j/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 360, in main\r\n raw_datasets[\"train\"] = load_dataset(\r\n ^^^^^^^^^^^^^\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/load.py\", line 2153, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 954, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 1717, in _download_and_prepare\r\n super()._download_and_prepare(\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 1049, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 1555, in _prepare_split\r\n for job_id, done, content in self._prepare_split_single(\r\n File \"/home/j/venvs/pycur/lib/python3.11/site-packages/datasets/builder.py\", line 1712, in _prepare_split_single\r\n raise DatasetGenerationError(\"An error occurred while generating the dataset\") from e\r\ndatasets.builder.DatasetGenerationError: An error occurred while generating the dataset\r\n", "@jaggzh Hi, I actually came around with a fix for this, wasn't that easy to solve since there were a lot of hidden pitfalls in the code, and it's quite hacky, but I was able to download the full dataset.\r\n\r\nI just didn't create a PR for it yet since I was too lazy to create a fork and change my local repo's origin. 😅 \r\nLet me try to do this tonight, I'll give you a ping once it's up.\r\n\r\nEDIT: And no, what I wrote above about adding a param to the download config does NOT solve it apparently. A code fix is required here.", "@jaggzh PR is up: https://github.com/huggingface/datasets/pull/6380\r\n\r\n🤞 on approval for merge to the main repo.", "@mariosasko Can you re-open this? We really need some better diagnostics output, at the least, to locate which files are contributing, some checksum output, etc. I can't even tell if this is a mozilla...py issue or huggingface datasets or ....", "@RuntimeRacer \r\nBeautiful, thank you so much. I patched with your PR and am re-running now.\r\n(I'm running this script: https://github.com/huggingface/transformers/blob/main/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py)\r\nOkay, actually it failed; so now I'm running with verification_mode='all_checks' added to the load_data() call and it's re-running now. Wish me luck.\r\n(Note: It's generating checksums; I don't see an option that handles anything between basic_checks and all_checks -- Something checking dl'ed files' lengths would be a good common fix I'd think; corruption is more rare nowadays than a short file (although maybe your patch helps prevent that in the first place.) :}", "@RuntimeRacer \r\nNo luck. Sigh.\r\n[Edit: My tmux copy didn't get some data. That was weird. I'm adding in the initial part of the output:]\r\n```\r\nDownloading data files: 100%|██████████| 5/5 [00:00<00:00, 2190.69it/s]\r\nComputing checksums: 100%|██████████| 41/41 [11:39<00:00, 17.05s/it] Extracting data files: 100%|██████████| 5/5 [00:00<00:00, 12.37it/s]\r\nDownloading data files: 100%|██████████| 5/5 [00:00<00:00, 107.64it/s]\r\nExtracting data files: 100%|██████████| 5/5 [00:00<00:00, 3149.82it/s]\r\nReading metadata...: 948736it [00:03, 243227.36it/s]s/s]\r\n...\r\n```\r\n```\r\n...\r\nReading metadata...: 252599it [00:01, 249267.71it/s]xamples/s]\r\nGenerating invalidated split: 60130 examples [00:31, 1916.33 examples/s]\r\nTraceback (most recent call last):\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 1676, in _prepare_split_single\r\nfor key, record in generator:\r\nFile \"/home/j/.cache/huggingface/modules/datasets_modules/datasets/mozilla-foundation--common_voice_11_0/3f27acf10f303eac5b6fbbbe02495aeddb46ecffdb0a2fe3507fcfbf89094631/common_voice_11_0.py\", line 195, in _generate_examples\r\nresult[\"audio\"] = {\"path\": path, \"bytes\": file.read()}\r\n^^^^^^^^^^^\r\nFile \"/usr/lib/python3.11/tarfile.py\", line 687, in read\r\nraise ReadError(\"unexpected end of data\")\r\ntarfile.ReadError: unexpected end of data\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\nFile \"/home/j/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 627, in <module>\r\nmain()\r\nFile \"/home/j/src/transformers/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py\", line 360, in main\r\nraw_datasets[\"train\"] = load_dataset(\r\n^^^^^^^^^^^^^\r\nFile \"/home/j/src/py/datasets/src/datasets/load.py\", line 2153, in load_dataset\r\nbuilder_instance.download_and_prepare(\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 954, in download_and_prepare\r\nself._download_and_prepare(\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 1717, in _download_and_prepare\r\nsuper()._download_and_prepare(\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 1049, in _download_and_prepare\r\nself._prepare_split(split_generator, **prepare_split_kwargs)\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 1555, in _prepare_split\r\nfor job_id, done, content in self._prepare_split_single(\r\nFile \"/home/j/src/py/datasets/src/datasets/builder.py\", line 1712\r\n```", "I'm unable to reproduce this error. Based on https://github.com/psf/requests/issues/4956, newer releases of `urllib3` check the returned content length by default, so perhaps updating `requests` and `urllib3` to the latest versions (`pip install -U requests urllib3`) and loading the dataset with `datasets.load_dataset(\"xtreme\", \"udpos.English\", download_config=datasets.DownloadConfig(resume_download=True))` (re-run when it fails to resume the download) can fix the issue.", "@jaggzh I think you will need to re-download the whole dataset with my patched code. Files which have already been downloaded and marked as complete by the broken downloader won't be detected even on re-run (I described that in the PR).\r\nI also had to download reazonspeech, which is over 1TB, twice. 🙈 \r\nFor re-download, you need to manually delete the dataset files from your local machine's huggingface download cache.\r\n\r\n@mariosasko Not sure how you tested it, but it's not an issue in `requests` or `urllib`. The problem is the huggingface downloader, which generates a nested download thread for the actual download I think.\r\nThe issue I had with the reazonspeech dataset (https://huggingface.co/datasets/reazon-research/reazonspeech/tree/main) basically was, that it started downloading a part, but sometimes the connection would 'starve' and only continue with a few kilobytes, and eventually stop receiving any data at all.\r\nSometimes it would even recover during the download and finish properly.\r\nHowever, if it did not recover, the request would hit the really generous default timeout (which is 100 seconds I think), however the exception thrown by the failure inside `urllib`, isn't captured or handled by the upper level downloader code of the `datasets` library.\r\n`datasets` even has a retry mechanism, which would continue interrupted downloads if they have the `.incomplete` suffix, which isn't cleared if, for example, a manual `CTRL+C` is sent by the user to the python process.\r\nBut: If it runs into that edge case I described above (TL;DR: connection starves after minutes + timeout exception which isn't captured), the cache downloader will consider the download as successful and remove the `.incomplete` suffix nevertheless, leaving the archive file in a corrupted state.\r\n\r\nHonestly, I spent hours on trying to figure out what was even going on and why the retry mechanics of the cache downloader didn't work at all.\r\nBut it is indeed an issue caused by the download process itself not receiving any info about actual content size and filesize size on disk of the archive to be downloaded, thus, having no direct control in case something fails on the request level.\r\n\r\nIMHO, this requires a major refactor of the way this part of the downloader works.\r\nYet I was able to quick-fix it by adding some synthetic Exception handling and explicit retry-handling in the code, als done in my PR.", "@RuntimeRacer \r\nUgh. It took a day. I'm seeing if I can get some debug code in here to examine the files myself. (I'm not sure why checksum tests would fail, so, yeah, I think you're right -- this stuff needs some work. Going through ipdb right now to try to get some idea of what's going on in the code).", "@RuntimeRacer Data can only be appended to the `.incomplete` files if `load_dataset` is called with `download_config=DownloadConfig(resume_download=True)`. \r\n\r\nWhere exactly does this exception happen (in the code)? The error stack trace would help a lot.", "@mariosasko I do not have a trace of this exception nor do I know which type it is. I am honestly not even sure if an exception is thrown, or the process just aborts without error.\r\n\r\n> @RuntimeRacer Data can only be appended to the .incomplete files if load_dataset is called with download_config=DownloadConfig(resume_download=True).\r\n\r\nWell, I think I did a very clear explaination of the issue in the PR I shared, and the description above, but maybe I wasn't precise enough. Let me try to explain once more:\r\n\r\nWhat you mention here is the \"normal\" case, if the process is aborted. In this case, there will be files with `.incomplete` suffix, which the cache downloader can continue to download. That is correct.\r\n\r\nBUT: What I am talking about all the time is an edge case: if the download step crashes / timeouts internally, the cache downloader will NOT be aware of this, and REMOVES the `.incomplete` suffix.\r\nIt does NOT know that the file is incomplete when the `http_get` function returns and will remove the `.incomplete` suffix in any case once `http_get` returns.\r\nBut the problem is that `http_get` returns without failure, even if the download failed.\r\nAnd this is still a problem even with latest `urllib` and `requests` library.\r\n", "@RuntimeRacer Updating `urllib3` and `requests` to the latest versions fixes the issue explained in this [blog](https://blog.petrzemek.net/2018/04/22/on-incomplete-http-reads-and-the-requests-library-in-python/) post. \r\n\r\nHowever, the issue explained above seems more similar to [this](https://stackoverflow.com/questions/52731196/python-3-6-5-requests-with-streaming-getting-stuck-in-iter-content-even-if-chun) one. To address it, we can reduce the default timeout to 10 seconds (btw, this was the initial value, but it was causing problems for some users) and expose a config variable so that users can easily control it. Additionally, we can re-run `http_get` similarly to https://github.com/huggingface/huggingface_hub/pull/1766 when the connection/timeout error happens to make the logic even more robust. Would this work for you? The last part is what you did in the PR, right?\r\n\r\n@jaggzh From all the datasets mentioned in this issue, `xtreme` is the only one that stores the data file checksums in the metadata. So, the checksum check has no effect when enabled for the rest of the datasets.", "(I don't have any .incomplete files, just the extraction errors.)\r\nI was going through the code to try to relate filenames to the hex/hash files, but realized I might not need to.\r\nSo instead I coded up a script in bash to examine the tar files for validity (had an issue with bash subshells not adding to my array so I had cgpt recode it in perl).\r\n\r\n```perl\r\n#!/usr/bin/perl\r\nuse strict;\r\nuse warnings;\r\n\r\n# Initialize the array to store tar files\r\nmy @tars;\r\n\r\n# Open the current directory\r\nopendir(my $dh, '.') or die \"Cannot open directory: $!\";\r\n\r\n# Read files in the current directory\r\nwhile (my $f = readdir($dh)) {\r\n # Skip files ending with lock, json, or py\r\n next if $f =~ /\\.(lock|json|py)$/;\r\n\r\n # Use the `file` command to determine the type of file\r\n my $ft = `file \"$f\"`;\r\n\r\n # If it's a tar archive, add it to the list\r\n if ($ft =~ /tar archive/) {\r\n push @tars, $f;\r\n }\r\n}\r\n\r\nclosedir($dh);\r\n\r\nprint \"Final Tars count: \" . scalar(@tars) . \"\\n\";\r\n\r\n# Iterate over the tar files and check them\r\nforeach my $i (0 .. $#tars) {\r\n my $f = $tars[$i];\r\n printf '%d/%d ', $i+1, scalar(@tars);\r\n \r\n # Use `ls -lgG` to list the files, similar to the original bash script\r\n system(\"ls -lgG '$f'\");\r\n\r\n # Check the integrity of the tar file\r\n my $errfn = \"/tmp/$f.tarerr\";\r\n if (system(\"tar tf '$f' > /dev/null 2> '$errfn'\") != 0) {\r\n print \" BAD $f\\n\";\r\n print \" ERR: \";\r\n system(\"cat '$errfn'\");\r\n }\r\n\r\n # Remove the error file if it exists\r\n unlink $errfn if -e $errfn;\r\n}\r\n```\r\n\r\nThis found one hash file that errored in the tar extraction, and one small tmp* file that also was supposedly a tar and was erroring. I removed those two and re-data loaded.. it grabbed just what it needed and I'm on my way. Yay!\r\n\r\nSo... is there a way for the datasets api to get file sizes? That would be a very easy and fast test, leaving checksum slowdowns for extra-messed-up situations.\r\n\r\n", "> @RuntimeRacer Updating `urllib3` and `requests` to the latest versions fixes the issue explained in this [blog](https://blog.petrzemek.net/2018/04/22/on-incomplete-http-reads-and-the-requests-library-in-python/) post.\r\n> \r\n> However, the issue explained above seems more similar to [this](https://stackoverflow.com/questions/52731196/python-3-6-5-requests-with-streaming-getting-stuck-in-iter-content-even-if-chun) one. To address it, we can reduce the default timeout to 10 seconds (btw, this was the initial value, but it was causing problems for some users) and expose a config variable so that users can easily control it. Additionally, we can re-run `http_get` similarly to [huggingface/huggingface_hub#1766](https://github.com/huggingface/huggingface_hub/pull/1766) when the connection/timeout error happens to make the logic even more robust. Would this work for you? The last part is what you did in the PR, right?\r\n> \r\n> @jaggzh From all the datasets mentioned in this issue, `xtreme` is the only one that stores the data file checksums in the metadata. So, the checksum check has no effect when enabled for the rest of the datasets.\r\n\r\n@mariosasko Well if you look at my commit date, you will see that I run into this problem still in October. The blog post you mention and the update in the pull request for `urllib` was from July: https://github.com/psf/requests/issues/4956#issuecomment-1648632935\r\n\r\nBut yeah the [issue on StackOverflow](https://stackoverflow.com/questions/52731196/python-3-6-5-requests-with-streaming-getting-stuck-in-iter-content-even-if-chun) you mentioned seems like that's the source issue I was running into there.\r\nI experimented with timeouts, but changing them didn't help to resolve the issue of the starving connection unfortunately.\r\nHowever, https://github.com/huggingface/huggingface_hub/pull/1766 seems like that could be working; it's very similar to my change. So yeah I think this would fix it probably.\r\n\r\nAlso I can confirm the checksum option did not work for [reazonspeech](https://huggingface.co/datasets/reazon-research/reazonspeech/tree/main) as well. So maybe it's a double edge case that only occurs for some datasets. 🤷‍♂️ ", "Also, the hf urls to files -- while I can't see a way of getting a listing from the hf site side -- do include the file size in the http header response. So we do have a quick way of just verifying lengths for resume. (This message may not be interesting to you all).\r\n\r\nFirst, a json clip (mozilla-foundation___common_voice_11_0/en/11.0.0/3f27acf10f303eac5b6fbbbe02495aeddb46ecffdb0a2fe3507fcfbf89094631/dataset_info.json):\r\n\r\n* I don't know how specific this .json is to mozilla common voice\r\n* Note that *dataset_size* is not the dataset size :) DatasetInfo class docs indicate it might be their \"combined size in bytes of the Arrow tables for all splits.\"\r\n* *num_bytes*: does match the individual file size though, and matches the http header (further down)\r\n```\r\n{\r\n \"builder_name\" : \"common_voice_11_0\",\r\n...\r\n \"config_name\" : \"en\",\r\n \"dataset_name\" : \"common_voice_11_0\",\r\n \"dataset_size\" : 1680793952,\r\n...\r\n \"download_checksums\" : {\r\n...\r\n \"https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/resolve/main/audio/en/invalidated/en_invalidated_3.tar\" : {\r\n \"checksum\" : null,\r\n \"num_bytes\" : 2110853120\r\n },\r\n...\r\n```\r\n\r\n```bash\r\n~/.cache/huggingface/datasets/downloads$ ls -lgG b45f82cb87bab2c35361857fcd46042ab658b42c37dc9a455248c2866c9b8f40* | cut -c 14-\r\n```\r\n```\r\n2110853120 Nov 1 16:28 b45f82cb87bab2c35361857fcd46042ab658b42c37dc9a455248c2866c9b8f40\r\n148 Nov 1 16:28 b45f82cb87bab2c35361857fcd46042ab658b42c37dc9a455248c2866c9b8f40.json\r\n0 Nov 1 16:07 b45f82cb87bab2c35361857fcd46042ab658b42c37dc9a455248c2866c9b8f40.lock\r\n```\r\n\r\n* Note the -L to follow redirects. Two headers are below:\r\n\r\n```bash\r\n$ curl -I -L https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/resolve/main/audio/en/invalidated/en_invalidated_3.tar\r\n```\r\n```\r\nHTTP/2 302 \r\ncontent-type: text/plain; charset=utf-8\r\ncontent-length: 1215\r\nlocation: https://cdn-lfs.huggingface.co/repos/00/ce/00ce867b4ae70bd23a10b60c32a8626d87b2666fc088ad03f86b94788faff554/984086fc250badece2992e8be4d7c4430f7c1208fb8bf37dc7c4aecdc803b220?response-content-disposition=attachment%3B+filename*%3DUTF-8%27%27en_invalidated_3.tar%3B+filename%3D%22en_invalidated_3.tar%22%3B&response-content-type=application%2Fx-tar&Expires=1699389040&Policy=eyJTdGF0ZW1lbnQiOlt7IkNvbmRpdGlvbiI6eyJEYXRlTGVzc1RoYW4iOnsiQVdTOkVwb2NoVGltZSI6MTY5OTM4OTA0MH19LCJSZXNvdXJjZSI6Imh0dHBzOi8vY2RuLWxmcy5odWdnaW5nZmFjZS5jby9yZXBvcy8wMC9jZS8wMGNlODY3YjRhZTcwYmQyM2ExMGI2MGMzMmE4NjI2ZDg3YjI2NjZmYzA4OGFkMDNmODZiOTQ3ODhmYWZmNTU0Lzk4NDA4NmZjMjUwYmFkZWNlMjk5MmU4YmU0ZDdjNDQzMGY3YzEyMDhmYjhiZjM3ZGM3YzRhZWNkYzgwM2IyMjA%7EcmVzcG9uc2UtY29udGVudC1kaXNwb3NpdGlvbj0qJnJlc3BvbnNlLWNvbnRlbnQtdHlwZT0qIn1dfQ__&Signature=WYc32e75PqbKSAv3KTpG86ooFT6oOyDDQpCt1i2B8gVS10J3qvpZlDmxaBgnGlCCl7SRiAvhIQctgwooNtWbUeDqK3T4bAo0-OOrGCuVi-%7EKWUBcoHce7nHWpl%7Ex9ubHS%7EFoYcGB2SCEqh5fIgGjNV-VKRX6TSXkRto5bclQq4VCJKHufDsJ114A1V4Qu%7EYiRIWKG4Gi93Xv4OFhyWY0uqykvP5c0x02F%7ELX0m3WbW-eXBk6Fw2xnV1XLrEkdR-9Ax2vHqMYIIw6yV0wWEc1hxE393P9mMG1TNDj%7EXDuCoOaA7LbrwBCxai%7Ew2MopdPamTXyOia5-FnSqEdsV29v4Q__&Key-Pair-Id=KVTP0A1DKRTAX\r\ndate: Sat, 04 Nov 2023 20:30:40 GMT\r\nx-powered-by: huggingface-moon\r\nx-request-id: Root=1-6546a9f0-5e7f729d09bdb38e35649a7e\r\naccess-control-allow-origin: https://huggingface.co\r\nvary: Origin, Accept\r\naccess-control-expose-headers: X-Repo-Commit,X-Request-Id,X-Error-Code,X-Error-Message,ETag,Link,Accept-Ranges,Content-Range\r\nx-repo-commit: 23b4059922516c140711b91831aa3393a22e9b80\r\naccept-ranges: bytes\r\nx-linked-size: 2110853120\r\nx-linked-etag: \"984086fc250badece2992e8be4d7c4430f7c1208fb8bf37dc7c4aecdc803b220\"\r\nx-cache: Miss from cloudfront\r\nvia: 1.1 f31a6426ebd75ce4393909b12f5cbdcc.cloudfront.net (CloudFront)\r\nx-amz-cf-pop: LAX53-P4\r\nx-amz-cf-id: BcYMFcHVcxPome2IjAvx0ZU90G41QlNI_HEHDGDqCQaEPvrOsnsGXw==\r\n\r\nHTTP/2 200 \r\ncontent-type: application/x-tar\r\ncontent-length: 2110853120\r\ndate: Sat, 04 Nov 2023 20:19:35 GMT\r\nlast-modified: Fri, 18 Nov 2022 15:08:22 GMT\r\netag: \"acac28988e2f7e73b68e865179fbd008\"\r\nx-amz-storage-class: INTELLIGENT_TIERING\r\nx-amz-version-id: LgTuOcd9FGN4JnAXp26O.1v2VW42GPtF\r\ncontent-disposition: attachment; filename*=UTF-8''en_invalidated_3.tar; filename=\"en_invalidated_3.tar\";\r\naccept-ranges: bytes\r\nserver: AmazonS3\r\nx-cache: Hit from cloudfront\r\nvia: 1.1 d07c8167eda81d307ca96358727f505e.cloudfront.net (CloudFront)\r\nx-amz-cf-pop: LAX50-P5\r\nx-amz-cf-id: 6oNZg_V8U1M_JXsMHQAPuRmDfxbY2BnMUWcVH0nz3VnfEZCzF5lgkQ==\r\nage: 666\r\ncache-control: public, max-age=604800, immutable, s-maxage=604800\r\nvary: Origin\r\n\r\n```\r\n" ]

