Datasets:

---

Francesco-A

/

github-issues_huggingface-datasets

like 0

[]

[]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

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

[]

[]

[ "I just tried on my side and I didn't encounter your problem.\r\nApparently the script doesn't generate all the examples on your side.\r\n\r\nCan you provide the version of `nlp` you're using ?\r\nCan you try to clear your cache and re-run the code ?", "I updated it, that was it, thanks!", "Hello, I am facing the same problem... how do you clear the huggingface cache?", "Hi ! The cache is at ~/.cache/huggingface\r\nYou can just delete this folder if needed :)" ]

[]

[ "Hi,\r\n\r\nthanks for the report, but this is a duplicate of #2052. ", "@mariosasko \r\nThank you for your quick response! Following #2052, I've fixed the problem." ]

[]

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

[]

[ "I have an error on windows:\r\n```python\r\naiohttp.client_exceptions.ClientConnectorCertificateError: Cannot connect to host moon-staging.huggingface.co:443 ssl:True [SSLCertVerificationError: (1, '[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: certificate has expired (_ssl.c:1131)')]\r\n```\r\nat the `fsspec` call in `xglob`:\r\n```python\r\nfs, *_ = fsspec.get_fs_token_paths(urlpath, storage_options=storage_options)\r\n```\r\n\r\nLooks like the windows CI has an SSL issue... ", "I can reproduce it on my windows machine. On linux it works fine though", "I'm just skipping the windows test for now", "The Windows CI failure seems unrelated to this PR\r\n```python\r\nERROR tests/test_arrow_dataset.py::test_dummy_dataset_serialize_s3\r\n```" ]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "The `test_dataset_cards` method is failing with the error:\r\n\r\n```\r\nif error_messages:\r\n> raise ValueError(\"\\n\".join(error_messages))\r\nE ValueError: The following issues have been found in the dataset cards:\r\nE YAML tags:\r\nE The following typing errors are found: {'annotations_creators': \"(Expected `typing.List` with length > 0. Found value of type: `<class 'list'>`, with length: 0.\\n)\\nOR\\n(Expected `typing.Dict` with length > 0. Found value of type: `<class 'list'>`, with length: 0.\\n)\"}\r\n```\r\n\r\nAdding `found` as annotation creators." ]

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

[ "This issue was original reported at https://github.com/huggingface/transformers/issues/14931 and It seems like this issue also occur with other AutoClass like AutoFeatureExtractor.", "Thanks for moving the issue here !\r\n\r\nI wasn't able to reproduce the issue on my env (the hashes stay the same):\r\n```\r\n- `transformers` version: 1.15.0\r\n- `tokenizers` version: 0.10.3\r\n- `datasets` version: 1.18.1\r\n- `dill` version: 0.3.4\r\n- Platform: Linux-4.19.0-18-cloud-amd64-x86_64-with-debian-10.11\r\n- Python version: 3.7.10\r\n- PyArrow version: 6.0.1\r\n```\r\nHowever I was able to reproduce it on Google Colab (the hashes end up different):\r\n```\r\n- `transformers` version: 1.15.0\r\n- `tokenizers` version: 0.10.3\r\n- `datasets` version: 1.18.1\r\n- `dill` version: 0.3.4\r\n- Platform: Linux-5.4.144+-x86_64-with-Ubuntu-18.04-bionic\r\n- Python version: 3.7.12\r\n- PyArrow version: 3.0.0\r\n```\r\nI'll investigate why it doesn't work properly on Google Colab :)", "I found the issue: the tokenizer has something inside it that changes.\r\n\r\nBefore the call, `tokenizer._tokenizer.truncation` is None, and after the call it changes to this for some reason:\r\n```\r\n{'max_length': 512, 'strategy': 'longest_first', 'stride': 0}\r\n```\r\n\r\nDoes anybody know why calling the tokenizer would change its state this way ? cc @Narsil @SaulLu maybe ?", "`tokenizer.encode(..)` does not accept argument like max_length, strategy or stride.\r\n\r\nIn `tokenizers` you have to modify the tokenizer state by setting various `TruncationParams` (and/or `PaddingParams`).\r\nHowever, since this is modifying the state, you need to mutably borrow the tokenizer (a rust concept). The key principle is that there can ever be only 1 mutable borrow at a time during the span of the tokenizer lifecycle.\r\n\r\nBecause of this, if `transformers` blindly set `TruncationParams` and `PaddingParams` on every call, it would cause the tokenizer to crash (or make the various threads accessing it hang, which is not necessarily better).\r\n\r\nIn order to avoid that, we decided to handle it this way : https://github.com/huggingface/transformers/pull/12550 . \r\n\r\nWhich should explain the state of the tokenizer being modified (hence its hash).\r\n\r\nNow for a temporary solution, simply encoding once with the tokenizer should give it it's proper hash (since by default the tokenizer doesn't have this state, looks at the first encoding call, and creates it).\r\n\r\nWe could try and set these 2 dicts at initialization time, but it wouldn't work if a user modified the tokenizer state later\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(..)\r\ntokenizer.truncation_side = \"left\"\r\n# Now we have a difference between `tokenizer._tokenizer.truncation` and `tokenizer.truncation_side`\r\n```\r\nIf we wanted to fix it correctly it would mean mapping every assignation to it's proper location on `tokenizer.{padding/truncation}`\r\n\r\nI think it's important to note that we cannot guarantee a tokenizer' hash remains the same if *any* of those parameters are modified through the `.map` function.\r\n\r\nEdit: Another option would be to override the default __hash__ function, but I don't know if there's a sound implementation that could fit.", "Thanks a lot for the explanation !\r\nI think if we set these 2 dicts at initialization time it would be amazing already\r\n\r\nShall we open an issue in `transformers` to ask for these dictionaries to be set when the tokenizer is instantiated ?\r\n\r\n> Edit: Another option would be to override the default hash function, but I don't know if there's a sound implementation that could fit.\r\n\r\nIn `datasets` we can easily have custom hashing for objects of the other HF libraries if we want. For example we ignore the cache some tokenizers have. However in this specific case it touches parameters that may change the behavior of the tokenizer itself. I'm not sure the logic that determines how a tokenizer behaves should be in `datasets`", "A hack we could have in the `datasets` lib would be to call the tokenizer before hashing it in order to set all its parameters correctly - but it sounds a lot like a hack and I'm not sure this can work in the long run", "Fully agree with everything you said. \r\n\r\nI think the best course of action is creating an issue in `transformers`. I can start the work on this.\r\nI think the code changes are fairly simple. Making a sound test + not breaking other stuff might be different :D", "It should be noted that this problem also occurs in other AutoClasses, such as AutoFeatureExtractor, so I don't think handling it in Datasets is a long-term practice either.", "> I think the best course of action is creating an issue in `transformers`. I can start the work on this.\r\n\r\n@Narsil Hi, I reopen this issue in `transformers` https://github.com/huggingface/transformers/issues/14931", "Here is @Narsil comment from https://github.com/huggingface/transformers/issues/14931#issuecomment-1074981569\r\n> # TL;DR\r\n> Call the function once on a dummy example beforehand will fix it.\r\n> \r\n> ```python\r\n> tokenizer(\"Some\", \"test\", truncation=True)\r\n> ```\r\n> \r\n> # Long answer\r\n> If I remember the last status, it's hard doing anything, since the call itself\r\n> \r\n> ```python\r\n> tokenizer(example[\"sentence1\"], example[\"sentence2\"], truncation=True)\r\n> ```\r\n> \r\n> will modify the tokenizer. It's the `truncation=True` that modifies the tokenizer to put it into truncation mode if you will. Calling the tokenizer once with that argument would fix the cache.\r\n> \r\n> Finding a fix that :\r\n> \r\n> * Doesn't imply a huge chunk of work on `tokenizers` (with potential loss of performance, and breaking backward compatibility)\r\n> * Doesn't imply `datasets` running a first pass of the loop\r\n> * Doesn't imply `datasets` looking at the map function itself\r\n> * Uses a sound `hash` for this object in `datasets`.\r\n> \r\n> is IIRC impossible for this use case.\r\n> \r\n> I can explain a bit more why the first option is not desirable.\r\n> \r\n> In order to \"fix\" this for tokenizers, we would need to make `tokenizer(..)` purely without side effects. This means that the \"options\" of tokenization (like `truncation` and `padding` at least) would have\r\n", "For me this workaround only works if I don't pass the `num_proc=X` argument to `datasets.map`" ]

