Datasets:

---

zongxiao

/

github-issues

like 0

[]

[]

[ "Looks good to me :-) \r\n\r\nJust would prefer to get rid of the `_DYNAMICALLY_IMPORTED_MODULE` attribute and replace it by a `get_imported_module()` function. Maybe there is something I'm not seeing here though - what do you think? ", "> * I'm not sure I understand the general organization of checksums. I see we have a checksum folder with potentially several checksum files but I also see that checksum files can potentially contain several checksums. Could you explain a bit more how this is organized?\r\n\r\nIt should look like this:\r\nsquad/\r\n├── squad.py/\r\n└── urls_checksums/\r\n...........└── checksums.txt\r\n\r\nIn checksums.txt, the format is one line per (url, size, checksum)\r\n\r\nI don't have a strong opinion between `urls_checksums/checksums.txt` or directly `checksums.txt` (not inside the `urls_checksums` folder), let me know what you think.\r\n\r\n\r\n> * Also regarding your comment on checksum files for \"canonical\" datasets. I understand we can just create these with `nlp-cli test` and then upload them manually to our S3, right?\r\n\r\nYes you're right", "Update of the commands:\r\n\r\n- nlp-cli test \\<dataset\\> : Run download_and_prepare and verify checksums\r\n * --name \\<name\\> : run only for the name\r\n * --all_configs : run for all configs\r\n * --save_checksums : instead of verifying checksums, compute and save them\r\n * --ignore_checksums : don't do checksums verification\r\n\r\n- nlp-cli upload \\<dataset_folder\\> : Upload a dataset\r\n * --upload_checksums : compute and upload checksums for uploaded files\r\n\r\nTODO:\r\n- don't overwrite checksums files on S3, to let the user upload a dataset in several steps if needed\r\n\r\nQuestion:\r\n- One idea from @patrickvonplaten : shall we upload checksums everytime we upload files ? (and therefore remove the upload_checksums parameter)", "Ok, ready to merge, then @lhoestq ?", "Yep :)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "For conflicts, I think the mention hint \"This should be modified because it mentions ...\" is missing.", "Looks great!" ]

[]

[]

[]

[]

[]

[ "Awesome!" ]

[]

[]

[ "LGTM" ]

[]

[]

[]

[]

[]

[ "Looks great to me! ", "The new download manager is ready. I removed the old folder and I fixed a few remaining dependencies.\r\nI tested it on squad and a few others from the dataset folder and it works fine.\r\n\r\nThe only impact of these changes is that it breaks the `download_and_prepare` script that was used to register the checksums when we create a dataset, as the checksum logic is not implemented.\r\n\r\nLet me know if you have remarks", "Ok merged it (a bit fast for you to update the copyright, now I see that. but it's ok, we'll do a pass on these doc/copyright before releasing anyway)", "Actually two additional things here @lhoestq (I merged too fast sorry, let's make a new PR for additional developments):\r\n- I think we can remove some dependencies now (e.g. `promises`) in setup.py, can you have a look?\r\n- also, I think we can remove the boto3 dependency like here: https://github.com/huggingface/transformers/pull/3968" ]

[]

[]

[ "> I think I'm fine with this.\r\n> \r\n> The alternative would be to host a small subset of the dataset on the S3 together with the testing script. But I think having all (test file creation + actual tests) in one file is actually quite convenient.\r\n> \r\n> Good for me!\r\n> \r\n> One question though, will we have to create one test file for each of the 100+ datasets or could we make some automatic conversion from tfds dataset test files?\r\n\r\nI think if we go the way shown in the PR we would have to create one test file for each of the 100+ datasets. \r\n\r\nAs far as I know the tfds test files all rely on the user having created a special download folder structure in `tensorflow-datasets/tensorflow_datasets/testing/test_data/fake_examples`. \r\n\r\nMy hypothesis was: \r\nBecasue, we don't want to work with PRs, no `dataset_script` is going to be in the official repo, so no `dataset_script_test` can be in the repo either. Therefore we can also not have any \"fake\" test folder structure in the repo. \r\n\r\n**BUT:** As you mentioned @thom, we could have a fake data structure on AWS. To add a test the user has to upload multiple small test files when uploading his data set script. \r\n\r\nSo for a cli this could look like:\r\n`python nlp-cli upload <data_set_script> --testfiles <relative path to test file 1> <relative path to test file 2> ...` \r\n\r\nor even easier if the user just creates the dataset folder with the script inside and the testing folder structure, then the API could look like:\r\n\r\n`python nlp-cli upload <path/to/dataset/folder>`\r\n\r\nand the dataset folder would look like\r\n```\r\nsquad\r\n- squad.py\r\n- fake_data # this dir would have to have the exact same structure we get when downloading from the official squad data url\r\n```\r\n\r\nThis way I think we wouldn't even need any test files at all for each dataset script. For special datasets like `c4` or `wikipedia` we could then allow to optionally upload another test script. \r\nWe just assume that this is our downloaded `url` and check all functionality from there. \r\n\r\nThinking a bit more about this solution sounds a) much less work and b) even easier for the user.\r\n\r\nA small problem I see here though:\r\n1) What do we do when the depending on the config name the downloaded folder structure is very different? I think for each dataset config name we should have one test, which could correspond to one \"fake\" folder structure on AWS\r\n\r\n@thomwolf What do you think? I would actually go for this solution instead now.\r\n@mariamabarham You have written many more tfds dataset scripts and tests than I have - what do you think? \r\n\r\n", "Regarding the tfds tests, I don't really see a point in keeping them because:\r\n\r\n1) If you provide a fake data structure, IMO there is no need for each dataset to have an individual test file because (I think) most datasets have the same functions `_split_generators` and `_generate_examples` for which you can just test the functionality in a common test file. For special functions like these beam / pipeline functionality you probably need an extra test file. But @mariamabarham I think you have seen more than I have here as well \r\n\r\n2) The dataset test design is very much intertwined with the download manager design and contains a lot of code. I would like to seperate the tests into a) tests for downloading in general b) tests for post download data set pre-processing. Since we are going to change the download code anyways quite a lot, my plan was to focus on b) first. ", "I like the idea of having a fake data folder on S3. I have seen datasets with nested compressed files structures that would be tedious to generate with code. And for users it is probably easier to create a fake data folder by taking a subset of the actual data, and then upload it as you said.", "> > I think I'm fine with this.\r\n> > The alternative would be to host a small subset of the dataset on the S3 together with the testing script. But I think having all (test file creation + actual tests) in one file is actually quite convenient.\r\n> > Good for me!\r\n> > One question though, will we have to create one test file for each of the 100+ datasets or could we make some automatic conversion from tfds dataset test files?\r\n> \r\n> I think if we go the way shown in the PR we would have to create one test file for each of the 100+ datasets.\r\n> \r\n> As far as I know the tfds test files all rely on the user having created a special download folder structure in `tensorflow-datasets/tensorflow_datasets/testing/test_data/fake_examples`.\r\n> \r\n> My hypothesis was:\r\n> Becasue, we don't want to work with PRs, no `dataset_script` is going to be in the official repo, so no `dataset_script_test` can be in the repo either. Therefore we can also not have any \"fake\" test folder structure in the repo.\r\n> \r\n> **BUT:** As you mentioned @thom, we could have a fake data structure on AWS. To add a test the user has to upload multiple small test files when uploading his data set script.\r\n> \r\n> So for a cli this could look like:\r\n> `python nlp-cli upload <data_set_script> --testfiles <relative path to test file 1> <relative path to test file 2> ...`\r\n> \r\n> or even easier if the user just creates the dataset folder with the script inside and the testing folder structure, then the API could look like:\r\n> \r\n> `python nlp-cli upload <path/to/dataset/folder>`\r\n> \r\n> and the dataset folder would look like\r\n> \r\n> ```\r\n> squad\r\n> - squad.py\r\n> - fake_data # this dir would have to have the exact same structure we get when downloading from the official squad data url\r\n> ```\r\n> \r\n> This way I think we wouldn't even need any test files at all for each dataset script. For special datasets like `c4` or `wikipedia` we could then allow to optionally upload another test script.\r\n> We just assume that this is our downloaded `url` and check all functionality from there.\r\n> \r\n> Thinking a bit more about this solution sounds a) much less work and b) even easier for the user.\r\n> \r\n> A small problem I see here though:\r\n> \r\n> 1. What do we do when the depending on the config name the downloaded folder structure is very different? I think for each dataset config name we should have one test, which could correspond to one \"fake\" folder structure on AWS\r\n> \r\n> @thomwolf What do you think? I would actually go for this solution instead now.\r\n> @mariamabarham You have written many more tfds dataset scripts and tests than I have - what do you think?\r\n\r\nI'm agreed with you just one thing, for some dataset like glue or xtreme you may have multiple datasets in it. so I think a good way is to have one main fake folder and a subdirectory for each dataset inside", "> Regarding the tfds tests, I don't really see a point in keeping them because:\r\n> \r\n> 1. If you provide a fake data structure, IMO there is no need for each dataset to have an individual test file because (I think) most datasets have the same functions `_split_generators` and `_generate_examples` for which you can just test the functionality in a common test file. For special functions like these beam / pipeline functionality you probably need an extra test file. But @mariamabarham I think you have seen more than I have here as well\r\n> 2. The dataset test design is very much intertwined with the download manager design and contains a lot of code. I would like to seperate the tests into a) tests for downloading in general b) tests for post download data set pre-processing. Since we are going to change the download code anyways quite a lot, my plan was to focus on b) first.\r\n\r\nFor _split_generator, yes. But I'm not sure for _generate_examples because there is lots of things that should be taken into account such as feature names and types, data format (json, jsonl, csv, tsv,..)", "Sounds good to me!\r\n\r\nWhen testing, we could thus just override the prefix in the URL inside the download manager to have them point to the test directory on our S3.\r\n\r\nCc @lhoestq ", "Ok, here is a second draft for the testing structure. \r\n\r\nI think the big difficulty here is \"How can you generate tests on the fly from a given dataset name, *e.g.* `squad`\"?\r\n\r\nSo, this morning I did some research on \"parameterized testing\" and pure `unittest` or `pytest` didn't work very well. \r\nI found the lib https://github.com/wolever/parameterized, which works very nicely for our use case I think. \r\n@thomwolf - would it be ok to have a dependence on this lib for `nlp`? It seems like a light-weight lib to me. \r\n\r\nThis lib allows to add a `parameterization` decorator to a `unittest.TestCase` class so that the class can be instantiated for multiple different arguments (which are the dataset names `squad` etc. in our case).\r\n\r\nWhat I like about this lib is that one only has to add the decorator and the each of the parameterized tests are shown, like this: \r\n\r\n\r\n\r\nWith this structure we would only have to upload the dummy data for each dataset and would not require a specific testing file. \r\n\r\nWhat do you think @thomwolf @mariamabarham @lhoestq ?", "I think this is a nice solution.\r\n\r\nDo you think we could have the `parametrized` dependency in a `[test]` optional installation of `setup.py`? I would really like to keep the dependencies of the standard installation as small as possible. ", "> I think this is a nice solution.\r\n> \r\n> Do you think we could have the `parametrized` dependency in a `[test]` optional installation of `setup.py`? I would really like to keep the dependencies of the standard installation as small as possible.\r\n\r\nYes definitely!", "UPDATE: \r\n\r\nThis test design is ready now. I added dummy data to S3 for the dataests: `squad, crime_and_punish, sentiment140` . The structure can be seen on `https://s3.console.aws.amazon.com/s3/buckets/datasets.huggingface.co/nlp/squad/dummy/?region=us-east-1&tab=overview` for `squad`. \r\n\r\nAll dummy data files have to be in .zip format and called `dummy_data.zip`. The zip file should thereby have the exact same folder structure one gets from downloading the \"real\" data url(s). \r\n\r\nTo show how the .zip file looks like for the added datasets, I added the folder `nlp/datasets/dummy_data` in this PR. I think we can leave for the moment so that people can see better how to add dummy data tests and later delete it like `nlp/datasets/nlp`." ]

