Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "Hi ! We plan to add streaming features in the future.\r\n\r\nThis should allow to load a dataset instantaneously without generating the arrow table. The trade-off is that accessing examples from a streaming dataset must be done in an iterative way, and with an additional (but hopefully minor) overhead.\r\nWhat do you think about this ?\r\n\r\nIf you have ideas or suggestions of what you expect from such features as a user, feel free to share them, this is really valuable to us !", "People mainly want this feature either because it takes too much time too make arrow tables, or they occupy too much memory on the disk. I think both the problem can be solved if we provide arrow tables themselves on datasets hub. Can we do this currently @lhoestq ? \r\n", "@lhoestq I think the ```try_from_hf_gcs``` provide the same functionality. What all datasets are available on HF GCS? Are all the datasets on huggingFace datasets hub are made available on GCS, automatically?", "Only datasets like wikipedia, wiki40b, wiki_dpr and natural questions are available already processed on the HF google storage. This is used to download directly the arrow file instead of building it from the original data files.", "@lhoestq How can we make sure that the data we upload on HuggingFace hub is available in form of preprocessed arrow files ?", "We're still working on this :) This will be available soon\r\nUsers will be able to put their processed arrow files on the Hub", "Hi! You can now use `Dataset.push_to_hub` to store preprocessed files on the Hub.\r\n\r\nAnd to avoid downloading preprocessed files, you can use streaming by setting `streaming=True` in `load_dataset`." ]

[ "Hi ! We are using the [datasets-tagging app](https://github.com/huggingface/datasets-tagging) to select the tags to add.\r\n\r\nWe are also adding the full list of tags in #2107 \r\nThis covers multilinguality, language_creators, licenses, size_categories and task_categories.\r\n\r\nIn general if you want to add a tag that doesn't exist (for example for a custom license) you must make it start with `other-` and then a custom tag name.\r\n\r\nedit (@theo-m) if you ever find yourself resorting to adding an `other-*` tag, please do ping us somewhere so we can think about adding it to the \"official\" list :)", "@lhoestq hmm - ok thanks for the answer.\r\nTo be honest I am not sure if this issue can be closed now.\r\nI just wanted to point out that this should either be documented or linked in the documentation.\r\nIf you feel like it is (will be) please just close this.", "We're still working on the validation+documentation in this.\r\nFeel free to keep this issue open till we've added them", "@lhoestq what is the status on this? Did you add documentation?", "Hi ! There's the tagging app at https://huggingface.co/datasets/tagging/ that you can use.\r\nIt shows the list of all the tags you can use.\r\n\r\nIt is based on all the tag sets defined in this folder:\r\nhttps://github.com/huggingface/datasets/tree/master/src/datasets/utils/resources", "@lhoestq is there something like this form Models?", "I don't think so. Feel free to take a look at the tags of other models (example [here](https://huggingface.co/bert-base-uncased/blob/main/README.md)). But we should definitely have some docs or an app to write the tags. Feel free to open an issue in the `transformers` repo or in the `huggingface_hub` repo so we can discuss this", "When modifying a README file, the Hub now displays a special UI with allowed values (see https://huggingface.co/docs/datasets/main/en/upload_dataset#create-a-dataset-card)." ]

[ "+1 on this request" ]

[ "Hi,\r\n\r\nthis bug is related to `Dataset.{remove_columns, rename_column, flatten}` not propagating the change to the schema metadata when the info features are updated, so this line is the culprit:\r\n```python\r\ncommon_voice_train = common_voice_train.remove_columns(['client_id', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'])\r\n\r\n``` \r\nThe order is important because the resulting dataset inherits the schema metadata of the first dataset passed to the `concatenate_datasets(...)` function (`pa.concat_tables` [docs](https://arrow.apache.org/docs/python/generated/pyarrow.concat_tables.html)). I'll try to fix this ASAP." ]

[ "Hi !\r\n\r\nThis is actually supported ! but not yet in `from_pandas`.\r\nYou can use `from_dict` for now instead:\r\n```python\r\nfrom datasets import Dataset, Array2D, Features, Value\r\nimport pandas as pd\r\nimport numpy as np\r\n\r\ndataset = {\r\n 'bbox': [\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]])\r\n ],\r\n 'input_ids': [1, 2, 3, 4]\r\n}\r\ndataset = Dataset.from_dict(dataset)\r\n```\r\n\r\nThis will work but to use it with the torch formatter you must specify the `Array2D` feature type in order to tell the shape:\r\n```python\r\nfrom datasets import Dataset, Array2D, Features, Value\r\nimport pandas as pd\r\nimport numpy as np\r\n\r\ndataset = {\r\n 'bbox': [\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]]),\r\n np.array([[1,2,3,4],[1,2,3,4],[1,2,3,4]])\r\n ],\r\n 'input_ids': [1, 2, 3, 4]\r\n}\r\ndataset = Dataset.from_dict(dataset, features=Features({\r\n \"bbox\": Array2D(shape=(3, 4), dtype=\"int64\"),\r\n \"input_ids\": Value(\"int64\")\r\n}))\r\ndataset.set_format(\"torch\")\r\nprint(dataset[0]['bbox'])\r\n# tensor([[1, 2, 3, 4],\r\n# [1, 2, 3, 4],\r\n# [1, 2, 3, 4]])\r\n```\r\nIf you don't specify the `Array2D` feature type, then the inferred type will be Sequence(Sequence(Value(\"int64\"))) and therefore the torch formatter will return list of tensors", "Thanks for the explanation. \r\nWith my original DataFrame, I did\r\n```\r\ndataset = dataset.to_dict(\"list\")\r\n```\r\nand then the rest of the transformation from dictionary works just fine." ]

[ "Hi ! Thanks for reporting.\r\nWe're indeed using `jiwer` to compute the WER.\r\n\r\nMaybe instead of calling `jiwer.wer` once for all the preditions/references we can compute the WER iteratively to avoid memory issues ? I'm not too familial with `jiwer` but this must be possible.\r\n\r\nCurrently the code to compute the WER is defined here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/349ac4398a3bcae6356f14c5754483383a60e8a4/metrics/wer/wer.py#L93-L94", "Hi,\r\n\r\nI've just pushed a pull request that is related to this issue https://github.com/huggingface/datasets/pull/2169. It's not iterative, but it should avoid memory errors. It's based on the editdistance python library. An iterative implementation should be as easy as storing scores and words stepwise and dividing at the end. ", "I see, this was solved by other thread. Ok, let me know if you want to switch the implementation for any reason :)", "Thanks for diving into this anyway ^^'\r\nAs you said this actually got solved a few days ago", "Someone created an issue https://github.com/jitsi/jiwer/issues/40 at jiwer which shows that this is still a problem in the current version. Would be curious to figure out how this can be fixed by jiwer... :) I assume that it runs of out memory because it's trying to compute the WER over (too many) test samples?", "Hi !\r\n\r\nIt's computed iteratively so not sure what could go wrong\r\n\r\nhttps://github.com/huggingface/datasets/blob/8afd0ba8c27800a55ea69d9fcd702dc97d9c16d8/metrics/wer/wer.py#L100-L106\r\n\r\n@NiklasHoltmeyer what version of `datasets` are you running ?\r\n", "One possible explanation might be that it is the user who is passing all the sentences in a single element to `wer.compute`?\r\n\r\nAs current implementation iterates over the elements of `predictions` and `references`, this can be problematic if `predictions` and `references` contain a single huge element each. \r\n\r\nThis could be the case, for example, with a single string with all sentences:\r\n```python\r\nresult[\"predicted\"] = \"One sentence. Other sentence.\"\r\n```\r\nor with a __double__ nested list of sentence lists\r\n```python\r\nresult[\"predicted\"] = [[ [\"One sentence.\"], [\"Other sentence\"] ]]\r\n```\r\n\r\nThe user should check the dimensions of the data structure passed to `predictions` and `references`.", "Hi all,\r\n\r\nin my case I was using and older version of datasets and, as @albertvillanova points out, passing the full list of sentences for the metric calculation. The problem was in the way jiwer implements WER, as it tries to compute WER for the full list at once instead of doing it element-wise. I think that with the latest implementation of datasets, or by using the alternative WER function that I've contributed on this [pull request](https://github.com/huggingface/datasets/pull/2169) there shouldn't be memory errors.", "@lhoestq i was using Datasets==1.5.0 with 1.6.1 it worked (atleast the first run) but 1.5.0 is not compatible with my preprocessing. i cant save my dataset to a parquet file while using the latest datasets version\r\n\r\n-> \r\n```\r\n File \"../preprocess_dataset.py\", line 132, in <module>\r\n pq.write_table(train_dataset.data, f'{resampled_data_dir}/{data_args.dataset_config_name}.train.parquet')\r\n File \"/usr/local/lib/python3.8/dist-packages/pyarrow/parquet.py\", line 1674, in write_table\r\n writer.write_table(table, row_group_size=row_group_size)\r\n File \"/usr/local/lib/python3.8/dist-packages/pyarrow/parquet.py\", line 588, in write_table\r\n self.writer.write_table(table, row_group_size=row_group_size)\r\nTypeError: Argument 'table' has incorrect type (expected pyarrow.lib.Table, got ConcatenationTable)\r\n``` \r\n\r\nif i do \r\n```\r\nimport pyarrow.parquet as pq\r\n...\r\n...\r\npq.write_table(train_dataset.data, 'train.parquet')\r\npq.write_table(eval_dataset.data, 'eval.parquet')\r\n```\r\n\r\nwhile using 1.6.1. and its working with 1.5.0\r\n", "Hi ! You can pass dataset.data.table instead of dataset.data to pq.write_table", "This seems to be working so far! Thanks!" ]

[ "Hi @XuhuiZhou, thanks for reporting this issue. \r\n\r\nIndeed, the old links are no longer valid (404 Not Found error), and the script must be updated with the new links to Google Drive.", "It would be nice to update the urls indeed !\r\n\r\nTo do this, you just need to replace the urls in `iwslt2017.py` and then update the dataset_infos.json file with\r\n```\r\ndatasets-cli test ./datasets/iwslt2017 --all_configs --save_infos --ignore_verifications\r\n```", "Is this a command to update my local files or fix the file Github repo in general? (I am not so familiar with the datasets-cli command here)\r\n\r\nI also took a brief look at the **Sharing your dataset** section, looks like I could fix that locally and push it to the repo? I guess we are \"canonical\" category?", "This command will update your local file. Then you can open a Pull Request to push your fix to the github repo :)\r\nAnd yes you are right, it is a \"canonical\" dataset, i.e. a dataset script defined in this github repo (as opposed to dataset repositories of users on the huggingface hub)", "Hi, thanks for the answer. \r\n\r\nI gave a try to the problem today. But I encountered an upload error: \r\n\r\n```\r\ngit push -u origin fix_link_iwslt\r\nEnter passphrase for key '/home2/xuhuizh/.ssh/id_rsa': \r\nERROR: Permission to huggingface/datasets.git denied to XuhuiZhou.\r\nfatal: Could not read from remote repository.\r\n\r\nPlease make sure you have the correct access rights\r\nand the repository exists.\r\n```\r\n\r\nAny insight here? \r\n\r\nBy the way, when I run the datasets-cli command, it shows the following error, but does not seem to be the error coming from `iwslt.py`\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/home2/xuhuizh/anaconda3/envs/UMT/bin/datasets-cli\", line 33, in <module>\r\n sys.exit(load_entry_point('datasets', 'console_scripts', 'datasets-cli')())\r\n File \"/home2/xuhuizh/projects/datasets/src/datasets/commands/datasets_cli.py\", line 35, in main\r\n service.run()\r\n File \"/home2/xuhuizh/projects/datasets/src/datasets/commands/test.py\", line 141, in run\r\n try_from_hf_gcs=False,\r\n File \"/home2/xuhuizh/projects/datasets/src/datasets/builder.py\", line 579, in download_and_prepare\r\n dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n File \"/home2/xuhuizh/projects/datasets/src/datasets/builder.py\", line 639, in _download_and_prepare\r\n self.info.download_checksums, dl_manager.get_recorded_sizes_checksums(), \"dataset source files\"\r\n File \"/home2/xuhuizh/projects/datasets/src/datasets/utils/info_utils.py\", line 32, in verify_checksums\r\n raise ExpectedMoreDownloadedFiles(str(set(expected_checksums) - set(recorded_checksums)))\r\ndatasets.utils.info_utils.ExpectedMoreDownloadedFiles: {'https://wit3.fbk.eu/archive/2017-01-trnmted//texts/DeEnItNlRo/DeEnItNlRo/DeEnItNlRo-DeEnItNlRo.tgz'}\r\n```", "Hi ! To create a PR on this repo your must fork it and create a branch on your fork. See how to fork the repo [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md#start-by-preparing-your-environment).\r\nAnd to make the command work without the `ExpectedMoreDownloadedFiles` error, you just need to use the `--ignore_verifications` flag.", "Hi @XuhuiZhou,\r\n\r\nAs @lhoestq has well explained, you need to fork HF's repository, create a feature branch in your fork, push your changes to it and then open a Pull Request to HF's upstream repository. This is so because at HuggingFace Datasets we follow a development model called \"Fork and Pull Model\". You can find more information here:\r\n- [Understanding the GitHub flow](https://guides.github.com/introduction/flow/)\r\n- [Forking Projects](https://guides.github.com/activities/forking/)\r\n\r\nAlternatively, if you find all these steps too complicated, you can use the GitHub official command line tool: [GitHub CLI](https://cli.github.com/). Once installed, in order to create a Pull Request, you only need to use this command:\r\n```shell\r\ngh pr create --web\r\n```\r\nThis utility will automatically create the fork, push your changes and open a Pull Request, under the hood." ]

[ "Hi @LifaSun, thanks for reporting this issue.\r\n\r\nSometimes, GitHub has some connectivity problems. Could you confirm that the problem persists?", "@albertvillanova Thanks! It works well now. " ]

[ "dupe of #1992" ]

[ "Hi ! This error happens when you use `map` in batched mode and then your function doesn't return the same number of values per column.\r\n\r\nIndeed since you're using `map` in batched mode, `prepare_validation_features` must take a batch as input (i.e. a dictionary of multiple rows of the dataset), and return a batch.\r\n\r\nHowever it seems like `tokenized_examples` doesn't have the same number of elements in each field. One field seems to have `1180` elements while `candidate_attention_mask` only has `1178`." ]

[ "cc @philschmid ", "Hey @sivakhno,\r\n\r\nhow does your `requirements.txt` look like to install the `datasets` library and which version of it are you running? Can you try to install `datasets>=1.4.0`", "Hi @philschmid - thanks for suggestion. I am using `datasets==1.4.1`. \r\nI have also tried using `torch=1.6.0` (docker `763104351884.dkr.ecr.eu-central-1.amazonaws.com/pytorch-training:1.6.0-gpu-py3 `), but the error is the same. ", "Could paste the code you use the start your training job and the fine-tuning script you run? ", "@sivakhno this should be now fixed in `datasets>=1.5.0`. ", "@philschmid Recently released tensorflow-macos seems to be missing. ", "I've created a PR to add this. " ]

[ "Hi ! Thanks for reporting.\r\nThis looks like a bug, could you try to provide a minimal code example that reproduces the issue ? This would be very helpful !\r\n\r\nOtherwise I can try to run the wav2vec2 code above on my side but probably not this week..", "```\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset('glue', 'mrpc', split='train')\r\n\r\n\r\nupdated_dataset = dataset.map(lambda example: {'sentence1': 'My sentence: ' + example['sentence1']}, num_proc=4)\r\n\r\n```", "\r\n\r\n\r\n\r\n\r\nI was able to copy some of the shell \r\nThis is repeating every half second\r\nWin 10, Anaconda with python 3.8, datasets installed from main branche\r\n```\r\n\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 287, in _fixup_main_from_path\r\n _check_not_importing_main()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 116, in spawn_main\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 134, in _check_not_importing_main\r\n main_content = runpy.run_path(main_path,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 265, in run_path\r\n exitcode = _main(fd, parent_sentinel)\r\n raise RuntimeError('''\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 125, in _main\r\nRuntimeError:\r\n An attempt has been made to start a new process before the\r\n current process has finished its bootstrapping phase.\r\n\r\n This probably means that you are not using fork to start your\r\n child processes and you have forgotten to use the proper idiom\r\n in the main module:\r\n\r\n if __name__ == '__main__':\r\n freeze_support()\r\n ...\r\n\r\n The \"freeze_support()\" line can be omitted if the program\r\n is not going to be frozen to produce an executable. return _run_module_code(code, init_globals, run_name,\r\n prepare(preparation_data)\r\n\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 97, in _run_module_code\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 236, in prepare\r\n _run_code(code, mod_globals, init_globals,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 87, in _run_code\r\n _fixup_main_from_path(data['init_main_from_path'])\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 287, in _fixup_main_from_path\r\n exec(code, run_globals)\r\n File \"F:\\Codes\\Python Apps\\asr\\test.py\", line 6, in <module>\r\n updated_dataset = dataset.map(lambda example: {'sentence1': 'My sentence: ' + example['sentence1']}, num_proc=4)\r\n main_content = runpy.run_path(main_path,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\arrow_dataset.py\", line 1370, in map\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 265, in run_path\r\n with Pool(num_proc, initargs=(RLock(),), initializer=tqdm.set_lock) as pool:\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\context.py\", line 119, in Pool\r\n return _run_module_code(code, init_globals, run_name,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 97, in _run_module_code\r\n _run_code(code, mod_globals, init_globals,\r\n return Pool(processes, initializer, initargs, maxtasksperchild,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 87, in _run_code\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 212, in __init__\r\n exec(code, run_globals)\r\n File \"F:\\Codes\\Python Apps\\asr\\test.py\", line 6, in <module>\r\n self._repopulate_pool()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 303, in _repopulate_pool\r\n updated_dataset = dataset.map(lambda example: {'sentence1': 'My sentence: ' + example['sentence1']}, num_proc=4)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\arrow_dataset.py\", line 1370, in map\r\n return self._repopulate_pool_static(self._ctx, self.Process,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 326, in _repopulate_pool_static\r\n with Pool(num_proc, initargs=(RLock(),), initializer=tqdm.set_lock) as pool:\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\context.py\", line 119, in Pool\r\n w.start()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\process.py\", line 121, in start\r\n return Pool(processes, initializer, initargs, maxtasksperchild,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 212, in __init__\r\n self._popen = self._Popen(self)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\context.py\", line 327, in _Popen\r\n self._repopulate_pool()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 303, in _repopulate_pool\r\n return Popen(process_obj)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\popen_spawn_win32.py\", line 45, in __init__\r\n return self._repopulate_pool_static(self._ctx, self.Process,\r\n prep_data = spawn.get_preparation_data(process_obj._name)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\pool.py\", line 326, in _repopulate_pool_static\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 154, in get_preparation_data\r\n _check_not_importing_main()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 134, in _check_not_importing_main\r\n w.start()\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\process.py\", line 121, in start\r\n raise RuntimeError('''\r\nRuntimeError:\r\n An attempt has been made to start a new process before the\r\n current process has finished its bootstrapping phase.\r\n\r\n This probably means that you are not using fork to start your\r\n child processes and you have forgotten to use the proper idiom\r\n in the main module:\r\n\r\n if __name__ == '__main__':\r\n freeze_support()\r\n ...\r\n```", "Thanks this is really helpful !\r\nI'll try to reproduce on my side and come back to you", "if __name__ == '__main__':\r\n\r\n\r\nThis line before calling the map function stops the error but the script still repeats endless", "Indeed you needed `if __name__ == '__main__'` since accoding to [this stackoverflow post](https://stackoverflow.com/a/18205006):\r\n\r\n> On Windows the subprocesses will import (i.e. execute) the main module at start. You need to insert an if __name__ == '__main__': guard in the main module to avoid creating subprocesses recursively.\r\n\r\nRegarding the hanging issue, can you try to update `dill` and `multiprocess` ?", "It's already on the newest version", "```\r\nTraceback (most recent call last):\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\shutil.py\", line 791, in move\r\n os.rename(src, real_dst)\r\nFileExistsError: [WinError 183] Eine Datei kann nicht erstellt werden, wenn sie bereits vorhanden ist: 'D:\\\\huggingfacecache\\\\common_voice\\\\de\\\\6.1.0\\\\0041e06ab061b91d0a23234a2221e87970a19cf3a81b20901474cffffeb7869f\\\\tmpx9fl_jg8' -> 'D:\\\\huggingfacecache\\\\common_voice\\\\de\\\\6.1.0\\\\0041e06ab061b91d0a23234a2221e87970a19cf3a81b20901474cffffeb7869f\\\\cache-9b4f203a63742dfc.arrow'\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 116, in spawn_main\r\n exitcode = _main(fd, parent_sentinel)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 125, in _main\r\n prepare(preparation_data)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 236, in prepare\r\n _fixup_main_from_path(data['init_main_from_path'])\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\multiprocess\\spawn.py\", line 287, in _fixup_main_from_path\r\n main_content = runpy.run_path(main_path,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 265, in run_path\r\n return _run_module_code(code, init_globals, run_name,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 97, in _run_module_code\r\n _run_code(code, mod_globals, init_globals,\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\runpy.py\", line 87, in _run_code\r\n exec(code, run_globals)\r\n File \"F:\\Codes\\Python Apps\\asr\\cvtrain.py\", line 243, in <module>\r\n common_voice_train = common_voice_train.map(remove_special_characters, remove_columns=[\"sentence\"])\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\arrow_dataset.py\", line 1339, in map\r\n return self._map_single(\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\arrow_dataset.py\", line 203, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\fingerprint.py\", line 337, in wrapper\r\n out = func(self, *args, **kwargs)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\site-packages\\datasets\\arrow_dataset.py\", line 1646, in _map_single\r\n shutil.move(tmp_file.name, cache_file_name)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\shutil.py\", line 805, in move\r\n copy_function(src, real_dst)\r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\shutil.py\", line 435, in copy2\r\n copyfile(src, dst, follow_symlinks=follow_symlinks)\r\n 0%| | 0/27771 [00:00<?, ?ex/s] \r\n File \"C:\\Users\\flozi\\anaconda3\\envs\\wav2vec\\lib\\shutil.py\", line 264, in copyfile\r\n with open(src, 'rb') as fsrc, open(dst, 'wb') as fdst:\r\nOSError: [Errno 22] Invalid argument: 'D:\\\\huggingfacecache\\\\common_voice\\\\de\\\\6.1.0\\\\0041e06ab061b91d0a23234a2221e87970a19cf3a81b20901474cffffeb7869f\\\\cache-9b4f203a63742dfc.arrow'\r\n```\r\n\r\nI was adding freeze support before calling the mapping function like this\r\nif __name__ == '__main__':\r\n freeze_support()\r\n dataset.map(....)", "Usually OSError of an arrow file on windows means that the file is currently opened as a dataset object, so you can't overwrite it until the dataset object falls out of scope.\r\nCan you make sure that there's no dataset object that loaded the `cache-9b4f203a63742dfc.arrow` file ?", "Now I understand\r\nThe error occures because the script got restarted in another thread, so the object is already loaded.\r\nStill don't have an idea why a new thread starts the whole script again" ]