[]

[ "Thanks ! Good catch :)\r\n\r\nTo fix the CI you can do:\r\n1 - rebase from master\r\n2 - then run `make style` as specified in [CONTRIBUTING.md](https://github.com/huggingface/nlp/blob/master/CONTRIBUTING.md) ?" ]

[]

[ "This has been addressed in https://github.com/huggingface/datasets/pull/4415. Closing." ]

[]

[]

[]

[]

[]

[ "Sure, feel free to open a PR." ]

[]

[]

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

[]

[]

[]

[]

[]

[]

[ "Just found out about https://github.com/huggingface/datasets/blob/master/datasets/kilt_tasks/README.md\r\n(Not very clear in https://huggingface.co/datasets/kilt_tasks links to http://github.com/huggingface/datasets/datasets/kilt_tasks/README.md which is dead, closing the issue though :))" ]

[]

[ "I tried this with `ade_corpus_v2` dataset. `ade_corpus_v2-train.arrow` (downloaded dataset) and `cache-25d41a4d3c2d8a25.arrow` (ran a mapping function on the dataset) both had file permission with octal value of `0644`. " ]

[]

[ "Looks good to me, we can ping the authors for more info later. And yes apply `other-task` labels liberally, we can sort them out later :) \r\n\r\nLooks ready to merge when you're ready @joeddav " ]

[]

[ "Hi @ggdupont !\r\nHave you had a chance to take a look at my suggestions ?\r\nFeel free to ping me if you have questions or when you're ready for a review", "> Hi @ggdupont !\r\n> Have you had a chance to take a look at my suggestions ?\r\n> Feel free to ping me if you have questions or when you're ready for a review\r\n\r\nYes I did, just busy period (and no time on weekend right now ;-) )", "With some delay, reduced the dummy data and had t rebase", "Thanks !\r\n\r\nIt looks like the rebase messed up the github diff for this PR (2.000+ files changed)\r\nCould you create another branch and another PR please ?", "Cleaned PR: https://github.com/huggingface/datasets/pull/1850" ]

[]

