Datasets:

---

dongxiaoxia194

/

github-issues

like 0

[]

[]

[ "Unfortunately, I'm unable to reproduce this error. Can you share the reproducer?", "```\r\nds = datasets.Dataset.from_dict(dict(a=[i for i in range(100)]))\r\nds.map(lambda item: dict(b=item['a'] * 2), cache_file_name=\"/tmp/whatever-fn\") # this worked\r\nds.map(lambda item: dict(b=item['a'] * 2), cache_file_name=\"/tmp/whatever-folder/filename\") # this failed\r\nds.map(lambda item: dict(b=item['a'] * 2), cache_file_name=\"/tmp/whatever-folder/\") # this failed\r\n\r\n\r\nFileNotFoundError: [Errno 2] No such file or directory: '/tmp/whatever-folder/tmp1_izxvoo'\r\n```\r\n\r\nIt will fail if the filename parents do not exists. If we have `os.makedirs(\"/tmp/whatever-folder\")`, then it worked.\r\n\r\nMaybe add the `mkdir -p` into the map function?" ]

[]

[]

[]

[]

[]

[ "Hi ! The metrics in `datasets` are deprecated in favor of https://github.com/huggingface/evaluate\r\n\r\nYou can open a PR here instead: https://huggingface.co/spaces/evaluate-metric/squad_v2/tree/main" ]

[]

[]

[ "Hi! Parquet is the only format that supports complex/nested features such as `Translation`. So, this should work:\r\n```python\r\ntest_dataset = load_dataset(\"opus100\", name=\"en-fr\", split=\"test\")\r\n\r\n# Save with .to_parquet()\r\ntest_parquet_path = \"try_testset_save.parquet\"\r\ntest_dataset.to_parquet(test_parquet_path)\r\n\r\n# Load dataset from the Parquet\r\nloaded_dataset = load_dataset(\"parquet\", data_files=test_parquet_path)\r\nprint(test_dataset_fromfile[0][\"translation\"])\r\nprint(test_dataset_fromfile[0][\"translation\"][\"en\"])\r\n```", "Indeed this works great, thank you !" ]

[]

[]

[ "I am facing similar issue while reading a csv file from s3. Wondering if somebody has found a workaround. ", "same thing happened to other formats like parquet", "I am facing similar issue while reading a parquet file from s3.\r\ni try with every version between 2.14 to 2.16.1 but it dosen't work ", "Re-define the DownloadConfig might work:\r\n\r\n```\r\nclass ReviseDownloadConfig(DownloadConfig):\r\n def __post_init__(self, use_auth_token):\r\n if use_auth_token != \"deprecated\":\r\n warnings.warn(\r\n \"'use_auth_token' was deprecated in favor of 'token' in version 2.14.0 and will be removed in 3.0.0.\\n\"\r\n f\"You can remove this warning by passing 'token={use_auth_token}' instead.\",\r\n FutureWarning,\r\n )\r\n self.token = use_auth_token\r\n\r\n def copy(self):\r\n return self.__class__(**{k: copy.deepcopy(v) for k, v in self.__dict__.items()})\r\n\r\ndownloadconfig = ReviseDownloadConfig()\r\n```\r\n", "> Re-define the DownloadConfig might work:\r\n> \r\n> ```\r\n> class ReviseDownloadConfig(DownloadConfig):\r\n> def __post_init__(self, use_auth_token):\r\n> if use_auth_token != \"deprecated\":\r\n> warnings.warn(\r\n> \"'use_auth_token' was deprecated in favor of 'token' in version 2.14.0 and will be removed in 3.0.0.\\n\"\r\n> f\"You can remove this warning by passing 'token={use_auth_token}' instead.\",\r\n> FutureWarning,\r\n> )\r\n> self.token = use_auth_token\r\n> ```\r\nThis seemed to work for me.\r\n", "use pandas and then convert to `Dataset`", "I am currently facing the same issue while using a custom loading script with files located in a remote S3 instance. I was using the `download_custom` functionality but now it is deprecated mentioning that I should use the native S3 loading, which is not working. \r\n\r\nAs stated before, the library forces the existence of a `hf` key in the `storage_options` variable, which is **not** accepted by `s3fs` : \r\n\r\n```python\r\n.../site-packages/s3fs/core.py\", line 516, in set_session\r\n self.session = aiobotocore.session.AioSession(**self.kwargs)\r\nTypeError: __init__() got an unexpected keyword argument 'hf'.\r\n````\r\n\r\nMeanwhile, if my `storage_options` var stays like:\r\n```python\r\n{'key': '...',\r\n 'secret': '...',\r\n 'client_kwargs': {'endpoint_url': '...'}}\r\n```\r\nit works alright. " ]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[ { "login": "JochenSiegWork", "id": 135010976, "node_id": "U_kgDOCAwaoA", "avatar_url": "https://avatars.githubusercontent.com/u/135010976?v=4", "gravatar_id": "", "url": "https://api.github.com/users/JochenSiegWork", "html_url": "https://github.com/JochenSiegWork", "followers_url": "https://api.github.com/users/JochenSiegWork/followers", "following_url": "https://api.github.com/users/JochenSiegWork/following{/other_user}", "gists_url": "https://api.github.com/users/JochenSiegWork/gists{/gist_id}", "starred_url": "https://api.github.com/users/JochenSiegWork/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/JochenSiegWork/subscriptions", "organizations_url": "https://api.github.com/users/JochenSiegWork/orgs", "repos_url": "https://api.github.com/users/JochenSiegWork/repos", "events_url": "https://api.github.com/users/JochenSiegWork/events{/privacy}", "received_events_url": "https://api.github.com/users/JochenSiegWork/received_events", "type": "User", "site_admin": false } ]