[ "Hi ! Thanks for reporting.\r\n\r\nCurrently there's no way to specify this.\r\n\r\nWhen loading/processing a dataset, the arrow file is written using a temporary file. Then once writing is finished, it's moved to the cache directory (using `shutil.move` [here](https://github.com/huggingface/datasets/blob/f6b8251eb975f66a568356d2a40d86442c03beb9/src/datasets/arrow_dataset.py#L1646))\r\n\r\nThat means it keeps the permissions specified by the `tempfile.NamedTemporaryFile` object, i.e. `-rw-------` instead of `-rw-r--r--`. Improving this could be a nice first contribution to the library :)", "Hi @lhoestq,\r\nI looked into this and yes you're right. The `NamedTemporaryFile` is always created with mode 0600, which prevents group from reading the file. Should we change the permissions of `tmp_file.name` [here](https://github.com/huggingface/datasets/blob/f6b8251eb975f66a568356d2a40d86442c03beb9/src/datasets/arrow_dataset.py#L1871) and [here](https://github.com/huggingface/datasets/blob/f6b8251eb975f66a568356d2a40d86442c03beb9/src/datasets/arrow_dataset.py#L1590), post creation to 0644 inorder for group and others to read it?", "Good idea :) we could even update the permissions after the file has been moved by shutil.move [here](https://github.com/huggingface/datasets/blob/f6b8251eb975f66a568356d2a40d86442c03beb9/src/datasets/arrow_dataset.py#L1899) and [here](https://github.com/huggingface/datasets/blob/f6b8251eb975f66a568356d2a40d86442c03beb9/src/datasets/arrow_dataset.py#L1646) actually.\r\nApparently they set the default 0600 for temporary files for security reasons, so let's update the umask only after the file has been moved", "Would it be possible to actually set the umask based on a user provided argument? For example, a popular usecase my team has is using a shared file-system for processing datasets. This may involve writing/deleting other files, or changing filenames, which a -rw-r--r-- wouldn't fix. ", "Note that you can get the cache files of a dataset with the `cache_files` attributes.\r\nThen you can `chmod` those files and all the other cache files in the same directory.\r\n\r\nMoreover we can probably keep the same permissions after each transform. This way you just need to set the permissions once after doing `load_dataset` for example, and then all the new transformed cached files will have the same permissions.\r\nWhat do you think ?", "This means we'll check the permissions of other `cache_files` already created for a dataset before setting permissions for new `cache_files`?", "You can just check the permission of `dataset.cache_files[0]` imo", "> This way you just need to set the permissions once after doing load_dataset for example, and then all the new transformed cached files will have the same permissions.\r\n\r\nI was referring to this. Ensuring that newly generated `cache_files` have the same permissions", "Yes exactly\r\n\r\nI imagine users can first do `load_dataset`, then chmod on the arrow files. After that all the new cache files could have the same permissions as the first arrow files. Opinions on this ?", "Sounds nice but I feel this is a sub-part of the approach mentioned by @siddk. Instead of letting the user set new permissions by itself first and then making sure newly generated files have same permissions why don't we ask the user initially only what they want? What are your thoughts?", "Yes sounds good. Should this be a parameter in `load_dataset` ? Or an env variable ? Or use the value of `os.umask` ?", "Ideally it should be a parameter in `load_dataset` but I'm not sure how important it is for the users (considering only important things should go into `load_dataset` parameters)", "I think it's fairly important; for context, our team uses a shared file-system where many folks run experiments based on datasets that are cached by other users.\r\n\r\nFor example, I might start a training run, downloading a dataset. Then, a couple of days later, a collaborator using the same repository might want to use the same dataset on the same shared filesystem, but won't be able to under the default permissions.\r\n\r\nBeing able to specify directly in the top-level `load_dataset()` call seems important, but an equally valid option would be to just inherit from the running user's `umask` (this should probably be the default anyway).\r\n\r\nSo basically, argument that takes a custom set of permissions, and by default, use the running user's umask!", "Maybe let's start by defaulting to the user's umask !\r\nDo you want to give it a try @bhavitvyamalik ?", "Yeah sure! Instead of using default `0o644` should I first extract umask of current user and then use `os.umask` on it? We can do it inside `Dataset` class so that all folders/files created during the call use running user's umask\r\n\r\n", "You can get the umask using `os.umask` and then I guess you can just use `os.chmod` as in your previous PR, but with the right permissions depending on the umask.", "FWIW, we have this issue with other caches - e.g. `transformers` model files. So probably will need to backport this into `transformers` as well.\r\n\r\nthanks @thomwolf for the pointer.", "Hi @stas00,\r\nFor this should we use the same umask code in the respective model directory inside `TRANSFORMERS_CACHE`?", "That sounds very right to me, @bhavitvyamalik ", "The cluster I am working on does not allow me to change the permission of the files with os.chmod. I was wondering if there is any workaround for this? My cache is in a GCP bucket and I can't change file permissions once I mount it.", "@vmurahari3 what error do you have exactly ?", "I get a permission denied error on https://github.com/huggingface/datasets/blob/b8363e0539c6f0cb5de49af32962cf2eb4c47395/src/datasets/arrow_dataset.py#L2799. I suspect I don't have permissions to change group permissions. I am mounting a GCP bucket through [gcsfuse](https://github.com/GoogleCloudPlatform/gcsfuse). ", "What @lhoestq is asking for is the full multi-line traceback - it's almost never enough to show the last line - a full stack is needed to get the context. Thank you!\r\n\r\nI wonder if a workaround is to try/except and then issue a warning if this fails?", "Hello, I'm working on a project with a very similar setup to the one mentioned by @siddk, namely we have a shared cache directory in the team that we wish to use to avoid redundant dataset downloads. However, we're hitting `PermissionError` when one member of the team tries to reload a dataset that was downloaded by another team member (stack trace below).\r\n\r\nConcretely, we first create a shared directory and give everyone read, write, execute permissions as follows (I know this isn't best practice 🤫):\r\n\r\n```shell\r\nmkdir shared-folder\r\nchmod -R 777 shared-folder\r\n```\r\n\r\nWe then set the following in our `.bashrc` profiles:\r\n\r\n```\r\n# Hugging Face caches\r\nexport HUGGINGFACE_HUB_CACHE=/path/to/shared-folder\r\nexport HF_DATASETS_CACHE=/path/to/shared-folder\r\n\r\n# For shared access to the shared-folder directory\r\numask 000\r\n```\r\n\r\nNow, running e.g. `load_dataset(\"emotion\")` the first time works (as expected), but when another team member tries to load from cache we get something like:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 1759, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 1496, in load_dataset_builder\r\n dataset_module = dataset_module_factory(\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 1218, in dataset_module_factory\r\n raise e1 from None\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 1193, in dataset_module_factory\r\n ).get_module()\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 903, in get_module\r\n local_path = self.download_loading_script()\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/load.py\", line 871, in download_loading_script\r\n return cached_path(file_path, download_config=download_config)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/utils/file_utils.py\", line 210, in cached_path\r\n output_path = ExtractManager(cache_dir=download_config.cache_dir).extract(\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/utils/extract.py\", line 42, in extract\r\n extractor_format = self.extractor.infer_extractor_format(input_path)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/utils/extract.py\", line 287, in infer_extractor_format\r\n if extractor.is_extractable(path, magic_number=magic_number):\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/site-packages/datasets/utils/extract.py\", line 84, in is_extractable\r\n return tarfile.is_tarfile(path)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/tarfile.py\", line 2517, in is_tarfile\r\n t = open(name)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/tarfile.py\", line 1632, in open\r\n return func(name, \"r\", fileobj, **kwargs)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/tarfile.py\", line 1698, in gzopen\r\n fileobj = GzipFile(name, mode + \"b\", compresslevel, fileobj)\r\n File \"~/miniconda3/envs/llama-nol/lib/python3.10/gzip.py\", line 174, in __init__\r\n fileobj = self.myfileobj = builtins.open(filename, mode or 'rb')\r\nPermissionError: [Errno 13] Permission denied: '/path/to/shared-folder/downloads/4e7db366b1ea045d0faa083a2e47ac87326ad8e653f894763b0982c2a1e94078.cc96367835404d4195bf75b2602e6dbbfd2da9288170fc0b2298fc0e376ff52a.py'\r\n```\r\n\r\nIf I understand this comment from @bhavitvyamalik:\r\n\r\n> Yeah sure! Instead of using default `0o644` should I first extract umask of current user and then use `os.umask` on it? We can do it inside `Dataset` class so that all folders/files created during the call use running user's umask\r\n\r\nthe goal was to infer the `umask` of the user profile, but perhaps I misunderstood and the true solution is to chmod the `cache_files` as @lhoestq suggests above - is that correct?\r\n\r\ncc @natolambert @nazneenrajani @edbeeching ", "Python files are stored in the modules caches, that you can modify by setting `HF_MODULES_CACHE` as well and set appropriate permissions\r\n\r\nThey are in different locations because the `HF_MODULES_CACHE` is added to the python path to be able to import the dataset scripts :)", "Thanks a lot for the tip @lhoestq ! I didn't know about this extra cache - thanks :)" ]

[ "@lhoestq Adding \"_\" to the class labels in the dataset script will fix the issue.\r\n\r\nThe bigger issue IMO is that the data files are in conll format, but the examples are tokens, not sentences.", "Hi ! Thanks for reporting @adzcodez \r\n\r\n\r\n> @lhoestq Adding \"_\" to the class labels in the dataset script will fix the issue.\r\n> \r\n> The bigger issue IMO is that the data files are in conll format, but the examples are tokens, not sentences.\r\n\r\nYou're right: \"_\" should be added to the list of labels, and the examples must be sequences of tokens, not singles tokens.\r\n", "@lhoestq Can you please label this issue with the \"good first issue\" label? I'm not sure I'll find time to fix this.\r\n\r\nTo resolve it, the user should:\r\n1. add `\"_\"` to the list of labels\r\n2. transform the udpos subset to the conll format (I think the preprocessing logic can be borrowed from [the original repo](https://github.com/google-research/xtreme/blob/58a76a0d02458c4b3b6a742d3fd4ffaca80ff0de/utils_preprocess.py#L187-L204))\r\n3. update the dummy data\r\n4. update the dataset info\r\n5. [optional] add info about the data fields structure of the udpos subset to the dataset readme", "I tried fixing this issue, but its working fine in the dev version : \"1.6.2.dev0\"\r\n\r\nI think somebody already fixed it. ", "Hi,\r\n\r\nafter #2326, the lines with pos tags equal to `\"_\"` are filtered out when generating the dataset, so this fixes the KeyError described above. However, the udpos subset should be in the conll format i.e. it should yield sequences of tokens and not single tokens, so it would be great to see this fixed (feel free to borrow the logic from [here](https://github.com/google-research/xtreme/blob/58a76a0d02458c4b3b6a742d3fd4ffaca80ff0de/utils_preprocess.py#L187-L204) if you decide to work on this). ", "Closed by #2466." ]

[ "@skyprince999 as you authored the PR for this dataset, any comments?", "This has been fixed in #2064 by @mariosasko (thanks again !)\r\n\r\nThe fix is available on the master branch and we'll do a new release very soon :)" ]

[ "Hi! You can either save the TF dataset to one of the formats supported by datasets (`parquet`, `csv`, `json`, ...) or pass a generator function to `Dataset.from_generator` that yields its examples." ]

[ "@lhoestq I also deleted the cache and redownload the file and still the same issue, I appreciate any help on this. thanks ", "Here please find the minimal code to reproduce the issue @lhoestq note this only happens with MT5TokenizerFast\r\n\r\n```\r\nfrom datasets import load_dataset\r\nfrom transformers import MT5TokenizerFast\r\n\r\ndef get_tokenized_dataset(dataset_name, dataset_config_name, tokenizer):\r\n datasets = load_dataset(dataset_name, dataset_config_name, script_version=\"master\")\r\n column_names = datasets[\"train\"].column_names\r\n text_column_name = \"translation\"\r\n def process_dataset(datasets):\r\n def process_function(examples):\r\n lang = \"fr\"\r\n return {\"src_texts\": [example[lang] for example in examples[text_column_name]]}\r\n datasets = datasets.map(\r\n process_function,\r\n batched=True,\r\n num_proc=None,\r\n remove_columns=column_names,\r\n load_from_cache_file=True,\r\n )\r\n return datasets\r\n datasets = process_dataset(datasets)\r\n text_column_name = \"src_texts\"\r\n column_names = [\"src_texts\"]\r\n def tokenize_function(examples):\r\n return tokenizer(examples[text_column_name], return_special_tokens_mask=True)\r\n tokenized_datasets = datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n num_proc=None,\r\n remove_columns=column_names,\r\n load_from_cache_file=True\r\n )\r\n\r\nif __name__ == \"__main__\":\r\n tokenizer_kwargs = {\r\n \"cache_dir\": None,\r\n \"use_fast\": True,\r\n \"revision\": \"main\",\r\n \"use_auth_token\": None\r\n }\r\n tokenizer = MT5TokenizerFast.from_pretrained(\"google/mt5-small\", **tokenizer_kwargs)\r\n get_tokenized_dataset(dataset_name=\"opus100\", dataset_config_name=\"en-fr\", tokenizer=tokenizer)\r\n~ \r\n```", "as per https://github.com/huggingface/tokenizers/issues/626 this looks like to be the tokenizer bug, I therefore, reported it there https://github.com/huggingface/tokenizers/issues/626 and I am closing this one." ]

[ "Hi\r\nYou can rename the arrow file and update the name in `state.json`", "I tried this way, but when there is a mapping process to the dataset, it again uses a random cache name. atm, I am trying to use the following method by setting an exact cache file,\r\n\r\n```\r\n dataset_with_embedding =csv_dataset.map(\r\n partial(self.embed, ctx_encoder=ctx_encoder, ctx_tokenizer=self.context_tokenizer),\r\n batched=True,\r\n batch_size=1,\r\n features=new_features,\r\n cache_file_name=cache_arrow_path,\r\n load_from_cache_file=False\r\n )\r\n```\r\nSo here we set a cache_file_name , after this it uses the same file name when saving again and again. ", "I'm not sure I understand your issue, can you elaborate ?\r\n\r\n`cache_file_name` is indeed an argument you can set to specify the cache file that will be used for the processed dataset. By default the file is named with something like `cache-<fingerprint>.arrow` where the fingerprint is a hash.", "Let's say I am updating a set of embedding in a dataset that is around 40GB inside a training loop every 500 steps (Ex: calculating the embeddings in updated ctx_encoder in RAG and saving it to the passage path). So when we use **dataset_object.save_to_disk('passage_path_directory')** it will save the new dataset object every time with a random file name, especially when we do some transformations to dataset objects such as map or shards. This way, we keep collecting unwanted files that will eventually eat up all the disk space. \r\n\r\nBut if we can save the dataset object every time by a single name like **data_shard_1.arrow**, it will automatically remove the previous file and save the new one in the same directory. I found the above-mentioned code snippet useful to complete this task. \r\n\r\nIs this clear?" ]

[ "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!" ]

[ "Hi,\r\n\r\nthis was fixed by #1995, so you can wait for the next release or install the package directly from the master branch with the following command: \r\n```bash\r\npip install git+https://github.com/huggingface/datasets\r\n```", "Ty!" ]

[ "@lhoestq - We could simply use the \"general\" json dataset for this no? ", "Sure you can use the json loader\r\n```python\r\ndata_files = {\"train\": \"path/to/your/train_data.json\", \"test\": \"path/to/your/test_data.json\"}\r\ntrain_dataset = load_dataset(\"json\", data_files=data_files, split=\"train\")\r\ntest_dataset = load_dataset(\"json\", data_files=data_files, split=\"test\")\r\n```\r\n\r\nYou just need to make sure that the data contain the paths to the audio files.\r\nIf not, feel free to use `.map()` to add them.", "Many thanks! that was what I was looking for. " ]

[]

[ "I think faiss automatically sets the number of threads to use to build the index.\r\nCan you check how many CPU cores are being used when you build the index in `use_own_knowleldge_dataset` as compared to this script ? Are there other programs running (maybe for rank>0) ?", "Hi,\r\n I am running the add_faiss_index during the training process of the RAG from the master process (rank 0). But at the exact moment, I do not run any other process since I do it in every 5000 training steps. \r\n \r\n I think what you say is correct. It depends on the number of CPU cores. I did an experiment to compare the time taken to finish the add_faiss_index process on use_own_knowleldge_dataset.py vs the training loop thing. The training loop thing takes 40 mins more. It might be natural right? \r\n \r\n \r\n at the moment it uses around 40 cores of a 96 core machine (I am fine-tuning the entire process). ", "Can you try to set the number of threads manually ?\r\nIf you set the same number of threads for both the `use_own_knowledge_dataset.py` and RAG training, it should take the same amount of time.\r\nYou can see how to set the number of thread in the faiss wiki: https://github.com/facebookresearch/faiss/wiki/Threads-and-asynchronous-calls", "Ok, I will report the details too soon. I am the first one on the list and currently add_index being computed for the 3rd time in the loop. Actually seems like the time is taken to complete each interaction is the same, but around 1 hour more compared to running it without the training loop. A the moment this takes 5hrs and 30 mins. If there is any way to faster the process, an end-to-end rag will be perfect. So I will also try out with different thread numbers too. \r\n\r\n\r\n", "@lhoestq on a different note, I read about using Faiss-GPU, but the documentation says we should use it when the dataset has the ability to fit into the GPU memory. Although this might work, in the long-term this is not that practical for me.\r\n\r\nhttps://github.com/matsui528/faiss_tips", "@lhoestq \r\n\r\nHi, I executed the **use_own_dataset.py** script independently and ask a few of my friends to run their programs in the HPC machine at the same time. \r\n\r\n Once there are so many other processes are running the add_index function gets slows down naturally. So basically the speed of the add_index depends entirely on the number of CPU processes. Then I set the number of threads as you have mentioned and got actually the same time for RAG training and independat running. So you are correct! :) \r\n\r\n \r\n Then I added this [issue in Faiss repostiary](https://github.com/facebookresearch/faiss/issues/1767). I got an answer saying our current **IndexHNSWFlat** can get slow for 30 million vectors and it would be better to use alternatives. What do you think?", "It's a matter of tradeoffs.\r\nHSNW is fast at query time but takes some time to build.\r\nA flat index is flat to build but is \"slow\" at query time.\r\nAn IVF index is probably a good choice for you: fast building and fast queries (but still slower queries than HSNW).\r\n\r\nNote that for an IVF index you would need to have an `nprobe` parameter (number of cells to visit for one query, there are `nlist` in total) that is not too small in order to have good retrieval accuracy, but not too big otherwise the queries will take too much time. From the faiss documentation:\r\n> The nprobe parameter is always a way of adjusting the tradeoff between speed and accuracy of the result. Setting nprobe = nlist gives the same result as the brute-force search (but slower).\r\n\r\nFrom my experience with indexes on DPR embeddings, setting nprobe around 1/4 of nlist gives really good retrieval accuracy and there's no need to have a value higher than that (or you would need to brute-force in order to see a difference).", "@lhoestq \r\n\r\nThanks a lot for sharing all this prior knowledge. \r\n\r\nJust asking what would be a good nlist of parameters for 30 million embeddings?", "When IVF is used alone, nlist should be between `4*sqrt(n)` and `16*sqrt(n)`.\r\nFor more details take a look at [this section of the Faiss wiki](https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index#how-big-is-the-dataset)", "Thanks a lot. I was lost with calling the index from class and using faiss_index_factory. ", "@lhoestq Thanks a lot for the help you have given to solve this issue. As per my experiments, IVF index suits well for my case and it is a lot faster. The use of this can make the entire RAG end-to-end trainable lot faster. So I will close this issue. Will do the final PR soon. " ]