[ "Thanks for reporting and thanks @mariosasko for fixing ! We just did a patch release `2.10.1` with the fix" ]

[ "Hi ! No the cache is not transportable in general. It will work on a shared filesystem if you use the same python environment, but not across machines/os/environments.\r\n\r\nIn particular, reloading cached datasets does work, but reloading cached processed datasets (e.g. from `map`) may not work. This is because some hashes used by caching are based on pickle dumps of the function you pass to `map`.\r\n\r\nFinally you may copy the cache to another machine, but all the `cached-*.arrow` files are unlikely to be reloaded.", "OK good to know. Thanks @lhoestq !" ]

[ "Hi @manuaero \r\n\r\nThank you for your interest in the COYO dataset.\r\n\r\nOur dataset provides the img-url and alt-text in the form of a parquet, so to utilize the coyo dataset you will need to download it directly.\r\n\r\nWe provide a [guide](https://github.com/kakaobrain/coyo-dataset/blob/main/download/README.md) to download, so check it out.\r\n\r\nThank you." ]

[ "Thanks for reporting!" ]

[ "Hi! I've merged a PR to use `int32` instead of `int8` for `reddit_scores`, so it should work now.\r\n\r\n" ]

[ "Duplicated issue." ]

[ "Hi! Indeed it would be useful to support this. PyArrow natively supports schema-level and column-level metadata, so implementing this should be straightforward. The API I have in mind would work as follows:\r\n```python\r\ncol_feature = Value(\"string\", metadata=\"Some column-level metadata\")\r\n\r\nfeatures = Features({\"col\": col_feature}, metadata=\"Some schema-level metadata\")\r\n```\r\n\r\nWDYT?", "Sorry for the late reply, \r\nYes, I think this is the most straight-forward approach with the things that we already have.\r\n\r\n", "@mariosasko Let me know how I can help.", "Hi, is this feature to be implemented in the near future? It would be really nice if that would be the case! ", "Hi, I also need this feature for tell my customer if any of the feature is encrypted with a certain key. " ]

