Datasets:

---

abhishek-shrm

/

github-issues

like 0

[]

[]

[ "found a mirror: https://storage.googleapis.com/seldon-datasets/sentence_polarity_v1/rt-polaritydata.tar.gz", "fixed in #484 ", "Closing this one. Thanks again @jxmorris12 for taking care of this :)" ]

[]

[ { "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 } ]

[ "This comes from an overflow in pyarrow's array.\r\nIt is stuck inside the loop that reduces the batch size to avoid the overflow.\r\nI'll take a look", "I created a PR to fix the issue.\r\nIt was due to an overflow check that handled badly an empty list.\r\n\r\nYou can try the changes by using \r\n```\r\n!pip install git+https://github.com/huggingface/nlp.git@fix-bad-type-in-overflow-check\r\n```\r\n\r\nAlso I noticed that the first 1000 examples have an empty list in the `title_urls` field. The feature type inference in `.map` will consider it `null` because of that, and it will crash when it encounter the next example with a `title_urls` that is not empty.\r\n\r\nTherefore to fix that, what you can do for now is increase the writer batch size so that the feature inference will take into account at least one example with a non-empty `title_urls`:\r\n\r\n```python\r\n# default batch size is 1_000 and it's not enough for feature type inference because of empty lists\r\nvalid_dataset = valid_dataset.map(make_input_target, writer_batch_size=3_000) \r\n```\r\n\r\nI was able to run the frozen cell with these changes.", "@lhoestq Perfect and thank you very much!!\r\nClose the issue.", "@lhoestq mapping the function `make_input_target` was passed by your fixing.\r\n\r\nHowever, there is another error in the final step of `valid_dataset.map(convert_to_features, batched=True)`\r\n\r\n`ArrowInvalid: Could not convert Thepiratebay.vg with type str: converting to null type`\r\n(The [same colab notebook above with new error message](https://colab.research.google.com/drive/14wttOTv3ky74B_c0kv5WrbgQjCF2fYQk?usp=sharing#scrollTo=5sRrJ3_C8rLt))\r\n\r\nDo you have some ideas? (I am really sorry I could not debug it by myself since I never used `pyarrow` before) \r\nNote that `train_dataset.map(convert_to_features, batched=True)` can be run successfully even though train_dataset is 27x bigger than `valid_dataset` so I believe the problem lies in some field of `valid_dataset` again .", "I got this issue too and fixed it by specifying `writer_batch_size=3_000` in `.map`.\r\nThis is because Arrow didn't expect `Thepiratebay.vg` in `title_urls `, as all previous examples have empty lists in `title_urls `", "I am clear now . Thank so much again Quentin!" ]

[]

[]

[ "Not sure why the AWS test is failing - perhaps I made too many concurrent CI builds 😢. Can someone please rerun the CI to check the error is not on my end?", "I pushed an improved docstring and the unit tests now pass, which suggests the previous failure on AWS was simply a timeout error. \r\n\r\nFor some reason the docs are now failing to build, but does not seem related to my changes:\r\n```\r\nWarning, treated as error:\r\n/home/circleci/nlp/src/nlp/dataset_dict.py:docstring of nlp.DatasetDict.filter:27:Inline interpreted text or phrase reference start-string without end-string.\r\nmake: *** [Makefile:20: html] Error 2\r\n```\r\n\r\nAny ideas what's going wrong?", "The build_doc fail has been fixed on master.\r\nIt was due to the latest update of sphinx that has some issues, so I pinned the previous version for now.", "I noticed that you also changed the Apache Beam `open` to also use utf-8. However it doesn't have an `encoding` parameter.\r\nTherefore you should ignore lines like\r\n\r\n```python\r\nbeam.io.filesystems.FileSystems.open(filepath)\r\n```\r\n\r\nI guess you could add a rule to your regex to only include the `open` call that have a space right before it.", "Good catch @lhoestq! Your suggestion to match on `open(...)` with a whitespace was a great idea - it allowed me to simplify the regexp considerably 😄.\r\n\r\nI fixed the Apache Beam false positives and also caught a few problems in `json.load()`, e.g.\r\n```python\r\nrelation_name_map = json.load(open(rel_info), encoding='utf-8')\r\n```\r\n\r\nI've tested that the new regexp doesn't reintroduce these false positives, so I think the PR is ready for another review.", "Ok to merge this @lhoestq ?" ]

[]

[]

[ "Looks good to me as well but we'll want to add a test indeed.\r\nYou can add one if you have time @TevenLeScao.\r\nOtherwise, we'll do it when we are back with Quentin. ", "I fixed it in #494 " ]

[]

[]

[ "Really nice !\r\nThanks for adding this one.\r\n\r\nI noticed that there are some '-' that are left in the description in the middle of some workds. It migh come from copy-pasting the pdf paper. ex: `im-provement`. Could you fix that please ?", "@lhoestq \r\nLinebreaks have been removed! Note that there are still a few compound words that are hyphenated intentionally. ", "I think you just need to rebase from master to fix the CI :)", "Yes I made the mistake of simply merging master into this branch. A rebase seems to be neater :) Although all the commits ended up being added twice. I assume you just squash them into a single one on merge anyways?", "Yes indeed they'll be squashed" ]

[]

[]

[ "Nevermind, I restarted my python session and it worked fine...\r\n\r\n---\r\n\r\nI had an authentification error, and I authenticated from another terminal. After that, no more error but it was not working. Restarting the sessions makes it work :)" ]

[]

[]