[ "Hi !\r\nCan you share more information about the features of your dataset ? You can get them by printing `my_dataset.features`\r\nCan you also share the code of your `map` function ?", "It is actually just the tokenized `wikipedia` dataset with `input_ids`, `attention_mask`, etc, with one extra column which is a list of integers. The `text` column is removed during tokenization.\r\n\r\n```\r\ndef add_len_and_seq(example):\r\n end_idx = example['input_ids'].index(SEP)\r\n example['actual_len'] = end_idx-1\r\n seq_len = len(example['input_ids'])\r\n \r\n\r\n example['seq'] = [PAD_ID] + [np.uint8(example['some_integer'])]*(end_idx-1) + [PAD_ID]*(seq_len-end_idx)\r\n \r\n return example\r\n```\r\n", "Is `PAD_ID` a python integer ? You need all the integers in `example['seq']` to have the same type.\r\nDoes this work if you remove the `np.uint8` and use python integers instead ?", "yup I casted it to `np.uint8` outside the function where it was defined. It was originally using python integers.", "Strangely, even when I manually created `np.arrays` of specific `dtypes`, the types in the final `dataset_info.json` that gets written are still `int64`.\r\n\r\nUpdate: I tried creating lists of `int8`s and got the same result.", "Yes this is a known issue: #625 \r\nWe're working on making the precision kept for numpy :)\r\nTo specify the precision of the integers, currently one needs to specify the output features with `.map(..., features=output_features)`", "Do you know what step is taking forever in the code ?\r\nWhat happens if you interrupt the execution of the dataset loading ?", "After a synchronous discussion, we found that the cache file sizes have an enormous effect on the loading speed: smaller cache files result in faster load times. `num_proc` controls the number of cache files that are being written and is inversely proportional to the individual file size. In other words, increase `num_proc` for smaller cache files :)\r\n\r\nMaybe this can be highlighted somewhere in the docs." ]

[]

[ "Thanks for reporting, @cheaterHy.\r\n\r\nThe cause of this issue is a misalignment between the names of the repo (`goes-mrms`, with hyphen) and its Python loading scrip file (`goes_mrms.py`, with underscore).\r\n\r\nI've opened an Issue discussion in their repo: https://huggingface.co/datasets/openclimatefix/goes-mrms/discussions/1" ]

[]

[ "we're using quail 1.3 now thanks.\r\nclosing this one" ]

[]

[ "looks like your PR includes changes about many other files than the ones for menyo 20k mt\r\nCan you create another branch and another PR please ?", "Yes, I will" ]

[]

[ "Could you provide the error details. It is difficult to debug otherwise. Also try other config. `ha` is not a valid." ]

[]

[ "It's strange, after installing `nlp` on Colab, the `pyarrow` version seems fine from `pip` but not from python :\r\n\r\n```python\r\nimport pyarrow\r\n\r\n!pip show pyarrow\r\nprint(\"version = {}\".format(pyarrow.__version__))\r\n```\r\n\r\n> Name: pyarrow\r\nVersion: 0.17.0\r\nSummary: Python library for Apache Arrow\r\nHome-page: https://arrow.apache.org/\r\nAuthor: None\r\nAuthor-email: None\r\nLicense: Apache License, Version 2.0\r\nLocation: /usr/local/lib/python3.6/dist-packages\r\nRequires: numpy\r\nRequired-by: nlp, feather-format\r\n> \r\n> version = 0.14.1", "Ok I just had to restart the runtime after installing `nlp`. After restarting, the version of `pyarrow` is fine." ]

[]

[]

[]

[ "Hi ! PyTorch may hang when calling `load_state_dict()` in a subprocess. To fix that, set the multiprocessing start method to \"spawn\". Since `datasets` uses `multiprocess`, you should do:\r\n\r\n```python\r\n# Required to avoid issues with pytorch (otherwise hangs during load_state_dict in multiprocessing)\r\nimport multiprocess.context as ctx\r\nctx._force_start_method('spawn')\r\n```\r\n\r\nAlso make sure to run your main code in `if __name__ == \"__main__\":` to avoid issues with python multiprocesing", "Thanks!", "@lhoestq Hello, I also encountered this problem but maybe with another reason. Here is my code:\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir, model_max_length=training_args.model_max_length)\r\ndata = load_dataset(\"json\", data_files=data_args.train_file, cache_dir=data_args.data_cache_dir)\r\ndef func(samples):\r\n # main operation\r\n for sentence_value in samples:\r\n sentence_ids = tokenizer.encode(sentence_value, add_special_tokens=False, max_length=tokenizer.model_max_length, truncation=True)\r\n ... ...\r\ntrain_data = data[\"train\"].shuffle().map(func, num_proc=os.cpu_count())\r\n```\r\nIt hangs after the progress reaches 100%. Could you help me point out the reason?", "@SkyAndCloud your issue doesn't seem related to the original post - could you open a new issue and provide more details ? (size of the dataset, number of cpus, how much time it took to run, `datasets` version)", "@lhoestq Hi, I just solved this problem. Because the input is extremely long and the tokenizer requests a large amount of memory, which leads to a OOM error and may eventually causes the hang. I didn't filter those too-long sentences because I thought `tokenizer` would stop once the length exceeds the `max_length`. However, it actually firstly complete the tokenization of entire sentence and then truncate it." ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "Hum there seems to be some issues with tasks in test:\r\n - some tasks don't fit anything in `tasks.json`. Do I remove them in `task_categories`?\r\n - some tasks should exist, typically `visual-question-answering` (https://github.com/huggingface/datasets/blame/9f2ff14673cac1f1ad56d80221a793f5938b68c7/src/datasets/utils/resources/tasks.json#L195) yet the exception is failing on me. I'm guessing it's because my `master` is not up-to-date. However this means that the testing only tests my branch instead of the one merged with master?\r\n \r\n cc @mariosasko @lhoestq ", "> some tasks don't fit anything in tasks.json. Do I remove them in task_categories?\r\n\r\nYou can keep them, but add `other-` as a prefix to those tasks to make the CI ignore it\r\n\r\n> some tasks should exist, typically visual-question-answering (https://github.com/huggingface/datasets/blame/9f2ff14673cac1f1ad56d80221a793f5938b68c7/src/datasets/utils/resources/tasks.json#L195) yet the exception is failing on me. I'm guessing it's because my master is not up-to-date. However this means that the testing only tests my branch instead of the one merged with master?\r\n\r\nFeel free to merge upstream/master into your branch ;)\r\n\r\nEDIT: actually I just noticed you've already done this, thanks !", "After offline discussions: will keep that image essentially it's necessary as I have a mapping that creates a mapping between url and local path (images are downloaded via a zip file) and dummy data needs to store that dummy image. The issue is when I read an annotation, I get a url, compute the local path, and basically I assume the local path exists since I've extracted all the images ... This isn't true if dummy data doesn't have all the images, so instead I've added a script that \"fixes\" the dummy data after using the CLI, it essentially adds the dummy image in the zip corresponding to the url." ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