[ "Also encountering this issue for every dataset I try to stream! Installed datasets from main:\r\n```\r\n- `datasets` version: 2.10.1.dev0\r\n- Platform: macOS-13.1-arm64-arm-64bit\r\n- Python version: 3.9.13\r\n- PyArrow version: 10.0.1\r\n- Pandas version: 1.5.2\r\n```\r\n\r\nRepro:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nspigi = load_dataset(\"kensho/spgispeech\", \"dev\", split=\"validation\", streaming=True, use_auth_token=True)\r\nsample = next(iter(spigi))\r\n```\r\n\r\n<details>\r\n<summary> Traceback </summary>\r\n\r\n```python\r\n---------------------------------------------------------------------------\r\nClientResponseError Traceback (most recent call last)\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/implementations/http.py:407, in HTTPFileSystem._info(self, url, **kwargs)\r\n 405 try:\r\n 406 info.update(\r\n--> 407 await _file_info(\r\n 408 self.encode_url(url),\r\n 409 size_policy=policy,\r\n 410 session=session,\r\n 411 **self.kwargs,\r\n 412 **kwargs,\r\n 413 )\r\n 414 )\r\n 415 if info.get(\"size\") is not None:\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/implementations/http.py:792, in _file_info(url, session, size_policy, **kwargs)\r\n 791 async with r:\r\n--> 792 r.raise_for_status()\r\n 794 # TODO:\r\n 795 # recognise lack of 'Accept-Ranges',\r\n 796 # or 'Accept-Ranges': 'none' (not 'bytes')\r\n 797 # to mean streaming only, no random access => return None\r\n\r\nFile ~/venv/lib/python3.9/site-packages/aiohttp/client_reqrep.py:1005, in ClientResponse.raise_for_status(self)\r\n 1004 self.release()\r\n-> 1005 raise ClientResponseError(\r\n 1006 self.request_info,\r\n 1007 self.history,\r\n 1008 status=self.status,\r\n 1009 message=self.reason,\r\n 1010 headers=self.headers,\r\n 1011 )\r\n\r\nClientResponseError: 403, message='Forbidden', url=URL('[https://cdn-lfs.huggingface.co/repos/e2/89/e28905247d6f48bb4edad5baf9b1bb4158e897a13fdf18bf3b8ee89ff8387ab8/46eca7431a7b6bad344bf451800e5b10cea1dd168f26d1027a6d9eb374b7fac3?response-content-disposition=attachment%3B+filename*%3DUTF-8''dev.csv%3B+filename%3D%22dev.csv%22%3B&response-content-type=text/csv&Expires=1677494732&Policy=eyJTdGF0ZW1lbnQiOlt7IlJlc291cmNlIjoiaHR0cHM6Ly9jZG4tbGZzLmh1Z2dpbmdmYWNlLmNvL3JlcG9zL2UyLzg5L2UyODkwNTI0N2Q2ZjQ4YmI0ZWRhZDViYWY5YjFiYjQxNThlODk3YTEzZmRmMThiZjNiOGVlODlmZjgzODdhYjgvNDZlY2E3NDMxYTdiNmJhZDM0NGJmNDUxODAwZTViMTBjZWExZGQxNjhmMjZkMTAyN2E2ZDllYjM3NGI3ZmFjMz9yZXNwb25zZS1jb250ZW50LWRpc3Bvc2l0aW9uPSomcmVzcG9uc2UtY29udGVudC10eXBlPXRleHQlMkZjc3YiLCJDb25kaXRpb24iOnsiRGF0ZUxlc3NUaGFuIjp7IkFXUzpFcG9jaFRpbWUiOjE2Nzc0OTQ3MzJ9fX1dfQ__&Signature=EzQB9f7xPckvqfFB6LzcyR-wzTnQCqtPDdWtQUzZ3QJ-gY-IHG5mxQITJgMr1nVTbJZrPmGAaDngMcPFUfSQa8RmCqYH~dZl-UGE8CO4neKNUT1DvA2WEvLDS4WaAJ3SN-9rX0uFb03~c1QS78cIgIRboYvf6ugKiJz86Bd7Vs~tcp201JFR0A6jIMseqApOnkb9d8dHMP3Ny~F6gO3Qf2QpEWM-QsDIyw2Kz2QV55nq8TsDpRYZCZo50~WwD~73Hej0PoDhEA1K37d19pa0CQhkaN-gjCrbT9xLabbvhJWa~ZkWcMdD0teCgjYqv1wKyvFXDAxukxLGEc7OBXVbYw__&Key-Pair-Id=KVTP0A1DKRTAX](https://cdn-lfs.huggingface.co/repos/e2/89/e28905247d6f48bb4edad5baf9b1bb4158e897a13fdf18bf3b8ee89ff8387ab8/46eca7431a7b6bad344bf451800e5b10cea1dd168f26d1027a6d9eb374b7fac3?response-content-disposition=attachment%3B+filename*%3DUTF-8%27%27dev.csv%3B+filename%3D%22dev.csv%22%3B&response-content-type=text/csv&Expires=1677494732&Policy=eyJTdGF0ZW1lbnQiOlt7IlJlc291cmNlIjoiaHR0cHM6Ly9jZG4tbGZzLmh1Z2dpbmdmYWNlLmNvL3JlcG9zL2UyLzg5L2UyODkwNTI0N2Q2ZjQ4YmI0ZWRhZDViYWY5YjFiYjQxNThlODk3YTEzZmRmMThiZjNiOGVlODlmZjgzODdhYjgvNDZlY2E3NDMxYTdiNmJhZDM0NGJmNDUxODAwZTViMTBjZWExZGQxNjhmMjZkMTAyN2E2ZDllYjM3NGI3ZmFjMz9yZXNwb25zZS1jb250ZW50LWRpc3Bvc2l0aW9uPSomcmVzcG9uc2UtY29udGVudC10eXBlPXRleHQlMkZjc3YiLCJDb25kaXRpb24iOnsiRGF0ZUxlc3NUaGFuIjp7IkFXUzpFcG9jaFRpbWUiOjE2Nzc0OTQ3MzJ9fX1dfQ__&Signature=EzQB9f7xPckvqfFB6LzcyR-wzTnQCqtPDdWtQUzZ3QJ-gY-IHG5mxQITJgMr1nVTbJZrPmGAaDngMcPFUfSQa8RmCqYH~dZl-UGE8CO4neKNUT1DvA2WEvLDS4WaAJ3SN-9rX0uFb03~c1QS78cIgIRboYvf6ugKiJz86Bd7Vs~tcp201JFR0A6jIMseqApOnkb9d8dHMP3Ny~F6gO3Qf2QpEWM-QsDIyw2Kz2QV55nq8TsDpRYZCZo50~WwD~73Hej0PoDhEA1K37d19pa0CQhkaN-gjCrbT9xLabbvhJWa~ZkWcMdD0teCgjYqv1wKyvFXDAxukxLGEc7OBXVbYw__&Key-Pair-Id=KVTP0A1DKRTAX)')\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nFileNotFoundError Traceback (most recent call last)\r\nCell In[5], line 4\r\n 1 from datasets import load_dataset\r\n 3 spigi = load_dataset(\"kensho/spgispeech\", \"dev\", split=\"validation\", streaming=True)\r\n----> 4 sample = next(iter(spigi))\r\n\r\nFile ~/datasets/src/datasets/iterable_dataset.py:937, in IterableDataset.__iter__(self)\r\n 934 yield from self._iter_pytorch(ex_iterable)\r\n 935 return\r\n--> 937 for key, example in ex_iterable:\r\n 938 if self.features:\r\n 939 # `IterableDataset` automatically fills missing columns with None.\r\n 940 # This is done with `_apply_feature_types_on_example`.\r\n 941 yield _apply_feature_types_on_example(\r\n 942 example, self.features, token_per_repo_id=self._token_per_repo_id\r\n 943 )\r\n\r\nFile ~/datasets/src/datasets/iterable_dataset.py:113, in ExamplesIterable.__iter__(self)\r\n 112 def __iter__(self):\r\n--> 113 yield from self.generate_examples_fn(**self.kwargs)\r\n\r\nFile ~/.cache/huggingface/modules/datasets_modules/datasets/kensho--spgispeech/5fbf75dd9ef795a9b5a673457d2cbaf0b8fa0de8fb62acbd1da338d83a41e2f0/spgispeech.py:186, in Spgispeech._generate_examples(self, local_extracted_archive_paths, archives, meta_path)\r\n 183 dict_keys = [\"wav_filename\", \"wav_filesize\", \"transcript\"]\r\n 185 logging.info(\"Reading metadata...\")\r\n--> 186 with open(meta_path, encoding=\"utf-8\") as f:\r\n 187 csvreader = csv.DictReader(f, delimiter=\"|\")\r\n 188 metadata = {x[\"wav_filename\"]: dict((k, x[k]) for k in dict_keys) for x in csvreader}\r\n\r\nFile ~/datasets/src/datasets/streaming.py:70, in extend_module_for_streaming.<locals>.wrap_auth.<locals>.wrapper(*args, **kwargs)\r\n 68 @wraps(function)\r\n 69 def wrapper(*args, **kwargs):\r\n---> 70 return function(*args, use_auth_token=use_auth_token, **kwargs)\r\n\r\nFile ~/datasets/src/datasets/download/streaming_download_manager.py:495, in xopen(file, mode, use_auth_token, *args, **kwargs)\r\n 493 kwargs = {**kwargs, **new_kwargs}\r\n 494 try:\r\n--> 495 file_obj = fsspec.open(file, mode=mode, *args, **kwargs).open()\r\n 496 except ValueError as e:\r\n 497 if str(e) == \"Cannot seek streaming HTTP file\":\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/core.py:135, in OpenFile.open(self)\r\n 128 def open(self):\r\n 129 \"\"\"Materialise this as a real open file without context\r\n 130 \r\n 131 The OpenFile object should be explicitly closed to avoid enclosed file\r\n 132 instances persisting. You must, therefore, keep a reference to the OpenFile\r\n 133 during the life of the file-like it generates.\r\n 134 \"\"\"\r\n--> 135 return self.__enter__()\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/core.py:103, in OpenFile.__enter__(self)\r\n 100 def __enter__(self):\r\n 101 mode = self.mode.replace(\"t\", \"\").replace(\"b\", \"\") + \"b\"\r\n--> 103 f = self.fs.open(self.path, mode=mode)\r\n 105 self.fobjects = [f]\r\n 107 if self.compression is not None:\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/spec.py:1106, in AbstractFileSystem.open(self, path, mode, block_size, cache_options, compression, **kwargs)\r\n 1104 else:\r\n 1105 ac = kwargs.pop(\"autocommit\", not self._intrans)\r\n-> 1106 f = self._open(\r\n 1107 path,\r\n 1108 mode=mode,\r\n 1109 block_size=block_size,\r\n 1110 autocommit=ac,\r\n 1111 cache_options=cache_options,\r\n 1112 **kwargs,\r\n 1113 )\r\n 1114 if compression is not None:\r\n 1115 from fsspec.compression import compr\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/implementations/http.py:346, in HTTPFileSystem._open(self, path, mode, block_size, autocommit, cache_type, cache_options, size, **kwargs)\r\n 344 kw[\"asynchronous\"] = self.asynchronous\r\n 345 kw.update(kwargs)\r\n--> 346 size = size or self.info(path, **kwargs)[\"size\"]\r\n 347 session = sync(self.loop, self.set_session)\r\n 348 if block_size and size:\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/asyn.py:113, in sync_wrapper.<locals>.wrapper(*args, **kwargs)\r\n 110 @functools.wraps(func)\r\n 111 def wrapper(*args, **kwargs):\r\n 112 self = obj or args[0]\r\n--> 113 return sync(self.loop, func, *args, **kwargs)\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/asyn.py:98, in sync(loop, func, timeout, *args, **kwargs)\r\n 96 raise FSTimeoutError from return_result\r\n 97 elif isinstance(return_result, BaseException):\r\n---> 98 raise return_result\r\n 99 else:\r\n 100 return return_result\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/asyn.py:53, in _runner(event, coro, result, timeout)\r\n 51 coro = asyncio.wait_for(coro, timeout=timeout)\r\n 52 try:\r\n---> 53 result[0] = await coro\r\n 54 except Exception as ex:\r\n 55 result[0] = ex\r\n\r\nFile ~/venv/lib/python3.9/site-packages/fsspec/implementations/http.py:420, in HTTPFileSystem._info(self, url, **kwargs)\r\n 417 except Exception as exc:\r\n 418 if policy == \"get\":\r\n 419 # If get failed, then raise a FileNotFoundError\r\n--> 420 raise FileNotFoundError(url) from exc\r\n 421 logger.debug(str(exc))\r\n 423 return {\"name\": url, \"size\": None, **info, \"type\": \"file\"}\r\n\r\nFileNotFoundError: https://huggingface.co/datasets/kensho/spgispeech/resolve/main/data/meta/dev.csv\r\n```\r\n</details>", "Hi ! We're investigating this issue, sorry for the inconvenience", "This has been resolved ! Thanks for reporting", "Wow, thanks for the very quick fix!", "This problem now appears again, this time with an underlying HTTP 502 status code:\r\n\r\n```\r\naiohttp.client_exceptions.ClientResponseError: 502, message='Bad Gateway', url=URL('https://huggingface.co/datasets/allenai/c4/resolve/1ddc917116b730e1859edef32896ec5c16be51d0/en/c4-validation.00002-of-00008.json.gz')\r\n```", "Re-executing a minute later, the underlying cause is an HTTP 403 status code, as reported yesterday:\r\n\r\n```\r\naiohttp.client_exceptions.ClientResponseError: 403, message='Forbidden', url=URL('https://cdn-lfs.huggingface.co/datasets/allenai/c4/4bf6b248b0f910dcde2cdf2118d6369d8208c8f9515ec29ab73e531f380b18e2?response-content-disposition=attachment%3B+filename*%3DUTF-8''c4-validation.00002-of-00008.json.gz%3B+filename%3D%22c4-validation.00002-of-00008.json.gz%22%3B&response-content-type=application/gzip&Expires=1677571273&Policy=eyJTdGF0ZW1lbnQiOlt7IlJlc291cmNlIjoiaHR0cHM6Ly9jZG4tbGZzLmh1Z2dpbmdmYWNlLmNvL2RhdGFzZXRzL2FsbGVuYWkvYzQvNGJmNmIyNDhiMGY5MTBkY2RlMmNkZjIxMThkNjM2OWQ4MjA4YzhmOTUxNWVjMjlhYjczZTUzMWYzODBiMThlMj9yZXNwb25zZS1jb250ZW50LWRpc3Bvc2l0aW9uPSomcmVzcG9uc2UtY29udGVudC10eXBlPWFwcGxpY2F0aW9uJTJGZ3ppcCIsIkNvbmRpdGlvbiI6eyJEYXRlTGVzc1RoYW4iOnsiQVdTOkVwb2NoVGltZSI6MTY3NzU3MTI3M319fV19&Signature=WW42NOKkLuX~xVB1QfbkqzdvGo2AOXpgbF3PjTXy6iKd~ffilr1N9ScPXfvTXqy5yvdhJg1G0xJy1zYtUjGAL8GEx3Av-0vIhpWMGYTM8XKEU5gYA9qt30oVtNph6TkTYSABrsYTaj-hzQL9WCgyapmjvG69ETMh4wj44r2rcbk4T3j0l6l4u76Gh~lyRSll3aK4qycdUwcyL7FECDu~0W1mJIJwKkCrWHhSpHJSshb-0ElwG71pq4eyQ5g2uxHdK6JbRF7loxUpRQQJ1vlk0EHXdw0wTMaQ9tqHy6xcrQd8Ep0Yvx3tUD8MR0vWOcbQKnL6LwPQByc8tkChlpjnig__&Key-Pair-Id=KVTP0A1DKRTAX')\r\n```", "I'm facing the same problem. Interestingly using `wget` I can download the file. ", "It's been resolved again ;)", "> It's been resolved again ;)\r\n\r\nI'm experiencing the same issue when trying to load this dataset, `FileNotFoundError: https://huggingface.co/datasets/allenai/c4/resolve/1ddc917116b730e1859edef32896ec5c16be51d0/realnewslike/c4-train.00000-of-00512.json.gz`", "Experiencing the same issues as above : `FileNotFoundError: https://huggingface.co/datasets/allenai/c4/resolve/1ddc917116b730e1859edef32896ec5c16be51d0/en/c4-train.00000-of-01024.json.gz\r\nIf the repo is private or gated, make sure to log in with `huggingface-cli login`.`\r\n\r\nHave made sure to login as well, issue persists.", "> Experiencing the same issues as above : `FileNotFoundError: https://huggingface.co/datasets/allenai/c4/resolve/1ddc917116b730e1859edef32896ec5c16be51d0/en/c4-train.00000-of-01024.json.gz If the repo is private or gated, make sure to log in with `huggingface-cli login`.`\r\n> \r\n> Have made sure to login as well, issue persists.\r\n\r\nI meet the same issue", "I meet the same issue" ]