[]

[]

[]

[]

[]

[ "I think this can be quickly reproduced. \r\nI use `black, version 19.10b0`. \r\n\r\nWhen running: \r\n`black nlp/src/arrow_reader.py` \r\nit gives me: \r\n\r\n```\r\nerror: cannot format /home/patrick/hugging_face/nlp/src/nlp/arrow_reader.py: cannot use --safe with this file; failed to parse source file. AST error message: invalid syntax (<unknown>, line 78)\r\nOh no! 💥 💔 💥\r\n1 file failed to reformat.\r\n```\r\n\r\nThe line in question is: \r\nhttps://github.com/huggingface/nlp/blob/6922a16705e61f9e31a365f2606090b84d49241f/src/nlp/arrow_reader.py#L78\r\n\r\nWhat is weird is that the trainer file in `transformers` has more or less the same syntax and black does not fail there: \r\nhttps://github.com/huggingface/transformers/blob/cb3c2212c79d7ff0a4a4e84c3db48371ecc1c15d/src/transformers/trainer.py#L95\r\n\r\nI googled quite a bit about black & typing hints yesterday and didn't find anything useful. \r\nAny ideas @thomwolf @julien-c @LysandreJik ?", "> I think this can be quickly reproduced.\r\n> I use `black, version 19.10b0`.\r\n> \r\n> When running:\r\n> `black nlp/src/arrow_reader.py`\r\n> it gives me:\r\n> \r\n> ```\r\n> error: cannot format /home/patrick/hugging_face/nlp/src/nlp/arrow_reader.py: cannot use --safe with this file; failed to parse source file. AST error message: invalid syntax (<unknown>, line 78)\r\n> Oh no! 💥 💔 💥\r\n> 1 file failed to reformat.\r\n> ```\r\n> \r\n> The line in question is:\r\n> https://github.com/huggingface/nlp/blob/6922a16705e61f9e31a365f2606090b84d49241f/src/nlp/arrow_reader.py#L78\r\n> \r\n> What is weird is that the trainer file in `transformers` has more or less the same syntax and black does not fail there:\r\n> https://github.com/huggingface/transformers/blob/cb3c2212c79d7ff0a4a4e84c3db48371ecc1c15d/src/transformers/trainer.py#L95\r\n> \r\n> I googled quite a bit about black & typing hints yesterday and didn't find anything useful.\r\n> Any ideas @thomwolf @julien-c @LysandreJik ?\r\n\r\nOk I found the problem. It was the one Julien mentioned and has nothing to do with this line. Black's error message is a bit misleading here, I guess", "Ok, just had to remove the python 2 syntax comments `# type`. \r\n\r\nGood to merge for me now @thomwolf " ]

[]

[]

[ "Also added to an assert statement that if a dict is returned by function, all values of `dicts` are `lists`", "Wait to merge until `make style` is set in place.", "Updated the assert statements. Played around with multiple cases and it should be good now IMO. " ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Yes!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Yes looks good to me.\r\nNote that we may refactor quite strongly the `info.py` to make it a lot simpler (it's very complicated for basically a dictionary of info I think)", "No, problem ^^ It might just be a temporary fix :)", "Fixed." ]

[]

[]

[ "To fix this I propose to modify only the file `arrow_reader.py` with few updates. First update, the following method:\r\n```python\r\ndef _make_file_instructions_from_absolutes(\r\n name,\r\n name2len,\r\n absolute_instructions,\r\n):\r\n \"\"\"Returns the files instructions from the absolute instructions list.\"\"\"\r\n # For each split, return the files instruction (skip/take)\r\n file_instructions = []\r\n num_examples = 0\r\n for abs_instr in absolute_instructions:\r\n length = name2len[abs_instr.splitname]\r\n if not length:\r\n raise ValueError(\r\n 'Split empty. This might means that dataset hasn\\'t been generated '\r\n 'yet and info not restored from GCS, or that legacy dataset is used.')\r\n filename = filename_for_dataset_split(\r\n dataset_name=name,\r\n split=abs_instr.splitname,\r\n filetype_suffix='arrow')\r\n from_ = 0 if abs_instr.from_ is None else abs_instr.from_\r\n to = length if abs_instr.to is None else abs_instr.to\r\n num_examples += to - from_\r\n single_file_instructions = [{\"filename\": filename, \"skip\": from_, \"take\": to - from_}]\r\n file_instructions.extend(single_file_instructions)\r\n return FileInstructions(\r\n num_examples=num_examples,\r\n file_instructions=file_instructions,\r\n )\r\n```\r\nBecomes:\r\n```python\r\ndef _make_file_instructions_from_absolutes(\r\n name,\r\n name2len,\r\n absolute_instructions,\r\n):\r\n \"\"\"Returns the files instructions from the absolute instructions list.\"\"\"\r\n # For each split, return the files instruction (skip/take)\r\n file_instructions = []\r\n num_examples = 0\r\n for abs_instr in absolute_instructions:\r\n length = name2len[abs_instr.splitname]\r\n ## Delete the if not length and the raise\r\n filename = filename_for_dataset_split(\r\n dataset_name=name,\r\n split=abs_instr.splitname,\r\n filetype_suffix='arrow')\r\n from_ = 0 if abs_instr.from_ is None else abs_instr.from_\r\n to = length if abs_instr.to is None else abs_instr.to\r\n num_examples += to - from_\r\n single_file_instructions = [{\"filename\": filename, \"skip\": from_, \"take\": to - from_}]\r\n file_instructions.extend(single_file_instructions)\r\n return FileInstructions(\r\n num_examples=num_examples,\r\n file_instructions=file_instructions,\r\n )\r\n```\r\n\r\nSecond update the following method:\r\n```python\r\ndef _read_files(files, info):\r\n \"\"\"Returns Dataset for given file instructions.\r\n\r\n Args:\r\n files: List[dict(filename, skip, take)], the files information.\r\n The filenames contain the absolute path, not relative.\r\n skip/take indicates which example read in the file: `ds.slice(skip, take)`\r\n \"\"\"\r\n pa_batches = []\r\n for f_dict in files:\r\n pa_table: pa.Table = _get_dataset_from_filename(f_dict)\r\n pa_batches.extend(pa_table.to_batches())\r\n pa_table = pa.Table.from_batches(pa_batches)\r\n ds = Dataset(arrow_table=pa_table, data_files=files, info=info)\r\n return ds\r\n```\r\nBecomes:\r\n```python\r\ndef _read_files(files, info):\r\n \"\"\"Returns Dataset for given file instructions.\r\n\r\n Args:\r\n files: List[dict(filename, skip, take)], the files information.\r\n The filenames contain the absolute path, not relative.\r\n skip/take indicates which example read in the file: `ds.slice(skip, take)`\r\n \"\"\"\r\n pa_batches = []\r\n for f_dict in files:\r\n pa_table: pa.Table = _get_dataset_from_filename(f_dict)\r\n pa_batches.extend(pa_table.to_batches())\r\n ## we modify the table only if there are some batches\r\n if pa_batches:\r\n pa_table = pa.Table.from_batches(pa_batches)\r\n ds = Dataset(arrow_table=pa_table, data_files=files, info=info)\r\n return ds\r\n```\r\n\r\nWith these two updates it works now. @thomwolf are you ok with this changes?", "Yes sounds good to me!\r\nDo you want to make a PR? or I can do it as well", "Fixed." ]