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

[]

[]

[ "Thanks for reporting, I opened a PR to fix that", "One the PR is merged the fix will be available in the next release of `datasets`.\r\n\r\nIf you don't want to wait the next release you can still load the script from the master branch with\r\n\r\n```python\r\nload_dataset(\"dane\", script_version=\"master\")\r\n```", "If you have other questions feel free to reopen :) " ]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "OK, thanks @lhoestq. I'll close this PR, and come back to it with @stevhliu once we work on https://github.com/huggingface/datasets-server/issues/568" ]

[]

[]

[]

[]

[]

[ "Hi ! Can you try adding the `s3://` prefix ?\r\n```python\r\nf\"s3://{s3_bucket}/{s3_dir}/{dataset_name}\"\r\n```", "Ugh, yeah that was it. Thank you!" ]

[]

[]

[ { "color": "fef2c0", "default": false, "description": "", "id": 3287858981, "name": "streaming", "node_id": "MDU6TGFiZWwzMjg3ODU4OTgx", "url": "https://api.github.com/repos/huggingface/datasets/labels/streaming" } ]

[ "The dataset cannot be loaded, be it in normal or streaming mode.\r\n\r\n```python\r\n>>> import datasets\r\n>>> dataset=datasets.load_dataset(\"hans\", split=\"train\", streaming=True)\r\n>>> next(iter(dataset))\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 497, in __iter__\r\n for key, example in self._iter():\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 494, in _iter\r\n yield from ex_iterable\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 87, in __iter__\r\n yield from self.generate_examples_fn(**self.kwargs)\r\n File \"/home/slesage/.cache/huggingface/modules/datasets_modules/datasets/hans/1bbcb735c482acd54f2e118074b59cfd2bf5f7a5a285d4d540d1e632216672ac/hans.py\", line 121, in _generate_examples\r\n for idx, line in enumerate(open(filepath, \"rb\")):\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/spec.py\", line 1595, in __next__\r\n out = self.readline()\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/spec.py\", line 1592, in readline\r\n return self.readuntil(b\"\\n\")\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/spec.py\", line 1581, in readuntil\r\n self.seek(start + found + len(char))\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 676, in seek\r\n raise ValueError(\"Cannot seek streaming HTTP file\")\r\nValueError: Cannot seek streaming HTTP file\r\n>>> dataset=datasets.load_dataset(\"hans\", split=\"train\", streaming=False)\r\nDownloading and preparing dataset hans/plain_text (download: 29.51 MiB, generated: 30.34 MiB, post-processed: Unknown size, total: 59.85 MiB) to /home/slesage/.cache/huggingface/datasets/hans/plain_text/1.0.0/1bbcb735c482acd54f2e118074b59cfd2bf5f7a5a285d4d540d1e632216672ac...\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/load.py\", line 1687, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 605, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1104, in _download_and_prepare\r\n super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 694, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1087, in _prepare_split\r\n for key, record in logging.tqdm(\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/tqdm/std.py\", line 1180, in __iter__\r\n for obj in iterable:\r\n File \"/home/slesage/.cache/huggingface/modules/datasets_modules/datasets/hans/1bbcb735c482acd54f2e118074b59cfd2bf5f7a5a285d4d540d1e632216672ac/hans.py\", line 121, in _generate_examples\r\n for idx, line in enumerate(open(filepath, \"rb\")):\r\nValueError: readline of closed file\r\n```\r\n\r\n", "Hi! I've opened a PR that should make this dataset stremable. You can test it as follows:\r\n```python\r\nfrom datasets import load_dataset\r\ndset = load_dataset(\"hans\", split=\"train\", streaming=True, revision=\"49decd29839c792ecc24ac88f861cbdec30c1c40\")\r\n```\r\n\r\n@severo The current script doesn't throw an error in normal mode (only in streaming mode) on my local machine or in Colab. Can you update your installation of `datasets` and see if that fixes the issue?", "Thanks for this. It works well, thanks! The dataset viewer is using https://github.com/huggingface/datasets/releases/tag/2.0.0, I'm eager to upgrade to 2.0.1 😉" ]

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

[ "Hi @RylanSchaeffer, thanks for reporting.\r\n\r\nI'm sorry, but I was not able to reproduce your problem.\r\n\r\nNormally, the reason for this type of error is that, during your download of the data files, this was not fully complete.\r\n\r\nCould you please try to load the dataset again but forcing its redownload? Please use:\r\n```python\r\ndataset = load_dataset(\"openwebtext\", download_mode=\"FORCE_REDOWNLOAD\")\r\n```\r\n\r\nLet me know if the problem persists.", "I close this issue for the moment. Feel free to re-open it again if the problem persists." ]