[ "Hi ! To help me understand the situation, can you print the values of `load_from_disk(PATH_DATA_CLS_A)['train']._indices_data_files` and `load_from_disk(PATH_DATA_CLS_B)['train']._indices_data_files` ?\r\nThey should both have a path to an arrow file\r\n\r\nAlso note that from #2025 concatenating datasets will no longer have such restrictions.", "Sure, thanks for the fast reply!\r\n\r\nFor dataset A: `[{'filename': 'drive/MyDrive/data_target_task/dataset_a/train/cache-4797266bf4db1eb7.arrow'}]`\r\nFor dataset B: `[]`\r\n\r\nNo clue why for B it returns nothing. `PATH_DATA_CLS_B` is exactly the same in `save_to_disk` and `load_from_disk`... Also I can verify that the folder physically exists under 'drive/MyDrive/data_target_task/dataset_b/'", "In the next release you'll be able to concatenate any kinds of dataset (either from memory or from disk).\r\n\r\nFor now I'd suggest you to flatten the indices of the A and B datasets. This will remove the indices mapping and you will be able to concatenate them. You can flatten the indices with\r\n```python\r\ndataset = dataset.flatten_indices()\r\n```", "Indeed this works. Not the most elegant solution, but it does the trick. Thanks a lot! " ]

[ "Hi ! Thanks for reporting.\r\n\r\nTo update the dataset_infos.json you can run:\r\n```\r\ndatasets-cli test ./datasets/doc2dial --all_configs --save_infos --ignore_verifications\r\n```", "Fixed by #2041, thanks again @songfeng !" ]

[ "Solved!" ]

[ "Dear @lhoestq for wikipedia dataset I also get the same error, I greatly appreciate if you could have a look into this dataset as well. Below please find the command to reproduce the error:\r\n\r\n```\r\ndataset = load_dataset(\"wikipedia\", \"20200501.bg\")\r\nprint(dataset)\r\n```\r\n\r\nYour library is my only chance to be able training the models at scale and I am grateful for your help.\r\n\r\n", "Hi @dorost1234,\r\nTry installing this library first, `pip install 'apache-beam[gcp]' --use-feature=2020-resolver` followed by loading dataset like this using beam runner.\r\n\r\n`dataset = load_dataset(\"wiki40b\", \"cs\", beam_runner='DirectRunner')`\r\n\r\n I also read in error stack trace that:\r\n\r\n> Trying to generate a dataset using Apache Beam, yet no Beam Runner or PipelineOptions() has been provided in `load_dataset` or in the builder arguments. For big datasets it has to run on large-scale data processing tools like Dataflow, Spark, etc.\r\n\r\nWorked perfectly fine after this (Ignore these warnings)\r\n\r\n\r\n\r\n", "For wikipedia dataset, looks like the files it's looking for are no longer available. For `bg`, I checked [here](https://dumps.wikimedia.org/bgwiki/). For this I think `dataset_infos.json` for this dataset has to made again? You'll have to load this dataset also using beam runner.\r\n\r\n", "Hello @dorost1234,\r\n\r\nIndeed, Wikipedia datasets need a lot of preprocessing and this is done using Apache Beam. That is the reason why it is required that you install Apache Beam in order to preform this preprocessing.\r\n\r\nFor some specific default parameters (English Wikipedia), Hugging Face has already preprocessed the dataset for you (and it is stored in the cloud). That is the reason why you do not get the error for English: the preprocessing is already done by HF and you just get the preprocessed dataset; Apache Beam is not required in that case.", "Hi\nI really appreciate if huggingface can kindly provide preprocessed\ndatasets, processing these datasets require sufficiently large resources\nand I do not have unfortunately access to, and perhaps many others too.\nthanks\n\nOn Fri, Mar 12, 2021 at 9:04 AM Albert Villanova del Moral <\n***@***.***> wrote:\n\n> Hello @dorost1234 <https://github.com/dorost1234>,\n>\n> Indeed, Wikipedia datasets need a lot of preprocessing and this is done\n> using Apache Beam. That is the reason why it is required that you install\n> Apache Beam in order to preform this preprocessing.\n>\n> For some specific default parameters (English Wikipedia), Hugging Face has\n> already preprocessed the dataset for you (and it is stored in the cloud).\n> That is the reason why you do not get the error for English: the\n> preprocessing is already done by HF and you just get the preprocessed\n> dataset; Apache Beam is not required in that case.\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/datasets/issues/2035#issuecomment-797310899>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMXACFQZAGMK4VGXRETTDHDI3ANCNFSM4ZA5R2UA>\n> .\n>\n", "Hi everyone\r\nthanks for the helpful pointers, I did it as @bhavitvyamalik suggested, for me this freezes on this command for several hours, \r\n\r\n`Downloading and preparing dataset wiki40b/cs (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /users/dara/cache/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f...\r\n`\r\n\r\nDo you know how long this takes? Any specific requirements the machine should have? like very large memory or so? @lhoestq \r\n\r\nthanks \r\n\r\n\r\n", "HI @dorost1234, \r\nThe dataset size is 631.84 MiB so depending on your internet speed it'll take some time. You can monitor your internet speed meanwhile to see if it's downloading the dataset or not (use `nload` if you're using linux/mac to monitor the same). In my case it took around 3-4 mins. Since they haven't used `download_and_extract` here that's why there's no download progress bar.", "Hi\r\nthanks, my internet speed should be good, but this really freezes for me, this is how I try to get this dataset:\r\n\r\n`from datasets import load_dataset\r\ndataset = load_dataset(\"wiki40b\", \"cs\", beam_runner='DirectRunner')`\r\n\r\nthe output I see if different also from what you see after writing this command:\r\n\r\n`Downloading and preparing dataset wiki40b/cs (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /users/dara/cache/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f...`\r\n\r\ndo you have any idea why it might get freezed? anything I am missing @lhoestq @bhavitvyamalik. Do I need maybe to set anything special for apache-beam? \r\n\r\nthanks a lot \r\n\r\nOn Tue, Mar 16, 2021 at 9:03 AM Bhavitvya Malik ***@***.***>\r\nwrote:\r\n\r\n> HI @dorost1234 <https://github.com/dorost1234>,\r\n> The dataset size is 631.84 MiB so depending on your internet speed it'll\r\n> take some time. You can monitor your internet speed meanwhile to see if\r\n> it's downloading the dataset or not (use nload if you're using linux/mac\r\n> to monitor the same). In my case it took around 3-4 mins. Since they\r\n> haven't used download_and_extract here that's why there's no download\r\n> progress bar.\r\n>\r\n> —\r\n> You are receiving this because you were mentioned.\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/2035#issuecomment-800044303>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AS37NMQIHNNLM2LGG6QKZ73TD4GDJANCNFSM4ZA5R2UA>\r\n> .\r\n>\r\n", "I tried this on another machine (followed the same procedure I've mentioned above). This is what it shows (during the freeze period) for me:\r\n```\r\n>>> dataset = load_dataset(\"wiki40b\", \"cs\", beam_runner='DirectRunner')\r\nDownloading: 5.26kB [00:00, 1.23MB/s] \r\nDownloading: 1.40kB [00:00, 327kB/s] \r\nDownloading and preparing dataset wiki40b/cs (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/bhavitvya/.cache/huggingface/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f...\r\nWARNING:apache_beam.internal.gcp.auth:Unable to find default credentials to use: The Application Default Credentials are not available. They are available if running in Google Compute Engine. Otherwise, the environment variable GOOGLE_APPLICATION_CREDENTIALS must be defined pointing to a file defining the credentials. See https://developers.google.com/accounts/docs/application-default-credentials for more information.\r\nConnecting anonymously.\r\nWARNING:apache_beam.io.tfrecordio:Couldn't find python-snappy so the implementation of _TFRecordUtil._masked_crc32c is not as fast as it could be.\r\n```\r\nAfter around 10 minutes, here's the loading of dataset:\r\n```\r\n100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:16<00:00, 16.42s/sources]\r\n100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1.12sources/s]\r\n100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1.14sources/s]\r\nDataset wiki40b downloaded and prepared to /home/bhavitvya/.cache/huggingface/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f. Subsequent calls will reuse this data.\r\n```", "Hi\r\nI honestly also now tried on another machine and nothing shows up after\r\nhours of waiting. Are you sure you have not set any specific setting? maybe\r\ngoogle cloud which seems it is used here, needs some credential setting?\r\nthanks for any suggestions on this\r\n\r\nOn Tue, Mar 16, 2021 at 10:02 AM Bhavitvya Malik ***@***.***>\r\nwrote:\r\n\r\n> I tried this on another machine (followed the same procedure I've\r\n> mentioned above). This is what it shows (during the freeze period) for me:\r\n>\r\n> >>> dataset = load_dataset(\"wiki40b\", \"cs\", beam_runner='DirectRunner')\r\n> Downloading: 5.26kB [00:00, 1.23MB/s]\r\n> Downloading: 1.40kB [00:00, 327kB/s]\r\n> Downloading and preparing dataset wiki40b/cs (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/bhavitvya/.cache/huggingface/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f...\r\n> WARNING:apache_beam.internal.gcp.auth:Unable to find default credentials to use: The Application Default Credentials are not available. They are available if running in Google Compute Engine. Otherwise, the environment variable GOOGLE_APPLICATION_CREDENTIALS must be defined pointing to a file defining the credentials. See https://developers.google.com/accounts/docs/application-default-credentials for more information.\r\n> Connecting anonymously.\r\n> WARNING:apache_beam.io.tfrecordio:Couldn't find python-snappy so the implementation of _TFRecordUtil._masked_crc32c is not as fast as it could be.\r\n>\r\n> After around 10 minutes, here's the loading of dataset:\r\n>\r\n> 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:16<00:00, 16.42s/sources]\r\n> 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1.12sources/s]\r\n> 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1.14sources/s]\r\n> Dataset wiki40b downloaded and prepared to /home/bhavitvya/.cache/huggingface/datasets/wiki40b/cs/1.1.0/063778187363ffb294896eaa010fc254b42b73e31117c71573a953b0b0bf010f. Subsequent calls will reuse this data.\r\n>\r\n> —\r\n> You are receiving this because you were mentioned.\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/2035#issuecomment-800081772>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AS37NMX6A2ZTRZUIIZVFRCDTD4NC3ANCNFSM4ZA5R2UA>\r\n> .\r\n>\r\n", "Closing as `apache-beam`/`tensorflow_datasets` are no longer needed to load `wikipedia`/`wiki40b`." ]

[ "Actually table.filter returns a new table in memory, which can fill users RAM.\r\n\r\nTherefore it's not a good solution if we want to keep supporting bigger than RAM datastes" ]

[ "Hi @miyamonz \r\nThanks for investigating this issue, good job !\r\nIt would be awesome to integrate your fix in the library, could you open a pull request ?", "OK! I'll send it later." ]

[ "In your code `dataset2` doesn't contain the \"embeddings\" column, since it is created from the pandas DataFrame with columns \"text\" and \"label\".\r\n\r\nTherefore when you call `dataset2[embeddings_name]`, you get a `KeyError`.\r\n\r\nIf you want the \"embeddings\" column back, you can create `dataset2` with\r\n```python\r\ndataset2 = load_from_disk(dataset_filename)\r\n```\r\nwhere `dataset_filename` is the place where you saved you dataset with the embeddings in the first place.", "Ok in that case HF should fix their misleading example at https://huggingface.co/docs/datasets/faiss_and_ea.html#adding-a-faiss-index \r\n\r\nI copy-pasted it here.\r\n\r\n> When you are done with your queries you can save your index on disk:\r\n> \r\n> ```python\r\n> ds_with_embeddings.save_faiss_index('embeddings', 'my_index.faiss')\r\n> ```\r\n> Then reload it later:\r\n> \r\n> ```python\r\n> ds = load_dataset('crime_and_punish', split='train[:100]')\r\n> ds.load_faiss_index('embeddings', 'my_index.faiss')\r\n> ```", "Hi !\r\n\r\nThe code of the example is valid.\r\nAn index is a search engine, it's not considered a column of a dataset.\r\nWhen you do `ds.load_faiss_index(\"embeddings\", 'my_index.faiss')`, it attaches an index named \"embeddings\" to the dataset but it doesn't re-add the \"embeddings\" column. You can list the indexes of a dataset by using `ds.list_indexes()`.\r\n\r\nIf I understand correctly by reading this example you thought that it was re-adding the \"embeddings\" column.\r\nThis looks misleading indeed, and we should add a note to make it more explicit that it doesn't store the column that was used to build the index.\r\n\r\nFeel free to open a PR to suggest an improvement on the documentation if you want to contribute :)", "> If I understand correctly by reading this example you thought that it was re-adding the \"embeddings\" column.\r\nYes. I was trying to use the dataset in RAG and it complained that the dataset didn't have the right columns. No problems when loading the dataset with `load_from_disk` and then doing `load_faiss_index`\r\n\r\nWhat I learned was\r\n1. column and index are different\r\n2. loading the index does not create a column\r\n3. the column is not needed to be able to use the index\r\n4. RAG needs both the embeddings column and the index\r\n\r\nIf I can come up with a way to articulate this in the right spot in the docs, I'll open a PR" ]

[ "Hi,\r\n\r\nActually, the error occurs due to these two lines:\r\n```python\r\nraw_dataset.set_format('torch',columns=['img','label'])\r\nraw_dataset = raw_dataset.rename_column('img','image')\r\n```\r\n`Dataset.rename_column` doesn't update the `_format_columns` attribute, previously defined by `Dataset.set_format`, with a new column name which is why this new column is missing in the output.", "Hi @mariosasko,\n\nThanks for opening a PR on this :)\nWhy does the old name also disappear?", "I just merged a @mariosasko 's PR that fixes this issue.\r\nIf it happens again, feel free to re-open :)" ]

[ "Hi,\r\n\r\nThis is a bug so thanks for reporting it. `Dataset.__setstate__` is the problem, which is called when `Dataset.rename_column` tries to copy the dataset with `copy.deepcopy(self)`. This only happens if the `split` arg in `load_dataset` was defined as `ReadInstruction`.\r\n\r\nTo overcome this issue, use the named splits API (for now):\r\n```python\r\ntrain_ds, test_ds = load_dataset(\r\n path='csv', \r\n delimiter='\\t', \r\n data_files=text_files, \r\n split=['train[:90%]', 'train[-10%:]'],\r\n)\r\n\r\ntrain_ds = train_ds.rename_column('sentence', 'text')\r\n```", "This has been fixed in #2043 , thanks @mariosasko \r\nThe fix is available on master and we'll do a new release soon :)\r\n\r\nfeel free to re-open if you still have issues" ]

[ "Hi,\r\n\r\nCan you give us a minimal reproducible example? This [part](https://huggingface.co/docs/datasets/master/processing.html#controling-the-cache-behavior) of the docs explains how to control caching." ]

[ "What's the issue exactly ?\r\n\r\nGiven an `upstream` remote repository with url `https://github.com/huggingface/datasets.git`, you can totally rebase from `upstream/master`.\r\n\r\nIt's mentioned at the beginning how to add the `upstream` remote repository\r\n\r\nhttps://github.com/huggingface/datasets/blob/987df6b4e9e20fc0c92bc9df48137d170756fd7b/ADD_NEW_DATASET.md#L10-L14", "~~What difference is there with the default `origin` remote that is set when you clone the repo?~~ I've just understood that this applies to **forks** of the repo 🤡 " ]

[ "I'm not able to reproduce on my side.\r\nCan you provide the full stacktrace please ?\r\nWhat version of `python` and `dill` do you have ? Which OS are you using ?", "```\r\nco_filename = '<ipython-input-2-e0383a102aae>', returned_obj = [0]\r\n \r\n def create_ipython_func(co_filename, returned_obj):\r\n def func():\r\n return returned_obj\r\n \r\n code = func.__code__\r\n> code = CodeType(*[getattr(code, k) if k != \"co_filename\" else co_filename for k in code_args])\r\nE TypeError: an integer is required (got type bytes)\r\n\r\ntests/test_caching.py:152: TypeError\r\n```\r\n\r\nPython 3.8.8 \r\ndill==0.3.1.1\r\n", "I managed to reproduce. This comes from the CodeType init signature that is different in python 3.8.8\r\nI opened a PR to fix this test\r\nThanks !" ]

[ "Hi @theo-m !\r\n\r\nA few lines above this line, you'll find that the `_split_generators` method returns a list of `SplitGenerator`s objects:\r\n\r\n```python\r\ndatasets.SplitGenerator(\r\n name=datasets.Split.VALIDATION,\r\n # These kwargs will be passed to _generate_examples\r\n gen_kwargs={\r\n \"filepath\": os.path.join(data_dir, \"dev.jsonl\"),\r\n \"split\": \"dev\",\r\n },\r\n),\r\n```\r\n\r\nNotice the `gen_kwargs` argument passed to the constructor of `SplitGenerator`: this dict will be unpacked as keyword arguments to pass to the `_generat_examples` method (in this case the `filepath` and `split` arguments).\r\n\r\nLet me know if that helps!", "Oh ok I hadn't made the connection between those two, will offer a tweak to the comment and the template then - thanks!" ]

[ "So maybe a summary here: \r\nIf I could fit a large model with batch_size = X into memory, is there a way I could train this model for huge datasets with keeping setting the same? thanks ", "The `datastets` library doesn't load datasets into memory. Therefore you can load a dataset that is terabytes big without filling up your RAM.\r\n\r\nThe only thing that's loaded into memory during training is the batch used in the training step.\r\nSo as long as your model works with batch_size = X, then you can load an even bigger dataset and it will work as well with the same batch_size.\r\n\r\nNote that you still have to take into account that some batches take more memory than others, depending on the texts lengths. If it works for a batch with batch_size = X and with texts of maximum length, then it will work for all batches.\r\n\r\nIn your case I guess that there are a few long sentences in the dataset. For those long sentences you get a memory error on your GPU because they're too long. By passing `max_train_samples` you may have taken a subset of the dataset that only contain short sentences. That's probably why in your case it worked only when you set `max_train_samples`.\r\nI'd suggest you to reduce the batch size so that the batches with long sentences can be loaded on the GPU.\r\n\r\nLet me know if that helps or if you have other questions" ]