[ "Hi ! This issue comes from the fact that `map()` with `num_proc>1` shards the dataset in multiple chunks to be processed (one per process) and merges them. The DatasetInfos of each chunk are then merged together, but for some fields like `dataset_name` it's not been implemented and default to None.\r\n\r\nThe DatasetInfo merge is defined here, in case you'd like to contribute an improvement: \r\n\r\nhttps://github.com/huggingface/datasets/blob/d2e0034122a788015c0834a72e6c6279e7ecbac5/src/datasets/info.py#L269-L270", "#self-assign" ]

[]

[]

[ "@lhoestq\r\nCan you provide me with some ideas？", "Hi ! What's the error ?", "@lhoestq \r\n```\r\nTraceback (most recent call last):\r\n File \"/home/dongzf/miniconda3/envs/dataset_ai/lib/python3.11/runpy.py\", line 198, in _run_module_as_main\r\n return _run_code(code, main_globals, None,\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/dongzf/miniconda3/envs/dataset_ai/lib/python3.11/runpy.py\", line 88, in _run_code\r\n exec(code, run_globals)\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/__main__.py\", line 39, in <module>\r\n cli.main()\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py\", line 430, in main\r\n run()\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher/../../debugpy/../debugpy/server/cli.py\", line 284, in run_file\r\n runpy.run_path(target, run_name=\"__main__\")\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py\", line 321, in run_path\r\n return _run_module_code(code, init_globals, run_name,\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py\", line 135, in _run_module_code\r\n _run_code(code, mod_globals, init_globals,\r\n File \"/home/dongzf/.vscode/extensions/ms-python.python-2023.22.1/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydevd_bundle/pydevd_runpy.py\", line 124, in _run_code\r\n exec(code, run_globals)\r\n File \"/mnt/sda/code/dataset_ai/dataset_ai/example/test.py\", line 83, in <module>\r\n data = xnumpy_fromfile(current_dir, download_config=config,dtype=numpy.float32,)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/mnt/sda/code/dataset_ai/dataset_ai/src/datasets/download/streaming_download_manager.py\", line 765, in xnumpy_fromfile\r\n return np.fromfile(xopen(filepath_or_buffer, \"rb\", download_config=download_config).read(), *args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\nValueError: embedded null byte\r\n```", " not add read() \r\nthe error is \r\n\r\nreturn np.fromfile(xopen(filepath_or_buffer, \"rb\", download_config=download_config), *args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\nio.UnsupportedOperation: fileno", "xopen return obj do not have fileno function\r\nI don't know why?", "I used this method to read point cloud data in the script\r\n\r\n\r\n```python\r\nwith open(velodyne_filepath,\"rb\") as obj:\r\n velodyne_data = numpy.frombuffer(obj.read(), dtype=numpy.float32).reshape([-1, 4])\r\n```" ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[ "Seems like I just used the old code which did not have `keep_in_memory=True` argument, sorry.\r\n\r\nAlthough i encountered a different problem – at 97% my python process just hung for around 11 minutes with no logs (when running dataset.map without `keep_in_memory=True` over around 3 million of dataset samples)...", "Can you open a new issue and provide a bit more details ? What kind of map operations did you run ?", "Hey. I will try to find some free time to describe it.\r\n\r\n(can't do it now, cause i need to reproduce it myself to be sure about everything, which requires spinning a new Azuree VM, copying a huge dataset to drive from network disk for a long time etc...)", "@lhoestq loading dataset like this does not spawn 50 python processes:\r\n\r\n```\r\ndatasets.load_dataset(\"/preprocessed_2256k/train\", num_proc=50)\r\n```\r\n\r\nI have 64 vCPU so i hoped it could speed up the dataset loading...\r\n\r\nMy dataset onlly has images and metadata.csv with text column alongside image file path column", "now noticed\r\n```\r\n'Setting num_proc from 50 back to 1 for the train split to disable multiprocessing as it only contains one shard\r\n```\r\n\r\nAny way to work around this?", "@lhoestq thanks, [this helped](https://github.com/huggingface/datasets/blob/9d6d16117a30ba345b0236407975f701c5b288d4/src/datasets/arrow_dataset.py#L1053)\r\n\r\n" ]