[ "Thanks for reporting this issue\r\n\r\nThere was a bug where numpy arrays would get returned instead of tensorflow tensors.\r\nThis is fixed on master.\r\n\r\nI tried to re-run the colab and encountered this error instead:\r\n\r\n```\r\nAttributeError: 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'to_tensor'\r\n```\r\n\r\nThis is because the dataset returns a Tensor and not a RaggedTensor.\r\nBut I think we should always return a RaggedTensor unless the length of the sequence is fixed (it that case they can be stack into a Tensor).", "Hi, I got another error (on Colab):\r\n\r\n```python\r\n# You can read a few attributes of the datasets before loading them (they are python dataclasses)\r\nfrom dataclasses import asdict\r\n\r\nfor key, value in asdict(datasets[6]).items():\r\n print('👉 ' + key + ': ' + str(value))\r\n\r\n---------------------------------------------------------------------------\r\n\r\nTypeError Traceback (most recent call last)\r\n\r\n<ipython-input-6-b8ace6c227a2> in <module>()\r\n 2 from dataclasses import asdict\r\n 3 \r\n----> 4 for key, value in asdict(datasets[6]).items():\r\n 5 print('👉 ' + key + ': ' + str(value))\r\n\r\n/usr/local/lib/python3.6/dist-packages/dataclasses.py in asdict(obj, dict_factory)\r\n 1008 \"\"\"\r\n 1009 if not _is_dataclass_instance(obj):\r\n-> 1010 raise TypeError(\"asdict() should be called on dataclass instances\")\r\n 1011 return _asdict_inner(obj, dict_factory)\r\n 1012 \r\n\r\nTypeError: asdict() should be called on dataclass instances\r\n```", "Indeed we'll update the cola with the new release coming up this week." ]

[]

[]

[ "> Also, a little out of my depth there, but would there be a way to have the default pip install checklist command not require mysql and mariadb to be installed? Feels like that might be a source of confusion for users.\r\n\r\nI removed the pattern dependency, mysql is not a requirement anymore. I'm not sure where mariadb is coming from. " ]

[]

[]

[ "Cool thanks, I made those changes. LMK if you think it's ready for merge.", "Ok to merge for me" ]

[]

[]

[ "The `data_dir` parameter has been removed. Now the error is `ValueError: Config name is missing`\r\n\r\nAs mentioned in #470 I think we can have one test with the first config of BUILDER_CONFIGS, and another test that runs all of the configs in BUILDER_CONFIGS", "This was fixed in #527 \r\n\r\nClosing this one, but feel free to re-open if you have other questions" ]

[]

[]

[]

[]

[]

[ "This PR was already approved by @lhoestq in #456 . This one just make style to remove some typos" ]

[]

[]

[]

[]

[]

[ "Ok I tried to add the dummy dataset (I actually modified the dummy_data command to generate them for me because it was too painful to do that manually).\r\n\r\nThe dummy_data test seems to work:\r\n```bash\r\nRUN_SLOW=1 pytest tests/test_dataset_common.py::LocalDatasetTest::test_load_dataset_all_configs_iwslt2017\r\n```\r\n\r\nHowever the test on the full data fails, because the `**config_kwargs` don't include `pair, multilingual`.\r\nI could add a default parameter for the Config (but that feels broken, how can one config be the \"default\" ?). If I do I still have errors, saying that something within the downloader is a directory so I'm not sure where that comes from.\r\n\r\nI can share my auto_zip dummy data code if you want (I tried to keep it clean). [Edit: it's [here](https://github.com/Narsil/nlp/tree/auto_zip)]. \r\nThe way it works is that it just keeps X line from the beginning of the original files, and Y lines at the end. It's good enough for my usage, but I guess it could work for most data files out there (as long as they're real text and not binary format)", "The slow test doesn't support dataset that require config parameters that don't have default values.\r\n\r\nTo improve that we can replace it by two tests:\r\n- one test that loads the default config (it can simply be the first config of the config lists for example)\r\n- one tests that iterate over all configs and load them all one by one\r\n\r\nBy using the configs inside the builder config lists, there is no need to instantiate new configs, so the missing parameter error doesn't happen.\r\n\r\nDoes that sound good to you ?", "Seems fair.\r\nHowever I'm unsure what I should do ?\r\n\r\nShould I wait for #527 to pass and rebase and the command will be the same ?\r\nShould I update something ?", "I think everything is fine on your side. Thanks for adding this dataset :)\r\n\r\nI think it's better to wait for the slow test to be updated if you don't mind.\r\n", "Sure ! :)", "Thanks for fixing the isort/black changes :)\r\nFeel free to merge if it's good for you @Narsil " ]

[]

[]