[ "Adding to the previous information, I think `torch.utils.data.DataLoader` is doing some conversion. \r\nWhat I tried:\r\n```python\r\ntrain_dataset = load_dataset('mnist')\r\n```\r\nI don't use any `map` or `set_format` or any `transform`. I use this directly, and try to load batches using the `DataLoader` with batch size 2, I get an output like this for the `image`:\r\n\r\n```\r\n[[tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor([0, 0]), tensor...\r\n```\r\nFor `label`, it works fine:\r\n```\r\ntensor([7, 6])\r\n```\r\nNote that I didn't specify conversion to torch tensors anywhere.\r\n\r\nBasically, there are two problems here:\r\n1. `dataset.map` doesn't return tensor type objects, even though it uses the transforms, the grayscale conversion in transform was done, but the output was lists only.\r\n2. The `DataLoader` performs its own conversion, which may be not desired.\r\n\r\nI understand that we can't change `DataLoader` because it is a torch functionality, however, is there a way we can handle image data to allow using it with torch `DataLoader` and `torchvision` properly?\r\n\r\nI think if the `image` was a torch tensor (N,H,W,C), or a list of torch tensors (H,W,C), before it is passed to `DataLoader`, then we might not face this issue. ", "What's the feature types of your new dataset after `.map` ?\r\n\r\nCan you try with adding `features=` in the `.map` call in order to set the \"image\" feature type to `Array2D` ?\r\nThe default feature type is lists of lists, we've not implemented shape verification to use ArrayXD instead of nested lists yet", "Hi @lhoestq\r\n\r\nRaw feature types are like this:\r\n```\r\nImage:\r\n<class 'list'> 60000 #(type, len)\r\n<class 'list'> 28\r\n<class 'list'> 28\r\n<class 'int'>\r\nLabel:\r\n<class 'list'> 60000\r\n<class 'int'>\r\n```\r\nInside the `prepare_feature` method with batch size 100000 , after processing, they are like this:\r\n\r\nInside Prepare Train Features\r\n```\r\nImage:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nAfter map, the feature type are like this:\r\n```\r\nImage:\r\n<class 'list'> 60000\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'list'> 28\r\n<class 'float'>\r\nLabel:\r\n<class 'list'> 60000\r\n<class 'int'>\r\n```\r\n\r\nAfter dataloader with batch size 2, the batch features are like this:\r\n```\r\nImage:\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\n<hr>\r\n\r\nWhen I was setting the format of `train_dataset` to 'torch' after mapping - \r\n```\r\nImage:\r\n<class 'list'> 60000\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nCorresponding DataLoader batch:\r\n```\r\nFrom DataLoader batch features\r\nImage:\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nI will check with features and get back.\r\n\r\n\r\n\r\n", "Hi @lhoestq\r\n\r\n# Using Array3D\r\nI tried this:\r\n```python\r\nfeatures = datasets.Features({\r\n \"image\": datasets.Array3D(shape=(1,28,28),dtype=\"float32\"),\r\n \"label\": datasets.features.ClassLabel(names=[\"0\", \"1\", \"2\", \"3\", \"4\", \"5\", \"6\", \"7\", \"8\", \"9\"]),\r\n })\r\ntrain_dataset = raw_dataset.map(prepare_features, features = features,batched=True, batch_size=10000)\r\n```\r\nand it didn't fix the issue.\r\n\r\nDuring the `prepare_train_features:\r\n```\r\nImage:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nAfter the `map`:\r\n\r\n```\r\nImage:\r\n<class 'list'> 60000\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'list'> 28\r\n<class 'float'>\r\nLabel:\r\n<class 'list'> 60000\r\n<class 'int'>\r\n```\r\nFrom the DataLoader batch:\r\n```\r\nImage:\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\nIt is the same as before.\r\n\r\n---\r\n\r\nUsing `datasets.Sequence(datasets.Array2D(shape=(28,28),dtype=\"float32\"))` gave an error during `map`:\r\n\r\n```python\r\nArrowNotImplementedError Traceback (most recent call last)\r\n<ipython-input-95-d28e69289084> in <module>()\r\n 3 \"label\": datasets.features.ClassLabel(names=[\"0\", \"1\", \"2\", \"3\", \"4\", \"5\", \"6\", \"7\", \"8\", \"9\"]),\r\n 4 })\r\n----> 5 train_dataset = raw_dataset.map(prepare_features, features = features,batched=True, batch_size=10000)\r\n\r\n15 frames\r\n/usr/local/lib/python3.7/dist-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc)\r\n 446 num_proc=num_proc,\r\n 447 )\r\n--> 448 for k, dataset in self.items()\r\n 449 }\r\n 450 )\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/dataset_dict.py in <dictcomp>(.0)\r\n 446 num_proc=num_proc,\r\n 447 )\r\n--> 448 for k, dataset in self.items()\r\n 449 }\r\n 450 )\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint)\r\n 1307 fn_kwargs=fn_kwargs,\r\n 1308 new_fingerprint=new_fingerprint,\r\n-> 1309 update_data=update_data,\r\n 1310 )\r\n 1311 else:\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)\r\n 202 }\r\n 203 # apply actual function\r\n--> 204 out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n 205 datasets: List[\"Dataset\"] = list(out.values()) if isinstance(out, dict) else [out]\r\n 206 # re-apply format to the output\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)\r\n 335 # Call actual function\r\n 336 \r\n--> 337 out = func(self, *args, **kwargs)\r\n 338 \r\n 339 # Update fingerprint of in-place transforms + update in-place history of transforms\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, update_data)\r\n 1580 if update_data:\r\n 1581 batch = cast_to_python_objects(batch)\r\n-> 1582 writer.write_batch(batch)\r\n 1583 if update_data:\r\n 1584 writer.finalize() # close_stream=bool(buf_writer is None)) # We only close if we are writing in a file\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py in write_batch(self, batch_examples, writer_batch_size)\r\n 274 typed_sequence = TypedSequence(batch_examples[col], type=col_type, try_type=col_try_type)\r\n 275 typed_sequence_examples[col] = typed_sequence\r\n--> 276 pa_table = pa.Table.from_pydict(typed_sequence_examples)\r\n 277 self.write_table(pa_table, writer_batch_size)\r\n 278 \r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/table.pxi in pyarrow.lib.Table.from_pydict()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/array.pxi in pyarrow.lib.asarray()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/array.pxi in pyarrow.lib.array()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/array.pxi in pyarrow.lib._handle_arrow_array_protocol()\r\n\r\n/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py in __arrow_array__(self, type)\r\n 95 out = pa.ExtensionArray.from_storage(type, pa.array(self.data, type.storage_dtype))\r\n 96 else:\r\n---> 97 out = pa.array(self.data, type=type)\r\n 98 if trying_type and out[0].as_py() != self.data[0]:\r\n 99 raise TypeError(\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/array.pxi in pyarrow.lib.array()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/array.pxi in pyarrow.lib._sequence_to_array()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()\r\n\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/error.pxi in pyarrow.lib.check_status()\r\n\r\nArrowNotImplementedError: extension\r\n```", "# Convert raw tensors to torch format\r\nStrangely, converting to torch tensors works perfectly on `raw_dataset`:\r\n```python\r\nraw_dataset.set_format('torch',columns=['image','label'])\r\n```\r\nTypes:\r\n```\r\nImage:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nUsing this for transforms:\r\n```python\r\ndef prepare_features(examples):\r\n images = []\r\n labels = []\r\n for example_idx, example in enumerate(examples[\"image\"]):\r\n if transform is not None:\r\n images.append(transform(\r\n examples[\"image\"][example_idx].numpy()\r\n ))\r\n else:\r\n images.append(examples[\"image\"][example_idx].numpy())\r\n labels.append(examples[\"label\"][example_idx])\r\n output = {\"label\":labels, \"image\":images}\r\n return output\r\n```\r\n\r\nInside `prepare_train_features`:\r\n```\r\nImage:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nAfter `map`:\r\n```\r\nImage:\r\n<class 'list'> 60000\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'>\r\n```\r\nDataLoader batch:\r\n\r\n```\r\nImage:\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\n\r\n---\r\n\r\n## Using `torch` format:\r\n```\r\nImage:\r\n<class 'list'> 60000\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'>\r\n```\r\nDataLoader batches:\r\n\r\n```\r\nImage:\r\n<class 'list'> 1\r\n<class 'list'> 28\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\n\r\n---\r\n## Using the features - `Array3D`:\r\n\r\n```\r\nImage:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'list'> 10000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nAfter `map`:\r\n```\r\nImage:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 60000\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nAfter DataLoader `batch`:\r\n```\r\nImage:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'> 1\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'> 28\r\n<class 'torch.Tensor'>\r\nLabel:\r\n<class 'torch.Tensor'> 2\r\n<class 'torch.Tensor'>\r\n```\r\n\r\nThe last one works perfectly.\r\n\r\n\r\n\r\nI wonder why this worked, and others didn't.\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "Concluding, the way it works right now is:\r\n\r\n1. Converting raw dataset to `torch` format.\r\n2. Use the transform and apply using `map`, ensure the returned values are tensors. \r\n3. When mapping, use `features` with `image` being `Array3D` type.", "What the dataset returns depends on the feature type.\r\nFor a feature type that is Sequence(Sequence(Sequence(Value(\"uint8\")))), a dataset formatted as \"torch\" return lists of lists of tensors. This is because the lists lengths may vary.\r\nFor a feature type that is Array3D on the other hand it returns one tensor. This is because the size of the tensor is fixed and defined bu the Array3D type.", "Okay, that makes sense.\r\nRaw images are list of Array2D, hence we get a single tensor when `set_format` is used. But, why should I need to convert the raw images to `torch` format when `map` does this internally?\r\n\r\nUsing `Array3D` did not work with `map` when raw images weren't `set_format`ted to torch type.", "I understand that `map` needs to know what kind of output tensors are expected, and thus converting the raw dataset to `torch` format is necessary. Closing the issue since it is resolved." ]

[ "This is expected to show this message to the user via stdout.\r\nThis way the users see it directly and can cancel the downloading if they want to.\r\nCould you elaborate why it would be better to have it in stderr instead of stdout ?", "@lhoestq, sorry for the late reply\r\n\r\nI completely understand why you decided to output a message that is always shown. The only problem is that the message is printed to the `stdout`. For example, if the user runs `python run_glue.py > log_file`, it will redirect `stdout` to the file named `log_file`, and the message will not be shown to the user.\r\n\r\nInstead, we should print this message to `stderr`. Even in the case of `python run_glue.py > log_file` only `stdout` is being redirected and so the message is always shown.", "We now log these messages to `stderr` (the built-in `logging` module's default behavior). " ]

[ "Why did you close this issue? How did you end up finding the evidence documents? I'm running into a similar issue with other KILT tasks." ]

[ "Hi @itsLuisa !\r\n\r\nCould you give us more information about the error you're getting, please?\r\nA copy-paste of the Traceback would be nice to get a better understanding of what is wrong :) ", "Hello @SBrandeis , this is it:\r\n```\r\nTraceback (most recent call last):\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 537, in incomplete_dir\r\n yield tmp_dir\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 578, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 656, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 982, in _prepare_split\r\n num_examples, num_bytes = writer.finalize()\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\arrow_writer.py\", line 297, in finalize\r\n self.write_on_file()\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\arrow_writer.py\", line 230, in write_on_file\r\n pa_array = pa.array(typed_sequence)\r\n File \"pyarrow\\array.pxi\", line 222, in pyarrow.lib.array\r\n File \"pyarrow\\array.pxi\", line 110, in pyarrow.lib._handle_arrow_array_protocol\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\arrow_writer.py\", line 97, in __arrow_array__\r\n out = pa.array(self.data, type=type)\r\n File \"pyarrow\\array.pxi\", line 305, in pyarrow.lib.array\r\n File \"pyarrow\\array.pxi\", line 39, in pyarrow.lib._sequence_to_array\r\n File \"pyarrow\\error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow\\error.pxi\", line 107, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowTypeError: Expected bytes, got a 'list' object\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"C:/Users/Luisa/Documents/Uni/WS 2020,21/Neural Networks/Final_Project/NN_Project/data_loading.py\", line 122, in <module>\r\n main()\r\n File \"C:/Users/Luisa/Documents/Uni/WS 2020,21/Neural Networks/Final_Project/NN_Project/data_loading.py\", line 111, in main\r\n dataset = datasets.load_dataset(\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\load.py\", line 740, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 586, in download_and_prepare\r\n self._save_info()\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\contextlib.py\", line 131, in __exit__\r\n self.gen.throw(type, value, traceback)\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\site-packages\\datasets\\builder.py\", line 543, in incomplete_dir\r\n shutil.rmtree(tmp_dir)\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\shutil.py\", line 740, in rmtree\r\n return _rmtree_unsafe(path, onerror)\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\shutil.py\", line 618, in _rmtree_unsafe\r\n onerror(os.unlink, fullname, sys.exc_info())\r\n File \"C:\\Users\\Luisa\\AppData\\Local\\Programs\\Python\\Python38\\lib\\shutil.py\", line 616, in _rmtree_unsafe\r\n os.unlink(fullname)\r\nPermissionError: [WinError 32] Der Prozess kann nicht auf die Datei zugreifen, da sie von einem anderen Prozess verwendet wird: 'C:\\\\Users\\\\Luisa\\\\.cache\\\\huggingface\\\\datasets\\\\sample\\\\default-20ee7d51a6a9454f\\\\0.0.0\\\\5fc4c3a355ea77ab446bd31fca5082437600b8364d29b2b95264048bd1f398b1.incomplete\\\\sample-train.arrow'\r\n\r\nProcess finished with exit code 1\r\n```", "Hi @itsLuisa, thanks for sharing the Traceback.\r\n\r\nYou are defining the \"id\" field as a `string` feature:\r\n```python\r\nclass Sample(datasets.GeneratorBasedBuilder):\r\n ...\r\n\r\n def _info(self):\r\n return datasets.DatasetInfo(\r\n features=datasets.Features(\r\n {\r\n \"id\": datasets.Value(\"string\"),\r\n # ^^ here\r\n \"tokens\": datasets.Sequence(datasets.Value(\"string\")),\r\n \"pos_tags\": datasets.Sequence(datasets.features.ClassLabel(names=[...])),\r\n[...]\r\n```\r\n\r\nBut in the `_generate_examples`, the \"id\" field is a list:\r\n```python\r\nids = list()\r\n```\r\n\r\nChanging:\r\n```python\r\n\"id\": datasets.Value(\"string\"),\r\n```\r\nInto:\r\n```python\r\n\"id\": datasets.Sequence(datasets.Value(\"string\")),\r\n```\r\n\r\nShould fix your issue.\r\n\r\nLet me know if this helps!", "It seems to be working now, thanks a lot for the help, @SBrandeis !", "Glad to hear it!\r\nI'm closing the issue" ]

[]

[ "Thanks for reporting! We'll fix that as soon as possible", "Actually soundfile is not a dependency of this dataset.\r\nThe error comes from a bug that was fixed in this commit: https://github.com/huggingface/datasets/pull/1767/commits/c304e63629f4453367de2fd42883a78768055532\r\nBasically the library used to consider the `import soundfile` in the docstring as a dependency, while it's just here as a code example.\r\n\r\nUpdating the viewer to the latest version of `datasets` should fix this issue\r\n", "Hi! The viewer at https://huggingface.co/datasets/arabic_speech_corpus works fine. Closing." ]

[ "@lhoestq I really appreciate if you could help me providiing processed datasets, I do not really have access to enough resources to run the apache-beam and need to run the codes on these datasets. Only en/de/fr currently works, but I need all the languages more or less. thanks ", "Hi @dorost1234, I think I can help you a little. I’ve processed some Wikipedia datasets (Spanish inclusive) using the HF/datasets library during recent research.\r\n\r\n@lhoestq Could you help me to upload these preprocessed datasets to Huggingface's repositories? To be more precise, I've built datasets from the following languages using the 20201201 dumps: Spanish, Portuguese, Russian, French, Japanese, Chinese, and Turkish. Process these datasets have high costs that most of the community can't afford. I think these preprocessed datasets I have could be helpful for someone without access to high-resource machines to process Wikipedia's dumps like @dorost1234\r\n\r\n", "Thank you so much @jonatasgrosman , I greatly appreciate your help with them. \r\nYes, I unfortunately does not have access to a good resource and need it for my\r\nresearch. I greatly appreciate @lhoestq your help with uploading the processed datasets in huggingface datasets. This would be really helpful for some users like me with not access to high-memory GPU resources.\r\n\r\nthank you both so much again.\r\n\r\nOn Sat, Mar 6, 2021 at 12:55 AM Jonatas Grosman <[email protected]>\r\nwrote:\r\n\r\n> Hi @dorost1234 <https://github.com/dorost1234>, I think I can help you a\r\n> little. I’ve processed some Wikipedia datasets (Spanish inclusive) using\r\n> the HF/datasets library during recent research.\r\n>\r\n> @lhoestq <https://github.com/lhoestq> Could you help me to upload these\r\n> preprocessed datasets to Huggingface's repositories? To be more precise,\r\n> I've built datasets from the following languages using the 20201201 dumps:\r\n> Spanish, Portuguese, Russian, French, Japanese, Chinese, and Turkish.\r\n> Process these datasets have high costs that most of the community can't\r\n> afford. I think these preprocessed datasets I have could be helpful for\r\n> someone without access to high-resource machines to process Wikipedia's\r\n> dumps like @dorost1234 <https://github.com/dorost1234>\r\n>\r\n> —\r\n> You are receiving this because you were mentioned.\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/1994#issuecomment-791798195>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AS37NMWMK5GFJFU3ACCJFUDTCFVNZANCNFSM4YUZIF4A>\r\n> .\r\n>\r\n", "Hi @dorost1234, so sorry, but looking at my files here, I figure out that I've preprocessed files using the HF/datasets for all the languages previously listed by me (Portuguese, Russian, French, Japanese, Chinese, and Turkish) except the Spanish (on my tests I've used the [wikicorpus](https://www.cs.upc.edu/~nlp/wikicorpus/) instead).\r\n\r\nOnly with the Spanish Wikipedia's dump, I had the same `KeyError: '000nbsp'` problem already reported here https://github.com/huggingface/datasets/issues/577\r\n\r\nSo nowadays, even with access to a high resource machine, you couldn't be able to get Wikipedia's Spanish data using the HF/datasets :(\r\n\r\n\r\n\r\n\r\n", "Thanks a lot for the information and help. This would be great to have\nthese datasets.\n@lhoestq <https://github.com/lhoestq> Do you know a way I could get\nsmaller amount of these data like 1 GBtype of each language to deal with\ncomputatioanl requirements? thanks\n\nOn Sat, Mar 6, 2021 at 5:36 PM Jonatas Grosman <[email protected]>\nwrote:\n\n> Hi @dorost1234 <https://github.com/dorost1234>, so sorry, but looking at\n> my files here, I figure out that I've preprocessed files using the\n> HF/datasets for all the languages previously listed by me (Portuguese,\n> Russian, French, Japanese, Chinese, and Turkish) except the Spanish (on my\n> tests I've used the wikicorpus <https://www.cs.upc.edu/~nlp/wikicorpus/>\n> instead).\n>\n> Only with the Spanish Wikipedia's dump, I had the same KeyError: '000nbsp'\n> problem already reported here #577\n> <https://github.com/huggingface/datasets/issues/577>\n>\n> So nowadays, even with access to a high resource machine, you couldn't be\n> able to get Wikipedia's Spanish data using the HF/datasets :(\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/datasets/issues/1994#issuecomment-791985546>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMWMO7WOHWLOROPD6Q3TCJKXPANCNFSM4YUZIF4A>\n> .\n>\n", "Hi ! As mentioned above the Spanish configuration have parsing issues from `mwparserfromhell`. I haven't tested with the latest `mwparserfromhell` >=0.6 though. Which version of `mwparserfromhell` are you using ?\r\n\r\n> @lhoestq Could you help me to upload these preprocessed datasets to Huggingface's repositories? To be more precise, I've built datasets from the following languages using the 20201201 dumps: Spanish, Portuguese, Russian, French, Japanese, Chinese, and Turkish. Process these datasets have high costs that most of the community can't afford. I think these preprocessed datasets I have could be helpful for someone without access to high-resource machines to process Wikipedia's dumps like @dorost1234\r\n\r\nThat would be awesome ! Feel free to ping me on slack so we can put the processed wikipedia files on google storage with the other ones we've already preprocessed.\r\n\r\n> Do you know a way I could get smaller amount of these data like 1 GBtype of each language to deal with computatioanl requirements? thanks\r\n\r\nI'd suggest to copy the [wikipedia.py](https://github.com/huggingface/datasets/blob/master/datasets/wikipedia/wikipedia.py) to a new script `custom_wikipedia.py` and modify it to only download and process only a subset of the raw data files.\r\nYou can for example replace [this line](https://github.com/huggingface/datasets/blob/64e59fc45ca2134218b3e42e83fddddbe840ff74/datasets/wikipedia/wikipedia.py#L446) by:\r\n```python\r\n if total_bytes >= (1 << 30): # stop if the total amount of data is >= 1GB\r\n break\r\n else:\r\n xml_urls.append(_base_url(lang) + fname)\r\n```\r\n\r\nThen you can load your custom wikipedia dataset with\r\n```python\r\nload_dataset(\"path/to/my/custom_wikipedia.py\", f\"{date}.{language}\")\r\n```", "Hi @lhoestq!\r\n\r\n> Hi ! As mentioned above the Spanish configuration have parsing issues from mwparserfromhell. I haven't tested with the latest mwparserfromhell >=0.6 though. Which version of mwparserfromhell are you using ?\r\n\r\nI'm using the latest mwparserfromhell version (0.6)\r\n\r\n> That would be awesome ! Feel free to ping me on slack so we can put the processed wikipedia files on google storage with the other ones we've already preprocessed.\r\n\r\nI'll ping you there 👍 ", "Thank you so much @jonatasgrosman and @lhoestq this would be a great help. I am really thankful to you both and to wonderful Huggingface dataset library allowing us to train models at scale." ]

[ "Hi ! That looks like a bug, can you provide some code so that we can reproduce ?\r\nIt's not supposed to update the original dataset", "Hi, I experimented with RAG. \r\n\r\nActually, you can run the [use_own_knowldge_dataset.py](https://github.com/shamanez/transformers/blob/rag-end-to-end-retrieval/examples/research_projects/rag/use_own_knowledge_dataset.py#L80). In the 80 you can save the dataset object to the disk with save_to_disk. Then in order to compute the embeddings in this use **load_from_disk**. \r\n\r\nThen finally save it. You can see the original dataset object (CSV after splitting also will be changed)\r\n\r\nOne more thing- when I save the dataset object with **save_to_disk** it name the arrow file with cache.... rather than using dataset. arrow. Can you add a variable that we can feed a name to save_to_disk function?", "@lhoestq I also found that cache in tmp directory gets updated after transformations. This is really problematic when using datasets interactively. Let's say we use the shards function to a dataset loaded with csv, atm when we do transformations to shards and combine them it updates the original csv cache. ", "I plan to update the save_to_disk method in #2025 so I can make sure the new save_to_disk doesn't corrupt your cache files.\r\nBut from your last message it looks like save_to_disk isn't the root cause right ?", "ok, one more thing. When we use save_to_disk there are two files other than .arrow. dataset_info.json and state.json. Sometimes most of the fields in the dataset_infor.json are null, especially when saving dataset objects. Anyways I think load_from_disk uses the arrow files mentioned in state.json right? ", "> Anyways I think load_from_disk uses the arrow files mentioned in state.json right?\r\n\r\nYes exactly", "Perfect. For now, I am loading the dataset from CSV in my interactive process and will wait until you make the PR!" ]

[ "Hi @hwijeen, you might want to look at issues #1796 and #1949. I think it could be something related to the I/O operations being performed.", "I see that many people are experiencing the same issue. Is this problem considered an \"official\" bug that is worth a closer look? @lhoestq", "Yes this looks like a bug. On my side I haven't managed to reproduce it but @theo-m has. We'll investigate this !", "Thank you for the reply! I would be happy to follow the discussions related to the issue.\r\nIf you do not mind, could you also give a little more explanation on my p.s.2? I am having a hard time figuring out why the single processing `map` uses all of my cores.\r\n@lhoestq @theo-m ", "Regarding your ps2: It depends what function you pass to `map`.\r\nFor example, fast tokenizers from `transformers` in Rust tokenize texts and parallelize the tokenization over all the cores.", "I am still experiencing this issue with datasets 1.9.0..\r\nHas there been a further investigation? \r\n<img width=\"442\" alt=\"image\" src=\"https://user-images.githubusercontent.com/29157715/126143387-8b5ddca2-a896-4e18-abf7-4fbf62a48b41.png\">\r\n", "Hi. Is there any update on this issue? I am desperately trying to decrease my times, and multiprocessing \"should\" be the solution, but it literally takes 5 times longer.", "Which version of `datasets` are you using ?", "Hi,\r\n\r\nI’m running into the same issue and trying to come up with a simple benchmark. \r\n\r\n# environment info\r\nI have a total of 80 CPUs.\r\n\r\n- `datasets` version: 2.4.0\r\n- Platform: Linux-4.18.0-305.65.1.el8_4.x86_64-x86_64-with-glibc2.28\r\n- Python version: 3.10.4\r\n- PyArrow version: 8.0.0\r\n- Pandas version: 1.4.3\r\n\r\n# How to reproduce\r\n\r\n```py\r\nIn [1]: from datasets import Dataset, set_caching_enabled \r\nIn [2]: import numpy as np \r\nIn [3]: set_caching_enabled(False) \r\nIn [4]: d = Dataset.from_dict({'foo': np.random.randn(1000,256)}) \r\nIn [9]: d.set_format('np')\r\nIn [14]: def sort(array): \r\n ...: np.sort(array) \r\n# multiprocessing disabled\r\nIn [19]: %%timeit \r\n ...: d.map(sort, input_columns='foo') \r\n78.8 ms ± 1.22 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) \r\n# multiprocessing enabled \r\nIn [27]: %%timeit \r\n ...: d.map(sort, input_columns='foo',num_proc=10) \r\n858 ms ± 45.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) \r\n```", "Spawning multiple processes has an overhead. For small datasets the processing is likely to be faster than spawning the processes and passing the data to them.\r\n\r\nEspecially since your dataset is in memory: the data has to be copied to the subprocesses.\r\nOn the other hand, datasets loaded from disk are much faster to reload from a subprocess thanks to memory mapping.", "Thanks for the clarifications! \r\n\r\nIndeed, when saving then loading the above dataset to disk, and increasing the number of rows to 10K or 100K, the performance gap narrows.\r\n\r\n```py\r\n# with 10000 rows\r\nIn [3]: %%timeit\r\n ...: d.map(sort, input_columns='foo')\r\n578 ms ± 5.89 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\r\nIn [4]: %%timeit \r\n ...: d.map(sort, input_columns='foo',num_proc=10) \r\n1.06 s ± 47.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\r\n\r\n# with 100000 rows\r\nIn [6]: %%timeit\r\n ...: d.map(sort, input_columns='foo')\r\n5.8 s ± 25.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\r\nIn [7]: %%timeit\r\n ...: d.map(sort, input_columns='foo',num_proc=10)\r\n7.23 s ± 154 ms per loop (mean ± std. dev. of 7 runs, 1 loop each\r\n```", "any updates on this issue? \r\nI'm using `datasets=2.12.0`. Adding `num_proc` to the mapping function makes it at least 5x slower than using a single process.", "What kind of function are you passing to `map` ? How many CPUs do you have and what did you set for `num_proc` ?", "Hello! Any solution for this? Thanks!" ]