[]

[]

[ "I also see the same error and get passed it by casting that line to float. \r\n\r\nso `for x in obj.detach().cpu().numpy()` becomes `for x in obj.detach().to(torch.float).cpu().numpy()`\r\n\r\nI got the idea from [this ](https://github.com/kohya-ss/sd-webui-additional-networks/pull/128/files) PR where someone was facing a similar issue (in a different repository). I guess numpy doesn't support bfloat16.\r\n\r\n" ]

[]

[]

[]

[ { "id": 1935892861, "node_id": "MDU6TGFiZWwxOTM1ODkyODYx", "url": "https://api.github.com/repos/huggingface/datasets/labels/documentation", "name": "documentation", "color": "0075ca", "default": true, "description": "Improvements or additions to documentation" } ]

[]

[ "In particular, I would like to have an example of how to replace the following configuration (from https://huggingface.co/docs/hub/datasets-manual-configuration#splits)\r\n\r\n```\r\n---\r\nconfigs:\r\n- config_name: default\r\n data_files:\r\n - split: train\r\n path: \"data/*.csv\"\r\n - split: test\r\n path: \"holdout/*.csv\"\r\n---\r\n```\r\n\r\nwith the `data_dir` field." ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[ "Maybe we can add a helper message like `Maybe try again using \"hf://path/without/resolve\"` if the path contains `/resolve/` ?\r\n\r\ne.g.\r\n\r\n```\r\nFileNotFoundError: Unable to find 'hf://datasets/HuggingFaceTB/eval_data/resolve/main/eval_data_context_and_answers.json'\r\nIt looks like you used parts of the URL of the file from the Hugging Face website, but you should remove the \"/resolve/<revision>\" part to have a valid `hf://` path.\r\nPlease try again using this path instead:\r\n hf://datasets/HuggingFaceTB/eval_data/eval_data_context_and_answers.json\r\n```\r\n\r\nand suggest `f\"hf://datasets/HuggingFaceTB/eval_data@{revision}/eval_data_context_and_answers.json\"` if revision != \"main\"\r\n\r\nEDIT: I think this message should also be raised from the `huggingface_hub`'s `HfFileSystem` implementation" ]

[]

[]

[ "It looks like a transient DNS issue. It should work fine now if you try again.\r\n\r\nThere is no parameter in load_dataset to skip failed downloads. In your case it would have skipped every single subsequent download until the DNS issue was resolved anyway." ]

[]

[]

[ "Concatenating datasets doesn't create any indices mapping - so flattening indices is not needed (unless you shuffle the dataset).\r\nCan you share the snippet of code you are using to merge your datasets and save them to disk ?" ]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[ "Related to #3113 ", "Conceptually, we can use xarray to load the netCDF file, then xarray -> pandas -> pyarrow.", "I'd still need to verify that such a conversion would be lossless, especially for multi-dimensional data." ]

[]

[]

[ "@sabman @pvl @kashif @vigsterkr ", "This is surely the same issue as https://discuss.huggingface.co/t/indexerror-invalid-key-16-is-out-of-bounds-for-size-0/14298/25 that comes from the `transformers` `Trainer`. You should add `remove_unused_columns=False` to `TrainingArguments`\r\n\r\nAlso check your logs: the `Trainer` should log the length of your dataset before training starts and it surely showed length=0.", "the same error \r\nIndexError: Invalid key: 22330 is out of bounds for size 0" ]

[]

[]

[ "@albertvillanova" ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[ "You can simply use a python dict as index:\r\n\r\n```python\r\n>>> from datasets import load_dataset\r\n>>> ds = load_dataset(\"BeIR/dbpedia-entity\", \"corpus\", split=\"corpus\")\r\n>>> index = {key: idx for idx, key in enumerate(ds[\"_id\"])}\r\n>>> ds[index[\"<dbpedia:Pikachu>\"]]\r\n{'_id': '<dbpedia:Pikachu>',\r\n 'title': 'Pikachu',\r\n 'text': 'Pikachu (Japanese: ピカチュウ) are a fictional species of Pokémon. Pokémon are fictional creatures that appear in an assortment of comic books, animated movies and television shows, video games, and trading card games licensed by The Pokémon Company, a Japanese corporation. The Pikachu design was conceived by Ken Sugimori.'}\r\n```", "Thanks for your reply. Yes, I can do that, but it is time-consuming to do that every time I launch the program (some datasets are extremely big). HF Datasets has a nice feature to support instant data loading and efficient random access via row ids. I'm curious if this beneficial feature could be further extended to custom data columns.\r\n", "+1 on the issue I think it would be extremely useful" ]

[]

[]