[]

[ "Thank you :)" ]

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

[]

[]

[ "This is due to the slow resolution of the data files: https://github.com/huggingface/datasets/issues/5537.\r\n\r\nWe plan to switch to `huggingface_hub`'s `HfFileSystem` soon to make the resolution faster (will be up to 20x faster once we merge https://github.com/huggingface/huggingface_hub/pull/1443)\r\n\r\n", "You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.", "> You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.\r\n\r\nThat's unrelated to the problem discussed in this issue. ", "> > You're right, when I try to parse more than 50GB of text data, I also get very slow, usually taking hours or even tens of hours.\r\n> \r\n> That's unrelated to the problem discussed in this issue.\r\n\r\nSorry, I misunderstood it." ]

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

[ "We raise the error “ FileNotFoundError: can’t find the dataset” mainly to follow best practice in security (otherwise users could be able to guess what private repositories users/orgs may have)\r\n\r\nWe can indeed reformulate this and add the \"If this is a private repository,...\" part !", "Resolved via https://github.com/huggingface/datasets/pull/4536" ]

[]

[]

[ { "color": "0e8a16", "default": false, "description": "Contribution to a dataset script", "id": 4564477500, "name": "dataset contribution", "node_id": "LA_kwDODunzps8AAAABEBBmPA", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20contribution" } ]

[ "_The documentation is not available anymore as the PR was closed or merged._", "We are closing PRs to dataset scripts because we are moving them to the Hub.\r\n\r\nThanks anyway.\r\n\r\n" ]

[]

[]

[]

[ "This can also mean one of the files was not downloaded correctly.\r\n\r\nWe log an erroneous file's name before raising the reader's error, so this is how you can find the problematic file. Then, you should delete it and call `load_dataset` again.\r\n\r\n(I checked all the uploaded files, and they seem to be valid Parquet files, so I don't think this is a bug on their side)\r\n" ]

[]

[]

[]

[ "This seems to be fixed in a non-released version. \r\n\r\nInstalling nlp from source\r\n```\r\ngit clone https://github.com/huggingface/nlp\r\ncd nlp\r\npip install .\r\n```\r\nsolves the issue. " ]

[ { "color": "2edb81", "default": false, "description": "A bug in a dataset script provided in the library", "id": 2067388877, "name": "dataset bug", "node_id": "MDU6TGFiZWwyMDY3Mzg4ODc3", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20bug" } ]

[ "The zest dataset url was changed (allenai/zest#3) and #2057 should resolve this.", "This has been fixed in #2057 by @matt-peters (thanks again !)\r\n\r\nThe fix is available on the master branch and we'll do a new release very soon :)", "Thanks @lhoestq and @matt-peters ", "I am closing this issue since its fixed!" ]

[]

[ "That's great!\r\n\r\nI think it would be nice to test \"deterministic-ness\" of datasets in CI if we can do it (should be left for future PR of course)\r\n\r\nHere is a possibility (maybe there are other ways to do it): given that we may soon have efficient and large-scale hashing (cf our discussion on versioning/tracability), we could incorporate a hash of the final Arrow Dataset to the `dataset.json` file and have a test on it as well as CI on a diversity of platform to test the hash (Win/Mac/Linux + various python/env).\r\nWhat do you think @lhoestq @patrickvonplaten?", "> That's great!\r\n> \r\n> I think it would be nice to test \"deterministic-ness\" of datasets in CI if we can do it (should be left for future PR of course)\r\n> \r\n> Here is a possibility (maybe there are other ways to do it): given that we may soon have efficient and large-scale hashing (cf our discussion on versioning/tracability), we could incorporate a hash of the final Arrow Dataset to the `dataset.json` file and have a test on it as well as CI on a diversity of platform to test the hash (Win/Mac/Linux + various python/env).\r\n> What do you think @lhoestq @patrickvonplaten?\r\n\r\nI think that's a great idea! The test should be a `RUN_SLOW` test, since it takes a considerable amount of time to download the dataset and generate the examples, but I think we should add some kind of hash test for each dataset.", "Really nice!!" ]

[]