[ "Hi ! Did you try to set the output format to pytorch ? (or tensorflow if you're using tensorflow)\r\nIt can be done with `dataset.set_format(\"torch\", columns=columns)` (or \"tensorflow\").\r\n\r\nNote that for pytorch, string columns can't be converted to `torch.Tensor`, so you have to specify in `columns=` the list of columns you want to keep (`input_ids` for example)", "Hello . Yes, I did set the output format as below for the two columns \r\n\r\n `train_dataset.set_format('torch',columns=['Text','Label'])`\r\n ", "I think you're having this issue because you try to format strings as pytorch tensors, which is not possible.\r\nIndeed by having \"Text\" in `columns=['Text','Label']`, you try to convert the text values to pytorch tensors.\r\n\r\nInstead I recommend you to first tokenize your dataset using a tokenizer from transformers. For example\r\n\r\n```python\r\nfrom transformers import BertTokenizer\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\n\r\ntrain_dataset.map(lambda x: tokenizer(x[\"Text\"]), batched=True)\r\ntrain_dataset.set_format(\"torch\", column=[\"input_ids\"])\r\n```\r\n\r\nAnother way to fix your issue would be to not set the format to pytorch, and leave the dataset as it is by default. In that case, the strings are returned normally when you get examples from your dataloader. It means that you would have to tokenize the examples in the training loop (or using a data collator) though.\r\n\r\nLet me know if you have other questions", "Hi, actually the thing is I am getting the same error and even after tokenizing them I am passing them through batch_encode_plus.\r\nI dont know what seems to be the problem is. I even converted it into 'pt' while passing them through batch_encode_plus but when I am evaluating my model , i am getting this error\r\n\r\n\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\n<ipython-input-145-ca218223c9fc> in <module>()\r\n----> 1 val_loss, predictions, true_val = evaluate(dataloader_validation)\r\n 2 val_f1 = f1_score_func(predictions, true_val)\r\n 3 tqdm.write(f'Validation loss: {val_loss}')\r\n 4 tqdm.write(f'F1 Score (Weighted): {val_f1}')\r\n\r\n6 frames\r\n/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataset.py in <genexpr>(.0)\r\n 160 \r\n 161 def __getitem__(self, index):\r\n--> 162 return tuple(tensor[index] for tensor in self.tensors)\r\n 163 \r\n 164 def __len__(self):\r\n\r\nTypeError: new(): invalid data type 'str' ", "> Hi, actually the thing is I am getting the same error and even after tokenizing them I am passing them through batch_encode_plus.\r\n> I dont know what seems to be the problem is. I even converted it into 'pt' while passing them through batch_encode_plus but when I am evaluating my model , i am getting this error\r\n> \r\n> TypeError Traceback (most recent call last)\r\n> in ()\r\n> ----> 1 val_loss, predictions, true_val = evaluate(dataloader_validation)\r\n> 2 val_f1 = f1_score_func(predictions, true_val)\r\n> 3 tqdm.write(f'Validation loss: {val_loss}')\r\n> 4 tqdm.write(f'F1 Score (Weighted): {val_f1}')\r\n> \r\n> 6 frames\r\n> /usr/local/lib/python3.6/dist-packages/torch/utils/data/dataset.py in (.0)\r\n> 160\r\n> 161 def **getitem**(self, index):\r\n> --> 162 return tuple(tensor[index] for tensor in self.tensors)\r\n> 163\r\n> 164 def **len**(self):\r\n> \r\n> TypeError: new(): invalid data type 'str'\r\n\r\nI got the same error and fix it .\r\nyou can check your input where there may be string contained.\r\nsuch as\r\n```\r\na = [1,2,3,4,'<unk>']\r\ntorch.tensor(a)\r\n```", "I didn't know tokenizers could return strings in the token ids. Which tokenizer are you using to get this @Doragd ?", "> I didn't know tokenizers could return strings in the token ids. Which tokenizer are you using to get this @Doragd ?\r\n\r\ni'm sorry that i met this issue in another place (not in huggingface repo). ", "@akhilkapil do you have strings in your dataset ? When you set the dataset format to \"pytorch\" you should exclude columns with strings as pytorch can't make tensors out of strings" ]

[]

[]

[ "Indeed. Solution 1 is the simplest.\r\n\r\nThis is actually a recurring problem.\r\nI think we should scan all the datasets with regexpr to fix the use of `open()` without encodings.\r\nAnd probably add a test in the CI to forbid using this in the future.", "I'm happy to tackle the broader problem - will open a PR when it's ready!", "That would be awesome!", "I've created a simple function that seems to do the trick:\r\n\r\n```python\r\ndef apply_encoding_on_file_open(filepath: str):\r\n \"\"\"Apply UTF-8 encoding for all instances where a non-binary file is opened.\"\"\"\r\n \r\n with open(filepath, 'r', encoding='utf-8') as input_file:\r\n regexp = re.compile(r\"\"\"\r\n (?!.*\\b(?:encoding|rb|wb|wb+|ab|ab+)\\b)\r\n (open)\r\n \\((.*)\\)\r\n \"\"\")\r\n input_text = input_file.read()\r\n match = regexp.search(input_text)\r\n \r\n if match:\r\n print('Found match!', match.group())\r\n # append utf-8 encoding to matching groups in-place\r\n output = regexp.sub(lambda m: m.group()[:-1]+', encoding=\"utf-8\")', input_text)\r\n with open(filepath, 'w', encoding='utf-8') as output_file:\r\n output_file.write(output)\r\n else:\r\n print(\"No match found!\")\r\n```\r\n\r\nThe regexp does a negative lookahead to avoid matching on cases where the encoding is already specified or when binary files are involved.\r\n\r\nFrom an implementation perspective:\r\n\r\n* Would it make sense to include this function in `nlp-cli` so that we can run something like\r\n```\r\nnlp-cli fix_encoding path/to/folder\r\n```\r\nand the command recursively fixes all files in the target?\r\n* What is the desired behaviour in the CI test? Here we could either have a simple script that we run as a `job` in the CI and raises an error if a missing encoding is detected. Alternatively we could incorporate this behaviour into the CLI and run that in the CI.\r\n\r\nPlease let me know what you prefer among the alternatives.\r\n", "I realised I was overthinking the problem, so decided to just run the regexp over the codebase and make the PR. In other words, we can ignore my comments about using the CLI 😸 " ]

[]

[]

[ "Thanks!" ]

[]

[]