[]

[]

[ "Yes! I see mostly two options for this:\r\n- a `Feature` approach like currently (but we might deprecate features)\r\n- wrapping in a smart way the Dictionary arrays of Arrow: https://arrow.apache.org/docs/python/data.html?highlight=dictionary%20encode#dictionary-arrays", "I would have a preference for the second bullet point.", "This should be accessible now as a feature in dataset.info.features (and even have the mapping methods).", "Perfect! Well done!!", "Hi,\r\nI hope we could get a better documentation.\r\nIt took me more than 1 hour to found this way to get the label information.", "Yes we are working on the doc right now, should be in the next release quite soon." ]

[]

[]

[ "Yes!\r\n- pandas will be a one-liner in `arrow_dataset`: https://arrow.apache.org/docs/python/generated/pyarrow.Table.html#pyarrow.Table.to_pandas\r\n- for Spark I have no idea. let's investigate that at some point", "For Spark it looks to be pretty straightforward as well https://spark.apache.org/docs/latest/sql-pyspark-pandas-with-arrow.html but looks to be having a dependency to Spark is necessary, then nevermind we can skip it", "Now Pandas is available." ]

[]

[]

[ "My first bug report ❤️\r\nLooking into this right now!", "Ok, there are some news, most good than bad :laughing: \r\n\r\nThe dataset script now became:\r\n```python\r\nimport csv\r\n\r\nimport nlp\r\n\r\n\r\nclass Bbc(nlp.GeneratorBasedBuilder):\r\n VERSION = nlp.Version(\"1.0.0\")\r\n\r\n def __init__(self, **config):\r\n self.train = config.pop(\"train\", None)\r\n self.validation = config.pop(\"validation\", None)\r\n super(Bbc, self).__init__(**config)\r\n\r\n def _info(self):\r\n return nlp.DatasetInfo(builder=self, description=\"bla\", features=nlp.features.FeaturesDict({\"id\": nlp.int32, \"text\": nlp.string, \"label\": nlp.string}))\r\n\r\n def _split_generators(self, dl_manager):\r\n return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={\"filepath\": self.train}),\r\n nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={\"filepath\": self.validation})]\r\n\r\n def _generate_examples(self, filepath):\r\n with open(filepath) as f:\r\n reader = csv.reader(f, delimiter=',', quotechar=\"\\\"\")\r\n lines = list(reader)[1:]\r\n\r\n for idx, line in enumerate(lines):\r\n yield idx, {\"id\": idx, \"text\": line[1], \"label\": line[0]}\r\n\r\n```\r\n\r\nAnd the dataset folder becomes:\r\n```\r\n.\r\n├── bbc\r\n│ ├── bbc.py\r\n│ └── data\r\n│ ├── test.csv\r\n│ └── train.csv\r\n```\r\nI can load the dataset by using the keywords arguments like this:\r\n```python\r\nimport nlp\r\ndataset = nlp.load(\"bbc\", builder_kwargs={\"train\": \"bbc/data/train.csv\", \"validation\": \"bbc/data/test.csv\"})\r\n```\r\n\r\nThat was the good part ^^ Because it took me some time to understand that the script itself is put in cache in `datasets/src/nlp/datasets/some-hash/bbc.py` which is very difficult to discover without checking the source code. It means that doesn't matter the changes you do to your original script it is taken into account. I think instead of doing a hash on the name (I suppose it is the name), a hash on the content of the script itself should be a better solution.\r\n\r\nThen by diving a bit in the code I found the `force_reload` parameter [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L50) but the call of this `load_dataset` method is done with the `builder_kwargs` as seen [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L166) which is ok until the call to the builder is done as the builder do not have this `force_reload` parameter. To show as example, the previous load becomes:\r\n```python\r\nimport nlp\r\ndataset = nlp.load(\"bbc\", builder_kwargs={\"train\": \"bbc/data/train.csv\", \"validation\": \"bbc/data/test.csv\", \"force_reload\": True})\r\n```\r\nRaises\r\n```\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/jplu/dev/jplu/datasets/src/nlp/load.py\", line 283, in load\r\n dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)\r\n File \"/home/jplu/dev/jplu/datasets/src/nlp/load.py\", line 170, in builder\r\n builder_instance = builder_cls(**builder_kwargs)\r\n File \"/home/jplu/dev/jplu/datasets/src/nlp/datasets/84d638d2a8ca919d1021a554e741766f50679dc6553d5a0612b6094311babd39/bbc.py\", line 12, in __init__\r\n super(Bbc, self).__init__(**config)\r\nTypeError: __init__() got an unexpected keyword argument 'force_reload'\r\n```\r\nSo yes the cache is refreshed with the new script but then raises this error.", "Ok great, so as discussed today, let's:\r\n- have a main dataset directory inside the lib with sub-directories hashed by the content of the file\r\n- keep a cache for downloading the scripts from S3 for now\r\n- later: add methods to list and clean the local versions of the datasets (and the distant versions on S3 as well)\r\n\r\nSide question: do you often use `builder_kwargs` for other things than supplying file paths? I was thinking about having a more easy to read and remember `data_files` argument maybe.", "Good plan!\r\n\r\nYes I do use `builder_kwargs` for other things such as:\r\n- dataset name\r\n- properties to know how to properly read a CSV file: do I have to skip the first line in a CSV, which delimiter is used, and the columns ids to use.\r\n- properties to know how to properly read a JSON file: which properties in a JSON object to read", "Done!" ]

[]

[]

[]

[]

[]

[]

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

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Seems like a regression from https://github.com/huggingface/datasets/pull/2843\r\n\r\nOr maybe it's an issue with the hosting. I don't think so, though, because https://www.dropbox.com/s/aox84z90nyyuikz/discovery.zip seems to work as expected\r\n\r\n", "Hi @severo, thanks for reporting.\r\n\r\nSome servers do not support HTTP range requests, and those are required to stream some file formats (like ZIP in this case).\r\n\r\nLet me try to propose a workaround. " ]

[]

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

[ "It looks like this is because of this behavior in pyarrow:\r\n```python\r\nimport pyarrow as pa\r\n\r\narr = pa.array([None, [0]])\r\nreconstructed_arr = pa.ListArray.from_arrays(arr.offsets, arr.values)\r\nprint(reconstructed_arr.to_pylist())\r\n# [[], [0]]\r\n```\r\n\r\nIt seems that `arr.offsets` can reconstruct the array properly, but an offsets array with null values can:\r\n```python\r\nfixed_offsets = pa.array([None, 0, 1])\r\nfixed_arr = pa.ListArray.from_arrays(fixed_offsets, arr.values)\r\nprint(fixed_arr.to_pylist())\r\n# [None, [0]]\r\n\r\nprint(arr.offsets.to_pylist())\r\n# [0, 0, 1]\r\nprint(fixed_offsets.to_pylist())\r\n# [None, 0, 1]\r\n```\r\nEDIT: this is because `arr.offsets` is not enough to reconstruct the array, we also need the validity bitmap", "The offsets don't have nulls because they don't include the validity bitmap from `arr.buffers()[0]`, which is used to say which values are null and which values are non-null.\r\n\r\nThough the validity bitmap also seems to be wrong:\r\n```python\r\nbin(int(arr.buffers()[0].hex(), 16))\r\n# '0b10'\r\n# it should be 0b110 - 1 corresponds to non-null and 0 corresponds to null, if you take the bits in reverse order\r\n```\r\n\r\nSo apparently I can't even create the fixed offsets array using this.\r\n\r\nIf I understand correctly it's always missing the 1 on the left, so I can add it manually as a hack to fix the issue until this is fixed in pyarrow EDIT: actually it may be more complicated than that\r\n\r\nEDIT2: actuall it's right, it corresponds to the validity bitmap of the array of logical length 2. So if we use the offsets array, the values array, and this validity bitmap it should be possible to reconstruct the array properly", "I created an issue on Apache Arrow's JIRA: https://issues.apache.org/jira/browse/ARROW-15837", "And another one: https://issues.apache.org/jira/browse/ARROW-15839", "FYI the behavior is the same with:\r\n- `datasets` version: 1.18.3\r\n- Platform: Linux-5.8.0-50-generic-x86_64-with-debian-bullseye-sid\r\n- Python version: 3.7.11\r\n- PyArrow version: 6.0.1\r\n\r\n\r\nbut not with:\r\n- `datasets` version: 1.8.0\r\n- Platform: Linux-4.18.0-305.40.2.el8_4.x86_64-x86_64-with-redhat-8.4-Ootpa\r\n- Python version: 3.7.11\r\n- PyArrow version: 3.0.0\r\n\r\ni.e. it outputs:\r\n```py\r\n0 [None, [0]]\r\n1 [None, [0]]\r\n2 [None, [0]]\r\n3 [None, [0]]\r\n```\r\n", "Thanks for the insights @PaulLerner !\r\n\r\nI found a way to workaround this issue for the code example presented in this issue.\r\n\r\nNote that empty lists will still appear when you explicitly `cast` a list of lists that contain None values like [None, [0]] to a new feature type (e.g. to change the integer precision). In this case it will show a warning that it happened. If you don't cast anything, then the None values will be kept as expected.\r\n\r\nLet me know what you think !", "Hi! I feel like I’m missing something in your answer, *what* is the workaround? Is it fixed in some `datasets` version?", "`pa.ListArray.from_arrays` returns empty lists instead of None values. The workaround I added inside `datasets` simply consists in not using `pa.ListArray.from_arrays` :)\r\n\r\nOnce this PR [here ](https://github.com/huggingface/datasets/pull/4282)is merged, we'll release a new version of `datasets` that currectly returns the None values in the case described in this issue\r\n\r\nEDIT: released :) but let's keep this issue open because it might happen again if users change the integer precision for example" ]