[ "Do you think this is trying to bring the dataset into memory and if I can avoid it to save on memory so it only brings a batch into memory? @lhoestq thank you", "It's not trying to bring the dataset into memory.\r\n\r\nActually, it's trying to memory map the dataset file, which is different. It allows to load large dataset files without filling up memory.\r\n\r\nWhat dataset did you use to get this error ?\r\nOn what OS are you running ? What's your python and pyarrow version ?", "Dear @lhoestq \r\nthank you so much for coming back to me. Please find info below:\r\n1) Dataset name: I used wikipedia with config 20200501.en\r\n2) I got these pyarrow in my environment:\r\npyarrow 2.0.0 <pip>\r\npyarrow 3.0.0 <pip>\r\n\r\n3) python version 3.7.10\r\n4) OS version \r\n\r\nlsb_release -a\r\nNo LSB modules are available.\r\nDistributor ID:\tDebian\r\nDescription:\tDebian GNU/Linux 10 (buster)\r\nRelease:\t10\r\nCodename:\tbuster\r\n\r\n\r\nIs there a way I could solve the memory issue and if I could run this model, I am using GeForce GTX 108, \r\nthanks \r\n", "I noticed that the error happens when loading the validation dataset.\r\nWhat value of `data_args.validation_split_percentage` did you use ?", "Dear @lhoestq \r\n\r\nthank you very much for the very sharp observation, indeed, this happens there, I use the default value of 5, I basically plan to subsample a part of the large dataset and choose it as validation set. Do you think this is bringing the data into memory during subsampling? Is there a way I could avoid this?\r\n\r\nThank you very much for the great help.\r\n\r\n\r\nOn Mon, Mar 8, 2021 at 11:28 AM Quentin Lhoest ***@***.***>\r\nwrote:\r\n\r\n> I noticed that the error happens when loading the validation dataset.\r\n> What value of data_args.validation_split_percentage did you use ?\r\n>\r\n> —\r\n> You are receiving this because you authored the thread.\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/1990#issuecomment-792655644>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AS37NMS337ZUJ7HGGVVCCR3TCSREFANCNFSM4YTYAQ2A>\r\n> .\r\n>\r\n", "Methods like `dataset.shard`, `dataset.train_test_split`, `dataset.select` etc. don't bring the dataset in memory. \r\nThe only time when samples are brought to memory is when you access elements via `dataset[0]`, `dataset[:10]`, `dataset[\"my_column_names\"]`.\r\n\r\nBut it's possible that trying to use those methods to build your validation set doesn't fix the issue since, if I understand correctly, the error happens when when the dataset arrow file is opened (just before the 5% percentage is applied).\r\n\r\nDid you try to reproduce this issue in a google colab ? This would be super helpful to investigate why this happened.\r\n\r\nAlso maybe you can try clearing your cache at `~/.cache/huggingface/datasets` and try again. If the arrow file was corrupted somehow, removing it and rebuilding may fix the issue." ]

[ "It seems that I get parsing errors for various fields in my data. For example now I get this:\r\n```\r\n File \"../../../models/tr-4.3.2/run_puppets.py\", line 523, in <module>\r\n main()\r\n File \"../../../models/tr-4.3.2/run_puppets.py\", line 249, in main\r\n datasets = load_dataset(\"csv\", data_files=data_files)\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/datasets/load.py\", line 740, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/datasets/builder.py\", line 572, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/datasets/builder.py\", line 650, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/datasets/builder.py\", line 1028, in _prepare_split\r\n writer.write_table(table)\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/datasets/arrow_writer.py\", line 292, in write_table\r\n pa_table = pa_table.cast(self._schema)\r\n File \"pyarrow/table.pxi\", line 1311, in pyarrow.lib.Table.cast\r\n File \"pyarrow/table.pxi\", line 265, in pyarrow.lib.ChunkedArray.cast\r\n File \"/dccstor/redrug_ier/envs/last-tr/lib/python3.8/site-packages/pyarrow/compute.py\", line 87, in cast\r\n return call_function(\"cast\", [arr], options)\r\n File \"pyarrow/_compute.pyx\", line 298, in pyarrow._compute.call_function\r\n File \"pyarrow/_compute.pyx\", line 192, in pyarrow._compute.Function.call\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 84, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Failed to parse string: https://www.netgalley.com/catalog/book/121872\r\n```", "Not sure if this helps, this is how I load my files (as in the sample scripts on transformers):\r\n\r\n```\r\n if data_args.train_file.endswith(\".csv\"):\r\n # Loading a dataset from local csv files\r\n datasets = load_dataset(\"csv\", data_files=data_files)\r\n```", "Since this worked out of the box in a few examples before, I wonder if it's some quoting issue or something else. ", "Hi @ioana-blue,\r\nCan you share a sample from your .csv? A dummy where you get this error will also help.\r\n\r\nI tried this csv:\r\n```csv\r\nfeature,label\r\n1.2,not nurse\r\n1.3,nurse\r\n1.5,surgeon\r\n```\r\nand the following snippet:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nd = load_dataset(\"csv\",data_files=['test.csv'])\r\n\r\nprint(d)\r\nprint(d['train']['label'])\r\n```\r\nand this works perfectly fine for me:\r\n```sh\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['feature', 'label'],\r\n num_rows: 3\r\n })\r\n})\r\n['not nurse', 'nurse', 'surgeon']\r\n```\r\nI'm sure your csv is more complicated than this one. But it is hard to tell where the issue might be without looking at a sample.", "I've had versions where it worked fain. For this dataset, I had all kind of parsing issues that I couldn't understand. What I ended up doing is strip all the columns that I didn't need and also make the label 0/1. \r\n\r\nI think one line that may have caused a problem was the csv version of this:\r\n\r\n```crawl-data/CC-MAIN-2017-47/segments/1510934806225.78/wet/CC-MAIN-20171120203833-20171120223833-00571.warc.wet.gz Rose Blakey is an aspiring journalist. She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead-end waitressing job. ^M ('Rose', '', 'Blakey') journalist F 38 journalist https://www.netgalley.com/catalog/book/121872 _ is desperate to escape the from the small Australian town in which _ lives. Rejection after rejection mean _ is stuck in what _ sees as a dead-end waitressing job. She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead-end waitressing job.```\r\n\r\nThe error I got in this case is this one: https://github.com/huggingface/datasets/issues/1989#issuecomment-790842771\r\n\r\nNote, this line was part of a much larger file and until this line I guess it was working fine. ", "Hi @ioana-blue,\r\n\r\nWhat is the separator you're using for the csv? I see there are only two commas in the given line, but they don't seem like appropriate points. Also, is this a string part of one line, or an entire line? There should also be a label, right?", "Sorry for the confusion, the sample above was from a tsv that was used to derive the csv. Let me construct the csv again (I had remove it). \r\n\r\nThis is the line in the csv - this is the whole line:\r\n```crawl-data/CC-MAIN-2017-47/segments/1510934806225.78/wet/CC-MAIN-20171120203833-20171120223833-00571.warc.wet.gz,Rose Blakey is an aspiring journalist. She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead,\"('Rose', '', 'Blakey')\",journalist,F,38,journalist,https://www.netgalley.com/catalog/book/121872,_ is desperate to escape the from the small Australian town in which _ lives. Rejection after rejection mean _ is stuck in what _ sees as a dead-end waitressing job., She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead-end waitressing job.```", "Hi,\r\nJust in case you want to use tsv directly, you can use the separator argument while loading the dataset.\r\n```python\r\nd = load_dataset(\"csv\",data_files=['test.csv'],sep=\"\\t\")\r\n```\r\n\r\nAdditionally, I don't face the issues with the following csv (same as the one you provided):\r\n\r\n```sh\r\nlink1,text1,info1,info2,info3,info4,info5,link2,text2,text3\r\ncrawl-data/CC-MAIN-2017-47/segments/1510934806225.78/wet/CC-MAIN-20171120203833-20171120223833-00571.warc.wet.gz,Rose Blakey is an aspiring journalist. She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead,\"('Rose', '', 'Blakey')\",journalist,F,38,journalist,https://www.netgalley.com/catalog/book/121872,_ is desperate to escape the from the small Australian town in which _ lives. Rejection after rejection mean _ is stuck in what _ sees as a dead-end waitressing job., She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead-end waitressing job.\r\n```\r\nOutput after loading:\r\n```sh\r\n{'link1': 'crawl-data/CC-MAIN-2017-47/segments/1510934806225.78/wet/CC-MAIN-20171120203833-20171120223833-00571.warc.wet.gz', 'text1': 'Rose Blakey is an aspiring journalist. She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead', 'info1': \"('Rose', '', 'Blakey')\", 'info2': 'journalist', 'info3': 'F', 'info4': 38, 'info5': 'journalist', 'link2': 'https://www.netgalley.com/catalog/book/121872', 'text2': '_ is desperate to escape the from the small Australian town in which _ lives. Rejection after rejection mean _ is stuck in what _ sees as a dead-end waitressing job.', 'text3': ' She is desperate to escape the from the small Australian town in which she lives. Rejection after rejection mean she is stuck in what she sees as a dead-end waitressing job.'}\r\n```\r\nCan you check once if the tsv works for you directly using the separator argument? The conversion from tsv to csv could create issues, I'm only guessing though.", "thanks for the tip. very strange :/ I'll check my datasets version as well. \r\n\r\nI will have more similar experiments soon so I'll let you know if I manage to get rid of this. ", "No problem at all. I thought I'd be able to solve this but I'm unable to replicate the issue :/" ]

[ "Hi ! Yes it's possible to use image data. There are already a few of them available (MNIST, CIFAR..)" ]

[ "It's reachable for the viewer and me, so I suppose it was down at that moment?" ]

[ "caching issue, seems to work again.." ]

[ "Dummy data generation is deprecated now. Closing." ]

[ "Hi,\r\n\r\nif you inspect the raw data, you can find there are 946 occurrences of `-DOCSTART- -X- -X- O` in the train split and `14041 + 946 = 14987`, which is exactly the number of sentences the authors report. `-DOCSTART-` is a special line that acts as a boundary between two different documents and is filtered out in our implementation.\r\n\r\n@lhoestq What do you think about including these lines? ([Link](https://github.com/flairNLP/flair/issues/1097) to a similar issue in the flairNLP repo)", "We should mention in the Conll2003 dataset card that these lines have been removed indeed.\r\n\r\nIf some users are interested in using these lines (maybe to recombine documents ?) then we can add a parameter to the conll2003 dataset to include them.\r\n\r\nBut IMO the default config should stay the current one (without the `-DOCSTART-` stuff), so that you can directly train NER models without additional preprocessing. Let me know what you think", "@lhoestq Yes, I agree adding a small note should be sufficient.\r\n\r\nCurrently, NLTK's `ConllCorpusReader` ignores the `-DOCSTART-` lines so I think it's ok if we do the same. If there is an interest in the future to use these lines, then we can include them.", "I added a mention of this in conll2003's dataset card:\r\nhttps://github.com/huggingface/datasets/blob/fc9796920da88486c3b97690969aabf03d6b4088/datasets/conll2003/README.md#conll2003\r\n\r\nEdit: just saw your PR @mariosasko (noticed it too late ^^)\r\nLet me take a look at it :)" ]

[ "@stas00 Mea culpa... May I fix this tomorrow morning?", "yes, of course, I reverted to the version before that and it works ;)\r\n\r\nbut since a new release was just made you will probably need to make a hotfix.\r\n\r\nand add the wmt to the tests?", "Sure, I will implement a regression test!", "@stas00 it is fixed. @lhoestq are you releasing the hot fix or would you prefer me to do it?", "I'll do a patch release for this issue early tomorrow.\r\n\r\nAnd yes we absolutly need tests for the wmt datasets: The missing tests for wmt are an artifact from the early development of the lib but now we have tools to generate automatically the dummy data used for tests :)", "still facing the same issue or similar:\r\nfrom datasets import load_dataset\r\nwtm14_test = load_dataset('wmt14',\"de-en\",cache_dir='./datasets')\r\n\r\n~.cache\\huggingface\\modules\\datasets_modules\\datasets\\wmt14\\43e717d978d2261502b0194999583acb874ba73b0f4aed0ada2889d1bb00f36e\\wmt_utils.py in _split_generators(self, dl_manager)\r\n758 # Extract manually downloaded files.\r\n759 manual_files = dl_manager.extract(manual_paths_dict)\r\n--> 760 extraction_map = dict(downloaded_files, **manual_files)\r\n761\r\n762 for language in self.config.language_pair:\r\n\r\nTypeError: type object argument after ** must be a mapping, not list" ]

[ "I sometimes also get this error with other languages of the same dataset:\r\n\r\n File \"/dara/libs/anaconda3/envs/code/lib/python3.7/site-packages/datasets-1.3.0-py3.7.egg/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 99, in pyarrow.lib.check_status\r\nOSError: Memory mapping file failed: Cannot allocate memory\r\n\r\n@lhoestq \r\n", "Hi ! Thanks for reporting\r\nSome wikipedia configurations do require the user to have `apache_beam` in order to parse the wikimedia data.\r\n\r\nOn the other hand regarding your second issue\r\n```\r\nOSError: Memory mapping file failed: Cannot allocate memory\r\n```\r\nI've never experienced this, can you open a new issue for this specific error and provide more details please ?\r\nFor example what script did you use to get this, what language did you use, what's your environment details (os, python version, pyarrow version).." ]

[ "Echo'ing this observation: I have a few datasets in the neighborhood of 2GB CSVs uncompressed, and when I use something like `Dataset.save_to_disk()` it's ~18GB on disk.\r\n\r\nIf this is unexpected behavior, would be happy to help run debugging as needed.", "Thanks @ioana-blue for pointing out this problem (and thanks also @justin-yan). You are right that current implementation of the datasets caching files take too much memory. We are definitely changing this and optimizing the defaults, so that the file sizes are considerably reduced. I will come back to you as soon as this is fixed.", "Thank you! Also I noticed that the files don't seem to be cleaned after the jobs finish. Last night I had only 3 jobs running, but the cache was still at 180GB. ", "And to clarify, it's not memory, it's disk space. Thank you!", "Hi ! As Albert said they can sometimes take more space that expected but we'll fix that soon.\r\n\r\nAlso, to give more details about caching: computations on a dataset are cached by default so that you don't have to recompute them the next time you run them.\r\n\r\nSo by default the cache files stay on your disk when you job is finished (so that if you re-execute it, it will be reloaded from the cache).\r\nFeel free to clear your cache after your job has finished, or disable caching using\r\n```python\r\nimport datasets\r\n\r\ndatasets.set_caching_enabled(False)\r\n```", "Thanks for the tip, this is useful. ", "Hi @ioana-blue, we have optimized Datasets' disk usage in the latest release v1.5.\r\n\r\nFeel free to update your Datasets version\r\n```shell\r\npip install -U datasets\r\n```\r\nand see if it better suits your needs.", "Thank you!" ]

[ "Hi ! `rename_column` has been added recently and will be available in the next release" ]

[ "Hi !\r\nAs far as I know not all faiss indexes can be computed in parallel and then merged. \r\nFor example [here](https://github.com/facebookresearch/faiss/wiki/Special-operations-on-indexes#splitting-and-merging-indexes) is is mentioned that only IndexIVF indexes can be merged.\r\nMoreover faiss already works using multithreading to parallelize the workload over your different CPU cores. You can find more info [here](https://github.com/facebookresearch/faiss/wiki/Threads-and-asynchronous-calls#internal-threading)\r\nSo I feel like the gains we would get by implementing a parallel `add_faiss_index` would not be that important, but let me know what you think.\r\n", "Actually, you are right. I also had the same idea. I am trying this in the context of end-ton-end retrieval training in RAG. So far I have parallelized the embedding re-computation within the training loop by using datasets shards. \r\n\r\nThen I was thinking of can I calculate the indexes for each shard and combined them with **concatenate** before I save.", "@lhoestq As you mentioned faiss is already using multiprocessing. I tried to do the add_index with faiss for a dataset object inside a RAY actor and the process became very slow... if fact it takes so much time. It is because a ray actor comes with a single CPU core unless we assign it more. I also tried assigning more cores but still running add_index in the main process is very fast. " ]