[ "The cast function is now called inside `features.encode_example`.\r\nI also added `encode_batch` that was missing.\r\n\r\nMoreover I used the cast function in `Dataset.map` to support torch/tensorflow tensors or numpy arrays inputs.\r\n\r\nThere are tests for tensors inputs in metrics and in .map", "I think we can merge" ]

[]

[]

[ "One note on features inference:\r\n\r\nif an arrow type is `struct of items` where each item is a `list`, then we return a `dict` in which each item is a `Sequence`.\r\nIt means that we don't use the Sequence <-> dict swap when we infer features.\r\n\r\nIt's fine because the swap is generally used in dataset scripts, in which features are defined (inferred features are discarded)", "If it's fine for you @thomwolf we can merge this one :) ", "Yes this is fine I think!" ]

[]

[]

[]

[]

[]

[ "I think the problem is related to `wiki_dpr` dataset which is making the circle CI failed as you can see:\r\n```\r\nFAILED tests/test_dataset_common.py::AWSDatasetTest::test_load_dataset_wiki_dpr\r\nFAILED tests/test_hf_gcp.py::TestDatasetOnHfGcp::test_script_synced_with_s3_wiki_dpr/dummy_psgs_w100_no_embeddings\r\nFAILED tests/test_hf_gcp.py::TestDatasetOnHfGcp::test_script_synced_with_s3_wiki_dpr/dummy_psgs_w100_with_nq_embeddings\r\nFAILED tests/test_hf_gcp.py::TestDatasetOnHfGcp::test_script_synced_with_s3_wiki_dpr/psgs_w100_no_embeddings\r\nFAILED tests/test_hf_gcp.py::TestDatasetOnHfGcp::test_script_synced_with_s3_wiki_dpr/psgs_w100_with_nq_embeddings\r\n\r\n```\r\nI'm facing the same issues with my last commits, I tried to rebase from master but it still not working. Maybe @lhoestq can help with.", "Hello, I am confused about the next steps I need to do. Did the forced merge solve the issue ?", "Hello :)\r\nI think you can just rebase from master and it should solve the CI error" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks @TevenLeScao for finding this issue", "Thanks @lhoestq for catching this ❤️" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Sorry one nit: Could we use named arguments for the call to BLEURT?\r\n\r\ni.e. \r\n scores = self.scorer.score(references=references, candidates=predictions)\r\n\r\n(i.e. so it is less bug prone)\r\n", "Following up on Ankur's comment---we are going to drop support for\npositional (not named) arguments in the future releases because it seems to\ncause bugs and confusion. I hope it doesn't create too much of a mess.\n\nLe jeu. 30 juil. 2020 à 10:44, ankparikh <[email protected]> a\nécrit :\n\n> Sorry one nit: Could we use named arguments for the call to BLEURT?\n>\n> i.e.\n> scores = self.scorer.score(references=references, candidates=predictions)\n>\n> (i.e. so it is less bug prone)\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/nlp/pull/455#issuecomment-666414514>, or\n> unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABTMRNGAN2PMECS5K4DIHJDR6GBMLANCNFSM4PL323FA>\n> .\n>\n", "> Following up on Ankur's comment---we are going to drop support for positional (not named) arguments in the future releases because it seems to cause bugs and confusion. I hope it doesn't create too much of a mess. Le jeu. 30 juil. 2020 à 10:44, ankparikh <[email protected]> a écrit :\r\n> […](#)\r\n> Sorry one nit: Could we use named arguments for the call to BLEURT? i.e. scores = self.scorer.score(references=references, candidates=predictions) (i.e. so it is less bug prone) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <[#455 (comment)](https://github.com/huggingface/nlp/pull/455#issuecomment-666414514)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABTMRNGAN2PMECS5K4DIHJDR6GBMLANCNFSM4PL323FA> .\r\n\r\nChanged @ankparikh @tsellam, thanks for taking a look!", "We should avoid positional arguments in metrics on our side as well. It's a dangerous source of errors indeed." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi ! Glad you managed to fix the version issue.\r\n\r\nThe command `\r\npython nlp-cli dummy_data datasets/guardian_authorship --save_infos --all_configs` is supposed to generate a json file `dataset_infos.json` next to your dataset script, but I can't see it in the PR.\r\nCan you make sure you have the json file on your side and that you have pushed it ?", "Done!", "Is there anything else that I should do? and would the new dataset be available via the NLP package now? ", "Sorry I forgot to merge this one ! Doing it now", "Thanks for the heads up ;)", "No worries, this is my first contribution to an online package, and I feel very proud it's part of this library :) Thank you very much!" ]

[]

[]

[ "I think this is the way to go but I’m afraid this might be a little slow. I was thinking that we could use a high quality very fast non crypto hash like xxhash for these stuff (hashing data files)", "Yep good idea, I'll take a look", "I tested the hashing speed [here](https://colab.research.google.com/drive/1hlhP84kLIHmOzMRQN1h8x10hKWpXXyud?usp=sharing).\r\nI was able to get 8x speed with `xxhashlib` (42ms vs 345ms for 100MiB of data).\r\nWhat do you think @thomwolf ?", "I added xxhash and some tests" ]

[]

[]

[]

[]

[]

[ "> Awesome !\r\n> Good job on parsing these files :O\r\n> \r\n> Do you think it would be hard to get the two other split configurations ?\r\n\r\nIt shouldn't be that hard, I think I can consider different config names for each split ", "> > Awesome !\r\n> > Good job on parsing these files :O\r\n> > Do you think it would be hard to get the two other split configurations ?\r\n> \r\n> It shouldn't be that hard, I think I can consider different config names for each split\r\n\r\nYes that would be perfect", "closing this PR and opening a new one to fix the circle CI problems" ]