[]

[ "Hi! \r\n\r\nYou need to pass an input json file explicitly as `data_files` to `load_dataset` to avoid this error:\r\n```python\r\n ds = load_dataset(\"json\", data_files=args.input_json)\r\n```\r\n\r\n", "Thanks it worked!" ]

[ "Hi, thanks for the feedback! Would it help to add a tip or note saying the dataset is gated and you need to accept the license before downloading it?", "The error is now more informative:\r\n```\r\nFileNotFoundError: Couldn't find a dataset script at /content/imagenet-1k/imagenet-1k.py or any data file in the same directory. Couldn't find 'imagenet-1k' on the Hugging Face Hub either: FileNotFoundError: Dataset 'imagenet-1k' doesn't exist on the Hub. If the repo is private or gated, make sure to log in with `huggingface-cli login`.\r\n```\r\n\r\n" ]

[ "Hi ! Oh good catch. I think the features should be passed to `IterableDataset.from_generator()` in `to_iterable_dataset()` indeed.\r\n\r\nSetting this as a good first issue if someone would like to contribute, otherwise we can take care of it :)", "#self-assign", "seems like the feature parameter is missing from `return IterableDataset.from_generator(Dataset._iter_shards, gen_kwargs={\"shards\": shards})` hence it defaults to None." ]

[ "Hi ! I think is in the scope of this other issue: to https://github.com/huggingface/datasets/issues/5281 " ]

[ "Hi ! The indices mapping is written in the same cachedirectory as your dataset.\r\n\r\nCan you run this to show your current cache directory ?\r\n```python\r\nprint(train_dataset.cache_files)\r\n```", "```\r\n[{'filename': '.../train/dataset.arrow'}, {'filename': '.../train/dataset.arrow'}]\r\n```\r\n\r\nThese are the actual paths where `.hf` files are stored. ", "I'm not aware of any `.hf` file ? What are you referring to ?\r\n\r\nAlso the error says \"Protocol unknown: parent\". Is there a chance you may have ended up with a path that contains this string `parent://` ?", "I figured out why the issue was occuring but don't know the long-term fix.\r\nThe dataset I was trying to shuffle was loaded from a saved file which had `::` delimiter in filename. When I try with the exact same file without `::` in filename, it works as expected.\r\nQuick fix is to not use colons in filename. But if this is expected behaviour, this should be clearly stated in the documentation.\r\nThanks for help @lhoestq " ]

[]

[ "Hi ! Can you make sure you set `HF_HOME` before importing `datasets` ?\r\n\r\nThen you can print\r\n```python\r\nprint(datasets.config.HF_CACHE_HOME)\r\nprint(datasets.config.HF_DATASETS_CACHE)\r\n```" ]

[ "Thanks for reporting, @wjfwzzc.\r\n\r\nI am transferring this issue to the corresponding dataset on the Hub: https://huggingface.co/datasets/bookcorpusopen/discussions/1", "Thank you. All fixes are done:\r\n- [x] https://huggingface.co/datasets/bookcorpusopen/discussions/2\r\n- [x] https://huggingface.co/datasets/the_pile/discussions/1\r\n- [x] https://huggingface.co/datasets/the_pile_books3/discussions/1\r\n- [x] https://huggingface.co/datasets/the_pile_openwebtext2/discussions/2\r\n- [x] https://huggingface.co/datasets/the_pile_stack_exchange/discussions/2" ]

[ "Running the script above on the branch https://github.com/huggingface/datasets/pull/5542 results in the expected behaviour:\r\n```\r\nNum chunks for original ds: 1\r\nOriginal ds save/load\r\nsave_to_disk -- RAM memory used: 0.671875 MB -- Total time: 0.255265 s\r\nload_from_disk -- RAM memory used: 42.796875 MB -- Total time: 0.014899 s\r\nNum chunks for original ds after reloading: 5000\r\n\r\nNum chunks for selected ds: 1\r\nflatten_indices -- RAM memory used: 42.546875 MB -- Total time: 23.735089 s\r\nNum chunks for selected ds after flattening: 5000\r\n\r\nSelected ds save/load\r\nsave_to_disk -- RAM memory used: 0.0 MB -- Total time: 0.287112 s\r\nload_from_disk -- RAM memory used: 38.84375 MB -- Total time: 0.014772 s\r\nNum chunks for selected ds after reloading: 5000\r\n```", "Wouahouh super cool @marioga thanks a lot!", "We just released `datasets==2.10.0` with this big improvement, thanks again @marioga " ]

[ "Hi! The `set_transform` does not apply a custom formatting transform on a single example but the entire batch, so the fixed version of your transform would look as follows:\r\n```python\r\nfrom datasets import load_dataset\r\nimport torch\r\n\r\ndataset = load_dataset(\"lambdalabs/pokemon-blip-captions\", split='train')\r\ndef t(batch):\r\n return {\"test\": torch.tensor([1] * len(batch[next(iter(batch))]))}\r\n \r\ndataset.set_transform(t)\r\nd_0 = dataset[0]\r\n```\r\n\r\nStill, the formatter's error message should mention that a dict of **sequences** is expected as the returned value (not just a dict) to make debugging easier.", "I can take this", "Fixed in #5553 ", "> Hi! The `set_transform` does not apply a custom formatting transform on a single example but the entire batch, so the fixed version of your transform would look as follows:\r\n> \r\n> ```python\r\n> from datasets import load_dataset\r\n> import torch\r\n> \r\n> dataset = load_dataset(\"lambdalabs/pokemon-blip-captions\", split='train')\r\n> def t(batch):\r\n> return {\"test\": torch.tensor([1] * len(batch[next(iter(batch))]))}\r\n> \r\n> dataset.set_transform(t)\r\n> d_0 = dataset[0]\r\n> ```\r\n> \r\n> Still, the formatter's error message should mention that a dict of **sequences** is expected as the returned value (not just a dict) to make debugging easier.\r\n\r\nok, will change it according to suggestion. Thanks for the reply!" ]

[ "Hi! `seaborn`'s `load_dataset` pulls datasets from [here](https://github.com/mwaskom/seaborn-data) and not from our Hub, so this issue is not related to our library in any way and should be reported in their repo instead." ]