[ "I solved it with `train_dataset.with_format(None)`\r\nBut then faced some more issues (which i later solved too).\r\n\r\nHuggingface does not seem to care, so I do. Here is an updated training script which saves a pre-processed (mapped) dataset to your local directory if you specify `--save_precomputed_data_dir=DIR_NAME`. Then use `--train_precomputed_data_dir` with the same dir to load it instead of `--dataset_name`.\r\n\r\n[Proper SDXL trainer code](https://github.com/kopyl/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py)\r\n[Notebook for pre-computing a dataset and saving locally](https://colab.research.google.com/drive/17Yo08hePx-NlHs99RecdeiNc8CQg4O7l?usp=sharing)\r\n\r\nExample:\r\n\r\n1st run (nothing is pre-computed yet):\r\n```\r\naccelerate launch train_text_to_image_sdxl.py \\\r\n --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0 \\\r\n --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \\\r\n --dataset_name=lambdalabs/pokemon-blip-captions \\\r\n --save_precomputed_data_dir=\"test-5\"\r\n```\r\n\r\n2nd run (the pre-computed dataset is saved to `test-5` directory):\r\n```\r\naccelerate launch train_text_to_image_sdxl.py \\\r\n --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0 \\\r\n --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \\\r\n --train_precomputed_data_dir test-5\r\n```\r\n\r\nThis way when you're gonna be using your pre-computed dataset you don't need to worry about re-mapping your dataset when you change an argument for your trainer script" ]

[]

[]

[ "The only way right now is to load with streaming=True", "This feature has been proposed for a long time. I'm looking forward to the implementation. On clusters `streaming=True` is not an option since we do not have Internet on compute nodes. See: https://github.com/huggingface/datasets/discussions/1896#discussioncomment-2359593" ]

[]

[]

[]

[]

[]

[ "I ran into the same problem when I used a server cluster (Slurm system managed) that couldn't load any of the huggingface datasets or models, but it worked on my laptop. I suspected some system configuration-related problem, but I had no idea. \r\nMy problems are consistent with [issue #2618](https://github.com/huggingface/datasets/issues/2618). All the huggingface-related libraries I use are the latest versions.\r\n\r\n", "> I ran into the same problem when I used a server cluster (Slurm system managed) that couldn't load any of the huggingface datasets or models, but it worked on my laptop. I suspected some system configuration-related problem, but I had no idea. My problems are consistent with [issue #2618](https://github.com/huggingface/datasets/issues/2618). All the huggingface-related libraries I use are the latest versions.\r\n\r\nhave you solved this issue yet? i met the same problem on server but everything works on laptop. I think maybe the filelock repo is contradictory with file system.", "I am having the same issue on a computing cluster but this works on my laptop as well. I instead have this error:\r\n`/home/.conda/envs/py10/lib/python3.10/site-packages/filelock/_unix.py\", line 43, in _acquire\r\n fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)\r\nOSError: [Errno 5] Input/output error`\r\n\r\nthe load_dataset command does not work on server for local or hosted hugging-face datasets, and I have tried for several files", "Same here. Is there any solution?", "In my case, `.cahce` was in a shared folder. Moving it into the user's home folder fixed the problem. #2618 for more details", "> In my case, `.cahce` was in a shared folder. Moving it into the user's home folder fixed the problem. #2618 for more details在我的情况下， `.cahce` 在一个共享文件夹中。将其移动到用户的主文件夹中解决了问题。 #2618 获取更多详细信息。\r\n\r\nCan you be more specific? thank." ]

[]

[]

[]

[]

[]

[ "I transferred from datasets-server, since the issue is more about `datasets` and the integration with `huggingface_hub`." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Also, once the README.md exists in the python environment it is used when loading another dataset in the same format (e.g. json) since it always resolves the path to the same directory.\r\nThe consequence here is any other dataset won't load because of infos mismatch.\r\nTo reproduce this aspect:\r\n1. Do a `load_datasets(...,save_infos=True)` with one dataset without a loading script\r\n2. Try to load another dataset without a loading script in the same format (e.g. json) but with a different schema " ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[ "Transferring repos as this is a datasets issue ", "I'm getting a similar issue even though I've already downloaded the dataset 😅 \r\n\r\n```\r\nhuggingface_hub.utils._errors.HfHubHTTPError: 429 Client Error: Too Many Requests for url: https://huggingface.co/api/datasets/HuggingFaceM4/WebSight\r\n```" ]

[]

[]

[ "Hello @kenfus, this is meant to be possible to do yes. Let me ping @lhoestq or @mariosasko from the `datasets` team (`huggingface_hub` is only the underlying library to download files from the Hub but here it looks more like a `datasets` problem). ", "Hi! https://github.com/huggingface/datasets/pull/6459 will fix this." ]

[]

[]

[]

[]

[]