[ "Hi !\r\n\r\nThe cache at `~/.cache/huggingface/metrics` stores the users data for metrics computations (hence the arrow files).\r\n\r\nHowever python modules (i.e. dataset scripts, metric scripts) are stored in `~/.cache/huggingface/modules/datasets_modules`.\r\n\r\nIn particular the metrics are cached in `~/.cache/huggingface/modules/datasets_modules/metrics/`\r\n\r\nFeel free to take a look at your cache and let me know if you find any issue that would help explaining why you had an issue with `rouge` with no connection. I'm doing some tests on my side to try to reproduce the issue you have\r\n", "Thank you for clarifying that the metrics files are to be found elsewhere, @lhoestq \r\n\r\n> The cache at ~/.cache/huggingface/metrics stores the users data for metrics computations (hence the arrow files).\r\n\r\ncould it be renamed to reflect that? otherwise it misleadingly suggests that it's the metrics. Perhaps `~/.cache/huggingface/metrics-user-data`?\r\n\r\nAnd there are so many `.lock` files w/o corresponding files under `~/.cache/huggingface/metrics/`. Why are they there? \r\n\r\nfor example after I wipe out the dir completely and do one training I end up with:\r\n```\r\n~/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.lock\r\n```\r\nwhat is that lock file locking when nothing is running?", "The lock files come from an issue with filelock (see comment in the code [here](https://github.com/benediktschmitt/py-filelock/blob/master/filelock.py#L394-L398)). Basically on unix there're always .lock files left behind. I haven't dove into this issue", "are you sure you need an external lock file? if it's a single purpose locking in the same scope you can lock the caller `__file__` instead, e.g. here is how one can `flock` the script file itself to ensure atomic printing:\r\n\r\n```\r\nimport fcntl\r\ndef printflock(*msgs):\r\n \"\"\" print in multiprocess env so that the outputs from different processes don't get interleaved \"\"\"\r\n with open(__file__, \"r\") as fh:\r\n fcntl.flock(fh, fcntl.LOCK_EX)\r\n try:\r\n print(*msgs)\r\n finally:\r\n fcntl.flock(fh, fcntl.LOCK_UN)\r\n```\r\n", "OK, this issue is not about caching but some internal conflict/race condition it seems, I have just run into it on my normal env:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 356, in _finalize\r\n self.data = Dataset(**reader.read_files([{\"filename\": f} for f in file_paths]))\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 236, in read_files\r\n pa_table = self._read_files(files, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 171, in _read_files\r\n pa_table: pa.Table = self._get_dataset_from_filename(f_dict, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 302, in _get_dataset_from_filename\r\n pa_table = ArrowReader.read_table(filename, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 322, in read_table\r\n stream = stream_from(filename)\r\n File \"pyarrow/io.pxi\", line 782, in pyarrow.lib.memory_map\r\n File \"pyarrow/io.pxi\", line 743, in pyarrow.lib.MemoryMappedFile._open\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 97, in pyarrow.lib.check_status\r\nFileNotFoundError: [Errno 2] Failed to open local file '/home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow'. Detail: [errno 2] No such file or directory\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"examples/seq2seq/run_seq2seq.py\", line 655, in <module>\r\n main()\r\n File \"examples/seq2seq/run_seq2seq.py\", line 619, in main\r\n test_results = trainer.predict(\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer_seq2seq.py\", line 121, in predict\r\n return super().predict(test_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer.py\", line 1706, in predict\r\n output = self.prediction_loop(\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer.py\", line 1813, in prediction_loop\r\n metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))\r\n File \"examples/seq2seq/run_seq2seq.py\", line 556, in compute_metrics\r\n result = metric.compute(predictions=decoded_preds, references=decoded_labels)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 388, in compute\r\n self._finalize()\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 358, in _finalize\r\n raise ValueError(\r\nValueError: Error in finalize: another metric instance is already using the local cache file. Please specify an experiment_id to avoid colision between distributed metric instances.\r\n```\r\n\r\nI'm just running `run_seq2seq.py` under DeepSpeed:\r\n\r\n```\r\nexport BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0,1 deepspeed --num_gpus=2 examples/seq2seq/run_seq2seq.py --model_name_or_path t5-small --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --do_predict --evaluation_strategy=steps --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --val_max_target_length 128 --warmup_steps 500 --max_train_samples 100 --max_val_samples 100 --max_test_samples 100 --dataset_name wmt16 --dataset_config ro-en --source_prefix \"translate English to Romanian: \" --deepspeed examples/tests/deepspeed/ds_config.json\r\n```\r\n\r\nIt finished the evaluation OK and crashed on the prediction part of the Trainer. But the eval / predict parts no longer run under Deepspeed, it's just plain ddp.\r\n\r\nIs this some kind of race condition? It happens intermittently - there is nothing else running at the same time.\r\n\r\nBut if 2 independent instances of the same script were to run at the same time it's clear to see that this problem would happen. Perhaps it'd help to create a unique hash which is shared between all processes in the group and use that as the default experiment id?\r\n", "When you're using metrics in a distributed setup, there are two cases:\r\n1. you're doing two completely different experiments (two evaluations) and the 2 metrics jobs have nothing to do with each other\r\n2. you're doing one experiment (one evaluation) but use multiple processes to feed the data to the metric.\r\n\r\nIn case 1. you just need to provide two different `experiment_id` so that the metrics don't collide.\r\nIn case 2. they must have the same experiment_id (or use the default one), but in this case you also need to provide the `num_processes` and `process_id`\r\n\r\nIf understand correctly you're in situation 2.\r\n\r\nIf so, you make sure that you instantiate the metrics with both the right `num_processes` and `process_id` parameters ?