[ "Hi !\r\n\r\nTo fix 1, an you try to run this code ?\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nload_dataset(\"squad\", download_mode=\"force_redownload\")\r\n```\r\nMaybe the file your downloaded was corrupted, in this case redownloading this way should fix your issue 1.\r\n\r\nRegarding your 2nd point, you're right that loading the raw json this way doesn't give you a dataset with the column \"context\", \"question\" and \"answers\". Indeed the squad format is a very nested format so you have to preprocess the data. You can do it this way:\r\n```python\r\ndef process_squad(examples):\r\n \"\"\"\r\n Process a dataset in the squad format with columns \"title\" and \"paragraphs\"\r\n to return the dataset with columns \"context\", \"question\" and \"answers\".\r\n \"\"\"\r\n out = {\"context\": [], \"question\": [], \"answers\":[]} \r\n for paragraphs in examples[\"paragraphs\"]: \r\n for paragraph in paragraphs: \r\n for qa in paragraph[\"qas\"]: \r\n answers = [{\"answer_start\": answer[\"answer_start\"], \"text\": answer[\"text\"].strip()} for answer in qa[\"answers\"]] \r\n out[\"context\"].append(paragraph[\"context\"].strip()) \r\n out[\"question\"].append(qa[\"question\"].strip()) \r\n out[\"answers\"].append(answers) \r\n return out\r\n\r\ndatasets = load_dataset(extension, data_files=data_files, field=\"data\")\r\ncolumn_names = datasets[\"train\"].column_names\r\n\r\nif set(column_names) == {\"title\", \"paragraphs\"}:\r\n datasets = datasets.map(process_squad, batched=True, remove_columns=column_names)\r\n```\r\n\r\nHope that helps :)", "Thks for quickly answering！\r\n### 1 I try the first way,but seems not work \r\n```\r\nTraceback (most recent call last):\r\n File \"examples/question-answering/run_qa.py\", line 503, in <module>\r\n main()\r\n File \"examples/question-answering/run_qa.py\", line 218, in main\r\n datasets = load_dataset(data_args.dataset_name, download_mode=\"force_redownload\")\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/load.py\", line 746, in load_dataset\r\n use_auth_token=use_auth_token,\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/builder.py\", line 573, in download_and_prepare\r\n dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/builder.py\", line 633, in _download_and_prepare\r\n self.info.download_checksums, dl_manager.get_recorded_sizes_checksums(), \"dataset source files\"\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/utils/info_utils.py\", line 39, in verify_checksums\r\n raise NonMatchingChecksumError(error_msg + str(bad_urls))\r\ndatasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json']\r\n```\r\n### 2 I try the second way,and run the examples/question-answering/run_qa.py,it lead to another bug orz..\r\n```\r\nTraceback (most recent call last):\r\n File \"examples/question-answering/run_qa.py\", line 523, in <module>\r\n main()\r\n File \"examples/question-answering/run_qa.py\", line 379, in main\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1120, in map\r\n update_data = does_function_return_dict(test_inputs, test_indices)\r\n File \"/home2/zhenggo1/anaconda3/envs/lpot/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1091, in does_function_return_dict\r\n function(*fn_args, indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)\r\n File \"examples/question-answering/run_qa.py\", line 339, in prepare_train_features\r\n if len(answers[\"answer_start\"]) == 0:\r\nTypeError: list indices must be integers or slices, not str\r\n```\r\n## may be the function prepare_train_features in run_qa.py need to fix,I think is that the prep\r\n```python\r\nfor i, offsets in enumerate(offset_mapping):\r\n # We will label impossible answers with the index of the CLS token.\r\n input_ids = tokenized_examples[\"input_ids\"][i]\r\n cls_index = input_ids.index(tokenizer.cls_token_id)\r\n\r\n # Grab the sequence corresponding to that example (to know what is the context and what is the question).\r\n sequence_ids = tokenized_examples.sequence_ids(i)\r\n\r\n # One example can give several spans, this is the index of the example containing this span of text.\r\n sample_index = sample_mapping[i]\r\n answers = examples[answer_column_name][sample_index]\r\n print(examples,answers)\r\n # If no answers are given, set the cls_index as answer.\r\n if len(answers[\"answer_start\"]) == 0:\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Start/end character index of the answer in the text.\r\n start_char = answers[\"answer_start\"][0]\r\n end_char = start_char + len(answers[\"text\"][0])\r\n\r\n # Start token index of the current span in the text.\r\n token_start_index = 0\r\n while sequence_ids[token_start_index] != (1 if pad_on_right else 0):\r\n token_start_index += 1\r\n\r\n # End token index of the current span in the text.\r\n token_end_index = len(input_ids) - 1\r\n while sequence_ids[token_end_index] != (1 if pad_on_right else 0):\r\n token_end_index -= 1\r\n\r\n # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).\r\n if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Otherwise move the token_start_index and token_end_index to the two ends of the answer.\r\n # Note: we could go after the last offset if the answer is the last word (edge case).\r\n while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:\r\n token_start_index += 1\r\n tokenized_examples[\"start_positions\"].append(token_start_index - 1)\r\n while offsets[token_end_index][1] >= end_char:\r\n token_end_index -= 1\r\n tokenized_examples[\"end_positions\"].append(token_end_index + 1)\r\n\r\n return tokenized_examples\r\n``` ", "## I have fixed it, @lhoestq \r\n### the first section change as you said and add [\"id\"]\r\n```python\r\ndef process_squad(examples):\r\n \"\"\"\r\n Process a dataset in the squad format with columns \"title\" and \"paragraphs\"\r\n to return the dataset with columns \"context\", \"question\" and \"answers\".\r\n \"\"\"\r\n # print(examples)\r\n out = {\"context\": [], \"question\": [], \"answers\":[],\"id\":[]} \r\n for paragraphs in examples[\"paragraphs\"]: \r\n for paragraph in paragraphs: \r\n for qa in paragraph[\"qas\"]: \r\n answers = [{\"answer_start\": answer[\"answer_start\"], \"text\": answer[\"text\"].strip()} for answer in qa[\"answers\"]] \r\n out[\"context\"].append(paragraph[\"context\"].strip()) \r\n out[\"question\"].append(qa[\"question\"].strip()) \r\n out[\"answers\"].append(answers) \r\n out[\"id\"].append(qa[\"id\"]) \r\n return out\r\ncolumn_names = datasets[\"train\"].column_names if training_args.do_train else datasets[\"validation\"].column_names\r\n# print(datasets[\"train\"].column_names)\r\nif set(column_names) == {\"title\", \"paragraphs\"}:\r\n datasets = datasets.map(process_squad, batched=True, remove_columns=column_names)\r\n# Preprocessing the datasets.\r\n# Preprocessing is slighlty different for training and evaluation.\r\nif training_args.do_train:\r\n column_names = datasets[\"train\"].column_names\r\nelse:\r\n column_names = datasets[\"validation\"].column_names\r\n# print(column_names)\r\nquestion_column_name = \"question\" if \"question\" in column_names else column_names[0]\r\ncontext_column_name = \"context\" if \"context\" in column_names else column_names[1]\r\nanswer_column_name = \"answers\" if \"answers\" in column_names else column_names[2]\r\n```\r\n### the second section\r\n```python\r\ndef prepare_train_features(examples):\r\n # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results\r\n # in one example possible giving several features when a context is long, each of those features having a\r\n # context that overlaps a bit the context of the previous feature.\r\n tokenized_examples = tokenizer(\r\n examples[question_column_name if pad_on_right else context_column_name],\r\n examples[context_column_name if pad_on_right else question_column_name],\r\n truncation=\"only_second\" if pad_on_right else \"only_first\",\r\n max_length=data_args.max_seq_length,\r\n stride=data_args.doc_stride,\r\n return_overflowing_tokens=True,\r\n return_offsets_mapping=True,\r\n padding=\"max_length\" if data_args.pad_to_max_length else False,\r\n )\r\n\r\n # Since one example might give us several features if it has a long context, we need a map from a feature to\r\n # its corresponding example. This key gives us just that.\r\n sample_mapping = tokenized_examples.pop(\"overflow_to_sample_mapping\")\r\n # The offset mappings will give us a map from token to character position in the original context. This will\r\n # help us compute the start_positions and end_positions.\r\n offset_mapping = tokenized_examples.pop(\"offset_mapping\")\r\n\r\n # Let's label those examples!\r\n tokenized_examples[\"start_positions\"] = []\r\n tokenized_examples[\"end_positions\"] = []\r\n\r\n for i, offsets in enumerate(offset_mapping):\r\n # We will label impossible answers with the index of the CLS token.\r\n input_ids = tokenized_examples[\"input_ids\"][i]\r\n cls_index = input_ids.index(tokenizer.cls_token_id)\r\n\r\n # Grab the sequence corresponding to that example (to know what is the context and what is the question).\r\n sequence_ids = tokenized_examples.sequence_ids(i)\r\n\r\n # One example can give several spans, this is the index of the example containing this span of text.\r\n sample_index = sample_mapping[i]\r\n answers = examples[answer_column_name][sample_index]\r\n # print(examples,answers,offset_mapping,tokenized_examples)\r\n # If no answers are given, set the cls_index as answer.\r\n if len(answers) == 0:#len(answers[\"answer_start\"]) == 0:\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Start/end character index of the answer in the text.\r\n start_char = answers[0][\"answer_start\"]\r\n end_char = start_char + len(answers[0][\"text\"])\r\n\r\n # Start token index of the current span in the text.\r\n token_start_index = 0\r\n while sequence_ids[token_start_index] != (1 if pad_on_right else 0):\r\n token_start_index += 1\r\n\r\n # End token index of the current span in the text.\r\n token_end_index = len(input_ids) - 1\r\n while sequence_ids[token_end_index] != (1 if pad_on_right else 0):\r\n token_end_index -= 1\r\n\r\n # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).\r\n if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Otherwise move the token_start_index and token_end_index to the two ends of the answer.\r\n # Note: we could go after the last offset if the answer is the last word (edge case).\r\n while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:\r\n token_start_index += 1\r\n tokenized_examples[\"start_positions\"].append(token_start_index - 1)\r\n while offsets[token_end_index][1] >= end_char:\r\n token_end_index -= 1\r\n tokenized_examples[\"end_positions\"].append(token_end_index + 1)\r\n return tokenized_examples\r\n```", "I'm glad you managed to fix run_qa.py for your case :)\r\n\r\nRegarding the checksum error, I'm not able to reproduce on my side.\r\nThis errors says that the downloaded file doesn't match the expected file.\r\n\r\nCould you try running this and let me know if you get the same output as me ?\r\n```python\r\nfrom datasets.utils.info_utils import get_size_checksum_dict\r\nfrom datasets import cached_path\r\n\r\nget_size_checksum_dict(cached_path(\"https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json\"))\r\n# {'num_bytes': 30288272, 'checksum': '3527663986b8295af4f7fcdff1ba1ff3f72d07d61a20f487cb238a6ef92fd955'}\r\n```", "I run the code,and it show below:\r\n```\r\n>>> from datasets.utils.info_utils import get_size_checksum_dict\r\n>>> from datasets import cached_path\r\n>>> get_size_checksum_dict(cached_path(\"https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json\"))\r\nDownloading: 30.3MB [04:13, 120kB/s]\r\n{'num_bytes': 30288272, 'checksum': '3527663986b8295af4f7fcdff1ba1ff3f72d07d61a20f487cb238a6ef92fd955'}\r\n```", "Alright ! So in this case redownloading the file with `download_mode=\"force_redownload\"` should fix it. Can you try using `download_mode=\"force_redownload\"` again ?\r\n\r\nNot sure why it didn't work for you the first time though :/" ]

[ "Hi ! The labels -1 correspond to the examples without gold labels in the original snli dataset.\r\nFeel free to remove these examples if you don't need them by using\r\n```python\r\ndata = data.filter(lambda x: x[\"label\"] != -1)\r\n```" ]

[ "Hi,\r\n\r\ntry `skiprows` instead. This part is not properly documented in the docs it seems.\r\n\r\n@lhoestq I'll fix this as part of a bigger PR that fixes typos in the docs." ]

[ "Never mind, I ran it again and it worked this time. Strange." ]

[ "You can pass the same `experiment_id` for all the metrics of the same group, and use another `experiment_id` for the other groups.\r\nMaybe we can add an environment variable that sets the default value for `experiment_id` ? What do you think ?", "Ah, you're absolutely correct, @lhoestq - it's exactly the equivalent of the shared secret. Thank you!" ]

[ "Hi\r\nnot sure how change the lable after creation, but this is an issue not dataset request. thanks ", "Hi ! Currently you have to use `map` . You can see an example of how to do it in this comment: https://github.com/huggingface/datasets/issues/853#issuecomment-727872188\r\n\r\nIn the future we'll add support for a more native way of adding a new column ;)" ]

[ "Hi @gchhablani :)\r\nThanks for proposing your help !\r\n\r\nI'll be doing a refactor of some parts related to filtering in the scope of https://github.com/huggingface/datasets/issues/1877\r\nSo I would first wait for this refactor to be done before working on the filtering. In particular because I plan to make things simpler to manipulate.\r\n\r\nYour feedback on this refactor would also be appreciated since it also aims at making the core code more accessible (basically my goal is that no one's ever \"having troubles getting started\" ^^)\r\n\r\nThis will be available in a few days, I will be able to give you more details at that time if you don't mind waiting a bit !", "Sure! I don't mind waiting. I'll check the refactor and try to understand what you're trying to do :)" ]

[ "These warnings are showed when there's a call to `.map` to say to the user that a dataset is reloaded from the cache instead of being recomputed.\r\nThey are warnings since we want to make sure the users know that it's not recomputed.", "Thank you for explaining the intention, @lhoestq \r\n\r\n1. Could it be then made more human-friendly? Currently the hex gibberish tells me nothing of what's really going on. e.g. the following is instructive, IMHO:\r\n\r\n```\r\nWARNING: wmt16/ro-en/train dataset was loaded from cache instead of being recomputed\r\nWARNING: wmt16/ro-en/validation dataset was loaded from cache instead of being recomputed\r\nWARNING: wmt16/ro-en/test dataset was loaded from cache instead of being recomputed\r\n```\r\nnote that it removes the not so useful hex info and tells the user instead which split it's referring to - but probably no harm in keeping the path if it helps the debug. But the key is that now the warning is telling me what it is it's warning me about.\r\n```\r\nWarning:Loading cache path\r\n```\r\non the other hand isn't telling what it is warning about.\r\n\r\nAnd I still suggest this is INFO level, otherwise you need to turn all 'using cache' statements to WARNING to be consistent. The user is most likely well aware the cache is used for models, etc. So this feels very similar.\r\n\r\n2. Should there be a way for a user to void warranty by having a flag - `I know I'm expecting the cached version to load if it's available - please do not warn me about it=True`\r\n\r\nTo explain the need: Warnings are a problem, they constantly take attention away because they could be the harbinger of a problem. Therefore I prefer not to have any warnings in the log, and if I get any I usually try to deal with those so that my log is clean. \r\n\r\nIt's less of an issue for somebody doing long runs. It's a huge issue for someone who does a new run every few minutes and on the lookout for any potential problems which is what I have been doing a lot of integrating DeepSpeed and other things. And since there are already problems to deal with during the integration it's nice to have a clean log to start with. \r\n\r\nI hope my need is not unreasonable and I was able to explain it adequately. \r\n\r\nThank you.", "Hey, any news about the issue? So many warnings when I'm really ok with the dataset not being recomputed :)" ]

[ "sorry my mistake, datasets were overwritten closing now, thanks a lot" ]

[ "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" ]

[ "Thanks for reporting ! I'm taking a look", "Index training was missing, I fixed it here: https://github.com/huggingface/datasets/commit/f5986c46323583989f6ed1dabaf267854424a521\r\n\r\nCan you try again please ?", "Works great 👍 I just put a minor comment on the commit, I think you meant to pass the `train_size` from the one obtained from the config.\r\n\r\nThanks for a quick response!" ]

[ "Thanks for the report !\r\n\r\nCurrently we don't have a way to let the user easily disable this behavior.\r\nHowever I agree that we should support stateful processing functions, ideally by removing `does_function_return_dict`.\r\n\r\nWe needed this function in order to know whether the `map` functions needs to write data or not. if `does_function_return_dict` returns False then we don't write anything.\r\n\r\nInstead of checking the output of the processing function outside of the for loop that iterates through the dataset to process it, we can check the output of the first processed example and at that point decide if we need to write data or not.\r\n\r\nTherefore it's definitely possible to fix this unwanted behavior, any contribution going into this direction is welcome :)", "Thanks @mariosasko for the PR!" ]

[ "Thanks for reporting and for all the details and suggestions.\r\n\r\nI'm totally in favor of having a HF_DATASETS_OFFLINE env variable to disable manually all the connection checks, remove retries etc.\r\n\r\nMoreover you may know that the use case that you are mentioning is already supported from `datasets` 1.3.0, i.e. you already can:\r\n- first load datasets and metrics from an instance with internet connection\r\n- then be able to reload datasets and metrics from another instance without connection (as long as the filesystem is shared)\r\n\r\nThis is already implemented, but currently it only works if the requests return a `ConnectionError` (or any error actually). Not sure why it would hang instead of returning an error.\r\n\r\nMaybe this is just a issue with the timeout value being not set or too high ?\r\nIs there a way I can have access to one of the instances on which there's this issue (we can discuss this offline) ?\r\n", "I'm on master, so using all the available bells and whistles already.\r\n\r\nIf you look at the common issues - it for example tries to look up files if they appear in `_PACKAGED_DATASETS_MODULES` which it shouldn't do.\r\n\r\n--------------\r\n\r\nYes, there is a nuance to it. As I mentioned it's firewalled - that is it has a network but making any calls outside - it just hangs in:\r\n\r\n```\r\nsin_addr=inet_addr(\"xx.xx.xx.xx\")}, [28->16]) = 0\r\nclose(5) = 0\r\nsocket(AF_INET, SOCK_STREAM|SOCK_CLOEXEC, IPPROTO_TCP) = 5\r\nconnect(5, {sa_family=AF_INET, sin_port=htons(3128), sin_addr=inet_addr(\"yy.yy.yy.yy\")}, 16^C) = ? ERESTARTSYS (To be restarted if SA_RESTART is set)\r\n```\r\nuntil it times out.\r\n\r\nThat's why we need to be able to tell the software that there is no network to rely on even if there is one (good for testing too).\r\n\r\nSo what I'm thinking is that this is a simple matter of pre-ambling any network call wrappers with:\r\n\r\n```\r\nif HF_DATASETS_OFFLINE:\r\n assert \"Attempting to make a network call under Offline mode\"\r\n```\r\n\r\nand then fixing up if there is anything else to fix to make it work.\r\n\r\n--------------\r\n\r\nOtherwise I think the only other problem I encountered is that we need to find a way to pre-cache metrics, for some reason it's not caching it and wanting to fetch it from online.\r\n\r\nWhich is extra strange since it already has those files in the `datasets` repo itself that is on the filesystem.\r\n\r\nThe workaround I had to do is to copy `rouge/rouge.py` (with the parent folder) from the datasets repo to the current dir - and then it proceeded.", "Ok understand better the hanging issue.\r\nI guess catching connection errors is not enough, we should also avoid all the hangings.\r\nCurrently the offline mode tests are only done by simulating an instant connection fail that returns an error, let's have another connection mock that hangs instead.\r\n\r\nI'll also take a look at why you had to do this for `rouge`.\r\n", "FWIW, I think instant failure on the behalf of a network call is the simplest solution to correctly represent the environment and having the caller to sort it out is the next thing to do, since here it is the case of having no functional network, it's just that the software doesn't know this is the case, because there is some network. So we just need to help it to bail out instantly rather than hang waiting for it to time out. And afterwards everything else you said.", "Update on this: \r\n\r\nI managed to create a mock environment for tests that makes the connections hang until timeout.\r\nI managed to reproduce the issue you're having in this environment.\r\n\r\nI'll update the offline test cases to also test the robustness to connection hangings, and make sure we set proper timeouts where it's needed in the code. This should cover the _automatic_ section you mentioned.", "Fabulous! I'm glad you were able to reproduce the issues, @lhoestq!", "I lost access to the firewalled setup, but I emulated it with:\r\n\r\n```\r\nsudo ufw enable\r\nsudo ufw default deny outgoing\r\n```\r\n(thanks @mfuntowicz)\r\n\r\nI was able to test `HF_DATASETS_OFFLINE=1` and it worked great - i.e. didn't try to reach out with it and used the cached files instead.\r\n\r\nThank you!" ]

[ "Facing the same issue for [Squad](https://huggingface.co/datasets/viewer/?dataset=squad) and [TriviaQA](https://huggingface.co/datasets/viewer/?dataset=trivia_qa) datasets as well.", "We just fixed the issue, thanks for reporting !" ]

[ "Dataset added in release [1.5.0](https://github.com/huggingface/datasets/releases/tag/1.5.0), I think I can close this." ]

[ "Hi !\r\nI guess you can add a dataset without the fields that must be kept anonymous, and then update those when the anonymity period is over.\r\nYou can also make the PR from an anonymous org.\r\nPinging @yjernite just to make sure it's ok", "Hello,\r\nI would prefer to do the reverse: adding a link to an anonymous paper without the people names/institution in the PR. Would it be conceivable ?\r\nCheers\r\n", "Sure, I think it's ok on our side", "Yup, sounds good!" ]

[ "Hi @JieyuZhao !\r\n\r\nYou can edit the dataset card of wino_bias to update the URL via a Pull Request. This would be really appreciated :)\r\n\r\nThe dataset card is the README.md file you can find at https://github.com/huggingface/datasets/tree/master/datasets/wino_bias\r\nAlso the homepage url is also mentioned in the wino_bias.py so feel free to update it there as well.\r\n\r\nYou can create a Pull Request directly from the github interface by editing the files you want and submit a PR, or from a local clone of the repository.\r\n\r\nThanks for noticing !" ]

[ "Hi thanks for reporting and for proposing a fix :)\r\n\r\nI just merged a fix, feel free to try it from the master branch !", "Closing since this has been fixed by #1923" ]

[ "upgraded to php 3.9.2 and it works!" ]

[ "Thanks for reporting ! This is a bug. For now feel free to set `ignore_verifications=False` in `load_dataset`.\r\nI'm working on a fix", "I just merged a fix :)\r\n\r\nWe'll do a patch release soon. In the meantime feel free to try it from the master branch\r\nThanks again for reporting !", "Closing since this has been fixed by #1925" ]

[ "@thomwolf @lhoestq can you guys please take a look and recommend some solution.", "am suspicious of this thing? what's the purpose of this? pickling and unplickling\r\n`self = pickle.loads(pickle.dumps(self))`\r\n\r\n```\r\n def save_to_disk(self, dataset_path: str, fs=None):\r\n \"\"\"\r\n Saves a dataset to a dataset directory, or in a filesystem using either :class:`datasets.filesystem.S3FileSystem` or any implementation of ``fsspec.spec.AbstractFileSystem``.\r\n\r\n Args:\r\n dataset_path (``str``): path (e.g. ``dataset/train``) or remote uri (e.g. ``s3://my-bucket/dataset/train``) of the dataset directory where the dataset will be saved to\r\n fs (Optional[:class:`datasets.filesystem.S3FileSystem`,``fsspec.spec.AbstractFileSystem``], `optional`, defaults ``None``): instance of :class:`datasets.filesystem.S3FileSystem` or ``fsspec.spec.AbstractFileSystem`` used to download the files from remote filesystem.\r\n \"\"\"\r\n assert (\r\n not self.list_indexes()\r\n ), \"please remove all the indexes using `dataset.drop_index` before saving a dataset\"\r\n self = pickle.loads(pickle.dumps(self))\r\n ```", "It's been 24 hours and sadly it's still running. With not a single byte written", "Tried finding the root cause but was unsuccessful.\r\nI am using lazy tokenization with `dataset.set_transform()`, it works like a charm with almost same performance as pre-compute.", "Hi ! This very probably comes from the hack you used.\r\n\r\nThe pickling line was added an a sanity check because save_to_disk uses the same assumptions as pickling for a dataset object. The main assumption is that memory mapped pyarrow tables must be reloadable from the disk. In your case it's not possible since you altered the pyarrow table.\r\nI would suggest you to rebuild a valid Dataset object from your new pyarrow table. To do so you must first save your new table to a file, and then make a new Dataset object from that arrow file.\r\n\r\nYou can save the raw arrow table (without all the `datasets.Datasets` metadata) by calling `map` with `cache_file_name=\"path/to/outut.arrow\"` and `function=None`. Having `function=None` makes the `map` write your dataset on disk with no data transformation.\r\n\r\nOnce you have your new arrow file, load it with `datasets.Dataset.from_file` to have a brand new Dataset object :)\r\n\r\nIn the future we'll have a better support for the fast filtering method from pyarrow so you don't have to do this very unpractical workaround. Since it breaks somes assumptions regarding the core behavior of Dataset objects, this is very discouraged.", "Thanks, @lhoestq for your response. Will try your solution and let you know." ]

[ "Hi ! :)\r\n\r\nThis looks like the same issue as https://github.com/huggingface/datasets/issues/1856 \r\nBasically google drive has quota issues that makes it inconvenient for downloading files.\r\n\r\nIf the quota of a file is exceeded, you have to wait 24h for the quota to reset (which is painful).\r\n\r\nThe error says that the checksum of the downloaded file doesn't match because google drive returns a text file with the \"Quota Exceeded\" error instead of the actual data file.", "Thanks @lhoestq! Yes, it seems back to normal after a couple of days." ]