[ "You can create a `pandas` DataFrame following [this](https://lakefs.io/data-version-control/dvc-using-python/) tutorial, and then convert this DataFrame to a `Dataset` with `datasets.Dataset.from_pandas`. For larger datasets (to memory map them), you can use `Dataset.from_generator` with a generator function that reads lakeFS files with `s3fs`.", "@mariosasko hello,\r\nThis can achieve and https://huggingface.co/datasets Does the same effect apply to the dataset? For example, downloading while using" ]

[]

[]

[ "~~You will see this error if the cache dir filepath contains relative `..` paths. Use `os.path.realpath(_CACHE_DIR)` before passing it to the `load_dataset` function.~~", "This is a real issue and not related to paths.", "Based on the StackOverflow answer, this causes the error to go away:\r\n```diff\r\ndiff --git a/table.py b/table.py\r\n--- a/table.py\t\r\n+++ b/table.py\t(date 1701824849806)\r\n@@ -47,7 +47,7 @@\r\n \r\n \r\n def _memory_mapped_record_batch_reader_from_file(filename: str) -> pa.RecordBatchStreamReader:\r\n- memory_mapped_stream = pa.memory_map(filename)\r\n+ memory_mapped_stream = pa.memory_map(filename, \"r+\")\r\n return pa.ipc.open_stream(memory_mapped_stream)\r\n```\r\nBut now loading the dataset goes very, very slowly, which is unexpected.", "I don't really comprehend what it is that `datasets` gave me when it downloaded the laion2B-en dataset, because nothing can seemingly read these 1024 .arrow files it is retrieving. Not `polars`, not `pyarrow`, it's not an `ipc` file, it's not a `parquet` file...", "Hi! \r\n\r\nInstead of generating one (potentially large) Arrow file, we shard the generated data into 500 MB shards because memory-mapping large Arrow files can be problematic on some systems. Maybe deleting the dataset's cache and increasing the shard size (controlled with the `datasets.config.MAX_SHARD_SIZE` variable; e.g. to \"4GB\") can fix the issue for you.\r\n\r\n> I don't really comprehend what it is that `datasets` gave me when it downloaded the laion2B-en dataset, because nothing can seemingly read these 1024 .arrow files it is retrieving. Not `polars`, not `pyarrow`, it's not an `ipc` file, it's not a `parquet` file...\r\n\r\nOur `.arrow` files are in the [Arrow streaming format](https://arrow.apache.org/docs/python/ipc.html#using-streams). To load them as a `polars` DataFrame, do the following:\r\n```python\r\ndf = pl.from_arrow(Dataset.from_from(path_to_arrow_file).data.table)\r\n```\r\n\r\nWe plan to switch to the IPC version eventually.\r\n", "Hmm, I have a feeling this works fine on Linux, and is a real bug for however `datasets` is doing the sharding on Windows. I will follow up, but I think this is a real bug." ]

[]

[]

[]

[]

[]