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

[]

[ "@mariosasko Can I take this up?", "This dataset is now available here: https://huggingface.co/datasets/deepmind/code_contests." ]

[]

[]

[ "@lhoestq \r\n\r\nThe model used by default (`moussaKam/frugalscore_tiny_bert-base_bert-score`) is a tiny model.\r\n\r\nI still want to make one modification before merging.\r\nI would like to load the model checkpoint once. Do you think it's a good idea if I load it in `_download_and_prepare`? In this case should the model name be the `self.config_name` or another variable say `self.model_name` ? ", "OK, I added a commit that loads the checkpoint in `_download_and_prepare`. Please let me know if it looks good. ", "@lhoestq is everything OK to merge? ", "I triggered the CI and it's failing, can you merge the `master` branch into yours ? It should fix the issues.\r\n\r\nAlso the doctest apparently raises an error because it outputs `{'scores': [0.6307542, 0.6449357]}` instead of `{'scores': [0.631, 0.645]}` - feel free to edit the code example in the docstring to round the scores, that should fix it", "@lhoestq hope it's OK now" ]

[ { "id": 1935892857, "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 3470211881, "node_id": "LA_kwDODunzps7O1zsp", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset-viewer", "name": "dataset-viewer", "color": "E5583E", "default": false, "description": "Related to the dataset viewer on huggingface.co" } ]

[ { "login": "severo", "id": 1676121, "node_id": "MDQ6VXNlcjE2NzYxMjE=", "avatar_url": "https://avatars.githubusercontent.com/u/1676121?v=4", "gravatar_id": "", "url": "https://api.github.com/users/severo", "html_url": "https://github.com/severo", "followers_url": "https://api.github.com/users/severo/followers", "following_url": "https://api.github.com/users/severo/following{/other_user}", "gists_url": "https://api.github.com/users/severo/gists{/gist_id}", "starred_url": "https://api.github.com/users/severo/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/severo/subscriptions", "organizations_url": "https://api.github.com/users/severo/orgs", "repos_url": "https://api.github.com/users/severo/repos", "events_url": "https://api.github.com/users/severo/events{/privacy}", "received_events_url": "https://api.github.com/users/severo/received_events", "type": "User", "site_admin": false } ]

[ "Yes, we decided to replace the encoded label with the corresponding label when possible in the dataset viewer. But\r\n1. maybe it's the wrong default\r\n2. we could find a way to show both (with a switch, or showing both ie. `0 (neutral)`).\r\n", "Hi @severo,\r\n\r\nThanks for clarifying. \r\n\r\nI think this default is a bit counterintuitive for the user. However, this is a personal opinion that might not be general. I think it is nice to have the actual (non-encoded) labels in the viewer. On the other hand, it would be nice to match what the user sees with what they get when they download a dataset. I don't know - I can see the difficulty of choosing a default :)\r\nMaybe having non-encoded labels as a default can be useful?\r\n\r\nAnyway, I think the issue has been addressed. Thanks a lot for your super-quick answer!\r\n\r\n ", "Thanks for the 👍 in https://github.com/huggingface/datasets/issues/3673#issuecomment-1029008349 @mariosasko @gary149 @pietrolesci, but as I proposed various solutions, it's not clear to me which you prefer. Could you write your preferences as a comment?\r\n\r\n_(note for myself: one idea per comment in the future)_", "As I am working with seq2seq, I prefer having the label in string form rather than numeric. So the viewer is fine and the underlying dataset should be \"decoded\" (from int to str). In this way, the user does not have to search for a mapping `int -> original name` (even though is trivial to find, I reckon). Also, encoding labels is rather easy.\r\n\r\nI hope this is useful", "I like the idea of \"0 (neutral)\". The label name can even be greyed to make it clear that it's not part of the actual item in the dataset, it's just the meaning.", "I like @lhoestq's idea of having grayed-out labels.", "Proposals by @gary149. Which one do you prefer? Please vote with the thumbs\r\n\r\n- 👍 \r\n\r\n \r\n\r\n- 👎 \r\n\r\n \r\n\r\n", "I like Option 1 better as it shows clearly what the user is downloading", "Thanks! ", "It's [live](https://huggingface.co/datasets/glue/viewer/cola/train):\r\n\r\n<img width=\"1126\" alt=\"Capture d’écran 2022-02-14 à 10 26 03\" src=\"https://user-images.githubusercontent.com/1676121/153836716-25f6205b-96af-42d8-880a-7c09cb24c420.png\">\r\n\r\nThanks all for the help to improve the UI!", "Love it ! thanks :)" ]

[]

[]

[]

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

[]

[]

[]

[]

[ "It's a first attempt at solving https://github.com/huggingface/datasets/issues/3013.", "I only kept these ones:\r\n```\r\n path: str,\r\n data_files: Optional[Union[Dict, List, str]] = None,\r\n download_config: Optional[DownloadConfig] = None,\r\n download_mode: Optional[GenerateMode] = None,\r\n revision: Optional[Union[str, Version]] = None,\r\n use_auth_token: Optional[Union[bool, str]] = None,\r\n **config_kwargs,\r\n```\r\n\r\nLet me know if it's better for you now !\r\n\r\n(note that there's no breaking change since the ones that are removed can be passed as config_kwargs if you really want)", "(https://github.com/huggingface/datasets/pull/3670/commits/5636911880ea4306c27c7f5825fa3f9427ccc2b6 and https://github.com/huggingface/datasets/pull/3670/commits/07c3f0800dd34dfebb9674ad46c67a907b08ded8 -> I has forgotten to update black in my venv)" ]

[]

[]

[ "Hi @patrickvonplaten - could you please advise whether this would be a welcomed change, and if so, who I consult regarding the unit-tests?", "I'd be happy with adding this change. @anton-l @lhoestq - what do you think?", "Cool ! I just fixed the tests by adding a dummy `validated.tsv` file in the dummy data archive of common_voice\r\n\r\nI wonder if you should separate the train/valid/test configuration from the validated/invalidated configuration of the splits ? \r\nIn particular having `validated` along with the train/valid/test splits could be a bit weird since it comprises them. We can do that if you think it makes more sense. Otherwise it's also good as it is right now :)\r\n", "Thanks! I think that there are 2 cases for using the validated partition: 1) trainset = {validated - dev - test}, dev and test as they come; 2) train, dev, and test sampled from validated manually with the desired ratios.\r\nIn either case, I think that it's quite a big change on the HF interface part, so could as well be taken care of in the client code. Or is it not? (In which case, what's the most compact way to implement this?)", "What's important IMO is to let the users as much flexibility as they need - so we try to not do too much regarding splits to not constrain users. So I guess the way it is right now is ok. Can you confirm that it's ok @patrickvonplaten and that it won't break some speech training script out there ?", "@lhoestq all split names are explicit in our example scripts, so this shouldn't break anything, feel free to merge :)\r\nI'll go ahead and add this to the official `mozilla-foundation` datasets as well ", "Good for me! This has no real down-sides IMO and surely won't break any training scripts." ]

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

[]

[ "Hi ! I wasn't able to reproduce the error, are you still experiencing this ? I tried calling `cast_column` on a string column containing paths.\r\n\r\nIf you manage to share a reproducible code example that would be perfect", "Hi,\r\n\r\nI think my team mate got this solved. Clolsing it for now and will reopen if I experience this again.\r\nThanks :) ", "Hi @R4ZZ3,\r\n\r\nIf it is not too much of a bother, can you please help me how to resolve this error? I am exactly getting the same error where I am going as per the documentation guideline:\r\n\r\n`my_audio_dataset = my_audio_dataset.cast_column(\"audio_paths\", Audio())`\r\n\r\nwhere `\"audio_paths\"` is a dataset column (feature) having strings of absolute paths to mp3 files of the dataset.\r\n\r\n", "I was having the same issue with this code:\r\n\r\n```\r\ndataset = dataset.map(\r\n lambda batch: {\"full_path\" : os.path.join(self.data_path, batch[\"path\"])},\r\n num_procs = 4\r\n)\r\nmy_audio_dataset = dataset.cast_column(\"full_path\", Audio(sampling_rate=16_000))\r\n```\r\n\r\nRemoving the \"num_procs\" argument fixed it somehow.\r\nUsing a mac with m1 chip", "Hi @Hubert-Bonisseur, I think this will be fixed by https://github.com/huggingface/datasets/pull/4614" ]

[]