[ "#self-assign", "You were right, if `self.dir_cache` is not None in glob, it is exactly the same as what is returned by find, at least for all the tests we have, and some extended evaluation I did across a random sample of about 1000 datasets. \r\n\r\nThanks for the nice hints, and let me know if this is not exactly what we want here!\r\n\r\nsee PR: https://github.com/huggingface/datasets/pull/5704\r\n\r\n", "I think we can make the data files resolution (significantly) faster in 2 steps:\r\n\r\n1. `glob` calls `find` (which in turn calls `ls`), so we need `find` to be fast, and this can be achieved by fetching all the entries in a single API call and avoiding calls to `ls`. Implementing this for `HfFileSystem.find` (the one in `huggingface_hub`) is on my TO-DO list.\r\n2. caching the repeated `find` calls in `_get_data_files_patterns` when the `data_files` patterns are not provided in `load_dataset`. To address this, we can introduce a `_resolve_single_pattern` function that would accept a filesystem object and a list of regex patterns to resolve. Then we can wrap this filesystem object in `_get_data_files_patterns` with an object that would cache the find calls before resolving the patterns with `_resolve_single_pattern`. (Feel free to suggest a cleaner implementation)\r\n\r\nWDYT?", "Good idea :) \r\n\r\nFor 2:\r\n\r\nThat would work ! It's also possible to have a FileSystem with a cache on `.find` and use it inside the resolver passed to `_get_data_files_patterns`. Right now they're pretty simple:\r\n\r\n```python\r\n# for remote repositories\r\nresolver = partial(_resolve_single_pattern_in_dataset_repository, dataset_info, base_path=base_path)\r\n# for local\r\nresolver = partial(_resolve_single_pattern_locally, base_path)\r\n```", "something like this maybe (with Quentin's reimplementation of `HfFilesystem.find`)?\r\n\r\n ```\r\n @lru_cache(max_size=None)\r\n def _find(self, path, maxdepth=None, withdirs=False, detail=False, **kwargs):\r\n```\r\n\r\nIn any case please let me know if I can help in any way!" ]

[ "Hi ! `enc` is not hashable:\r\n```python\r\nimport tiktoken\r\nfrom datasets.fingerprint import Hasher\r\n\r\nenc = tiktoken.get_encoding(\"gpt2\")\r\nHasher.hash(enc)\r\n# raises TypeError: cannot pickle 'builtins.CoreBPE' object\r\n```\r\nIt happens because it's not picklable, and because of that it's not possible to cache the result of `map`, hence the warning message.\r\n\r\nYou can find more details about caching here: https://huggingface.co/docs/datasets/about_cache\r\n\r\nYou can also provide your own unique hash in `map` if you want, with the `new_fingerprint` argument.\r\nOr disable caching using\r\n```python\r\nimport datasets\r\ndatasets.disable_caching()\r\n```", "@lhoestq Thank you for the explanation and advice. Will relay all of this to the repo where this (non)issue arose. \r\n\r\nGreat job with huggingface! ", "We made tiktoken tokenizers hashable in #5552, which is included in today's release `datasets==2.10.0`", "Just a heads up that when I'm trying to use TikToken along with the a given Dataset `.map()` method, I am still met with the following error :\r\n\r\n```\r\n File \"/opt/conda/lib/python3.8/site-packages/dill/_dill.py\", line 388, in save\r\n StockPickler.save(self, obj, save_persistent_id)\r\n File \"/opt/conda/lib/python3.8/pickle.py\", line 578, in save\r\n rv = reduce(self.proto)\r\nTypeError: cannot pickle 'builtins.CoreBPE' object\r\n```\r\n\r\nMy current environment is running datasets v2.10.0.", "cc @mariosasko ", "@lhoestq @edhenry I am also seeing this, do you have any suggested solution?", "With which `datasets` version ? Can you try to udpate ?", "@lhoestq @edhenry I am on datasets version `'2.12.0'. I see the same `TypeError: cannot pickle 'builtins.CoreBPE' object` that others are seeing.", "I am able to reproduce this on datasets 2.14.2. The `datasets.disable_caching()` doesn't work around it.\r\n\r\n@lhoestq - you might want to reopen this issue. Because of this issue folks won't be able run Karpathy's NanoGPT :(.", "update： temporarily solved the problem by setting\r\n```\r\n--preprocess_num_workers 1\r\n```\r\n\r\n-------------\r\nI have met the same problem, here is my env:\r\n```\r\ndatasets 2.14.4\r\ntransformers 4.31.0\r\ntiktoken 0.4.0\r\ntorch 1.13.1\r\n```", "@mengban I cannot reproduce the issue even with these versions installed. It would help if you could provide info about your system and the `pip list` output.", "@mariosasko Please take a look at this\r\n```python\r\nfrom typing import Any\r\nfrom datasets import Dataset\r\nimport tiktoken\r\n\r\ndataset = Dataset.from_list([{\"n\": str(i)} for i in range(20)])\r\nenc = tiktoken.get_encoding(\"gpt2\")\r\n\r\n\r\nclass A:\r\n tokenizer = enc #tiktoken.get_encoding(\"gpt2\")\r\n\r\n def __call__(self, example) -> Any:\r\n ids = self.tokenizer.encode(example[\"n\"])\r\n example[\"len\"] = len(ids)\r\n return example\r\n\r\na = A()\r\n\r\ndef process(example):\r\n ids = a.tokenizer.encode(example[\"n\"])\r\n example[\"len\"] = len(ids)\r\n return example\r\n\r\n# success\r\ntokenized = dataset.map(process, desc=\"tiktoken\", num_proc=2)\r\n\r\n# raise TypeError: cannot pickle 'builtins.CoreBPE' object\r\ntokenized = dataset.map(a, desc=\"tiktoken\", num_proc=2)\r\n```\r\n\r\npip list\r\n```\r\ndatasets 2.14.4\r\ntiktoken 0.4.0\r\n```", "Thanks @maxwellzh! Our `Hasher` works with this snippet, but the problem is running multiprocessing with a non-serializable `tiktoken.Encoding` object.\r\n\r\nInserting the following code before the `map` should fix this:\r\n```python\r\nimport copyreg\r\n\r\ndef pickle_Encoding(enc):\r\n return (functools.partial(tiktoken.core.Encoding, enc.name, pat_str=enc._pat_str, mergeable_ranks=enc._mergeable_ranks, special_tokens=enc._special_tokens), ())\r\n\r\ncopyreg.pickle(tiktoken.core.Encoding, pickle_Encoding)\r\n```\r\n\r\nBut the best fix would be implementing `__reduce__` for `tiktoken.Encoding` or `tiktoken.CoreBPE`. If I find time, I'll try to fix this in the `tiktoken` repo.", "I think the right way to fix this would be to have new tokenizer instance for each process. This applies to many other tokenizers that don't support multi-process or have bugs. To do this, first define tokenizer factory class like this:\r\n\r\n```\r\n class TikTokenFactory:\r\n def __init__(self):\r\n self._enc = None\r\n self.eot_token = None\r\n\r\n def encode_ordinary(self, text):\r\n if self._enc is None:\r\n self._enc = tiktoken.get_encoding(\"gpt2\")\r\n self.eot_token = self._enc.eot_token\r\n return self._enc.encode_ordinary(text)\r\n```\r\n\r\nNow use this in `.map()` like this:\r\n\r\n```\r\n # tokenize the dataset\r\n tokenized = dataset.map(\r\n partial(process, TikTokenFactory()),\r\n remove_columns=['text'],\r\n desc=\"tokenizing the splits\",\r\n num_proc=max(1, cpu_count()//2),\r\n )\r\n```\r\n\r\nA full working example is here: https://github.com/sytelus/nanoGPT/blob/refactor/nanogpt_common/hf_data_prepare.py" ]

[ "Hi! This code works for me locally or in Colab. What's the output of `python -c \"import pyarrow as pa; print(pa.__version__)\"` when you run it inside your environment?", "Thanks for looking into this!\r\nThe output of `python -c \"import pyarrow as pa; print(pa.__version__)\"` is:\r\n```\r\n11.0.0\r\n```\r\n\r\nI did the following to setup the environment:\r\n```\r\nconda create -n datasets_debug python=3.9\r\nconda activate datasets_debug\r\npip install datasets==2.9.0\r\n```\r\n\r\nI just tested this on another machine (Ubuntu 18.04.6 LTS) with the same result as mentioned in the issue description.\r\n" ]

[ "Hi! You can get this behavior by specifying `stratify_by_column=\"label\"` in `train_test_split`.\r\n\r\nThis is the full example:\r\n```python\r\nimport numpy as np\r\nfrom datasets import Dataset, ClassLabel\r\n\r\ndata = [\r\n {'label': 0, 'text': \"example1\"},\r\n {'label': 1, 'text': \"example2\"},\r\n {'label': 1, 'text': \"example3\"},\r\n {'label': 1, 'text': \"example4\"},\r\n {'label': 0, 'text': \"example5\"},\r\n {'label': 1, 'text': \"example6\"},\r\n {'label': 2, 'text': \"example7\"},\r\n {'label': 2, 'text': \"example8\"}\r\n]\r\n\r\nfor _ in range(10):\r\n data_set = Dataset.from_list(data)\r\n data_set = data_set.cast_column(\"label\", ClassLabel(num_classes=3))\r\n data_set = data_set.train_test_split(test_size=0.5, stratify_by_column=\"label\")\r\n unique_labels_train = np.unique(data_set[\"train\"][:][\"label\"])\r\n unique_labels_test = np.unique(data_set[\"test\"][:][\"label\"])\r\n assert len(unique_labels_train) >= len(unique_labels_test) \r\n```\r\n" ]

[]