[ "Hi, thanks for reporting! https://github.com/huggingface/datasets/pull/6459 will fix this." ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[ { "id": 1935892857, "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "There is a typo in one of the file names - `data/edf.csv` should be renamed to `data/def.csv` 🙂. ", "wow, I reviewed it twice to avoid being ashamed like that, but... I didn't notice the typo.\r\n\r\n---\r\n\r\nBesides this: do you think we would be able to improve the error message to make this clearer?" ]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[ "Thank you, @severo ! I would be more than happy to help in any way I can. I am not familiar with this repo's codebase, but I would be eager to contribute. :)\r\n\r\nFor the preview in Datasets Hub, I think it makes sense to just display the geospatial column as text. If there were a dataset loader, though, I think it should be able to support the geospatial components. Geopandas is probably the most user-friendly interface for that. I'm not sure if it's currently relevant in the context of geoparquet, but I think the pyogrio driver is faster than fiona.\r\n\r\nBut the whole gdal dependency thing can be a real pain. If anything, it would need to be an optional dependency. Maybe it would be best if the loader tries importing relevant geospatial libraries, and in the event of an ImportError, falls back to text for the geometry column.\r\n\r\nPlease let me know if I can be of assistance, and thanks again for creating this Issue. :)", "Just hitting into this same issue too showing GeoParquet files in Datasets Viewer. I tried to implement a custom reader for GeoParquet in https://huggingface.co/datasets/weiji14/clay_vector_embeddings/discussions/1, but it seems like HuggingFace has disabled datasets with custom loading scripts from using the dataset viewer according to https://discuss.huggingface.co/t/dataset-repo-requires-arbitrary-python-code-execution/59346 :frowning_face: \r\n\r\n\r\n\r\nI'm thinking now if there's a way to simply map files with GeoParquet extensions (*.gpq, *.geoparquet, etc) to use the Parquet reader. Maybe we could allowlist these geoparquet file extensions at https://github.com/huggingface/datasets/blame/0caf91285116ec910f409e82cc6e1f4cff7496e3/src/datasets/packaged_modules/__init__.py#L30-L51? Having the table columns show up would be a quick win.\r\n\r\nLonger term though, it would certainly be nice if the WKB geometry columns could be displayed in a nicer form. Geopandas' [read_parquet](https://geopandas.org/en/v0.14.1/docs/reference/api/geopandas.read_parquet.html) function is supposedly faster than `pyogrio.read_dataframe` according to https://github.com/geopandas/geopandas/discussions/2724#discussioncomment-4606048, but there's also [`pyogrio.raw.read_arrow`](https://pyogrio.readthedocs.io/en/latest/api.html#pyogrio.raw.read_arrow) now that can read into a `pyarrow.Table` directly.", "Update: It looks like renaming the GeoParquet file to have a file extension of `*.parquet` works (see https://huggingface.co/datasets/weiji14/clay_vector_embeddings). HuggingFace's default parquet reader is able to read the GeoParquet file, though the geometry column is of an unknown type:\r\n\r\n\r\n\r\nI've opened a quick PR at #6508 to allow files with a `*.geoparquet` or `*.gpq` extension to be read using the default Parquet reader. Let's see how that goes :smile:", "@joshuasundance-swca, @weiji14, If I'm understanding this correctly, the code below wouldn't be recommended to due to dependency headaches? If that's the case, what solution would there be to see the geometry features for .gpq files in huggingfaceHub? \r\n\r\ncode for dataset_loader.py\r\n```\r\nimport geopandas as gpd\r\n# ... (other imports remain the same)\r\n\r\nclass ClayVectorEmbeddings(datasets.ArrowBasedBuilder):\r\n # ... (other parts of the class remain the same)\r\n\r\n def _info(self):\r\n # Read the GeoParquet file to get the schema for the 'geometry' feature\r\n gdf = gpd.read_file(\"path/to/your/geoparquet/file.gpq\") # Replace with your file path\r\n geometry_schema = str(gdf.geometry.dtype)\r\n\r\n return datasets.DatasetInfo(\r\n # This is the description that will appear on the datasets page.\r\n description=\"Clay Vector Embeddings in GeoParquet format.\",\r\n # This defines the different columns of the dataset and their types\r\n features=datasets.Features(\r\n {\r\n \"source_url\": datasets.Value(dtype=\"string\"),\r\n \"date\": datasets.Value(dtype=\"date32\"),\r\n \"embeddings\": datasets.Value(\"string\"),\r\n \"geometry\": datasets.Value(dtype=geometry_schema), # Use the schema read by GeoPandas\r\n # ... (other features)\r\n }\r\n ),\r\n )\r\n\r\n# ... (rest of the script remains the same)\r\n\r\n```", "Hi @mehrdad-es, I'm not sure if HuggingFace would be keen to add `geopandas` to HuggingFace Hub (maybe a question for @severo?). Having a geometry viewer would be an even bigger task, and if you're thinking of a map-viewer, it might involve some redesign of the website UI. Some of my colleagues are working on streamlining GeoParquet visualization from cloud-hosted instances like HuggingFace (see e.g. https://github.com/developmentseed/lonboard/issues/314), and we could definitely come up with something if there's interest.", "I've created https://github.com/huggingface/datasets-server/issues/2416 to discuss the possibility of supporting (vectorial) geospatial columns in the dataset viewer, or in the converted parquet files.\r\n\r\nAt the same time, it would be super interesting to see what is already possible to do with a Hugging Face dataset that hosts geospatial data. \r\n\r\n> Some of my colleagues are working on streamlining GeoParquet visualization from cloud-hosted instances like HuggingFace (see e.g. https://github.com/developmentseed/lonboard/issues/314), and we could definitely come up with something if there's interest.\r\n\r\nIt would be awesome to show this inside a [Space](https://huggingface.co/docs/hub/spaces)." ]

[]

[]

[ "Has anyone ever encountered this problem before?", "`split_dataset_by_node` doesn't give the exact same number of examples to each node in the case of iterable datasets, though it tries to be as equal as possible. In particular if your dataset is sharded and you have a number of shards that is a factor of the number of workers, then the shards will be evenly distributed among workers. If the shards don't contain the same number of examples, then some workers might end up with more examples than others.\r\n\r\nHowever if you use a Dataset you'll end up with the same amount of data, because we know the length of the dataset we can split it exactly where we want. Also Dataset objects don't load the full dataset in memory; instead it memory maps Arrow files from disk.", "> `split_dataset_by_node` doesn't give the exact same number of examples to each node in the case of iterable datasets, though it tries to be as equal as possible. In particular if your dataset is sharded and you have a number of shards that is a factor of the number of workers, then the shards will be evenly distributed among workers. If the shards don't contain the same number of examples, then some workers might end up with more examples than others.\r\n> \r\n> However if you use a Dataset you'll end up with the same amount of data, because we know the length of the dataset we can split it exactly where we want. Also Dataset objects don't load the full dataset in memory; instead it memory maps Arrow files from disk.\r\n\r\nThanks for your answer! I finally solve it by using the torch.distributed.algorithms.join.Join. I think maybe some rookie like me would face the same question the day after tomorrow hh.", "Great ! Maybe it can be worth having an example that we can include in the docs for other people, did you need anything else than the Join context manager used with the model and optimizer ?" ]