[ { "login": "polinaeterna", "id": 16348744, "node_id": "MDQ6VXNlcjE2MzQ4NzQ0", "avatar_url": "https://avatars.githubusercontent.com/u/16348744?v=4", "gravatar_id": "", "url": "https://api.github.com/users/polinaeterna", "html_url": "https://github.com/polinaeterna", "followers_url": "https://api.github.com/users/polinaeterna/followers", "following_url": "https://api.github.com/users/polinaeterna/following{/other_user}", "gists_url": "https://api.github.com/users/polinaeterna/gists{/gist_id}", "starred_url": "https://api.github.com/users/polinaeterna/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/polinaeterna/subscriptions", "organizations_url": "https://api.github.com/users/polinaeterna/orgs", "repos_url": "https://api.github.com/users/polinaeterna/repos", "events_url": "https://api.github.com/users/polinaeterna/events{/privacy}", "received_events_url": "https://api.github.com/users/polinaeterna/received_events", "type": "User", "site_admin": false } ]

[ "Note that torchaudio is maybe less practical to use for TF or JAX users.\r\nThis is not in the scope of this PR, but in the future if we manage to find a way to let the user control the decoding it would be nice", "> Note that torchaudio is maybe less practical to use for TF or JAX users. This is not in the scope of this PR, but in the future if we manage to find a way to let the user control the decoding it would be nice\r\n\r\n@lhoestq so maybe don't do this PR? :) if it doesn't work anyway with an opened file, only with path", "Yes as discussed offline there seems to be issues with torchaudio on opened files. Feel free to close this PR if it's better to stick with soundfile because of that", "We should be able to remove torchaudio, which has torch as a hard dependency, soon and use only soundfile for decoding: https://github.com/bastibe/python-soundfile/issues/252#issuecomment-1000246773 (opus + mp3 support is on the way)." ]

[]

[]

[ "@lhoestq I still have problems with processing `.opus` files with `soundfile` so I actually cannot fully check that it works but it should... Maybe this should be investigated in case of someone else would also have problems with that.\r\n\r\nAlso, as the data is in a private repo on the hub (before we come to a decision about audio data privacy), the needed checks cannot be done right now.", "@lhoestq I check the data redownloading for configs sharing the same languages, you were right: the data is downloaded once for each language. But samples are generated from scratch each time. Is it a supposed behavior? ", "> But samples are generated from scratch each time. Is it a supposed behavior?\r\n\r\nYea that's the way it works right now, because we generate one arrow file per configuration. Since changing the languages creates a new configuration, then it generates a new arrow file." ]

[]

[]

[]

[]

[]

[ "Cool thanks for giving it a try @anton-l ! \r\n\r\nWould be very much in favor of having \"real\" paths to the audio files again for non-streaming use cases. At the same time it would be nice to make the audio data loading script as understandable as possible so that the community can easily add audio datasets in the future by looking at this one as an example. Think if it's clear for a contributor how to add an audio datasets script that works for the standard non-streaming case while it is easy to extend it afterwards to a streaming dataset script, then this would be perfect", "@anton-l @patrickvonplaten @lhoestq Is it possible somehow to provide this logic inside the library instead of a loading script so that we don't need to completely rewrite all the scripts for audio datasets and users don't have to care about two different loading approaches in the same script? 🤔 ", "> @anton-l @patrickvonplaten @lhoestq Is it possible somehow to provide this logic inside the library instead of a loading script so that we don't need to completely rewrite all the scripts for audio datasets and users don't have to care about two different loading approaches in the same script? thinking\r\n\r\nNot sure @lhoestq - what do you think? \r\n\r\nNow that we've corrected the previous resampling bug, think this one here is of high importance. @lhoestq - what do you think how we should proceed here? ", "> @anton-l @patrickvonplaten @lhoestq Is it possible somehow to provide this logic inside the library instead of a loading script so that we don't need to completely rewrite all the scripts for audio datasets and users don't have to care about two different loading approaches in the same script? 🤔\r\n\r\nYes let's do this :)\r\n\r\nMaybe we can change the behavior of `DownloadManager.iter_archive` back to extracting the TAR archive locally, and return an iterable of (local path, file obj). And the `StreamingDownloadManager.iter_archive` can return an iterable of (relative path inside the archive, file obj) ?\r\n\r\nIn this case, a dataset would need to have something like this:\r\n```python\r\nfor path, f in files:\r\n yield id_, {\"audio\": {\"path\": path, \"bytes\": f.read() if not is_local_file(path) else None}}\r\n```\r\n\r\nAlternatively, we can allow this if we consider that `Audio.encode_example` sets the \"bytes\" field to `None` automatically if `path` is a local path:\r\n```python\r\nfor path, f in files:\r\n yield id_, {\"audio\": {\"path\": path, \"bytes\": f.read()}}\r\n```\r\nNote that in this case the file is read for nothing though (maybe it's not a big deal ?)\r\n\r\nLet me know if it sounds good to you and what you'd prefer !", "@lhoestq I'm very much in favor of your first aproach! With the full paths returned I think we won't even need to mess with `os.path.join` vs `\"/\".join()\"` and other local/streaming differences 👍 ", "@lhoestq I also like the idea and favor your first approach to avoid an unnecessary read and make yielding faster.", "Looks cool - thanks for working on this. I just feel strongly about `path` being an absolute `path` that exist and can be inspected in the non-streaming case :-) For streaming=True IMO it's absolutely fine if we only have access to the bytes", "Hi ! I started implementing this but I noticed that returning an absolute path is breaking for many datasets that do things like\r\n```python\r\nfor path, f in files:\r\n if path.startswith(data_dir):\r\n ...\r\n```\r\nso I think I will have to add a parameter to `iter_archive` like `extract_locally=True` to avoid the breaking change, does that sound good to you ?\r\n\r\nThis makes me also think that in streaming mode it could also return a local path too, if we think that writing and deleting temporary files on-the-fly while iterating over the streaming dataset is ok.", "@lhoestq I think it is a good idea to rollback to extracting the archives locally in non-streaming mode, as far as (as you mentioned) we do not store the bytes in the Arrow file for those cases to avoid \"doubling\" the disk space usage.\r\n\r\nOn the other hand, I don't like:\r\n- neither the possibility to avoid extracting locally in non-streaming: the behavior should be consistent; thus we always extract in non-streaming\r\n - which could be the criterium to decide whether an archive should or should not be extracted? Just because I want to make a condition on path.startswith?\r\n- nor the option to download/delete temporary files in streaming (see discussion here: https://github.com/huggingface/datasets/pull/3689#issuecomment-1032858345)\r\n\r\nUnfortunately, in order to fix the datasets that are breaking after the rollback, I would suggest fixing their scripts so that the paths are handled more robustly (considering that they can be absolute or relative).", "I agree with Albert, fixing all of the audio datasets isn't too big of a deal (yet). I can help with those if needed :)", "Ok cool ! I'm completely rolling it back then", "Alright I did the rollback and now you can get local paths :)\r\nFeel free to try it out and let me know if it's good for you", "I'll fix the CI tomorrow x)", "Ok according to the CI there around 60+ datasets to fix", "> fixing all of the audio datasets isn't too big of a deal (yet). I can help with those if needed :)\r\n\r\nI can help with them too :)\r\n", "Here is the full list to keep track of things:\r\n\r\n- [x] air_dialogue\r\n- [x] id_nergrit_corpus\r\n- [ ] id_newspapers_2018\r\n- [x] imdb\r\n- [ ] indic_glue\r\n- [ ] inquisitive_qg\r\n- [x] klue\r\n- [x] lama\r\n- [x] lex_glue\r\n- [ ] lm1b\r\n- [x] amazon_polarity\r\n- [ ] mac_morpho\r\n- [ ] math_dataset\r\n- [ ] md_gender_bias\r\n- [ ] mdd\r\n- [ ] assin\r\n- [ ] atomic\r\n- [ ] babi_qa\r\n- [ ] mlqa\r\n- [ ] mocha\r\n- [ ] blended_skill_talk\r\n- [ ] capes\r\n- [ ] cbt\r\n- [ ] newsgroup\r\n- [ ] cifar10\r\n- [ ] cifar100\r\n- [ ] norec\r\n- [ ] ohsumed\r\n- [ ] code_x_glue_cc_clone_detection_poj104\r\n- [x] openslr\r\n- [ ] orange_sum\r\n- [ ] paws\r\n- [ ] paws-x\r\n- [ ] cppe-5\r\n- [ ] polyglot_ner\r\n- [ ] dbrd\r\n- [ ] empathetic_dialogues\r\n- [ ] eraser_multi_rc\r\n- [ ] flores\r\n- [ ] flue\r\n- [ ] food101\r\n- [ ] py_ast\r\n- [ ] qasc\r\n- [ ] qasper\r\n- [ ] race\r\n- [ ] reuters21578\r\n- [ ] ropes\r\n- [ ] rotten_tomatoes\r\n- [x] vivos\r\n- [ ] wi_locness\r\n- [ ] wiki_movies\r\n- [ ] wikiann\r\n- [ ] wmt20_mlqe_task1\r\n- [ ] wmt20_mlqe_task2\r\n- [ ] wmt20_mlqe_task3\r\n- [ ] scicite\r\n- [ ] xsum\r\n- [ ] scielo\r\n- [ ] scifact\r\n- [ ] setimes\r\n- [ ] social_bias_frames\r\n- [ ] sogou_news\r\n- [x] speech_commands\r\n- [ ] ted_hrlr\r\n- [ ] ted_multi\r\n- [ ] tlc\r\n- [ ] turku_ner_corpus\r\n\r\n", "I'll do my best to fix as many as possible tomorrow :)", "the audio datasets are fixed if I didn't forget anything :)\r\n\r\nbtw what are we gonna do with the community ones that would be broken after the fix?", "Closing in favor of https://github.com/huggingface/datasets/pull/3736" ]