[ "We already have a ClassLabel type that does this kind of mapping between the label ids (integers) and actual label values (strings).\r\n\r\nI wonder if actually we should use the DictionaryType from Arrow and the Categorical type from pandas for the `datasets` ClassLabel feature type.\r\nCurrently ClassLabel corresponds to `pa.int64()` in pyarrow and `dtype('int64')` in pandas (so the label names are lost during conversions).\r\n\r\nWhat do you think ?", "Now that I've heard you explain ClassLabel, that makes a lot of sense! While DictionaryType for Arrow (I think) can have arbitrarily typed keys, so it won't cover all potential cases, pandas' Category is *probably* the most common use for that pyarrow type, and ClassLabel should match that perfectly?\r\n\r\nOther thoughts:\r\n\r\n- changing the resulting patype on ClassLabel might be backward-incompatible? I'm not totally sure if users of the `datasets` library tend to directly access the `patype` attribute (I don't think we really do, but we haven't been using it for very long yet).\r\n- would ClassLabel's dtype change to `dict[int64, string]`? It seems like in practice a ClassLabel (when not explicitly specified) would be constructed from the DictionaryType branch of `generate_from_arrow_type`, so it's not totally clear to me that anyone ever actually accesses/uses that dtype?\r\n- I don't quite know how `.int2str` and `.str2int` are used in practice - would those be kept? Perhaps the implementation might actually be substantially smaller if we can just delegate to pyarrow's dict methods?\r\n\r\nAnother idea that just occurred to me: add a branch in here to generate a ClassLabel if the dict key is int64 and the values are string: https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L932 , and then don't touch anything else.\r\n\r\nIn practice, I don't think this would be backward-incompatible in a way anyone would care about since the current behavior just throws an exception, and this way, we could support *reading* a pandas Categorical into a `Dataset` as a ClassLabel. I *think* from there, while it would require some custom glue it wouldn't be too hard to convert the ClassLabel into a pandas Category if we want to go back - I think this would improve on the current behavior without risking changing the behavior of ClassLabel in a backward-incompat way.\r\n\r\nThoughts? I'm not sure if this is overly cautious. Whichever approach you think is better, I'd be happy to take it on!\r\n", "I think we can first keep the int64 precision but with an arrow Dictionary for ClassLabel, and focus on the connection with arrow and pandas.\r\n\r\nIn this scope, I really like the idea of checking for the dictionary type:\r\n\r\n> Another idea that just occurred to me: add a branch in here to generate a ClassLabel if the dict key is int64 and the values are string: https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L932 , and then don't touch anything else.\r\n\r\nThis looks like a great start.\r\n\r\nThen as you said we'd have to add the conversion from classlabel to the correct arrow dictionary type. Arrow is already able to convert from arrow Dictionary to pandas Categorical so it should be enough.\r\n\r\nI can see two things that we must take case of to make this change backward compatible:\r\n- first we must still be able to load an arrow file with arrow int64 dtype and `datasets` ClassLabel type without crashing. This can be fixed by casting the arrow int64 array to an arrow Dictionary array on-the-fly when loading the table in the ArrowReader.\r\n- then we still have to return integers when accessing examples from a ClassLabel column. Currently it would return the strings values since it's based on the pandas behavior for converting from pandas to python/numpy. To do so we just have to adapt the python/numpy extractors in formatting.py (it takes care of converting an arrow table to a dictionary of python objects by doing arrow table -> pandas dataframe -> python dictionary)\r\n\r\nAny help on this matter is very much welcome :)" ]

[ "Thanks for reporting. This looks like a very bad issue. I'm looking into it", "I just merged a fix, we'll do a patch release soon. Thanks again for reporting, and sorry for the inconvenience.\r\nIn the meantime you can load `ALT` using `datasets` from the master branch", "Thanks!!! works perfectly in the bleading edge master version", "Closed by #1899" ]

[ "Thanks for reporting !\r\n\r\nYou're right, `string_to_arrow` should be able to take `\"timestamp[ns]\"` as input and return the right pyarrow timestamp type.\r\nFeel free to suggest a fix for `string_to_arrow` and open a PR if you want to contribute ! This would be very appreciated :)\r\n\r\nTo give you more context:\r\n\r\nAs you may know we define the features types of a dataset using the `Features` object in combination with feature types like `Value`. For example\r\n```python\r\nfeatures = Features({\r\n \"age\": Value(\"int32\")\r\n})\r\n```\r\nHowever under the hood we are actually using pyarrow to store the data, and so we have a mapping between the feature types of `datasets` and the types of pyarrow.\r\n\r\nFor example, the `Value` feature types are created from a pyarrow type with `Value(str(pa_type))`.\r\nHowever it looks like the conversion back to a pyarrow type doesn't work with `\"timestamp[ns]\"`.\r\nThis is the `string_to_arrow` function you highlighted that does this conversion, so we should fix that.\r\n\r\n", "Thanks for the clarification @lhoestq !\r\n\r\nThis may be a little bit of a stupid question, but I wanted to clarify one more thing before I took a stab at this:\r\n\r\nWhen the features get inferred, I believe they already have a pyarrow schema (https://github.com/huggingface/datasets/blob/master/src/datasets/arrow_dataset.py#L234).\r\n\r\nWe then convert it to a string (https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L778) only to convert it back into the arrow type (https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L143, and https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L35). Is there a reason for this round-trip?\r\n\r\nI'll open a PR later to add `timestamp` support to `string_to_arrow`, but I'd be curious to understand since it feels like there may be some opportunities to simplify!", "The objective in terms of design is to make it easy to create Features in a pythonic way. So for example we use a string to define a Value type.\r\nThat's why when inferring the Features from an arrow schema we have to find the right string definitions for Value types. I guess we could also have a constructor `Value.from_arrow_type` to avoid recreating the arrow type, but this could create silent errors if the pyarrow type doesn't have a valid mapping with the string definition. The \"round-trip\" is used to enforce that the ground truth is the string definition, not the pyarrow type, and also as a sanity check.\r\n\r\nLet me know if that makes sense ", "OK I think I understand now:\r\n\r\nFeatures are datasets' internal representation of a schema type, distinct from pyarrow's schema.\r\nValue() corresponds to pyarrow's \"primitive\" types (e.g. `int` or `string`, but not things like `list` or `dict`).\r\n`get_nested_type()` (https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L698) and `generate_from_arrow_type()` (https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L778) *should* be inverses of each other, and similarly, for the primitive values, `string_to_arrow()` and `Value.__call__` (https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L146) should be inverses of each other?\r\n\r\nThanks for taking the time to answer - I just wanted to make sure I understood before opening a PR so I'm not disrupting anything about how the codebase is expected to work!", "Yes you're totally right :)" ]

[ "Hi sam !\r\nIndeed we can expect the performances to be very close since both MMapIndexedDataset and the `datasets` implem use memory mapping. With memory mapping what determines the I/O performance is the speed of your hard drive/SSD.\r\n\r\nIn terms of performance we're pretty close to the optimal speed for reading text, even though I found recently that we could still slightly improve speed for big datasets (see [here](https://github.com/huggingface/datasets/issues/1803)).\r\n\r\nIn terms of number of examples and example sizes, the only limit is the available disk space you have.\r\n\r\nI haven't used `psrecord` yet but it seems to be a very interesting tool for benchmarking. Currently for benchmarks we only have github actions to avoid regressions in terms of speed. But it would be cool to have benchmarks with comparisons with other dataset tools ! This would be useful to many people", "Also I would be interested to know what data types `MMapIndexedDataset` supports. Is there some documentation somewhere ?", "no docs haha, it's written to support integer numpy arrays.\r\n\r\nYou can build one in fairseq with, roughly:\r\n```bash\r\n\r\nwget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip\r\nunzip wikitext-103-raw-v1.zip\r\nexport dd=$HOME/fairseq-py/wikitext-103-raw\r\n\r\nexport mm_dir=$HOME/mmap_wikitext2\r\nmkdir -p gpt2_bpe\r\nwget -O gpt2_bpe/encoder.json https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json\r\nwget -O gpt2_bpe/vocab.bpe https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe\r\nwget -O gpt2_bpe/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt\r\nfor SPLIT in train valid; do \\\r\n python -m examples.roberta.multiprocessing_bpe_encoder \\\r\n --encoder-json gpt2_bpe/encoder.json \\\r\n --vocab-bpe gpt2_bpe/vocab.bpe \\\r\n --inputs /scratch/stories_small/${SPLIT}.txt \\\r\n --outputs /scratch/stories_small/${SPLIT}.bpe \\\r\n --keep-empty \\\r\n --workers 60; \\\r\ndone\r\n\r\nmkdir -p $mm_dir\r\nfairseq-preprocess \\\r\n --only-source \\\r\n --srcdict gpt2_bpe/dict.txt \\\r\n --trainpref $dd/wiki.train.bpe \\\r\n --validpref $dd/wiki.valid.bpe \\\r\n --destdir $mm_dir \\\r\n --workers 60 \\\r\n --dataset-impl mmap\r\n```\r\n\r\nI'm noticing in my benchmarking that it's much smaller on disk than arrow (200mb vs 900mb), and that both incur significant cost by increasing the number of data loader workers. \r\nThis somewhat old [post](https://ray-project.github.io/2017/10/15/fast-python-serialization-with-ray-and-arrow.html) suggests there are some gains to be had from using `pyarrow.serialize(array).tobuffer()`. I haven't yet figured out how much of this stuff `pa.Table` does under the hood.\r\n\r\nThe `MMapIndexedDataset` bottlenecks we are working on improving (by using arrow) are:\r\n1) `MMapIndexedDataset`'s index, which stores offsets, basically gets read in its entirety by each dataloading process.\r\n2) we have separate, identical, `MMapIndexedDatasets` on each dataloading worker, so there's redundancy there; we wonder if there is a way that arrow can somehow dedupe these in shared memory.\r\n\r\nIt will take me a few hours to get `MMapIndexedDataset` benchmarks out of `fairseq`/onto a branch in this repo, but I'm happy to invest the time if you're interested in collaborating on some performance hacking." ]

[ "This was also mentioned in https://github.com/huggingface/datasets/issues/488 \r\n\r\nThe bucket where is data was stored seems to be unavailable now. Maybe we can change the URL to the ones in https://conferences.unite.un.org/uncorpus/en/downloadoverview ?", "Closing since this has been fixed by #1912" ]

[ "Yes that would be awesome. Not only the download speeds are awful, but also some files are missing.\r\nWe list all the URLs in the datasets/wmt19/wmt_utils.py so we can make a script to download them all and host on S3.\r\nAlso I think most of the materials are under the CC BY-NC-SA 3.0 license (must double check) so it should be possible to redistribute the data with no issues.\r\n\r\ncc @patrickvonplaten who knows more about the wmt scripts", "Yeah, the scripts are pretty ugly! A big refactor would make sense here...and I also remember that the datasets were veeery slow to download", "I'm downloading them.\r\nI'm starting with the ones hosted on http://data.statmt.org which are the slowest ones", "@lhoestq better to use our new git-based system than just raw S3, no? (that way we have built-in CDN etc.)", "Closing since the urls were changed to mirror urls in #1912 ", "Hi there! What about mirroring other datasets like [CCAligned](http://www.statmt.org/cc-aligned/) as well? All of them are really slow to download..." ]

[ "This is the current error thrown for missing datasets:\r\n```\r\nFileNotFoundError: Couldn't find a dataset script at C:\\Users\\Mario\\Desktop\\projects\\datasets\\missing_dataset\\missing_dataset.py or any data file in the same directory. Couldn't find 'missing_dataset' on the Hugging Face Hub either: FileNotFoundError: Dataset 'missing_dataset' doesn't exist on the Hub. If the repo is private, make sure you are authenticated with `use_auth_token=True` after logging in with `huggingface-cli login`.\r\n```\r\n\r\nSeems much more informative, so I think we can close this issue." ]

[ "I started working on this. My idea is to first add the pyarrow Table wrappers InMemoryTable and MemoryMappedTable that both implement what's necessary regarding copy/pickle. Then have another wrapper that takes the concatenation of InMemoryTable/MemoryMappedTable objects.\r\n\r\nWhat's important here is that concatenating two tables into one doesn't double the memory used (`total_allocated_bytes()` stays the same).", "Hi @lhoestq @albertvillanova,\r\n\r\nI checked the linked issues and PR, this seems like a great idea. Would you mind elaborating on the in-memory and memory-mapped datasets? \r\nBased on my understanding, it is something like this, please correct me if I am wrong:\r\n1. For in-memory datasets, we don't have any dataset files so the entire dataset is pickled to the cache during loading, and then whenever required it is unpickled .\r\n2. For on-disk/memory-mapped datasets, we have the data files provided, so they can be re-loaded from the paths, and only the file-paths are stored while pickling.\r\n\r\nIf this is correct, will the feature also handle pickling/unpickling of a concatenated dataset? Will this be cached?\r\n\r\nThis also leads me to ask whether datasets are chunked during pickling? \r\n\r\nThanks,\r\nGunjan", "Hi ! Yes you're totally right about your two points :)\r\n\r\nAnd in the case of a concatenated dataset, then we should reload each sub-table depending on whether it's in-memory or memory mapped. That means the dataset will be made of several blocks in order to keep track of what's from memory and what's memory mapped. This allows to pickle/unpickle concatenated datasets", "Hi @lhoestq\r\n\r\nThanks, that sounds nice. Can you explain where the issue of the double memory may arise? Also, why is the existing `concatenate_datasets` not sufficient for this purpose?", "Hi @lhoestq,\r\n\r\nWill the `add_item` feature also help with lazy writing (or no caching) during `map`/`filter`?", "> Can you explain where the issue of the double memory may arise?\r\n\r\nWe have to keep each block (in-memory vs memory mapped) separated in order to be able to reload them with pickle.\r\nOn the other hand we also need to have the full table from mixed in-memory and memory mapped data in order to iterate or extract data conveniently. That means that each block is accessible twice: once in the full table, and once in the separated blocks. But since pyarrow tables concatenation doesn't double the memory, then building the full table doesn't cost memory which is what we want :)\r\n\r\n> Also, why is the existing concatenate_datasets not sufficient for this purpose?\r\n\r\nThe existing `concatenate_datasets` doesn't support having both in-memory and memory mapped data together (there's no fancy block separation logic). It works for datasets fully in-memory or fully memory mapped but not a mix of the two.\r\n\r\n> Will the add_item feature also help with lazy writing (or no caching) during map/filter?\r\n\r\nIt will enable the implementation of the fast, masked filter from this discussion: https://github.com/huggingface/datasets/issues/1949\r\nHowever I don't think this will affect map." ]

[ "Thanks for reporting !\r\nThis is due to the changes made in the data files in the multiwoz repo: https://github.com/budzianowski/multiwoz/pull/59\r\nI'm opening a PR to update the checksums of the data files.", "I just merged the fix. It will be available in the new release of `datasets` later today.\r\nYou'll be able to get the new version with\r\n```\r\npip install --upgrade datasets\r\n```", "Hi, I still meet the error when loading the datasets after upgradeing datasets.\r\n\r\nraise NonMatchingChecksumError(error_msg + str(bad_urls))\r\ndatasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['https://github.com/budzianowski/multiwoz/raw/master/data/MultiWOZ_2.2/dialog_acts.json', 'https://github.com/budzianowski/multiwoz/raw/master/data/MultiWOZ_2.2/test/dialogues_001.json']", "This must be related to https://github.com/budzianowski/multiwoz/pull/72\r\nThose files have changed, let me update the checksums for this dataset.\r\n\r\nFor now you can use `ignore_verifications=True` in `load_dataset` to skip the checksum verification." ]

[ "Hi ! Indeed if you add a column to a formatted dataset, then the new dataset gets a new formatting in which:\r\n```\r\nnew formatted columns = (all columns - previously unformatted columns)\r\n```\r\nTherefore the new column is going to be formatted using the `torch` formatting.\r\n\r\nIf you want your new column to be unformatted you can re-run this line:\r\n```python\r\ndata.set_format(\"torch\", columns=[\"some_integer_column1\", \"some_integer_column2\"], output_all_columns=True)\r\n```", "Hi, thanks that solved my problem. Maybe mention that in the documentation. ", "Ok cool :) \r\nAlso I just did a PR to mention this behavior in the documentation", "Closed by #1888" ]

[ "Hi @alejandrocros it looks like an incompatibility with the current Trainer @sgugger \r\nIndeed currently the Trainer of `transformers` doesn't support a dataset with a transform\r\n\r\nIt looks like it comes from this line: https://github.com/huggingface/transformers/blob/f51188cbe74195c14c5b3e2e8f10c2f435f9751a/src/transformers/trainer.py#L442\r\n\r\nThis line sets the format to not return certain unused columns. But this has two issues:\r\n1. it forgets to also set the format_kwargs (this causes the error you got):\r\n```python\r\ndataset.set_format(type=dataset.format[\"type\"], columns=columns, format_kwargs=dataset.format[\"format_kwargs\"])\r\n```\r\n2. the Trainer wants to keep only the fields that are used as input for a model. However for a dataset with a transform, the output fields are often different from the columns fields. For example from a column \"text\" in the dataset, the strings can be transformed on-the-fly into \"input_ids\". If you want your dataset to only output certain fields and not other you must change your transform function.\r\n", "FYI that option can be removed with `remove_unused_columns = False` in your `TrainingArguments`, so there is a workaround @alexvaca0 while the fix in `Trainer` is underway.\r\n\r\n@lhoestq I think I will just use the line you suggested and if someone is using the columns that are removed in their transform they will need to change `remove_unused_columns` to `False`. We might switch the default of that argument in the next version if that proves too bug-proof.", "I've tried your solutions @sgugger @lhoestq and the good news is that it throws no error. However, TPU training is taking forever, in 1 hour it has only trained 1 batch of 8192 elements, which doesn't make much sense... Is it possible that \"on the fly\" tokenization of batches is slowing down TPU training to that extent?", "I'm pretty sure this is because of padding but @sgugger might know better", "I don't know what the value of `padding` is in your lines of code pasted above so I can't say for sure. The first batch will be very slow on TPU since it compiles everything, so that's normal (1 hour is long but 8192 elements is also large). Then if your batches are not of the same lengths, it will recompile everything at each step instead of using the same graph, which will be very slow, so you should double check you are using padding to make everything the exact same shape. ", "I have tried now on a GPU and it goes smooth! Amazing feature .set_transform() instead of .map()! Now I can pre-train my model without the hard disk limitation. Thanks for your work all HuggingFace team!! :clap: ", "In the end, to make it work I turned to A-100 gpus instead of TPUS, among other changes. Set_transform doesn't work as expected and slows down training very much even in GPUs, and applying map destroys the disk, as it multiplies by 100 the size of the data passed to it (due to inefficient implementation converting strings to int64 floats I guess). For that reason, I chose to use datasets to load the data as text, and then edit the Collator from Transformers to tokenize every batch it receives before processing it. That way, I'm being able to train fast, without memory breaks, without the disk being unnecessarily filled, while making use of GPUs almost all the time I'm paying for them (the map function over the whole dataset took ~15hrs, in which you're not training at all). I hope this info helps others that are looking for training a language model from scratch cheaply, I'm going to close the issue as the optimal solution I found after many experiments to the problem posted in it is explained above. ", "Great comment @alexvaca0 . I think that we could re-open the issue as a reformulation of why it takes so much space to save the arrow. Saving a 1% of oscar corpus takes more thank 600 GB (it breaks when it pass 600GB because it is the free memory that I have at this moment) when the full dataset is 1,3 TB. I have a 1TB M.2 NVMe disk that I can not train on because the saved .arrow files goes crazily big. If you can share your Collator I will be grateful. " ]

[ "Nevermind, this one is already available on the hub under the name `'wino_bias'`: https://huggingface.co/datasets/wino_bias" ]

[ "Hi @NielsRogge I would like to work on this dataset.\r\n\r\nThanks!", "Hi @udapy, are you working on this?" ]

[ "Hi @corticalstack ! Thanks for reporting. Indeed in the recent versions of Faiss we must use `getDevice` to check if the index in on GPU.\r\n\r\nI'm opening a PR", "I fixed this issue. It should work fine now.\r\nFeel free to try it out by installing `datasets` from source.\r\nOtherwise you can wait for the next release of `datasets` (in a few days)", "Thanks for such a quick fix and merge to master, pip installed git master, tested all OK" ]

[ "Hi ! We're in the process of switching the community datasets to git repos, exactly like what we're doing for models.\r\nYou can find an example here:\r\nhttps://huggingface.co/datasets/lhoestq/custom_squad/tree/main\r\n\r\nWe'll update the CLI in the coming days and do a new release :)\r\n\r\nAlso cc @julien-c maybe we can make improve the error message ?" ]