[]

[]

[ "You should use `datasets.load_dataset` instead of `nlp.load_dataset`, as the `nlp` package is outdated.\r\n\r\nIf switching to `datasets.load_dataset` doesn't fix the issue, sharing the JSON file (feel free to replace the data with dummy data) would be nice so that we can reproduce it ourselves." ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[ "We have a config variable that controls the batch size in `save_to_disk`:\r\n```python\r\nimport datasets\r\ndatasets.config.DEFAULT_MAX_BATCH_SIZE = <smaller_batch_size>\r\n...\r\nds.save_to_disk(...)\r\n```", "Thank you for your answer!\r\n\r\nFrom what I am reading in `https://github.com/huggingface/datasets/blob/2.14.5/src/datasets/arrow_dataset.py`, every function involved (`select`, `shard`, ...) has a default hardcoded batch size of 1000, as such:\r\n```python\r\ndef select(\r\n self,\r\n indices: Iterable,\r\n keep_in_memory: bool = False,\r\n indices_cache_file_name: Optional[str] = None,\r\n writer_batch_size: Optional[int] = 1000,\r\n new_fingerprint: Optional[str] = None,\r\n ) -> \"Dataset\":\r\n...\r\n```\r\nThen, `ArrowWriter` is instantiated with the specified `writer_batch_size`. In `ArrowWriter`, `writer_batch_size` is set to `datasets.config.DEFAULT_MAX_BATCH_SIZE` if it is `None`(https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_writer.py#L345C14-L345C31). However, in our case, it is already set to 1000 by \"parent\" methods, so it won't happen.\r\n\r\nNevertheless, due to this: \r\n```python\r\ndef _save_to_disk_single(job_id: int, shard: \"Dataset\", fpath: str, storage_options: Optional[dict]):\r\n batch_size = config.DEFAULT_MAX_BATCH_SIZE\r\n...\r\n```\r\nit seems to work. I will use it as such, but it should maybe be added to documentation? And maybe improved in next versions?" ]

[]

[]

[ "Hi! You can avoid the error by requesting only the `jsonl` files. `dataset = load_dataset(\"ai4privacy/pii-masking-200k\", data_files=[\"*.jsonl\"])`.\r\n\r\nOur data file inference does not filter out (incompatible) `json` files because `json` and `jsonl` use the same builder. Still, I think the inference should differentiate these extensions because it's safe to assume that loading them together will lead to an error. WDYT @lhoestq? ", "Raising an error if there is a mix of json and jsonl in the builder makes sense yea" ]

[]

[]

[]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[ { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "great idea IMO\r\n\r\nthis could be a `trust_remote_code=True` flag like in transformers. We could also default to loading the Parquet conversion rather than executing code (for dataset repos that have both)", "@julien-c that would be great!", "We added the `trust_remote_code` argument to `load_dataset()` in `datasets` 2.16:\r\n- in the future users will have to pass trust_remote_code=True to use datasets with a script\r\n- for now we just show a warning when a dataset script is used\r\n- we fallback on the Hugging Face Parquet exports when possible (to keep compatibility with old datasets with scripts)\r\n\r\nSo feel free to use `trust_remote_code=False` in the meantime to disable loading from dataset loading scripts :)" ]

[]

[]

[]

[]

[]