[]

[]

[ "Could you run `make style` to fix the code_quality fail ?\r\nYou'll need `black` and `isort` that you can install by doing `pip install -e .[quality]`", "Thanks @lhoestq\r\nI fixed the styling", "Thank you :)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "This issue was resolved by #447 " ]

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

[ { "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 } ]

[ "Same here !", "This is the only fix I could come up with without touching the repo's code.\r\n```python\r\nfrom nlp.builder import FORCE_REDOWNLOAD\r\ndataset = load_dataset('csv', data_file='./a.csv', download_mode=FORCE_REDOWNLOAD, version='0.0.1')\r\n```\r\nYou'll have to change the version each time you want to load a different csv file.\r\nIf you're willing to add a ```print```, you can go to ```nlp.load``` and add ```print(builder_instance.cache_dir)``` right before the ```return ds``` in the ```load_dataset``` method. It'll print the cache folder, and you'll just have to erase it (and then you won't need the change here above)." ]

[]

[]

[ "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. " ]

[ { "id": 2067401494, "node_id": "MDU6TGFiZWwyMDY3NDAxNDk0", "url": "https://api.github.com/repos/huggingface/datasets/labels/Dataset%20discussion", "name": "Dataset discussion", "color": "72f99f", "default": false, "description": "Discussions on the datasets" } ]

[]

[]

[]

[]

[ "This one is ready for review now", "I changed to using features only, instead of info.\r\nLet mw know if it sounds good to you now @thomwolf " ]

[]

[]

[]

[]

[]

[ "`DPRContextEncoder` and `DPRContextEncoderTokenizer` will be available in the next release of `transformers`.\r\n\r\nRight now you can experiment with it by installing `transformers` from the master branch.\r\nYou can also check the docs of DPR [here](https://huggingface.co/transformers/master/model_doc/dpr.html).\r\n\r\nMoreover all the indexing features will also be available in the next release of `nlp`.", "@lhoestq Thanks for the info ", "@lhoestq I tried installing transformer from the master branch. Python imports for DPR again didnt' work. Anyways, Looking forward to trying it in the next release of nlp ", "@nsankar have you tried with the latest version of the library?", "@yjernite it worked. Thanks" ]

[ { "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" } ]

[]

[ "Thanks Sam, we now have a very detailed tutorial and template on how to add a new dataset to the library. It typically take 1-2 hours to add one. Do you want to give it a try ?\r\nThe tutorial on writing a new dataset loading script is here: https://huggingface.co/nlp/add_dataset.html\r\nAnd the part on how to share a new dataset is here: https://huggingface.co/nlp/share_dataset.html", "Hi @sshleifer, I'm trying to add IWSLT using the link you provided but the download urls are not working. Only `[en, de]` pair is working. For others language pairs it throws a `404` error.\r\n\r\n" ]

[]

[]

[ "There is an interesting design question here (cc @lhoestq).\r\n\r\nI guess the labels form a closed set so we could also use a [nlp.ClassLabel](https://huggingface.co/nlp/package_reference/main_classes.html#nlp.ClassLabel) instead of a string. The differences will be mainly that:\r\n- the labels are stored as integers and thus ready for training a model\r\n- the string to int conversion methods are handled by the `nlp.ClassLabel` feature (see the [doc](https://huggingface.co/nlp/package_reference/main_classes.html#nlp.ClassLabel) and [here](https://huggingface.co/nlp/features.html) and [here](https://huggingface.co/nlp/quicktour.html#fine-tuning-a-deep-learning-model)).\r\n\r\nIn my opinion, storing the labels as integers instead of strings makes it:\r\n- slightly less readable when accessing a dataset example (e.g. with `dataset[0]`)\r\n- force you with a specific mapping from string to integers\r\n- more clear that there is a fixed and predefined list of labels\r\n- easier to list all the labels (directly visible in the features).\r\n\r\n=> overall I'm pretty neutral about using one or the other option (`nlp.string` or `nlp.ClassLabel`).\r\n\r\nNote that we can now rather easily convert from one to the other with the map function and something like:\r\n```python\r\ndataset = dataset.map(lambda x: x, features=nlp.Features({'labels': nlp.ClassLabel(MY_LABELS_NAMES)}))\r\ndataset = dataset.map(lambda x: {'labels': dataset.features['labels'].int2str(x['labels'])}, features=nlp.Features({'labels': nlp.Value('string')}))\r\n```\r\n^^ this could probably be made even simpler (in particular for the second case)", "I see. This is an interesting question.\r\nMaybe as the dataset doesn't provide the mapping we shouldn't force an arbitrary one, and keep them as strings ?\r\nMoreover for NER the labels are often different from a dataset to the other so it's probably good to keep strings (there is no conventional mapping).\r\nAlso as the column is called \"ner_tags\" (or \"langs\"), you can already assume that there is a fixed and predefined list of labels.", "Yes sounds good to me.\r\nThis make me wonder if we don’t want to have a default identity function in `map` so this method could also be used to easily cast features. What do you think?", "Yes sounds good. I also noticed that people use map with identity to write a dataset into a specified cache file." ]

[]

[]