[ "Hi ! This issue may be related to #996 \r\nThis comes probably from the Quota Exceeded error from Google Drive.\r\nCan you try again tomorrow and see if you still have the error ?\r\n\r\nOn my side I didn't get any error today with `load_dataset(\"amazon_polarity\")`", "+1 encountering this issue as well", "@lhoestq Hi! I encounter the same error when loading `yelp_review_full`.\r\n\r\n```\r\nfrom datasets import load_dataset\r\ndataset_yp = load_dataset(\"yelp_review_full\")\r\n```\r\n\r\nWhen you say the \"Quota Exceeded from Google drive\". Is this a quota from the dataset owner? or the quota from our (the runner) Google Drive?", "+1 Also encountering this issue", "> When you say the \"Quota Exceeded from Google drive\". Is this a quota from the dataset owner? or the quota from our (the runner) Google Drive?\r\n\r\nEach file on Google Drive can be downloaded only a certain amount of times per day because of a quota. The quota is reset every day. So if too many people download the dataset the same day, then the quota is likely to exceed.\r\nThat's a really bad limitations of Google Drive and we should definitely find another host for these dataset than Google Drive.\r\nFor now I would suggest to wait and try again later..\r\n\r\nSo far the issue happened with CNN DailyMail, Amazon Polarity and Yelp Reviews. \r\nAre you experiencing the issue with other datasets ? @calebchiam @dtch1997 ", "@lhoestq Gotcha, that is quite problematic...for what it's worth, I've had no issues with the other datasets I tried, such as `yelp_reviews_full` and `amazon_reviews_multi`.", "Same issue today with \"big_patent\", though the symptoms are slightly different.\r\n\r\nWhen running\r\n\r\n```py\r\nfrom datasets import load_dataset\r\nload_dataset(\"big_patent\", split=\"validation\")\r\n```\r\n\r\nI get the following\r\n`FileNotFoundError: Local file \\huggingface\\datasets\\downloads\\6159313604f4f2c01e7d1cac52139343b6c07f73f6de348d09be6213478455c5\\bigPatentData\\train.tar.gz doesn't exist`\r\n\r\nI had to look into `6159313604f4f2c01e7d1cac52139343b6c07f73f6de348d09be6213478455c5` (which is a file instead of a folder) and got the following:\r\n\r\n`<!DOCTYPE html><html><head><title>Google Drive - Quota exceeded</title><meta http-equiv=\"content-type\" content=\"text/html; charset=utf-8\"/><link href=&#47;static&#47;doclist&#47;client&#47;css&#47;4033072956&#45;untrustedcontent.css rel=\"stylesheet\" nonce=\"JV0t61Smks2TEKdFCGAUFA\"><link rel=\"icon\" href=\"//ssl.gstatic.com/images/branding/product/1x/drive_2020q4_32dp.png\"/><style nonce=\"JV0t61Smks2TEKdFCGAUFA\">#gbar,#guser{font-size:13px;padding-top:0px !important;}#gbar{height:22px}#guser{padding-bottom:7px !important;text-align:right}.gbh,.gbd{border-top:1px solid #c9d7f1;font-size:1px}.gbh{height:0;position:absolute;top:24px;width:100%}@media all{.gb1{height:22px;margin-right:.5em;vertical-align:top}#gbar{float:left}}a.gb1,a.gb4{text-decoration:underline !important}a.gb1,a.gb4{color:#00c !important}.gbi .gb4{color:#dd8e27 !important}.gbf .gb4{color:#900 !important}\r\n</style><script nonce=\"iNUHigT+ENVQ3UZrLkFtRw\"></script></head><body><div id=gbar><nobr><a target=_blank class=gb1 href=\"https://www.google.fr/webhp?tab=ow\">Search</a> <a target=_blank class=gb1 href=\"http://www.google.fr/imghp?hl=en&tab=oi\">Images</a> <a target=_blank class=gb1 href=\"https://maps.google.fr/maps?hl=en&tab=ol\">Maps</a> <a target=_blank class=gb1 href=\"https://play.google.com/?hl=en&tab=o8\">Play</a> <a target=_blank class=gb1 href=\"https://www.youtube.com/?gl=FR&tab=o1\">YouTube</a> <a target=_blank class=gb1 href=\"https://news.google.com/?tab=on\">News</a> <a target=_blank class=gb1 href=\"https://mail.google.com/mail/?tab=om\">Gmail</a> <b class=gb1>Drive</b> <a target=_blank class=gb1 style=\"text-decoration:none\" href=\"https://www.google.fr/intl/en/about/products?tab=oh\"><u>More</u> &raquo;</a></nobr></div><div id=guser width=100%><nobr><span id=gbn class=gbi></span><span id=gbf class=gbf></span><span id=gbe></span><a target=\"_self\" href=\"/settings?hl=en_US\" class=gb4>Settings</a> | <a target=_blank href=\"//support.google.com/drive/?p=web_home&hl=en_US\" class=gb4>Help</a> | <a target=_top id=gb_70 href=\"https://accounts.google.com/ServiceLogin?hl=en&passive=true&continue=https://drive.google.com/uc%3Fexport%3Ddownload%26id%3D1J3mucMFTWrgAYa3LuBZoLRR3CzzYD3fa&service=writely&ec=GAZAMQ\" class=gb4>Sign in</a></nobr></div><div class=gbh style=left:0></div><div class=gbh style=right:0></div><div class=\"uc-main\"><div id=\"uc-text\"><p class=\"uc-error-caption\">Sorry, you can&#39;t view or download this file at this time.</p><p class=\"uc-error-subcaption\">Too many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.</p></div></div><div class=\"uc-footer\"><hr class=\"uc-footer-divider\">&copy; 2021 Google - <a class=\"goog-link\" href=\"//support.google.com/drive/?p=web_home\">Help</a> - <a class=\"goog-link\" href=\"//support.google.com/drive/bin/answer.py?hl=en_US&amp;answer=2450387\">Privacy & Terms</a></div></body></html>`", "A similar issue arises when trying to stream the dataset\r\n\r\n```python\r\n>>> from datasets import load_dataset\r\n>>> iter_dset = load_dataset(\"amazon_polarity\", split=\"test\", streaming=True)\r\n>>> iter(iter_dset).__next__()\r\n\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n~\\lib\\tarfile.py in nti(s)\r\n 186 s = nts(s, \"ascii\", \"strict\")\r\n--> 187 n = int(s.strip() or \"0\", 8)\r\n 188 except ValueError:\r\n\r\nValueError: invalid literal for int() with base 8: 'e nonce='\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nInvalidHeaderError Traceback (most recent call last)\r\n~\\lib\\tarfile.py in next(self)\r\n 2288 try:\r\n-> 2289 tarinfo = self.tarinfo.fromtarfile(self)\r\n 2290 except EOFHeaderError as e:\r\n\r\n~\\lib\\tarfile.py in fromtarfile(cls, tarfile)\r\n 1094 buf = tarfile.fileobj.read(BLOCKSIZE)\r\n-> 1095 obj = cls.frombuf(buf, tarfile.encoding, tarfile.errors)\r\n 1096 obj.offset = tarfile.fileobj.tell() - BLOCKSIZE\r\n\r\n~\\lib\\tarfile.py in frombuf(cls, buf, encoding, errors)\r\n 1036\r\n-> 1037 chksum = nti(buf[148:156])\r\n 1038 if chksum not in calc_chksums(buf):\r\n\r\n~\\lib\\tarfile.py in nti(s)\r\n 188 except ValueError:\r\n--> 189 raise InvalidHeaderError(\"invalid header\")\r\n 190 return n\r\n\r\nInvalidHeaderError: invalid header\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nReadError Traceback (most recent call last)\r\n<ipython-input-5-6b9058341b2b> in <module>\r\n----> 1 iter(iter_dset).__next__()\r\n\r\n~\\lib\\site-packages\\datasets\\iterable_dataset.py in __iter__(self)\r\n 363\r\n 364 def __iter__(self):\r\n--> 365 for key, example in self._iter():\r\n 366 if self.features:\r\n 367 # we encode the example for ClassLabel feature types for example\r\n\r\n~\\lib\\site-packages\\datasets\\iterable_dataset.py in _iter(self)\r\n 360 else:\r\n 361 ex_iterable = self._ex_iterable\r\n--> 362 yield from ex_iterable\r\n 363\r\n 364 def __iter__(self):\r\n\r\n~\\lib\\site-packages\\datasets\\iterable_dataset.py in __iter__(self)\r\n 77\r\n 78 def __iter__(self):\r\n---> 79 yield from self.generate_examples_fn(**self.kwargs)\r\n 80\r\n 81 def shuffle_data_sources(self, seed: Optional[int]) -> \"ExamplesIterable\":\r\n\r\n~\\.cache\\huggingface\\modules\\datasets_modules\\datasets\\amazon_polarity\\56923eeb72030cb6c4ea30c8a4e1162c26b25973475ac1f44340f0ec0f2936f4\\amazon_polarity.py in _generate_examples(self, filepath, files)\r\n 114 def _generate_examples(self, filepath, files):\r\n 115 \"\"\"Yields examples.\"\"\"\r\n--> 116 for path, f in files:\r\n 117 if path == filepath:\r\n 118 lines = (line.decode(\"utf-8\") for line in f)\r\n\r\n~\\lib\\site-packages\\datasets\\utils\\streaming_download_manager.py in __iter__(self)\r\n 616\r\n 617 def __iter__(self):\r\n--> 618 yield from self.generator(*self.args, **self.kwargs)\r\n 619\r\n 620\r\n\r\n~\\lib\\site-packages\\datasets\\utils\\streaming_download_manager.py in _iter_from_urlpath(cls, urlpath, use_auth_token)\r\n 644 ) -> Generator[Tuple, None, None]:\r\n 645 with xopen(urlpath, \"rb\", use_auth_token=use_auth_token) as f:\r\n--> 646 yield from cls._iter_from_fileobj(f)\r\n 647\r\n 648 @classmethod\r\n\r\n~\\lib\\site-packages\\datasets\\utils\\streaming_download_manager.py in _iter_from_fileobj(cls, f)\r\n 624 @classmethod\r\n 625 def _iter_from_fileobj(cls, f) -> Generator[Tuple, None, None]:\r\n--> 626 stream = tarfile.open(fileobj=f, mode=\"r|*\")\r\n 627 for tarinfo in stream:\r\n 628 file_path = tarinfo.name\r\n\r\n~\\lib\\tarfile.py in open(cls, name, mode, fileobj, bufsize, **kwargs)\r\n 1603 stream = _Stream(name, filemode, comptype, fileobj, bufsize)\r\n 1604 try:\r\n-> 1605 t = cls(name, filemode, stream, **kwargs)\r\n 1606 except:\r\n 1607 stream.close()\r\n\r\n~\\lib\\tarfile.py in __init__(self, name, mode, fileobj, format, tarinfo, dereference, ignore_zeros, encoding, errors, pax_headers, debug, errorlevel, copybufsize)\r\n 1484 if self.mode == \"r\":\r\n 1485 self.firstmember = None\r\n-> 1486 self.firstmember = self.next()\r\n 1487\r\n 1488 if self.mode == \"a\":\r\n\r\n~\\lib\\tarfile.py in next(self)\r\n 2299 continue\r\n 2300 elif self.offset == 0:\r\n-> 2301 raise ReadError(str(e))\r\n 2302 except EmptyHeaderError:\r\n 2303 if self.offset == 0:\r\n\r\nReadError: invalid header\r\n\r\n```", "This error still happens, but for a different reason now: Google Drive returns a warning instead of the dataset.", "Met the same issue +1", "Hi ! Thanks for reporting. Google Drive changed the way to bypass the warning message recently.\r\n\r\nThe latest release `1.18.4` fixes this for datasets loaded in a regular way.\r\n\r\nWe opened a PR to fix this recently for streaming mode at #3843 - we'll do a new release once the fix is merged :)", "Fixed by:\r\n- #3787 \r\n- #3843" ]

[ "Hi @sshleifer.\r\n\r\nI am not sure of understanding the need of the `add_item` approach...\r\n\r\nBy just reading your \"Desired API\" section, I would say you could (nearly) get it with a 1-column Dataset:\r\n```python\r\ndata = {\"input_ids\": [np.array([4,4,2]), np.array([8,6,5,5,2]), np.array([3,3,31,5])]}\r\nds = Dataset.from_dict(data)\r\nassert (ds[\"input_ids\"][0] == np.array([4,4,2])).all()\r\n```", "Hi @sshleifer :) \r\n\r\nWe don't have methods like `Dataset.add_batch` or `Dataset.add_entry/add_item` yet.\r\nBut that's something we'll add pretty soon. Would an API that looks roughly like this help ? Do you have suggestions ?\r\n```python\r\nimport numpy as np\r\nfrom datasets import Dataset\r\n\r\ntokenized = [np.array([4,4,2]), np.array([8,6,5,5,2]), np.array([3,3,31,5])\r\n\r\n# API suggestion (not available yet)\r\nd = Dataset()\r\nfor input_ids in tokenized:\r\n d.add_item({\"input_ids\": input_ids})\r\n\r\nprint(d[0][\"input_ids\"])\r\n# [4, 4, 2]\r\n```\r\n\r\nCurrently you can define a dataset with what @albertvillanova suggest, or via a generator using dataset builders. It's also possible to [concatenate datasets](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=concatenate#datasets.concatenate_datasets).", "Your API looks perfect @lhoestq, thanks!" ]

[ "@patrickvonplaten Could you please help me with how the output text has to be represented in the data? TIMIT has Words, Phonemes and texts. Also has lot on info on the speaker and the dialect. Could you please help me? An example of how to arrange it would be super helpful!\r\n\r\n", "Hey @vrindaprabhu - sure I'll help you :-) Could you open a first PR for TIMIT where you copy-paste more or less the `librispeech_asr` script: https://github.com/huggingface/datasets/blob/28be129db862ec89a87ac9349c64df6b6118aff4/datasets/librispeech_asr/librispeech_asr.py#L93 (obviously replacing all the naming and links correctly...) and then you can list all possible outputs in the features dict: https://github.com/huggingface/datasets/blob/28be129db862ec89a87ac9349c64df6b6118aff4/datasets/librispeech_asr/librispeech_asr.py#L104 (words, phonemes should probably be of kind `datasets.Sequence(datasets.Value(\"string\"))` and texts I think should be of type `\"text\": datasets.Value(\"string\")`.\r\n\r\nWhen you've opened a first PR, I think it'll be much easier for us to take a look together :-) ", "I am sorry! I created the PR [#1903](https://github.com/huggingface/datasets/pull/1903#). Requesting your comments! CircleCI tests are failing, will address them along with your comments!" ]

[ "Hi ! You're right this can can useful.\r\nThis should be easy to add, so feel free to give it a try if you want to contribute :)\r\nI think we just need to add it to the _generate_examples method of the OpenSubtitles dataset builder [here](https://github.com/huggingface/datasets/blob/master/datasets/open_subtitles/open_subtitles.py#L103)", "Hey @lhoestq , absolutely yes! Just one question before I start implementing. The ids found in the zip file have this format: \r\n(the following is line `22497315` of the `ids` file of the `de-en` dump)\r\n\r\n\r\n`de/2017/7006210/7063319.xml.gz en/2017/7006210/7050201.xml.gz 335 339 340` (every space is actually a tab, aside from the space between `339` and `340`)\r\n\r\n\r\nWhere filenames encode the information like this: `lang/year/imdb_id/opensubtitles_id.xml.gz` whereas the numbers correspond to the sentence ids which are linked together (i.e. sentence `335` of the German subtitle corresponds to lines `339` and `340` of the English file)\r\n\r\nThat being said, do you think I should stick to the raw sentence id (and replace the current sequential id) or should I include more detailed metadata (or both things maybe)?\r\n\r\nGoing with raw ID is surely simpler, but including `year`, `imdbId` and `subtitleId` should save space as they're just integers; besides, any operation (like filtering or grouping) will be much easier if users don't have to manually parse the ids every time.\r\nAs for the language-specific sentenceIds, what could be the best option? A list of integers or a comma-separated string?\r\n\r\n**Note:** I did not find any official information about this encoding, but it appears to check out:\r\nhttps://www.imdb.com/title/tt7006210/, https://www.opensubtitles.org/en/subtitles/7063319 and https://www.opensubtitles.org/en/subtitles/7050201 all link to the same episode, so I guess (I hope!) it's correct.\r\n\r\n", "I like the idea of having `year`, `imdbId` and `subtitleId` as columns for filtering for example.\r\nAnd for the `sentenceIds` a list of integers is fine.", "Thanks for improving it @Valahaar :) ", "Something like this? (adapted from [here](https://github.com/huggingface/datasets/blob/master/datasets/open_subtitles/open_subtitles.py#L114))\r\n\r\n```python\r\nresult = (\r\n sentence_counter,\r\n {\r\n \"id\": str(sentence_counter),\r\n \"meta\": {\r\n \"year\": year,\r\n \"imdbId\": imdb_id,\r\n \"subtitleId\": {l1: l1_sub_id, l2: l2_sub_id},\r\n \"sentenceIds\": {l1: [... source_sids ...], l2: [... target_sids ...]},\r\n # or maybe src/tgt? I'd go with the first one for consistency with 'translation'\r\n \"subtitleId\": {\"src\": l1_sub_id, \"tgt\": l2_sub_id},\r\n \"sentenceIds\": {\"src\": [... source_sids ...], \"tgt\": [... target_sids ...]},\r\n },\r\n \"translation\": {l1: x, l2: y},\r\n },\r\n )\r\n```\r\nOr at top level, avoiding nesting into 'meta'?", "Merged in #1865, closing. Thanks :)" ]

[ "Hi @patrickvonplaten I would like to work on this dataset. \r\n\r\nThanks! ", "That's awesome! Actually, I just noticed that this dataset might become a bit too big!\r\n\r\nMuST-C is the main dataset used for IWSLT19 and should probably be added as a standalone dataset. Would you be interested also in adding `datasets/MuST-C` instead?\r\n\r\nDescription: \r\n_MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems for speech translation from English into several languages. For each target language, MuST-C comprises several hundred hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual transcriptions and translations._\r\n\r\nPaper: https://www.aclweb.org/anthology/N19-1202.pdf\r\n\r\nDataset: https://ict.fbk.eu/must-c/ (One needs to fill out a short from to download the data, but it's very easy).\r\n\r\nIt would be awesome if you're interested in adding this datates. I'm very happy to guide you through the PR! I think the easiest way to start would probably be to read [this README on how to add a dataset](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md) and open a PR. Think you can copy & paste some code from:\r\n\r\n- Librispeech_asr: https://github.com/huggingface/datasets/blob/master/datasets/librispeech_asr/librispeech_asr.py\r\n- Flores Translation: https://github.com/huggingface/datasets/blob/master/datasets/flores/flores.py\r\n\r\nThink all the rest can be handled on the PR :-) ", "Hi @patrickvonplaten \r\nI have tried downloading this dataset, but the connection seems to reset all the time. I have tried it via the browser, wget, and using gdown . But it gives me an error message. _\"The server is busy or down, pls try again\"_ (rephrasing the message here)\r\n\r\nI have completed adding 4 datasets in the previous data sprint (including the IWSLT dataset #1676 ) ...so just checking if you are able to download it at your end. Otherwise will write to the dataset authors to update the links. \r\n\r\n\r\n\r\n\r\n", "Let me check tomorrow! Thanks for leaving this message!", "cc @patil-suraj for notification ", "@skyprince999, I think I'm getting the same error you're getting :-/\r\n\r\n```\r\nSorry, you can't view or download this file at this time.\r\n\r\nToo many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.\r\n```\r\n\r\nIt would be great if you could write the authors to see whether they can fix it.\r\nAlso cc @lhoestq - do you think we could mirror the dataset? ", "Also there are huge those datasets. Think downloading MuST-C v1.2 amounts to ~ 1000GB... because there are 14 possible configs each around 60-70GB. I think users mostly will only use one of the 14 configs so that they would only need, in theory, will have to download ~60GB which is ok. But I think this functionality doesn't exist yet in `datasets` no? cc @lhoestq ", "> Also cc @lhoestq - do you think we could mirror the dataset?\r\n\r\nYes we can mirror it if the authors are fine with it. You can create a dataset repo on huggingface.co (possibly under the relevant org) and add the mirrored data files.\r\n\r\n> I think users mostly will only use one of the 14 configs so that they would only need, in theory, will have to download ~60GB which is ok. But I think this functionality doesn't exist yet in datasets no? cc @lhoestq\r\n\r\nIf there are different download links for each configuration we can make the dataset builder download only the files related to the requested configuration.", "I have written to the dataset authors, highlighting this issue. Waiting for their response. \r\n\r\nUpdate on 25th Feb: \r\nThe authors have replied back, they are updating the download link and will revert back shortly! \r\n\r\n```\r\nfirst of all thanks a lot for being interested in MuST-C and for building the data-loader.\r\n\r\nBefore answering your request, I'd like to clarify that the creation, maintenance, and expansion of MuST-c are not supported by any funded project, so this means that we need to find economic support for all these activities. This also includes permanently moving all the data to AWS or GCP. We are working at this with the goal of facilitating the use of MuST-C, but this is not something that can happen today. We hope to have some news ASAP and you will be among the first to be informed.\r\n\r\nI hope you understand our situation.\r\n```\r\n\r\n", "Awesome, actually @lhoestq let's just ask the authors if we should host the dataset no? They could just use our links then as well for their website - what do you think? Is it fine to use our AWS dataset storage also as external links? ", "Yes definitely. Shall we suggest them to create a dataset repository under their org on huggingface.co ? @julien-c \r\nThe dataset is around 1TB", "Sounds good! \r\n\r\nOrder of magnitude is storage costs ~$20 per TB per month (not including bandwidth). \r\n\r\nHappy to provide this to the community as I feel this is an important dataset. Let us know what the authors want to do!\r\n\r\n", "Great! @skyprince999, do you think you could ping the authors here or link to this thread? I think it could be a cool idea to host the dataset on our side then", "Done. They replied back, and they want to have a call over a meet/ skype. Is that possible ? \r\nBtw @patrickvonplaten you are looped in that email (_pls check you gmail account_) ", "Hello! Any news on this?", "@gegallego there were some concerns regarding dataset usage & attribution by a for-profit company, so couldn't take it forward. Also the download links were unstable. \r\nBut I guess if you want to test the fairseq benchmarks, you can connect with them directly for downloading the dataset. ", "Yes, that dataset is not easy to download... I had to copy it to my Google Drive and use `rsync` to be able to download it.\r\nHowever, we could add the dataset with a manual download, right?", "yes that is possible. I couldn't unfortunately complete this PR, If you would like to add it, please feel free to do it. " ]