[ "It looks like I may not be the first to encounter this: https://github.com/huggingface/datasets/issues/3172", "Adding some more information, it seems to occur more frequently with large (millions of samples) datasets.", "More information. My code is structured as (1) load (2) map (3) filter (4) filter. It was always the second filter that failed. Combining the two filters into one seems to reliably work.", "@lhoestq it'd be great if someone had a chance to look at this. I suspect it is impacting many users given the other issue that I linked.", "Hi ! Sorry for the late response. Was it happening after the first or the second filter ?\r\n\r\nIt looks like an issue with the garbage collector (which makes it random). Maybe datasets created with `filter` are not always handled properly ? cc @mariosasko", "It was after the second filter (and combining the two filters into one seemingly resolved it). I obviously haven't tried all settings to know that these details are causal, but it did work for me.", "Thanks, that's good to know.\r\n\r\nThe stacktrace suggests an issue when `del self._indices` is called, which happens when a filtered dataset falls out of scope. The indices are a PyArrow table memory mapped from disk, so I'm not quite sure how calling `del` on it can cause this issue. We do `del self._indices` to make sure the file on disk is not used anymore by the current process and avoid e.g. permission errors.\r\n\r\nHopefully we can find a way to reproduce this error, otherwise it will be quite hard to understand what happened", "Yeah, I have a reliable repro, but it is not even close to minimal and uses a dataset I can't share. Perhaps you could try getting close to my setting.\r\n\r\n(1) make a large (~20GB) jsonl with prompt/response pairs\r\n(2) load it on a linux machine (`dataset = load_dataset(...)`)\r\n(3) map a tokenizer to it, with multiprocessing (`tokenized_dataset = dataset.map(...)`)\r\n(4) filter it once based on something, with multiprocessing (`filtered_1 = tokenized_dataset.filter(...)`)\r\n(5) filter it again based on something, with multiprocessing (`filtered_2 = filtered_1.filter(...)`)\r\n\r\nI included the variable names just in case it is relevant that I was creating new datasets each time, not overwriting the same variable.", "@lhoestq I have another version of the repro that seems fairly reliably. I have lots of jsonl files, and I iteratively load each one with `load_dataset('json', data_files='path/to/my/file.jsonl', streaming=False, split='train')` and then `dataset.map(..., num_proc=<int>)`. This iteration hangs in a random place each time. So seems like there is a bug that hits with _some_ frequency.", "With `num_proc=None` it works fine.", "I am also having similar issue to #3172 when trying to tokenize the data. My dataset contains 10M samples. Is there anything that could be done without having to split up the processing into multiple datasets?" ]

[]

[]

[ "Hi! Can you make the above reproducer self-contained by adding code that generates the data?", "I managed a workaround eventually but I don't know what it was (I made a lot of changes to seq2seq). I'll try to include generating code in the future. (If I close, I don't know if you see it. Feel free to close; I'll re-open if I encounter it again (if I can))." ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" }, { "id": 2067376369, "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request", "name": "dataset request", "color": "e99695", "default": false, "description": "Requesting to add a new dataset" } ]

[]

[ "Hey @SauravMaheshkar ! Just responded to your email.\r\n\r\n_For transparency, copying part of my response here:_\r\nI agree, it would be really great to have this and other BenchMD datasets easily accessible on the hub.\r\n\r\nI think the main limiting factor is that the ChexPert dataset is currently hosted on the Stanford AIMI Shared Datasets website, with a license that does not permit redistribution IIRC. Thus, I believe we would need to create a [dataset loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script) that would check authentication with the Stanford AIMI site before downloading and extracting the data. \r\n\r\nI've started a HF dataset repo [here](https://huggingface.co/datasets/katielink/CheXpert), in case you want to collaborate on writing up this loading script! I'm also happy to take a stab when I have some more time next week.", "Hey @katielink I would love to try this out. Please guide me.", "Hi @katielink , I would also love to be on board and contribute to this loading script/project if it is still being developed. I'm interested because I personally would like to gain access to the CheXpert dataset and am facing some weird issues, so I'd like to sort it out for me, and potentially others. Please keep me updated and guide me on this as well!!!" ]

[]

[]

[]

[]

[]

[ "Most likely, the data file inference logic is the problem here.\r\n\r\nYou can run the following code to verify this:\r\n```python\r\nimport time\r\nfrom datasets.data_files import get_data_patterns\r\nstart_time = time.time()\r\nget_data_patterns(\"/path/to/img_dir\")\r\nend_time = time.time()\r\nprint(f\"Elapsed time: {end_time - start_time:.2f}s\")\r\n```\r\n \r\nWe plan to optimize this for the next version (or version after that). In the meantime, specifying the split patterns manually should give better performance:\r\n```python\r\nds = load_dataset(\"imagefolder\", data_files={\"train\": \"path/to/img_dir/train/**\", ...}, split=\"train\")\r\n```", "Hi, @mariosasko, you are right; data file inference logic is extremely slow.\r\n\r\nI have done a similar test, that is I modify the source code of datasets/load.py to measure the cost of two suspicious operations:\r\n```python\r\ndef get_module(self) -> DatasetModule:\r\n base_path = Path(self.data_dir or \"\").expanduser().resolve().as_posix()\r\n start = time.time()\r\n patterns = sanitize_patterns(self.data_files) if self.data_files is not None else get_data_patterns(base_path)\r\n print(f\"patterns: {time.time() - start}\")\r\n start = time.time()\r\n data_files = DataFilesDict.from_patterns(\r\n patterns,\r\n download_config=self.download_config,\r\n base_path=base_path,\r\n )\r\n print(f\"data_files: {time.time() - start}\")\r\n```\r\nIt gaves:\r\npatterns: 3062.2050700187683\r\ndata_files: 413.9576675891876\r\n\r\nThus, these two operations contribute to almost all of load time. What's going on in them?", "Furthermore, what's my current workaround about this problem? Should I save it by `save_to_disk()` and load dataset through `load_from_disk`?", "were you able to solve this issue?, I am facing the same issue" ]