[ "Indeed, we’ll make a new PyPi release next week to solve this. Cc @lhoestq ", "+1! this is the reason our tests are failing at [TextAttack](https://github.com/QData/TextAttack) \r\n\r\n(Though it's worth noting if we fixed the version number of pyarrow to 0.16.0 that would fix our problem too. But in this case we'll just wait for you all to update)", "Came to raise this issue, great to see other already have and it's being fixed so soon!\r\n\r\nAs an aside, since no one wrote this already, it seems like the version check only looks at the second part of the version number making sure it is >16, but pyarrow newest version is 1.0.0 so the second past is 0!", "> Indeed, we’ll make a new PyPi release next week to solve this. Cc @lhoestq\r\n\r\nYes definitely", "please fix this on pypi! @lhoestq ", "Is this issue fixed ?", "We’ll release the new version later today. Apologies for the delay.", "I just pushed the new version on pypi :)", "Thanks for the update." ]

[]

[]

[ "This was fixed in #434 \r\nWe'll do a release later this week to include this fix.\r\nThanks for reporting", "I dont know if the fix was made but the problem is still present : \r\nInstaled with pip : NLP 0.3.0 // pyarrow 1.0.0 \r\nOS : archlinux with kernel zen 5.8.5", "Yes it was fixed in `nlp>=0.4.0`\r\nYou can update with pip", "Sorry, I didn't got the updated version, all is now working perfectly thanks" ]

[]

[]

[ "Great, thanks!" ]

[]

[]

[ "Hi @ArneBinder, we have a few \"generic\" datasets which are intended to load data files with a predefined format:\r\n- csv: https://github.com/huggingface/nlp/tree/master/datasets/csv\r\n- json: https://github.com/huggingface/nlp/tree/master/datasets/json\r\n- text: https://github.com/huggingface/nlp/tree/master/datasets/text\r\n\r\nYou can find more details about this way to load datasets here in the documentation: https://huggingface.co/nlp/loading_datasets.html#from-local-files\r\n\r\nMaybe your brat loading script could be shared in a similar fashion?", "> Maybe your brat loading script could be shared in a similar fashion?\r\n\r\n@thomwolf that was also my first idea and I think I will tackle that in the next days. I separated the code and created a real abstract class `AbstractBrat` to allow to inherit from that (I've just seen that the dataset_loader loads the first non abstract class), now `Brat` is very similar in its functionality to https://github.com/huggingface/nlp/tree/master/datasets/text but inherits from `AbstractBrat`.\r\n\r\nHowever, it is still not clear to me how to add a specific dataset (as explained in https://huggingface.co/nlp/add_dataset.html) to your repo that uses this format/abstract class, i.e. re-using the `features` entry of the `DatasetInfo` object and `_generate_examples()`. Again, by doing so, the only remaining entries/functions to define would be `_DESCRIPTION`, `_CITATION`, `homepage` and `_URL` (which is all copy-paste stuff) and `_split_generators()`.\r\n \r\nIn a lack of better ideas, I tried sth like below, but of course it does not work outside `nlp` (`AbstractBrat` is currently defined in [datasets/brat.py](https://github.com/ArneBinder/nlp/blob/5e81fb8710546ee7be3353a7f02a3045e9a8351e/datasets/brat/brat.py)):\r\n```python\r\nfrom __future__ import absolute_import, division, print_function\r\n\r\nimport os\r\n\r\nimport nlp\r\n\r\nfrom datasets.brat.brat import AbstractBrat\r\n\r\n_CITATION = \"\"\"\r\n@inproceedings{lauscher2018b,\r\n title = {An argument-annotated corpus of scientific publications},\r\n booktitle = {Proceedings of the 5th Workshop on Mining Argumentation},\r\n publisher = {Association for Computational Linguistics},\r\n author = {Lauscher, Anne and Glava\\v{s}, Goran and Ponzetto, Simone Paolo},\r\n address = {Brussels, Belgium},\r\n year = {2018},\r\n pages = {40–46}\r\n}\r\n\"\"\"\r\n\r\n_DESCRIPTION = \"\"\"\\\r\nThis dataset is an extension of the Dr. Inventor corpus (Fisas et al., 2015, 2016) with an annotation layer containing \r\nfine-grained argumentative components and relations. It is the first argument-annotated corpus of scientific \r\npublications (in English), which allows for joint analyses of argumentation and other rhetorical dimensions of \r\nscientific writing.\r\n\"\"\"\r\n\r\n_URL = \"http://data.dws.informatik.uni-mannheim.de/sci-arg/compiled_corpus.zip\"\r\n\r\n\r\nclass Sciarg(AbstractBrat):\r\n\r\n VERSION = nlp.Version(\"1.0.0\")\r\n\r\n def _info(self):\r\n\r\n brat_features = super()._info().features\r\n return nlp.DatasetInfo(\r\n # This is the description that will appear on the datasets page.\r\n description=_DESCRIPTION,\r\n # nlp.features.FeatureConnectors\r\n features=brat_features,\r\n # If there's a common (input, target) tuple from the features,\r\n # specify them here. They'll be used if as_supervised=True in\r\n # builder.as_dataset.\r\n #supervised_keys=None,\r\n # Homepage of the dataset for documentation\r\n homepage=\"https://github.com/anlausch/ArguminSci\",\r\n citation=_CITATION,\r\n )\r\n\r\n def _split_generators(self, dl_manager):\r\n \"\"\"Returns SplitGenerators.\"\"\"\r\n # TODO: Downloads the data and defines the splits\r\n # dl_manager is a nlp.download.DownloadManager that can be used to\r\n # download and extract URLs\r\n dl_dir = dl_manager.download_and_extract(_URL)\r\n data_dir = os.path.join(dl_dir, \"compiled_corpus\")\r\n print(f'data_dir: {data_dir}')\r\n return [\r\n nlp.SplitGenerator(\r\n name=nlp.Split.TRAIN,\r\n # These kwargs will be passed to _generate_examples\r\n gen_kwargs={\r\n \"directory\": data_dir,\r\n },\r\n ),\r\n ]\r\n``` \r\n\r\nNevertheless, many thanks for tackling the dataset accessibility problem with this great library!", "As temporary fix I've created [ArneBinder/nlp-formats](https://github.com/ArneBinder/nlp-formats) (contributions welcome)." ]