[ "Thanks for reporting, @TJ-Solergibert.\r\n\r\nWe cannot access your Colab notebook: `There was an error loading this notebook. Ensure that the file is accessible and try again.`\r\nCould you please make it publicly accessible?\r\n", "I swear it's public, I've checked the settings and I've been able to open it in incognito mode.\r\n\r\nNotebook: https://colab.research.google.com/drive/1JCrS7FlGfu_kFqChMrwKZ_bpabnIMqbP?usp=sharing\r\n\r\nAnyway, this is the code to reproduce the error:\r\n\r\n```python3\r\nfrom datasets import ClassLabel\r\nfrom datasets import load_dataset\r\n\r\neuroparl_ds = load_dataset(\"tj-solergibert/Europarl-ST\")\r\n\r\nsource_lang = \"nl\"\r\nlanguages = list(europarl_ds[\"train\"][0][\"transcriptions\"].keys())\r\nClassLabels = ClassLabel(num_classes = len(languages), names = languages)\r\n\r\ndef map_label2id(example):\r\n example['dest_lang'] = ClassLabels.str2int(example['dest_lang'])\r\n return example\r\n\r\ndef unfold_transcriptions(example):\r\n for lang in languages:\r\n example[lang] = example[\"transcriptions\"][lang]\r\n return example\r\n\r\ndef unroll(batch, src_lang, dest_langs):\r\n source_t, dest_t, dest_l = [], [], []\r\n for lang in dest_langs: \r\n source_t += batch[src_lang]\r\n dest_t += batch[lang]\r\n dest_l += [lang]\r\n return_dict = {\"source_text\": source_t, \"dest_text\": dest_t, \"dest_lang\": dest_l}\r\n return return_dict\r\n\r\ndef preprocess_split(ds_split, src_lang):\r\n dest_langs = [x for x in languages if x != src_lang]\r\n\r\n ds_split = ds_split.map(unroll, fn_kwargs= {\"src_lang\": src_lang, \"dest_langs\": dest_langs}, batched = True, batch_size = 1, remove_columns= list(languages))\r\n ds_split = ds_split.filter(lambda x: x[\"source_text\"] != None and x[\"dest_text\"] != None) # Remove incomplete translations\r\n ds_split = ds_split.filter(lambda x: x[\"source_text\"] != \"None\" and x[\"dest_text\"] != \"None\")\r\n ds_split = ds_split.map(map_label2id) \r\n ds_split = ds_split.cast_column(\"dest_lang\", ClassLabels)\r\n return ds_split\r\n\r\ndef reset_cortas(example):\r\n for lang in languages:\r\n if isinstance(example[lang], str):\r\n if example[lang].isnumeric () or len(example[lang]) <= 5:\r\n example[lang] = \"None\"\r\n return example\r\n\r\ndef clean_dataset(dataset):\r\n # Remove columns\r\n dataset = dataset.remove_columns([\"original_speech\", \"original_language\", \"audio_path\", \"segment_start\", \"segment_end\"])\r\n # Unfold\r\n dataset = dataset.map(unfold_transcriptions, remove_columns = [\"transcriptions\"])\r\n dataset = dataset.map(reset_cortas)\r\n return dataset\r\n\r\nprocessed_europarl = clean_dataset(europarl_ds[\"test\"])\r\nnew_train_ds = preprocess_split(processed_europarl, 'nl')\r\n```", "Thanks, @TJ-Solergibert. I can access your notebook now. Maybe it was just a temporary issue.\r\n\r\nAt first sight, it seems something related to your data: maybe some of the examples do not have all the transcriptions for all the languages. Then, some of them are null when unrolled. And when trying to concatenate with the other rows containing strings, the cast issue is raised (the arrays to be concatenated have different types).\r\n\r\nDo you think this could be the case?", "See, in this example, \"nl\" and \"ro\" transcripts are null:\r\n```python\r\n>>> europarl_ds[\"test\"][:1]\r\n{'original_speech': ['− Señor Presidente, en primer lugar, quisiera felicitar al señor Seeber por el trabajo realizado, porque en su informe se recogen muchas de las preocupaciones manifestadas en esta'],\r\n 'original_language': ['es'],\r\n 'audio_path': ['es/audios/en.20081008.24.3-238.m4a'],\r\n 'segment_start': [0.6200000047683716],\r\n 'segment_end': [11.319999694824219],\r\n 'transcriptions': [{'de': '− Herr Präsident! Zunächst möchte ich Richard Seeber zu der von ihm geleisteten Arbeit gratulieren, denn sein Bericht greift viele der in diesem Haus zum Ausdruck gebrachten Anliegen',\r\n 'en': '− Mr President, firstly I would like to congratulate Mr Seeber on the work he has done, because his report picks up many of the concerns expressed in this',\r\n 'es': '− Señor Presidente, en primer lugar, quisiera felicitar al señor Seeber por el trabajo realizado, porque en su informe se recogen muchas de las preocupaciones manifestadas en esta',\r\n 'fr': '− Monsieur le Président, je voudrais tout d ’ abord féliciter M. Seeber pour le travail qu ’ il a effectué, parce que son rapport reprend beaucoup des inquiétudes exprimées au sein de cette',\r\n 'it': \"− Signor Presidente, mi congratulo innanzi tutto con l'onorevole Seeber per il lavoro svolto, perché la sua relazione accoglie molti dei timori espressi da quest'Aula\",\r\n 'nl': None,\r\n 'pl': '− Panie przewodniczący! Po pierwsze chciałabym pogratulować panu posłowi Seeberowi wykonanej pracy, ponieważ jego sprawozdanie podejmuje szereg podnoszonych w tej Izbie',\r\n 'pt': '− Senhor Presidente, começo por felicitar o senhor deputado Seeber pelo trabalho que desenvolveu em torno deste relatório, que retoma muitas das preocupações expressas nesta',\r\n 'ro': None}]}\r\n```\r\n```python\r\n>>> processed_europarl[0]\r\n{'de': '− Herr Präsident! Zunächst möchte ich Richard Seeber zu der von ihm geleisteten Arbeit gratulieren, denn sein Bericht greift viele der in diesem Haus zum Ausdruck gebrachten Anliegen',\r\n 'en': '− Mr President, firstly I would like to congratulate Mr Seeber on the work he has done, because his report picks up many of the concerns expressed in this',\r\n 'es': '− Señor Presidente, en primer lugar, quisiera felicitar al señor Seeber por el trabajo realizado, porque en su informe se recogen muchas de las preocupaciones manifestadas en esta',\r\n 'fr': '− Monsieur le Président, je voudrais tout d ’ abord féliciter M. Seeber pour le travail qu ’ il a effectué, parce que son rapport reprend beaucoup des inquiétudes exprimées au sein de cette',\r\n 'it': \"− Signor Presidente, mi congratulo innanzi tutto con l'onorevole Seeber per il lavoro svolto, perché la sua relazione accoglie molti dei timori espressi da quest'Aula\",\r\n 'nl': None,\r\n 'pl': '− Panie przewodniczący! Po pierwsze chciałabym pogratulować panu posłowi Seeberowi wykonanej pracy, ponieważ jego sprawozdanie podejmuje szereg podnoszonych w tej Izbie',\r\n 'pt': '− Senhor Presidente, começo por felicitar o senhor deputado Seeber pelo trabalho que desenvolveu em torno deste relatório, que retoma muitas das preocupações expressas nesta',\r\n 'ro': None}\r\n```", "You can fix this issue by forcing the cast of None to str by hand:\r\n- If you replace this line:\r\n```python\r\nsource_t += batch[src_lang]\r\n```\r\n- With this line (because the batch size is 1):\r\n```python\r\nsource_t += [str(batch[src_lang][0])]\r\n```\r\n- Or with this line (if the batch size were larger than 1):\r\n```python\r\nsource_t += [str(text) for text in batch[src_lang]]\r\n```", "Problem solved! Thanks @albertvillanova, now I have even increased the batch size and it's crazy fast :rocket: !" ]

[]

[]

[ "Hi! This behavior stems from these lines:\r\n\r\nhttps://github.com/huggingface/datasets/blob/b065547654efa0ec633cf373ac1512884c68b2e1/src/datasets/formatting/np_formatter.py#L45-L46\r\n\r\nI agree we should preserve the original type whenever possible and downcast explicitly with a warning.\r\n\r\n@lhoestq Do you remember why we need this \"default dtype\" logic in our formatters?", "I was also wondering why the default type logic is needed. Me just deleting it is probably too naive of a solution.", "Hmm I think the idea was to end up with the usual default precision for deep learning models - no matter how the data was stored or where it comes from.\r\n\r\nFor example in NLP we store tokens using an optimized low precision to save disk space, but when we set the format to `torch` we actually need to get `int64`. Although the need for a default for integers also comes from numpy not returning the same integer precision depending on your machine. Finally I guess we added a default for floats as well for consistency.\r\n\r\nI'm a bit embarrassed by this though, as a user I'd have expected to get the same precision indeed as well and get a zero copy view.", "Will you fix this or should I open a PR?", "Unfortunately removing it for integers is a breaking change for most `transformers` + `datasets` users for NLP (which is a common case). Removing it for floats is a breaking change for `transformers` + `datasets` for ASR as well. And it also is a breaking change for the other users relying on this behavior.\r\n\r\nTherefore I think that the only short term solution is for the user to provide `dtype=` manually and document better this behavior. We could also extend `dtype` to accept a value that means \"return the same dtype as the underlying storage\" and make it easier to do zero copy.", "@lhoestq It should be fine to remove this conversion in Datasets 3.0, no? For now, we can warn the user (with a log message) about the future change when the default type is changed.", "Let's see with the transformers team if it sounds reasonable ? We'd have to fix multiple example scripts though.\r\n\r\nIf it's not ok we can also explore keeping this behavior only for tokens and audio data.", "IMO being coupled with Transformers can lead to unexpected behavior when one tries to use our lib without pairing it with Transformers, so I think it's still important to \"fix\" this, even if it means we will need to update Transformers' example scripts afterward.\r\n", "Ideally let's update the `transformers` example scripts before the change :P", "For others that run into the same issue: A temporary workaround for me is this:\r\n```python\r\ndef numpy_transform(batch):\r\n return {key: np.asarray(val) for key, val in batch.items()}\r\n\r\ndataset = dataset.with_transform(numpy_transform)\r\n```", "This behavior (silent upcast from `int32` to `int64`) is also unexpected for the user in https://discuss.huggingface.co/t/standard-getitem-returns-wrong-data-type-for-arrays/62470/2", "Hi, I stumbled on a variation that upcasts uint8 to int64. I would expect the dtype to be the same as it was when I generated the dataset.\r\n\r\n```\r\nimport numpy as np\r\nimport datasets as ds\r\n\r\nfoo = np.random.randint(0, 256, size=(5, 10, 10), dtype=np.uint8)\r\n\r\nfeatures = ds.Features({\"foo\": ds.Array2D((10, 10), \"uint8\")})\r\ndataset = ds.Dataset.from_dict({\"foo\": foo}, features=features)\r\ndataset.set_format(\"torch\")\r\nprint(\"feature dtype:\", dataset.features[\"foo\"].dtype)\r\nprint(\"array dtype:\", dataset[\"foo\"].dtype)\r\n\r\n# feature dtype: uint8\r\n# array dtype: torch.int64\r\n```\r\n", "workaround to remove torch upcasting\r\n\r\n```\r\nimport datasets as ds\r\nimport torch\r\n\r\nclass FixedTorchFormatter(ds.formatting.TorchFormatter):\r\n def _tensorize(self, value):\r\n return torch.from_numpy(value)\r\n\r\n\r\nds.formatting._register_formatter(FixedTorchFormatter, \"torch\")\r\n```" ]

[ "Hi, thanks for noticing this! We can't just remove the cache control as this allows us to control where the arrow files generated by the ops are written (cached on disk if enabled or a temporary directory if disabled). The right way to address this inconsistency would be by having `load_from_cache_file=None` by default everywhere.", "Hi! Yes, this seems more plausible. I can implement that. One last thing is the type annotation `load_from_cache_file: bool = None`. Which I then would change to `load_from_cache_file: Optional[bool] = None`.", "PR #5515 ", "Yes, `Optional[bool]` is the correct type annotation and thanks for the PR." ]

[ "Hi! Let's not do this - renaming it would be a breaking change, and going through the deprecation cycle is only worth it if it improves user experience.", "Hi @mariosasko, ok it makes sense. Anyway, don't you think it's worth it at some point to start a deprecation cycle e.g. `fs` in `load_from_disk`? It doesn't affect user experience but it's for sure a bad practice IMO, but's up to you 😄 Feel free to close this issue otherwise!", "I don't think deprecating a param name in this particular instance is worth the hassle, so I'm closing the issue 🙂.", "Sure, makes sense @mariosasko thanks!" ]