\r\n\r\nIf they're not set, then the cache files of the two metrics collide it can cause issues. For example if one metric finishes before the other, then the cache file is deleted and the other metric gets a FileNotFoundError\r\nThere's more information in the [documentation](https://huggingface.co/docs/datasets/loading_metrics.html#distributed-setups) if you want\r\n\r\nHope that helps !", "Thank you for explaining that in a great way, @lhoestq \r\n\r\nSo the bottom line is that the `transformers` examples are broken since they don't do any of that. At least `run_seq2seq.py` just does `metric = load_metric(metric_name)`\r\n\r\nWhat test would you recommend to reliably reproduce this bug in `examples/seq2seq/run_seq2seq.py`?", "To give more context, we are just using the metrics for the `comput_metric` function and nothing else. Is there something else we can use that just applies the function to the full arrays of predictions and labels? Because that's all we need, all the gathering has already been done because the datasets Metric multiprocessing relies on file storage and thus does not work in a multi-node distributed setup (whereas the Trainer does).\r\n\r\nOtherwise, we'll have to switch to something else to compute the metrics :-(", "OK, it definitely leads to a race condition in how it's used right now. Here is how you can reproduce it - by injecting a random sleep time different for each process before the locks are acquired. \r\n```\r\n--- a/src/datasets/metric.py\r\n+++ b/src/datasets/metric.py\r\n@@ -348,6 +348,16 @@ class Metric(MetricInfoMixin):\r\n\r\n elif self.process_id == 0:\r\n # Let's acquire a lock on each node files to be sure they are finished writing\r\n+\r\n+ import time\r\n+ import random\r\n+ import os\r\n+ pid = os.getpid()\r\n+ random.seed(pid)\r\n+ secs = random.randint(1, 15)\r\n+ time.sleep(secs)\r\n+ print(f\"sleeping {secs}\")\r\n+\r\n file_paths, filelocks = self._get_all_cache_files()\r\n\r\n # Read the predictions and references\r\n@@ -385,7 +395,10 @@ class Metric(MetricInfoMixin):\r\n\r\n if predictions is not None:\r\n self.add_batch(predictions=predictions, references=references)\r\n+ print(\"FINALIZE START\")\r\n+\r\n self._finalize()\r\n+ print(\"FINALIZE END\")\r\n\r\n self.cache_file_name = None\r\n self.filelock = None\r\n```\r\n\r\nthen run with 2 procs: `python -m torch.distributed.launch --nproc_per_node=2`\r\n```\r\nexport BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 examples/seq2seq/run_seq2seq.py --model_name_or_path t5-small --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --do_predict --evaluation_strategy=steps --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --val_max_target_length 128 --warmup_steps 500 --max_train_samples 10 --max_val_samples 10 --max_test_samples 10 --dataset_name wmt16 --dataset_config ro-en --source_prefix \"translate English to Romanian: \"\r\n```\r\n\r\n```\r\n***** Running Evaluation *****\r\n Num examples = 10\r\n Batch size = 16\r\n 0%| | 0/1 [00:00<?, ?it/s]FINALIZE START\r\nFINALIZE START\r\nsleeping 11\r\nFINALIZE END\r\n100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.06s/it]\r\nsleeping 11\r\nTraceback (most recent call last):\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 368, in _finalize\r\n self.data = Dataset(**reader.read_files([{\"filename\": f} for f in file_paths]))\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 236, in read_files\r\n pa_table = self._read_files(files, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 171, in _read_files\r\n pa_table: pa.Table = self._get_dataset_from_filename(f_dict, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 302, in _get_dataset_from_filename\r\n pa_table = ArrowReader.read_table(filename, in_memory=in_memory)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/arrow_reader.py\", line 322, in read_table\r\n stream = stream_from(filename)\r\n File \"pyarrow/io.pxi\", line 782, in pyarrow.lib.memory_map\r\n File \"pyarrow/io.pxi\", line 743, in pyarrow.lib.MemoryMappedFile._open\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 97, in pyarrow.lib.check_status\r\nFileNotFoundError: [Errno 2] Failed to open local file '/home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow'. Detail: [errno 2] No such file or directory\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"examples/seq2seq/run_seq2seq.py\", line 645, in <module>\r\n main()\r\n File \"examples/seq2seq/run_seq2seq.py\", line 601, in main\r\n metrics = trainer.evaluate(\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer_seq2seq.py\", line 74, in evaluate\r\n return super().evaluate(eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer.py\", line 1703, in evaluate\r\n output = self.prediction_loop(\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer.py\", line 1876, in prediction_loop\r\n metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))\r\n File \"examples/seq2seq/run_seq2seq.py\", line 556, in compute_metrics\r\n result = metric.compute(predictions=decoded_preds, references=decoded_labels)\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 402, in compute\r\n self._finalize()\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 370, in _finalize\r\n raise ValueError(\r\nValueError: Error in finalize: another metric instance is already using the local cache file. Please specify an experiment_id to avoid colision between distributed metric instances.\r\n```", "I tried to adjust `run_seq2seq.py` and trainer to use the suggested dist env:\r\n```\r\n import torch.distributed as dist\r\n metric = load_metric(metric_name, num_process=dist.get_world_size(), process_id=dist.get_rank())\r\n```\r\nand in `trainer.py` added to call just for rank 0:\r\n```\r\n if self.is_world_process_zero() and self.compute_metrics is not None and preds is not None and label_ids is not None:\r\n metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))\r\n```\r\nand then the process hangs in a deadlock. \r\n\r\nHere is the tb:\r\n```\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/utils/filelock.py\", line 275 in acquire\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 306 in _check_all_processes_locks\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 501 in _init_writer\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 440 in add_batch\r\n File \"/mnt/nvme1/code/huggingface/datasets-master/src/datasets/metric.py\", line 397 in compute\r\n File \"examples/seq2seq/run_seq2seq.py\", line 558 in compute_metrics\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer.py\", line 1876 in prediction_loop\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer.py\", line 1703 in evaluate\r\n File \"/mnt/nvme1/code/huggingface/transformers-mp-pp/src/transformers/trainer_seq2seq.py\", line 74 in evaluate\r\n File \"examples/seq2seq/run_seq2seq.py\", line 603 in main\r\n File \"examples/seq2seq/run_seq2seq.py\", line 651 in <module>\r\n```\r\n\r\nBut this sounds right, since in the above diff I set up a distributed metric and only called one process - so it's blocking on waiting for other processes to do the same.