[]

[]

[ "Ok for this but I think we may want to use the general `filelock` method we are using at other places in the library instead of filecmp (in particular `filelock` take care of being an atomic operation which is safer for concurrent processes)", "Ok I see.\r\nWhy not use filelock in this case then ?", "I think we should 🙂", "Thanks for approving my patch.\n\nI agree that if copying is needed then some locking mechanism should be put in place. But, I don't think a file should be needlessly copied without a check. So I guess the flow should be, lock => copy if needed => unlock, and add locks wherever else that file is being accessed.\n\nI'll also add that my personal experience with filelock on a different project hasn't been that great, and on some occasions a process somehow got through the lock -- I've never gotten to the bottom of that but it tainted my view of that module. Perhaps it's been fixed (or I just miss used it), but thought you should know to take steps to test it." ]

[]

[]

[ "I was using a hack in `wiki_dpr` to download the index from GCS even for the configurations without the embeddings.\r\nHowever as GCS is something internal, I changed the logic to add a download step for indexes directly in the dataset script, using the `DownloadManager`.\r\n\r\nThis change was directly linked to the changes I did to take into account the split name in the post processing, so I included this change in this PR too.\r\n\r\nTo summarize:\r\n\r\nDataset builders can now implement\r\n- `_post_processing_resources(split)`: return a dict `resource_name -> resource_file_name`. It defines the additional resources such as indexes or arrow files that you need in post processing\r\n- `_download_post_processing_resources(split, resource_name, dl_manager))`: if some resources can be downloaded, you can use the download_manager to download them\r\n- `_post_process(dataset, resources_path)`: (main function for post processing) given a dataset, you can apply dataset transforms or add indexes. For resources that have been downloaded, you can load them. For the others, you can generate and save them. The paths to load/save resources are in `resources_path` which is a dictionary `resource_name -> resource_path`\r\n\r\nAbout the CI:\r\n```\r\n=========================== short test summary info ============================\r\nFAILED tests/test_dataset_common.py::AWSDatasetTest::test_load_dataset_wiki_dpr\r\n```\r\nIt fails because I changed the input of post processing functions (to include the split name)", "I started to add metadata in the DatasetInfo.\r\nNote that because there are new fields, **ALL the dataset_info[s].json generated after these changes won't be loadable from older versions of the lib**\r\n\r\nRight now it looks like this:\r\n```json\r\n \"post_processing_resources_checksums\": {\r\n \"train\": {\r\n \"embeddings_index\": {\r\n \"num_bytes\": 30720045,\r\n \"checksum\": \"b04fb4f4f3ab83b9d1b9f6f9eb236f1c04a9fd61bef7cee16b12df8ac911766a\"\r\n }\r\n }\r\n },\r\n \"post_processing_size\": 30720045,\r\n```", "Good point. Should we anticipate already that we may add other fields in the future and change the code to support the addition of new fields without breaking backward compatibility in the future?", "I added:\r\n- post processing features (inside a PostProcessedInfo object)\r\n- backward compatibility for dataset info\r\n- post processing tests (as_dataset and download_and_prepare) for map (change features), select (change number of elements) and add_faiss_index (add indexes)\r\nAnd I fixed a bug in `map` that I found thanks to the new tests\r\n\r\nNow I just have to move `post_processing_resources_checksums` to PostProcessedInfo as well and everything should be good :)\r\nEdit: done" ]

[]

[]

[ "I did the changes in the docstrings and I added a type check in each `DatasetDict` method to make sure all values are of type `Dataset`", "Awesome, do you mind adding these in the doc as well?", "I added it to the docs (processing + main classes)", "I'm trying to follow along with the following about datasets from the docs:\r\n\r\nhttps://huggingface.co/nlp/loading_datasets.html\r\nhttps://huggingface.co/nlp/processing.html\r\n\r\nHowever the train_test_split method no longer works as it is expecting a dataset, rather than a datsetdict. How would I got about splitting a CSV into a train and test set? \r\n\r\nI'm trying to utilize the Trainer() class, but am having trouble converting my data from a csv into dataset objects to pass in." ]

[]

[]

[ "Thanks @RachelKer for this PR.\r\n\r\nI think the dummy_data structure does not also match. In the `_split_generator` you have something like `os.path.join(downloaded_files[\"validation\"], lang+'_val.jsonl')` but in you dummy_data you have `os.path.join(downloaded_files[\"validation\"], lang+\"_val.zip\", lang+'_val.jsonl')`. I think ` jsonl` files should be directly in the `dummy_data` folder without the sub-folder \r\n\r\n@lhoestq ", "Hi @RachelKer :)\r\nThanks for adding MLSUM !\r\n\r\nTo fix the CI I think you just have to rebase from master", "Great, I think it is working now. Thanks :)", "It looks like your PR does tons of changes in other datasets. \r\nMaybe this is because of the merge from master ?", "Hmm, I see, sorry I messed up somewhere. Maybe it's easier if we close the pull request and I do another one ?", "Yea if it's easier for you feel free to re-open a PR" ]

[]

[]

[]

[]

[]

[]

[ { "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" } ]