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

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false }, { "login": "polinaeterna", "id": 16348744, "node_id": "MDQ6VXNlcjE2MzQ4NzQ0", "avatar_url": "https://avatars.githubusercontent.com/u/16348744?v=4", "gravatar_id": "", "url": "https://api.github.com/users/polinaeterna", "html_url": "https://github.com/polinaeterna", "followers_url": "https://api.github.com/users/polinaeterna/followers", "following_url": "https://api.github.com/users/polinaeterna/following{/other_user}", "gists_url": "https://api.github.com/users/polinaeterna/gists{/gist_id}", "starred_url": "https://api.github.com/users/polinaeterna/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/polinaeterna/subscriptions", "organizations_url": "https://api.github.com/users/polinaeterna/orgs", "repos_url": "https://api.github.com/users/polinaeterna/repos", "events_url": "https://api.github.com/users/polinaeterna/events{/privacy}", "received_events_url": "https://api.github.com/users/polinaeterna/received_events", "type": "User", "site_admin": false }, { "login": "anton-l", "id": 26864830, "node_id": "MDQ6VXNlcjI2ODY0ODMw", "avatar_url": "https://avatars.githubusercontent.com/u/26864830?v=4", "gravatar_id": "", "url": "https://api.github.com/users/anton-l", "html_url": "https://github.com/anton-l", "followers_url": "https://api.github.com/users/anton-l/followers", "following_url": "https://api.github.com/users/anton-l/following{/other_user}", "gists_url": "https://api.github.com/users/anton-l/gists{/gist_id}", "starred_url": "https://api.github.com/users/anton-l/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/anton-l/subscriptions", "organizations_url": "https://api.github.com/users/anton-l/orgs", "repos_url": "https://api.github.com/users/anton-l/repos", "events_url": "https://api.github.com/users/anton-l/events{/privacy}", "received_events_url": "https://api.github.com/users/anton-l/received_events", "type": "User", "site_admin": false }, { "login": "mariosasko", "id": 47462742, "node_id": "MDQ6VXNlcjQ3NDYyNzQy", "avatar_url": "https://avatars.githubusercontent.com/u/47462742?v=4", "gravatar_id": "", "url": "https://api.github.com/users/mariosasko", "html_url": "https://github.com/mariosasko", "followers_url": "https://api.github.com/users/mariosasko/followers", "following_url": "https://api.github.com/users/mariosasko/following{/other_user}", "gists_url": "https://api.github.com/users/mariosasko/gists{/gist_id}", "starred_url": "https://api.github.com/users/mariosasko/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/mariosasko/subscriptions", "organizations_url": "https://api.github.com/users/mariosasko/orgs", "repos_url": "https://api.github.com/users/mariosasko/repos", "events_url": "https://api.github.com/users/mariosasko/events{/privacy}", "received_events_url": "https://api.github.com/users/mariosasko/received_events", "type": "User", "site_admin": false } ]