[ "Update `datasets` or downgrade `huggingface-hub` ;)\r\n\r\nThe `huggingface-hub` lib did a breaking change a few months ago, and you're using an old version of `datasets` that does't support it", "Awesome thanks a lot! Everything works just fine with `datasets==2.9.0` :-) ", "Getting same error with latest versions.\r\n\r\n\r\n```shell\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\nCell In[99], line 1\r\n----> 1 dataset.push_to_hub(\"mirfan899/kids_phoneme_asr\")\r\n\r\nFile /opt/conda/lib/python3.10/site-packages/datasets/arrow_dataset.py:3538, in Dataset.push_to_hub(self, repo_id, split, private, token, branch, shard_size, embed_external_files)\r\n 3493 def push_to_hub(\r\n 3494 self,\r\n 3495 repo_id: str,\r\n (...)\r\n 3501 embed_external_files: bool = True,\r\n 3502 ):\r\n 3503 \"\"\"Pushes the dataset to the hub.\r\n 3504 The dataset is pushed using HTTP requests and does not need to have neither git or git-lfs installed.\r\n 3505 \r\n (...)\r\n 3536 ```\r\n 3537 \"\"\"\r\n-> 3538 repo_id, split, uploaded_size, dataset_nbytes = self._push_parquet_shards_to_hub(\r\n 3539 repo_id=repo_id,\r\n 3540 split=split,\r\n 3541 private=private,\r\n 3542 token=token,\r\n 3543 branch=branch,\r\n 3544 shard_size=shard_size,\r\n 3545 embed_external_files=embed_external_files,\r\n 3546 )\r\n 3547 organization, dataset_name = repo_id.split(\"/\")\r\n 3548 info_to_dump = self.info.copy()\r\n\r\nFile /opt/conda/lib/python3.10/site-packages/datasets/arrow_dataset.py:3474, in Dataset._push_parquet_shards_to_hub(self, repo_id, split, private, token, branch, shard_size, embed_external_files)\r\n 3472 shard.to_parquet(buffer)\r\n 3473 uploaded_size += buffer.tell()\r\n-> 3474 _retry(\r\n 3475 api.upload_file,\r\n 3476 func_kwargs=dict(\r\n 3477 path_or_fileobj=buffer.getvalue(),\r\n 3478 path_in_repo=path_in_repo(index),\r\n 3479 repo_id=repo_id,\r\n 3480 token=token,\r\n 3481 repo_type=\"dataset\",\r\n 3482 revision=branch,\r\n 3483 identical_ok=True,\r\n 3484 ),\r\n 3485 exceptions=HTTPError,\r\n 3486 status_codes=[504],\r\n 3487 base_wait_time=2.0,\r\n 3488 max_retries=5,\r\n 3489 max_wait_time=20.0,\r\n 3490 )\r\n 3491 return repo_id, split, uploaded_size, dataset_nbytes\r\n\r\nFile /opt/conda/lib/python3.10/site-packages/datasets/utils/file_utils.py:330, in _retry(func, func_args, func_kwargs, exceptions, status_codes, max_retries, base_wait_time, max_wait_time)\r\n 328 while True:\r\n 329 try:\r\n--> 330 return func(*func_args, **func_kwargs)\r\n 331 except exceptions as err:\r\n 332 if retry >= max_retries or (status_codes and err.response.status_code not in status_codes):\r\n\r\nFile /opt/conda/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py:120, in validate_hf_hub_args.<locals>._inner_fn(*args, **kwargs)\r\n 117 if check_use_auth_token:\r\n 118 kwargs = smoothly_deprecate_use_auth_token(fn_name=fn.__name__, has_token=has_token, kwargs=kwargs)\r\n--> 120 return fn(*args, **kwargs)\r\n\r\nTypeError: HfApi.upload_file() got an unexpected keyword argument 'identical_ok'\r\n```", "Feel free to update `datasets` and `huggingface-hub`, it should fix it :)", "I went ahead and upgraded both datasets and hub and still getting the same error\r\n", "Which version do you have ? It's been a while since it has been fixed", "huggingface 0.0.1\r\nhuggingface-hub 0.17.1\r\ndatasets 2.14.5\r\n\r\nstill has the issue!!", "I face the same issue even after upgrading :/" ]

[ "Hey, I'm a research engineer working on language modelling wanting to contribute to open source. I was wondering if I could give it a shot?", "Hi! This issue was fixed in https://github.com/huggingface/datasets/pull/4972, so please install `datasets>=2.5.0` to avoid it." ]

[]

[ "Hi ! `datasets` doesn't do batching - the PyTorch DataLoader does and is created by the `Trainer`. Do you pass other arguments to training_args with respect to data loading ?\r\n\r\nAlso we recently released `.to_iterable_dataset` that does pretty much what you implemented, but using contiguous shards to get a better speed:\r\n```python\r\nif use_iterable_dataset:\r\n num_shards = 100\r\n dataset = dataset.to_iterable_dataset(num_shards=num_shards)\r\n```", "This is the full set of training args passed. No training args were changed when switching dataset types.\r\n\r\n```python\r\ntraining_args = TrainingArguments(\r\n output_dir=\"./checkpoints\",\r\n overwrite_output_dir=True,\r\n num_train_epochs=1,\r\n per_device_train_batch_size=256,\r\n save_steps=2000,\r\n save_total_limit=4,\r\n prediction_loss_only=True,\r\n report_to='none',\r\n gradient_accumulation_steps=6,\r\n fp16=True,\r\n max_steps=60000,\r\n lr_scheduler_type='linear',\r\n warmup_ratio=0.1,\r\n logging_steps=100,\r\n weight_decay=0.01,\r\n adam_beta1=0.9,\r\n adam_beta2=0.98,\r\n adam_epsilon=1e-6,\r\n learning_rate=1e-4\r\n)\r\n```", "I think the issue comes from `transformers`: https://github.com/huggingface/transformers/issues/21444", "Makes sense. Given that it's a `transformers` issue and already being tracked, I'll close this out." ]

[ "This change seems to come from a few months ago in the PyTorch side. That's good news and it means we may not need to pass a batch_sampler as soon as we add `Dataset.__getitems__` to get the optimal speed :)\r\n\r\nThanks for reporting ! Would you like to open a PR to add `__getitems__` and remove this outdated documentation ?", "Yeah I figured this was the sort of thing that probably once worked. I can confirm that you no longer need the batch sampler, just `batch_size=n` in the `DataLoader`.\r\n\r\nI'll pass on the PR, I'm flat out right now, sorry." ]

[ "I update the `datatsets` version and it works." ]

[ "Hi ! To skip the verification step that checks if newer data exist, you can enable offline mode with `HF_DATASETS_OFFLINE=1`.\r\n\r\nAlthough I agree this step should be much faster for datasets hosted on the HF Hub - we could just compare the commit hash from the local data and the remote git repository. We're not been leveraging the git commit hashes, since the library was built before we even had git repositories for each dataset on HF.", "Thanks @lhoestq, for memory when I recorded those times I had `HF_DATASETS_OFFLINE` set." ]

[ "Hi! Instead of a single boolean, your filter function should return an iterable (of booleans) in the batched mode like so:\r\n```python\r\ntrain_dataset = train_dataset.filter(\r\n function=lambda batch: [image is not None for image in batch[\"image\"]], \r\n batched=True,\r\n batch_size=10)\r\n```\r\n\r\nPS: You can make this operation much faster by operating directly on the arrow data to skip the decoding part:\r\n```python\r\ntrain_dataset = train_dataset.with_format(\"arrow\")\r\ntrain_dataset = train_dataset.filter(\r\n function=lambda table: table[\"image\"].is_valid().to_pylist(), \r\n batched=True,\r\n batch_size=100)\r\ntrain_dataset = train_dataset.with_format(None)\r\n```", "Thank a lot!", "I hit the same issue and the error message isn't really clear on what's going wrong. It might be helpful to update the docs with a batched example." ]

[ "Hi! Sure, feel free to open a PR, so we can see the API you have in mind.", "I would like to give it a go! #self-assign", "Closing as `Dataset.map` can be used instead (see https://github.com/huggingface/datasets/pull/5533#issuecomment-1440571658 and https://github.com/huggingface/datasets/pull/5533#issuecomment-1446403263)" ]

[ "Hi! This is indeed a bug in our zero-copy logic.\r\n\r\nTo fix it, instead of the line:\r\nhttps://github.com/huggingface/datasets/blob/7cfac43b980ab9e4a69c2328f085770996323005/src/datasets/features/features.py#L702\r\n\r\nwe should have:\r\n```python\r\nreturn pa.types.is_primitive(pa_type) and not (pa.types.is_boolean(pa_type) or pa.types.is_temporal(pa_type))\r\n```", "@mariosasko submitted a small PR [here](https://github.com/huggingface/datasets/pull/5504)" ]

[ "Totally agree, the docs should be in sync with our code.\r\n\r\nIndeed to avoid confusing users, I think we should have updated the docs at the same time as this PR:\r\n- #5167", "@albertvillanova yeah sure I should have, but I forgot back then, sorry for that 😶", "No, @polinaeterna, nothing to be sorry about.\r\n\r\nMy comment was for all of us datasets team, as a reminder: when making a PR, but also when reviewing some other's PR, we should not forget to update the corresponding docstring and doc pages. It is something we can improve if we help each other in reminding about it... :hugs: ", "@polinaeterna I think we can close this issue now as we no longer use `torchaudio` for decoding." ]

[ "Assigned to myself and will get to it in the next week, but if someone finds this issue annoying and wants to submit a PR before I do, just ping me here and I'll reassign :). ", "I would like to be assigned to this issue, @nateraw . #self-assign" ]

[ "Hi @kradonneoh, thanks for reporting.\r\n\r\nPlease note that to work with audio datasets (and specifically with MP3 files) we have detailed installation instructions in our docs: https://huggingface.co/docs/datasets/installation#audio\r\n- one of the requirements is torchaudio<0.12.0\r\n\r\nLet us know if the problem persists after having followed them.", "I saw that and have followed it (hence the Expected Behavior section of the bug report). \r\n\r\nIs there no intention of updating to the latest version? It does limit the version of `torch` I can use, which isn’t ideal.", "@kradonneoh hey! actually with `ffmpeg4` loading of mp3 files should work, so this is a not expected behavior and we need to investigate it. It works on my side with `torchaudio==0.13` and `ffmpeg==4.2.7`. Which `torchaudio` version do you use?\r\n\r\n`datasets` should support decoding of mp3 files with `torchaudio` when its version is `>0.12` but as you noted it requires `ffmpeg>4`, we need to fix this in the documentation, thank you for pointing to this! \r\n\r\nBut according to your traceback it seems that it tries to use [`libsndfile`](https://github.com/libsndfile/libsndfile) backend for mp3 decoding. And `libsndfile` library supports mp3 decoding starting from version 1.1.0 which on Linux has to be compiled from source for now afaik. \r\n\r\nfyi - we are aiming at getting rid of `torchaudio` dependency at all by the next major library release in favor of `libsndfile` too.", "We now decode MP3 with `soundfile`, so I'm closing this issue" ]