[]

[ "Yes that's definitely something we plan to add ^^", "Yes, that would be nice. We could take a look at what tensorflow `tf.data` does under the hood for instance.", "So `tf.data.Dataset.map()` returns a `ParallelMapDataset` if `num_parallel_calls is not None` [link](https://github.com/tensorflow/tensorflow/blob/2b96f3662bd776e277f86997659e61046b56c315/tensorflow/python/data/ops/dataset_ops.py#L1623).\r\n\r\nThere, `num_parallel_calls` is turned into a tensor and and fed to `gen_dataset_ops.parallel_map_dataset` where it looks like tensorflow takes over.\r\n\r\nWe could start with something simple like a thread or process pool that `imap`s over some shards.\r\n ", "Multiprocessing was added in #552 . You can set the number of processes with `.map(..., num_proc=...)`. It also works for `filter`\r\n\r\nClosing this one, but feel free to reo-open if you have other questions", "@lhoestq Great feature implemented! Do you have plans to add it to official tutorials [Processing data in a Dataset](https://huggingface.co/docs/datasets/processing.html?highlight=save#augmenting-the-dataset)? It took me sometime to find this parallel processing api.", "Thanks for the heads up !\r\n\r\nI just added a paragraph about multiprocessing:\r\nhttps://huggingface.co/docs/datasets/master/processing.html#multiprocessing" ]

[]

[]

[ "Thanks for noticing ! This looks more reasonable indeed.\r\nFeel free to open a PR", "Hi @lhoestq \r\nI made the proposed changes to the `xtreme.py` script. I noticed that I also need to change the schema in the `dataset_infos.json` file. More specifically the `\"features\"` part of the PAN-X.LANG dataset:\r\n\r\n```json\r\n\"features\":{\r\n \"word\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n },\r\n \"ner_tag\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n },\r\n \"lang\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n }\r\n}\r\n```\r\nTo fit the code above the fields `\"word\"`, `\"ner_tag\"`, and `\"lang\"` would become `\"words\"`, `ner_tags\"` and `\"langs\"`. In addition the `dtype` should be changed from `\"string\"` to `\"list\"`.\r\n\r\n I made this changes but when trying to test this locally with `dataset = load_dataset(\"xtreme\", \"PAN-X.en\", data_dir='./data')` I face the issue that the `dataset_info.json` file is always overwritten by a downloaded version with the old settings, which then throws an error because the schema does not match. This makes it hard to test the changes locally. Do you have any suggestions on how to deal with that?\r\n", "Hi !\r\n\r\nYou have to point to your local script.\r\nFirst clone the repo and then:\r\n\r\n```python\r\ndataset = load_dataset(\"./datasets/xtreme\", \"PAN-X.en\")\r\n```\r\nThe \"xtreme\" directory contains \"xtreme.py\".\r\n\r\nYou also have to change the features definition in the `_info` method. You could use:\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": [nlp.Value(\"string\")],\r\n \"ner_tags\": [nlp.Value(\"string\")],\r\n \"langs\": [nlp.Value(\"string\")],\r\n})\r\n```\r\n\r\nHope this helps !\r\nLet me know if you have other questions.", "Thanks, I am making progress. I got a new error `NonMatchingSplitsSizesError ` (see traceback below), which I suspect is due to the fact that number of rows in the dataset changed (one row per word --> one row per sentence) as well as the number of bytes due to the slightly updated data structure. \r\n\r\n```python\r\nNonMatchingSplitsSizesError: [{'expected': SplitInfo(name='validation', num_bytes=1756492, num_examples=80536, dataset_name='xtreme'), 'recorded': SplitInfo(name='validation', num_bytes=1837109, num_examples=10000, dataset_name='xtreme')}, {'expected': SplitInfo(name='test', num_bytes=1752572, num_examples=80326, dataset_name='xtreme'), 'recorded': SplitInfo(name='test', num_bytes=1833214, num_examples=10000, dataset_name='xtreme')}, {'expected': SplitInfo(name='train', num_bytes=3496832, num_examples=160394, dataset_name='xtreme'), 'recorded': SplitInfo(name='train', num_bytes=3658428, num_examples=20000, dataset_name='xtreme')}]\r\n```\r\nI can fix the error by replacing the values in the `datasets_infos.json` file, which I tested for English. However, to update this for all 40 datasets manually is slightly painful. Is there a better way to update the expected values for all datasets?", "You can update the json file by calling\r\n```\r\nnlp-cli test ./datasets/xtreme --save_infos --all_configs\r\n```", "One more thing about features. I mentioned\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": [nlp.Value(\"string\")],\r\n \"ner_tags\": [nlp.Value(\"string\")],\r\n \"langs\": [nlp.Value(\"string\")],\r\n})\r\n```\r\n\r\nbut it's actually not consistent with the way we write datasets. Something like this is simpler to read and more consistent with the way we define datasets:\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": nlp.Sequence(nlp.Value(\"string\")),\r\n \"ner_tags\": nlp.Sequence(nlp.Value(\"string\")),\r\n \"langs\": nlp.Sequence(nlp.Value(\"string\")),\r\n})\r\n```\r\n\r\nSorry about that", "Closing this since PR #437 fixed the problem and has been merged to `master`. " ]

[]

[]

[]

[]

[]

[ "I had to make `SplitDict` serializable to be able to copy `DatasetInfo` objects properly.\r\nSerialization was also asked in #389 ", "One thing I forgot to say here, is that we also want to use the features arguments of `load_dataset` (which goes in the builder’s config) to override the default features of a dataset script." ]