[ "Having talked to @lhoestq, I see that this feature is no longer supported. \r\n\r\nI really don't think this was a good idea. It is a major breaking change and one for which we don't even have a working solution at the moment, which is bad for PyTorch as we don't want to force people to have `datasets` decode audio files automatically, but **really** bad for Tensorflow and Flax where we **currently cannot** even use `datasets` to load `.mp3` files - e.g. `common_voice` doesn't work anymore in a TF training script. Note this worked perfectly fine before making the change (think it was done [here](https://github.com/huggingface/datasets/pull/3290) no?)\r\n\r\nIMO, it's really important to think about a solution here and I strongly favor to make a difference here between loading a dataset in streaming mode and in non-streaming mode, so that in non-streaming mode the actual downloaded file is displayed. It's really crucial for people to be able to analyse the original files IMO when the dataset is not downloaded in streaming mode. \r\n\r\nThere are the following reasons why it is paramount to have access to the **original** audio file in my opinion (in non-streaming mode):\r\n- There are a wide variety of different libraries to load audio data with varying support on different platforms. For me it was quite clear that there is simply to single good library to load audio files for all platforms - so we have to leave the option to the user to decide which loading to use.\r\n- We had support for audio datasets a long time before streaming audio was possible. There were quite some versions where we advertised **everywhere** to load the audio from the path name (and there are many places where we still do even though it's not possible anymore). To give some examples:\r\n - Official example of TF Wav2Vec2: https://github.com/huggingface/transformers/blob/f427e750490b486944cc9be3c99834ad5cf78b57/src/transformers/models/wav2vec2/modeling_tf_wav2vec2.py#L1423 Wav2Vec2 is as important for speech as BERT is for NLP - so it's **very** important. The official example currently doesn't work and we don't even have a workaround for it for MP3 files at the moment. Same goes for Flax.\r\n - The most downloaded non-nlp checkpoint: https://huggingface.co/facebook/wav2vec2-base-960h#usage has a usage example which doesn't work anymore with the current datasets implementation. I'll update this now, but we have >1000 wav2vec2 checkpoints on the Hub and we can't update all the model cards.\r\n => This is a big breaking change with no current solution. For `transformers` breaking changes are one of the biggest complaints.\r\n- Similar to this we also shouldn't assume that there is only one resampling method for Audio. I think it's good to have one offered automatically by `datasets`, but we have to leave the user the freedom to choose her/his own resampling as well. Resampling can take very different filtering windows and other parameters which are currently somewhat hardcoded in `datasets`, which users might very well want to change.\r\n\r\n\r\n=> IMO, it's a **very** big priority to again have the correct absolute path in non-streaming mode. The other solution of providing a path-like object derived from the bytes stocked in the `.array` file is not nearly as user-friendly, but better than nothing. ", "Agree that we need to have access to the original sound files. Few days ago I was looking for these original files because I suspected there is bug in the audio resampling (confirmed in https://github.com/huggingface/datasets/issues/3662) and I want to do my own resampling to workaround the bug, which is now not possible anymore due to the unavailability of the original files.", "@patrickvonplaten \r\n> The other solution of providing a path-like object derived from the bytes stocked in the .array file is not nearly as user-friendly, but better than nothing\r\n\r\nJust to clarify, here you describe the approach that uses the `Audio.decode` attribute to access the underlying bytes?\r\n\r\n> The official example currently doesn't work and we don't even have a workaround for it for MP3 files at the moment\r\n\r\nI'd assume this is because we use `sox_io` as a backend for decoding. However, soon we should be able to use `soundfile`, which supports path-like objects, for MP3 (https://github.com/huggingface/datasets/pull/3667#issuecomment-1030090627).\r\n\r\nYour concern is reasonable, but there are situations where we can only serve bytes (see https://github.com/huggingface/datasets/pull/3685 for instance). IMO it makes sense to fix the affected datasets for now, but I don't think we should care too much whether we rely on local paths or bytes after soundfile adds support for MP3 as long as our examples work (shouldn't be too hard to update the `map_to_array` functions) and we properly document how to access the underlying path/bytes for custom decoding (via `ds.cast_column(\"audio\", Audio(decode=False))`).\r\n", "Related to this discussion: in https://github.com/huggingface/datasets/pull/3664#issuecomment-1031866858 I propose how we could change `iter_archive` to work for streaming and also return local paths (as it used too !). I'd love your opinions on this", "> @patrickvonplaten\r\n> \r\n> > The other solution of providing a path-like object derived from the bytes stocked in the .array file is not nearly as user-friendly, but better than nothing\r\n> \r\n> Just to clarify, here you describe the approach that uses the `Audio.decode` attribute to access the underlying bytes?\r\n\r\nYes! \r\n\r\n> \r\n> > The official example currently doesn't work and we don't even have a workaround for it for MP3 files at the moment\r\n> \r\n> I'd assume this is because we use `sox_io` as a backend for decoding. However, soon we should be able to use `soundfile`, which supports path-like objects, for MP3 ([#3667 (comment)](https://github.com/huggingface/datasets/pull/3667#issuecomment-1030090627)). \r\n> Your concern is reasonable, but there are situations where we can only serve bytes (see #3685 for instance). IMO it makes sense to fix the affected datasets for now, but I don't think we should care too much whether we rely on local paths or bytes after soundfile adds support for MP3 as long as our examples work (shouldn't be too hard to update the `map_to_array` functions) and we properly document how to access the underlying path/bytes for custom decoding (via `ds.cast_column(\"audio\", Audio(decode=False))`).\r\n\r\nYes this might be, but I highly doubt that `soundfile` is the go-to library for audio then. @anton-l and I have tried out a bunch of different audio loading libraries (`soundfile`, `librosa`, `torchaudio`, pure `ffmpeg`, `audioread`, ...). One thing that was pretty clear to me is that there is just no \"de-facto standard\" library and they all have pros and cons. None of the libraries really supports \"batch\"-ed audio loading. Some depend on PyTorch. `torchaudio` is 100x faster (really!) than `librosa's` fallback on MP3. `torchaudio` often has problems with multi-proessing, ... Also we should keep in mind that resampling is similarly not as simple as reading a text file. It's a pretty complex signal processing transform and people very well might want to use special filters, etc...at the moment we just hard-code `torchaudio's` or `librosa's` default filter when doing resampling.\r\n\r\n=> All this to say that we **should definitely** care about whether we rely on local paths or bytes IMO. We don't want to loose all users that are forced to use `datasets` decoding or resampling or have to built a very much not intuitive way of loading bytes into a numpy array. It's much more intuitive to be able to inspect a local file. I feel pretty strongly about this and am happy to also jump on a call. Keeping libraries flexible and lean as well as exposing internals is very important IMO (this philosophy has worked quite well so far with Transformers).\r\n\r\n", "Thanks a lot for the very detailed explanation. Now everything makes much more sense.", "From https://github.com/huggingface/datasets/pull/3736 the Common Voice dataset now gives access to the local audio files as before", "I understand the argument that it is bad to have a breaking change. How to deal with the introduction of breaking changes is a topic of its own and not sure how you want to deal with that (or is the policy this is never allowed, and there must be a `load_dataset_v2` or so if you really want to introduce a breaking change?).\r\n\r\nRegardless of whether it is a breaking change, however, I don't see the other arguments.\r\n\r\n> but **really** bad for Tensorflow and Flax where we **currently cannot** even use `datasets` to load `.mp3` files\r\n\r\nI don't exactly understand this. Why not?\r\n\r\nWhy does the HF dataset on-the-fly decoding mechanism not work? Why is it anyway specific to PyTorch or TensorFlow? Isn't this independent?\r\n\r\nBut even if you just provide the raw bytes to TF, on TF you could just use sth like `tfio.audio.decode_mp3` or `tf.audio.decode_ogg` or `tfio.audio.decode_flac`?\r\n\r\n> There are the following reasons why it is paramount to have access to the original audio file in my opinion ...\r\n\r\nI don't really understand the arguments (despite that it maybe breaks existing code). You anyway have the original audio files but it is just embedded in the dataset? I don't really know about any library which cannot also load the audio from memory (i.e. from the dataset).\r\n\r\nBtw, on librosa being slow for decoding audio files, I saw that as well, so we have this comment RETURNN:\r\n\r\n> Don't use librosa.load which internally uses audioread which would use Gstreamer as a backend which has multiple issues:\r\n> https://github.com/beetbox/audioread/issues/62\r\n> https://github.com/beetbox/audioread/issues/63\r\n> Instead, use PySoundFile (soundfile), which is also faster. See here for discussions:\r\n> https://github.com/beetbox/audioread/issues/64\r\n> https://github.com/librosa/librosa/issues/681\r\n\r\nResampling is also a separate aspect, which is also less straightforward and with different compromises between speed and quality. So there the different tradeoffs and different implementations can make a difference.\r\n\r\nHowever, I don't see how this is related to the question whether there should be the raw bytes inside the dataset or as separate local files.\r\n", "Thanks for your comments here @albertz - cool to get your input! \r\n\r\nAnswering a bit here between the lines:\r\n\r\n> I understand the argument that it is bad to have a breaking change. How to deal with the introduction of breaking changes is a topic of its own and not sure how you want to deal with that (or is the policy this is never allowed, and there must be a `load_dataset_v2` or so if you really want to introduce a breaking change?).\r\n> \r\n> Regardless of whether it is a breaking change, however, I don't see the other arguments.\r\n> \r\n> > but **really** bad for Tensorflow and Flax where we **currently cannot** even use `datasets` to load `.mp3` files\r\n> \r\n> I don't exactly understand this. Why not?\r\n\r\n> Why does the HF dataset on-the-fly decoding mechanism not work? Why is it anyway specific to PyTorch or TensorFlow? Isn't this independent?\r\n\r\nThe problem with decoding on the fly is that we currently rely on `torchaudio` for this now which relies on `torch` which is not necessarily something people would like to install when using `tensorflow` or `flax`. Therefore we cannot just rely on people using the decoding on the fly method. We just didn't find a library that is ML framework independent and fast enough for all formats. `torchaudio` is currently in our opinion by far the best here.\r\n\r\nSo for TF and Flax it's important that users can load audio files or bytes they way the want to - this might become less important if we find (or make) a good library with few dependencies that is fast for all kinds of platforms / use cases.\r\n\r\n\r\nNow the question is whether it's better to store audio data as a path to a file or as raw bytes I guess.\\\r\nMy main arguments for storing the audio data as a path to a file is pretty much all about users experience - I don't really expect our users to understand the inner workings of datasets:\r\n\r\n- 1. It's not straightforward to know which function to use to decode it - not all `load_audio(...)` or `read_audio(...)` work on raw bytes. E.g. Looking at https://pytorch.org/audio/stable/torchaudio.html?highlight=load#torchaudio.load one would not see directly how to load raw bytes . There are also some functions of other libraries which only work on files which would require the user to save the bytes as a file first before being able to load it.\r\n- 2. It's difficult to see which format the bytes are coming from (mp3, ogg, ...) - guess this could be remedied by adding the format to each sample though\r\n- 3. It is a bit scary IMO to see raw bytes for users. Overall, I think it's better to leave the data in it's raw form as this way it's much easier for people to play around with the audio files, less need to read docs because people don't worry about what happened to the audio files (are the bytes already resampled?)\r\n\r\nBut the argument that the audio should be loadable directly from memory is good - haven't thought about this too much. \r\nI guess it's still very much possible for the user to do this:\r\n\r\n```python\r\ndef save_as_bytes:\r\n batch[\"bytes\"] = read_in_bytes_from_file(batch[\"file\"])\\\r\n os.remove(batch[\"file\"])\r\n\r\nds = ds.map(save_as_bytes)\r\n\r\nds.save_to_disk(...)\r\n```\r\n\r\nGuess the question is more a bit about what should be the default case?", "> The problem with decoding on the fly is that we currently rely on `torchaudio` for this now which relies on `torch` which is not necessarily something people would like to install when using `tensorflow` or `flax`. Therefore we cannot just rely on people using the decoding on the fly method. We just didn't find a library that is ML framework independent and fast enough for all formats. `torchaudio` is currently in our opinion by far the best here.\r\n\r\nBut how is this relevant for this issue here? I thought this issue here is about having the (correct) path in the dataset or having raw bytes in the dataset.\r\n\r\nHow did TF users use it at all then? Or they just do not use on-the-fly decoding? I did not even notice this problem (maybe because I had `torchaudio` installed). But what do they use instead?\r\n\r\nBut as I outlined before, they could just use `tfio.audio.decode_flac` and co, where it would be more natural if you already provide the raw bytes.\r\n\r\n> Looking at https://pytorch.org/audio/stable/torchaudio.html?highlight=load#torchaudio.load one would not see directly how to load raw bytes\r\n\r\nI was not really familiar with `torchaudio`. It seems that they really don't provide an easy/direct API to operate on raw bytes. Which is very strange and unfortunate because as far as I can see, all the underlying backend libraries (e.g. soundfile) easily allow that. So I would say that this is the fault of `torchaudio` then. But despite, if you anyway use `torchaudio` with `soundfile` backend, why not just use `soundfile` directly. It's very simple to use and crossplatform.\r\n\r\nBut ok, now we are just discussing how to handle the on-the-fly decoding. I still think this is a separate issue and having raw bytes in the dataset instead of local files should just be fine as well.\r\n\r\n\r\n> It is a bit scary IMO to see raw bytes for users.\r\n\r\nI think nobody who writes code is scared by seeing the raw bytes content of a binary file. :)\r\n\r\n\r\n> I guess it's still very much possible for the user to do this:\r\n> \r\n> ```python\r\n> def save_as_bytes:\r\n> batch[\"bytes\"] = read_in_bytes_from_file(batch[\"file\"])\\\r\n> os.remove(batch[\"file\"])\r\n> \r\n> ds = ds.map(save_as_bytes)\r\n> \r\n> ds.save_to_disk(...)\r\n> ```\r\n\r\nIn https://github.com/huggingface/datasets/pull/4184#issuecomment-1105191639, you said/proposed that this `map` is not needed anymore and `save_to_disk` could do it automatically (maybe via some option)?\r\n\r\n> Guess the question is more a bit about what should be the default case?\r\n\r\nYea this is up to you. I'm happy as long as we can get it the way we want easily and this is a well supported use case. :)\r\n", "> In https://github.com/huggingface/datasets/pull/4184#issuecomment-1105191639, you said/proposed that this map is not needed anymore and save_to_disk could do it automatically (maybe via some option)?\r\n\r\nYes! Should be super easy now see discussion here: https://github.com/rwth-i6/i6_core/issues/257#issuecomment-1105494468\r\n\r\nThanks for the super useful input :-)", "Despite the comments that this has been fixed, I am finding the exact same problem is occurring again (with datasets version 2.3.2)", "> Despite the comments that this has been fixed, I am finding the exact same problem is occurring again (with datasets version 2.3.2)\r\n\r\nIt appears downgrading to torchaudio 0.11.0 fixed this problem.", "@DCNemesis, sorry which problem exactly is occuring again? Also cc @lhoestq @polinaeterna here", "@patrickvonplaten @lhoestq @polinaeterna I was unable to load audio from Common Voice using 🤗 with the current version of torchaudio, but downgrading to torchaudio 0.11.0 fixed it. This is probably more of a torch problem than a Hugging Face problem.", "@DCNemesis that's interesting, could you please share the error message if you still can access it? ", "@polinaeterna I believe it is the same exact error as above. It occurs on other .mp3 sources as well, but the problem is with torchaudio > 0.11.0. I've created a short colab notebook that reproduces the error, and the fix here: https://colab.research.google.com/drive/18wsuwdHwBPN3JkcnhEtk8MUYqF9swuWZ?usp=sharing", "Hi @DCNemesis,\r\n\r\nYour issue was slightly different from the original one in this issue page. Yours seems related to a change in the backend used by `torchaudio` (`ffmpeg` instead of `sox`). Refer to the issue page here:\r\n- #4776\r\n\r\nNormally, it should be circumvented with the patch made by @polinaeterna in:\r\n- #4923", "I think the original issue reported here was already fixed by:\r\n- #3736\r\n\r\nOtherwise, feel free to reopen." ]