\r\n\r\nSo one working solution is to leave:\r\n\r\n```\r\n metric = load_metric(metric_name)\r\n```\r\nalone, and only call `compute_metrics` from rank 0\r\n```\r\n if self.is_world_process_zero() and self.compute_metrics is not None and preds is not None and label_ids is not None:\r\n metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))\r\n```\r\n\r\nso we now no longer use the distributed env as far as `datasets` is concerned, it's just a single process.\r\n\r\nAre there any repercussions/side-effects to this proposed change in Trainer? If it always gathers all inputs on rank 0 then this is how it should have been done in first place - i.e. only run for rank 0. It appears that currently it was re-calculating the metrics on all processes on the same data just to throw the results away other than for rank 0. Unless I missed something.\r\n", "But no, since \r\n`\r\n metric = load_metric(metric_name)\r\n`\r\nis called for each process, the race condition is still there. So still getting:\r\n\r\n```\r\nValueError: Error in finalize: another metric instance is already using the local cache file. Please specify an experiment_id to avoid colision between distributed metric instances.\r\n```\r\n\r\ni.e. the only way to fix this is to `load_metric` only for rank 0, but this requires huge changes in the code and all end users' code.\r\n", "OK, here is a workaround that works. The onus here is absolutely on the user:\r\n\r\n```\r\ndiff --git a/examples/seq2seq/run_seq2seq.py b/examples/seq2seq/run_seq2seq.py\r\nindex 2a060dac5..c82fd83ea 100755\r\n--- a/examples/seq2seq/run_seq2seq.py\r\n+++ b/examples/seq2seq/run_seq2seq.py\r\n@@ -520,7 +520,11 @@ def main():\r\n\r\n # Metric\r\n metric_name = \"rouge\" if data_args.task.startswith(\"summarization\") else \"sacrebleu\"\r\n- metric = load_metric(metric_name)\r\n+ import torch.distributed as dist\r\n+ if dist.is_initialized():\r\n+ metric = load_metric(metric_name, num_process=dist.get_world_size(), process_id=dist.get_rank())\r\n+ else:\r\n+ metric = load_metric(metric_name)\r\n\r\n def postprocess_text(preds, labels):\r\n preds = [pred.strip() for pred in preds]\r\n@@ -548,12 +552,17 @@ def main():\r\n # Some simple post-processing\r\n decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)\r\n\r\n+ kwargs = dict(predictions=decoded_preds, references=decoded_labels)\r\n+ if metric_name == \"rouge\":\r\n+ kwargs.update(use_stemmer=True)\r\n+ result = metric.compute(**kwargs) # must call for all processes\r\n+ if result is None: # only process with rank-0 will return metrics, others None\r\n+ return {}\r\n+\r\n if metric_name == \"rouge\":\r\n- result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)\r\n # Extract a few results from ROUGE\r\n result = {key: value.mid.fmeasure * 100 for key, value in result.items()}\r\n else:\r\n- result = metric.compute(predictions=decoded_preds, references=decoded_labels)\r\n result = {\"bleu\": result[\"score\"]}\r\n\r\n prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]\r\n```\r\n\r\nThis is not user-friendly to say the least. And it's still wasteful as we don't need other processes to do anything.\r\n\r\nBut it solves the current race condition.\r\n\r\nClearly this calls for a design discussion as it's the responsibility of the Trainer to handle this and not user's. Perhaps in the `transformers` land?", "I don't see how this could be the responsibility of `Trainer`, who hasn't the faintest idea of what a `datasets.Metric` is. The trainer takes a function `compute_metrics` that goes from predictions + labels to metric results, there is nothing there. That computation is done on all processes \r\n\r\nThe fact a `datasets.Metric` object cannot be used as a simple compute function in a multi-process environment is, in my opinion, a bug in `datasets`. Especially since, as I mentioned before, the multiprocessing part of `datasets.Metric` has a deep flaw since it can't work in a multinode environment. So you actually need to do the job of gather predictions and labels yourself.\r\n\r\nThe changes you are proposing Stas are making the code less readable and also concatenate all the predictions and labels `number_of_processes` times I believe, which is not going to make the metric computation any faster.\r\n\r\n", "Right, to clarify, I meant it'd be good to have it sorted on the library side and not requiring the user to figure it out. This is too complex and error-prone and if not coded correctly the bug will be intermittent which is even worse.\r\n\r\nOh I guess I wasn't clear in my message - in no way I'm proposing that we use this workaround code - I was just showing what I had to do to make it work.\r\n\r\nWe are on the same page.\r\n\r\n> The changes you are proposing Stas are making the code less readable and also concatenate all the predictions and labels number_of_processes times I believe, which is not going to make the metric computation any faster.\r\n\r\nAnd yes, this is another problem that my workaround introduces. Thank you for pointing it out, @sgugger \r\n", "> The fact a datasets.Metric object cannot be used as a simple compute function in a multi-process environment is, in my opinion, a bug in datasets\r\n\r\nYes totally, this use case is supposed to be supported by `datasets`. And in this case there shouldn't be any collision between the metrics. I'm looking into it :)\r\nMy guess is that at one point the metric isn't using the right file name. It's supposed to use one with a unique uuid in order to avoid the collisions.", "I just opened #1966 to fix this :)\r\n@stas00 if have a chance feel free to try it !", "Thank you, @lhoestq - I will experiment and report back. \r\n\r\nedit: It works! Thank you", "Fixed in https://github.com/huggingface/datasets/pull/1966" ]

[]

[ "Hi @lhoestq,\r\n\r\nI have added the changes from the review." ]

[]

[]

[]

[]

[]

[]

[]

[ "\r\nThis wasn't a `datasets` problem, but `transformers`' and it was solved here https://github.com/huggingface/transformers/pull/11168\r\n" ]

[]