[ "Hi! The correct path is still `dill._dill.XXXX` in the latest release. What do you get when you run `python -c \"import dill; print(dill.__version__)\"` in your environment?", "`0.3.6` I feel like that's bad news, because it's probably not the issue.\r\n\r\nMy mistake, about the wrong path guess. I think I didn't notice that the first `dill` in the path isn't supposed to be included in the path specification in python.\r\n<img width=\"146\" alt=\"Screen Shot 2023-01-31 at 12 58 32 PM\" src=\"https://user-images.githubusercontent.com/35349273/215844209-74af6a8f-9bff-4c75-9495-44c658c8e9f7.png\">\r\n", "Hi, @avivbrokman, this issue you report appeared only with old versions of dill. See:\r\n- #288\r\n\r\nAre you sure you are in the right Python environment?\r\n- Please note that Jupyter (where I guess you get the error) may have multiple execution backends (IPython kernels) that might be different from the Python environment your are using to get the dill version\r\n - Have you run `import dill; print(dill.__version__)` in the same Jupyter/IPython that you were using when you got the error while executing `import datasets`?", "I'm using spyder, and I am still getting `0.3.6` for `dill`, so unfortunately #288 isn't applicable, I think. However, I found something odd that I believe is a clue: \r\n\r\n```\r\nimport inspect\r\nimport dill\r\n\r\ninspect.getfile(dill)\r\n>>> '/Users/avivbrokman/opt/anaconda3/lib/python3.9/site-packages/dill/__init__.py'\r\n```\r\n\r\nI checked out the directory, and there is no `dill` subdirectory within '/Users/avivbrokman/opt/anaconda3/lib/python3.9/site-packages/dill`, as there should be. Rather, `_dill.py` is in '/Users/avivbrokman/opt/anaconda3/lib/python3.9/site-packages/dill` itself. \r\n\r\n If I run `pip install dill` or `pip install --upgrade dill`, I get the message `Requirement already satisfied: dill in ./opt/anaconda3/lib/python3.9/site-packages (0.3.6)`. If I run `conda upgrade dill`, I get the message `Solving environment: failed with repodata from current_repodata.json, will retry with next repodata source.` a couple of times, followed by\r\n\r\n```\r\nSolving environment: failed\r\nSolving environment: / \r\nFound conflicts! Looking for incompatible packages.\r\n```\r\n\r\nAnd then terminal proceeds to list conflicts between different packages I have.\r\n\r\nThis is all very strange to me because I recently uninstalled and reinstalled `anaconda`.\r\n", "As I said above, I guess this is not a problem with `datasets`. I think you have different Python environments: one with the new dill version (the one you get while using pip) and other with the old dill version (the one where you get the AttributeError).\r\n\r\nYou should update `dill` in the Python environment you are using within spyder.\r\n\r\nPlease note that the `_dill` module is present in the `dill` package since their 2.8.0 version." ]

[ "Hi @omar-araboghli, thanks for your proposal.\r\n\r\nHave you tried to use \"csv\" loader instead of \"text\"? That already has a `sep` argument.", "Hi @albertvillanova, thanks for the quick response!\r\n\r\nIndeed, I have been trying to use `csv` instead of `text`. However I am still not able to define range of rows as one sequence, that is achievable with passing `sample_by='paragraph'` to the `TextConfig`\r\n\r\nFor instance, the below code\r\n\r\n```python\r\nimport datasets\r\n\r\ndataset = datasets.load_dataset(\r\n path='csv',\r\n data_files={'train': TRAINING_SET_PATH},\r\n sep='\\t',\r\n header=None,\r\n column_names=['tokens', 'pos_tags', 'chunk_tags', 'ner_tags']\r\n)\r\n```\r\n\r\nleads to \r\n\r\n```python\r\ndataset\r\n>>> DatasetDict({\r\n train: Dataset({\r\n features: ['tokens', 'pos_tags', 'chunk_tags', 'ner_tags'],\r\n num_rows: 62543\r\n })\r\n})\r\n\r\ndataset['train'][0]\r\n>>> {'tokens': 'Distribution',\r\n 'pos_tags': 'NN',\r\n 'chunk_tags': 'O',\r\n 'ner_tags': 'O'\r\n}\r\n```\r\nIs there a way to deal with multiple csv rows as one dataset instance, where each column is a sequence of those rows ?" ]

[ "Seems to work now, perhaps it was something internal with our university's network." ]

[ "I'd be happy to have a look, if nobody else has started working on this yet @lhoestq. \r\n\r\nIt seems to me that for the `arrow` format features are currently attached as metadata [in `datasets.arrow_writer`](https://github.com/huggingface/datasets/blob/5f810b7011a8a4ab077a1847c024d2d9e267b065/src/datasets/arrow_writer.py#L412) and retrieved from the metadata at `load_dataset` time using [`datasets.features.features.from_arrow_schema`](https://github.com/huggingface/datasets/blob/5f810b7011a8a4ab077a1847c024d2d9e267b065/src/datasets/features/features.py#L1602). \r\n\r\nThis will need to be replicated for `parquet` via calls to [this api](https://arrow.apache.org/docs/python/generated/pyarrow.parquet.write_metadata.html) from `io.parquet.ParquetWriter` and `io.parquet.ParquetReader` [respectively](https://github.com/huggingface/datasets/blob/5f810b7011a8a4ab077a1847c024d2d9e267b065/src/datasets/io/parquet.py#L104).\r\n\r\nAny other important considerations?\r\n", "Thanks @MFreidank ! That's correct :)\r\n\r\nReading the metadata to infer the features can be ideally done in the `parquet.py` file in `packaged_builder` when a parquet file is read. You can cast the arrow table to the schema you get from the features.arrow_schema", "#self-assign" ]

[ "Can I work on this issue? I am pretty new to this.", "Hi ! Sure :) you can comment `#self-assign` to assign yourself to this issue.\r\n\r\nI can give you some pointers to get started:\r\n\r\n`load_dataset` works roughly this way:\r\n1. it instantiate a dataset builder using `load_dataset_builder()`\r\n2. the builder download and prepare the dataset as Arrow files in the cache using `download_and_prepare()`\r\n3. the builder returns a Dataset object with `as_dataset()`\r\n\r\nOne way to approach this would be to implement `as_iterable_dataset()` in `builder.py`.\r\n\r\nAnd similarly to `as_dataset()`, you can use the `ArrowReader`. It has a `get_file_instructions()` method that can be helpful. It gives you the files to read as list of dictionaries with those keys: `filename`, `skip` and `take`.\r\n\r\nThe `skip` and `take` arguments are used in case the user wants to load a subset of the dataset, e.g.\r\n```python\r\nload_dataset(..., split=\"train[:10]\")\r\n```\r\n\r\nLet me know if you have questions or if I can help :)", "This use-case is a bit specific, and `load_dataset` already has enough parameters (plus, `streaming=True` also returns an iterable dataset, so we would have to explain the difference), so I think it would be better to add `IterableDataset.from_file` to the API (more flexible and aligned with the goal from https://github.com/huggingface/datasets/issues/3444) instead.", "> This use-case is a bit specific\r\n\r\nThis allows to use `datasets` for large scale training where map-style datasets are too slow and use too much memory in PyTorch. So I would still consider adding it.\r\n\r\nAlternatively we could add this feature one level bellow:\r\n```python\r\nbuilder = load_dataset_builder(...)\r\nbuilder.download_and_prepare()\r\nids = builder.as_iterable_dataset()\r\n```", "Yes, I see how this can be useful. Still, I think `Dataset.to_iterable` + `IterableDataset.from_file` would be much cleaner in terms of the API design (and more flexible since `load_dataset` can only access the \"initial\" (unprocessed) version of a dataset).\r\n\r\nAnd since it can be tricky to manually find the \"initial\" version of a dataset in the cache, maybe `load_dataset` could return an iterable dataset streamed from the cache if `streaming=True` and the cache is up-to-date. ", "> This allows to use datasets for large scale training where map-style datasets are too slow and use too much memory in PyTorch.\r\n\r\nI second that. e.g. In my last experiment Oscar-en uses 16GB RSS RAM per process and when using multiple processes the host quickly runs out cpu memory. ", ">And since it can be tricky to manually find the \"initial\" version of a dataset in the cache, maybe load_dataset could return an iterable dataset streamed from the cache if streaming=True and the cache is up-to-date.\r\n\r\nThis is exactly the need on JeanZay (HPC) - I have the dataset cache ready, but the compute node is offline, so making streaming work off a local cache would address that need.\r\n\r\nIf you will have a working POC I can be the tester. ", "> Yes, I see how this can be useful. Still, I think Dataset.to_iterable + IterableDataset.from_file would be much cleaner in terms of the API design (and more flexible since load_dataset can only access the \"initial\" (unprocessed) version of a dataset).\r\n\r\nI like `IterableDataset.from_file` as well. On the other hand `Dataset.to_iterable` first requires to load a Dataset object, which can take time depending on your hardware and your dataset size (sometimes 1h+).\r\n\r\n> And since it can be tricky to manually find the \"initial\" version of a dataset in the cache, maybe load_dataset could return an iterable dataset streamed from the cache if streaming=True and the cache is up-to-date.\r\n\r\nThat would definitely do the job. I was suggesting a different parameter just to make explicit the difference between\r\n- streaming from the raw data\r\n- streaming from the local cache\r\n\r\nBut I'd be fine with streaming from cache is the cache is up-to-date since it's always faster. We could log a message as usual to make it explicit that the cache is used", "> I was suggesting a different parameter just to make explicit the difference between\r\n\r\nMosaicML's `streaming` library does the same (tries to stream from the local cache if possible), so logging a message should be explicit enough :).", "Ok ! Sounds good then :)", "Hi Both! It has been a while since my first issue so I am gonna go for this one ! #self-assign", "#self-assign", "I like idea of `IterableDataset.from_file`. ", "https://github.com/huggingface/datasets/pull/5821 should be helpful to implement `IterableDataset.from_file`, since it defines a new ArrowExamplesIterable that takes an Arrow tables generator function (e.g. from a file) and can be used in an IterableDataset", "@lhoestq I have just started working on this issue. ", "@lhoestq Thank you for taking over.", "So what's recommanded usage of `IterableDataset.from_file` and `load_dataset`? How about I have multiple arrow files and `load_dataset` is often convenient to handle that.", "If you have multiple Arrow files you can load them using\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\ndata_files = {\"train\": [\"path/to/0.arrow\", \"path/to/1.arrow\", ..., \"path/to/n.arrow\"]}\r\n\r\nds = load_dataset(\"arrow\", data_files=data_files, streaming=True)\r\n```\r\n\r\nThis is equivalent to calling `IterableDataset.from_file` and `concatenate_datasets`." ]

[]

[ "@albertvillanova It looks like that issue has been fixed so I made a PR to unpin sqlalchemy! ", "The source issue:\r\n- https://github.com/pandas-dev/pandas/issues/40686\r\n\r\nhas been fixed:\r\n- https://github.com/pandas-dev/pandas/pull/48576\r\n\r\nThe fix was released yesterday (2023-04-03) only in `pandas-2.0.0`:\r\n- https://github.com/pandas-dev/pandas/releases/tag/v2.0.0\r\n\r\nbut it will not be back-ported to `pandas-1`:\r\n- https://github.com/pandas-dev/pandas/pull/48576#issuecomment-1466467159\r\n\r\nAlso note that `pandas-2.0.0` dropped support for Python 3.7:\r\n- https://github.com/pandas-dev/pandas/issues/41678\r\n- https://github.com/pandas-dev/pandas/pull/41989\r\n\r\nTherefore, we cannot unpin `sqlalchemy` until we drop support for Python 3.7 (these Python users cannot use `pandas-2`)." ]

[ "Hi ! In your code you only iterate on the Arrow buffers - you don't actually load the data as python objects. For a fair comparison, you can modify your code using:\r\n```diff\r\n- for _ in range(0, len(table), bsz):\r\n- _ = {k:table[k][_ : _ + bsz] for k in cols}\r\n+ for _ in range(0, len(table), bsz):\r\n+ _ = {k:table[k][_ : _ + bsz].to_pylist() for k in cols}\r\n```\r\n\r\nI re-ran your code and got a speed ratio of 1.00x and 1.02x", "Ah I see, datasets is implicitly making this conversion. Thanks for pointing that out!\r\n\r\nIf it's not too much, I would also suggest updating some of your docs with the same `.to_pylist()` conversion in the code snippet that follows [here](https://huggingface.co/course/chapter5/4?fw=pt#:~:text=let%E2%80%99s%20run%20a%20little%20speed%20test%20by%20iterating%20over%20all%20the%20elements%20in%20the%20PubMed%20Abstracts%20dataset%3A).", "This code snippet shows `datasets` code that reads the Arrow data as python objects already, there is no need to add to_pylist. Or were you thinking about something else ?" ]