[]

[]

[ "Thanks @lhoestq for finding the reason of incorrect resampling. This issue affects all languages which have sound files with different sampling rates such as Turkish and Luganda.", "@cahya-wirawan - do you know how many languages have different sampling rates in Common Voice? I'm quite surprised to see this for multiple languages actually", "@cahya-wirawan, I can reproduce the problem for Common Voice 7 for Turkish. Here a script you can use:\r\n\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nfrom datasets import load_dataset\r\nimport torchaudio\r\nfrom io import BytesIO\r\nfrom datasets import Audio\r\nfrom collections import Counter\r\nimport sys\r\n\r\nds_name = str(sys.argv[1])\r\nlang = str(sys.argv[2])\r\n\r\nds = load_dataset(ds_name, lang, split=\"train\", use_auth_token=True)\r\nds = ds.cast_column(\"audio\", Audio(decode=False))\r\n\r\nall_sampling_rates = []\r\n\r\n\r\ndef print_sampling_rate(x):\r\n x, sr = torchaudio.load(BytesIO(x[\"audio\"][\"bytes\"]), format=\"mp3\")\r\n all_sampling_rates.append(sr)\r\n\r\nds.map(print_sampling_rate)\r\n\r\n\r\nprint(Counter(all_sampling_rates))\r\n```\r\n\r\ncan be run with:\r\n\r\n```bash\r\npython run.py mozilla-foundation/common_voice_7_0 tr\r\n```\r\n\r\nFor CV 6.1 all samples seem to have the same audio", "It actually shows that many more samples are in 32kHz format than it 48kHz which is unexpected. Thanks a lot for flagging! Will contact Common Voice about this as well", "I only checked the CV 7.0 for Turkish, Luganda and Indonesian, they have audio files with difference sampling rates, and all of them are affected by this issue. Percentage of incorrect resampling as follow, Turkish: 9.1%, Luganda: 88.2% and Indonesian: 64.1%.\r\nI checked it using the original CV files. I check the original sampling rates and the length of audio array of each files and compare it with the length of audio array (and the sampling rate which is always 48kHz) from mozilla-foundation/common_voice_7_0 datasets. if the length of audio array from dataset is not equal to 48kHz/original sampling rate * length of audio array of the original audio file then it is affected,", "Ok wow, thanks a lot for checking this - you've found a pretty big bug :sweat_smile: It seems like **a lot** more datasets are actually affected than I original thought. We'll try to solve this as soon as possible and make an announcement tomorrow." ]

[]

[]

[ "The CI failure is only because of the datasets is missing some sections in their cards - we can ignore that since it's unrelated to this PR" ]

[]

[]

[]

[ { "id": 3470211881, "node_id": "LA_kwDODunzps7O1zsp", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset-viewer", "name": "dataset-viewer", "color": "E5583E", "default": false, "description": "Related to the dataset viewer on huggingface.co" } ]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Hi @thomas-happify, please note that the preview may take some time before rendering the data.\r\n\r\nI've seen it is already working.\r\n\r\nI close this issue. Please feel free to reopen it if the problem arises again." ]

[]

[]

[ "The error is now:\r\n\r\n```\r\nStatus code: 400\r\nException: Status400Error\r\nMessage: this dataset is not supported for now.\r\n```\r\n\r\nWe've disabled the dataset viewer for several big datasets like this one. We hope being able to reenable it soon.", "The list of splits cannot be obtained. cc @huggingface/datasets ", "```\r\nError code: SplitsNamesError\r\nException: SplitsNotFoundError\r\nMessage: The split names could not be parsed from the dataset config.\r\nTraceback: Traceback (most recent call last):\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 354, in get_dataset_config_info\r\n for split_generator in builder._split_generators(\r\n File \"/tmp/modules-cache/datasets_modules/datasets/bigscience--P3/12c0badfecad4564ecb8a6f81b5d0559656f269f08b13c59c93283f3a84134ba/P3.py\", line 154, in _split_generators\r\n data_dir = dl_manager.download_and_extract(_URLs)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 944, in download_and_extract\r\n return self.extract(self.download(url_or_urls))\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 907, in extract\r\n urlpaths = map_nested(self._extract, path_or_paths, map_tuple=True)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 393, in map_nested\r\n mapped = [\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 394, in <listcomp>\r\n _single_map_nested((function, obj, types, None, True, None))\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 346, in _single_map_nested\r\n return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 346, in <dictcomp>\r\n return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 346, in _single_map_nested\r\n return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 346, in <dictcomp>\r\n return {k: _single_map_nested((function, v, types, None, True, None)) for k, v in pbar}\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 330, in _single_map_nested\r\n return function(data_struct)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 912, in _extract\r\n protocol = _get_extraction_protocol(urlpath, use_auth_token=self.download_config.use_auth_token)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 402, in _get_extraction_protocol\r\n return _get_extraction_protocol_with_magic_number(f)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 367, in _get_extraction_protocol_with_magic_number\r\n magic_number = f.read(MAGIC_NUMBER_MAX_LENGTH)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 574, in read\r\n return super().read(length)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/spec.py\", line 1575, in read\r\n out = self.cache._fetch(self.loc, self.loc + length)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/caching.py\", line 377, in _fetch\r\n self.cache = self.fetcher(start, bend)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 111, in wrapper\r\n return sync(self.loop, func, *args, **kwargs)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 96, in sync\r\n raise return_result\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 53, in _runner\r\n result[0] = await coro\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 616, in async_fetch_range\r\n out = await r.read()\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/aiohttp/client_reqrep.py\", line 1036, in read\r\n self._body = await self.content.read()\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/aiohttp/streams.py\", line 375, in read\r\n block = await self.readany()\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/aiohttp/streams.py\", line 397, in readany\r\n await self._wait(\"readany\")\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/aiohttp/streams.py\", line 304, in _wait\r\n await waiter\r\n aiohttp.client_exceptions.ClientPayloadError: Response payload is not completed\r\n \r\n The above exception was the direct cause of the following exception:\r\n \r\n Traceback (most recent call last):\r\n File \"/src/services/worker/src/worker/responses/splits.py\", line 75, in get_splits_response\r\n split_full_names = get_dataset_split_full_names(dataset, hf_token)\r\n File \"/src/services/worker/src/worker/responses/splits.py\", line 35, in get_dataset_split_full_names\r\n return [\r\n File \"/src/services/worker/src/worker/responses/splits.py\", line 38, in <listcomp>\r\n for split in get_dataset_split_names(dataset, config, use_auth_token=hf_token)\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 404, in get_dataset_split_names\r\n info = get_dataset_config_info(\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 359, in get_dataset_config_info\r\n raise SplitsNotFoundError(\"The split names could not be parsed from the dataset config.\") from err\r\n datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.\r\n```" ]

[]

[]

[ "I'm impatient to see if it has an impact on the number of valid datasets for the dataset viewer. For the record, today:\r\n\r\n<img width=\"660\" alt=\"Capture d’écran 2022-02-01 à 14 32 19\" src=\"https://user-images.githubusercontent.com/1676121/151977579-b5a239d9-6662-4aeb-bfd1-eef6b8249991.png\">\r\n", "This is now available in `datasets` 1.18.3 :)", "I'm on it https://github.com/huggingface/datasets-preview-backend/issues/130\r\n", "The result:\r\n<img width=\"671\" alt=\"Capture d’écran 2022-02-03 à 23 45 55\" src=\"https://user-images.githubusercontent.com/1676121/152442169-bfdac643-9a00-4901-bfa7-1d60a1679d4b.png\">\r\n\r\nNot very different. Maybe it fixed issues in the community datasets... But I'm not 100% the two states are comparable (datasets have been created, or updated, meanwhile)" ]

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

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Hi @RensDimmendaal, \r\n\r\nI'm sorry but I can't reproduce your bug:\r\n```python\r\nIn [1]: from datasets import load_dataset\r\n ...: ds = load_dataset(\"subjqa\", \"electronics\")\r\nDownloading builder script: 9.15kB [00:00, 4.10MB/s] \r\nDownloading metadata: 17.7kB [00:00, 8.51MB/s] \r\nDownloading and preparing dataset subjqa/electronics (download: 10.86 MiB, generated: 3.01 MiB, post-processed: Unknown size, total: 13.86 MiB) to .../.cache/huggingface/datasets/subjqa/electronics/1.1.0/e5588f9298ff2d70686a00cc377e4bdccf4e32287459e3c6baf2dc5ab57fe7fd...\r\nDownloading data: 11.4MB [00:03, 3.50MB/s]\r\nDataset subjqa downloaded and prepared to .../.cache/huggingface/datasets/subjqa/electronics/1.1.0/e5588f9298ff2d70686a00cc377e4bdccf4e32287459e3c6baf2dc5ab57fe7fd. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 605.09it/s]\r\n\r\nIn [2]: ds\r\nOut[2]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\r\n num_rows: 1295\r\n })\r\n test: Dataset({\r\n features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\r\n num_rows: 358\r\n })\r\n validation: Dataset({\r\n features: ['domain', 'nn_mod', 'nn_asp', 'query_mod', 'query_asp', 'q_reviews_id', 'question_subj_level', 'ques_subj_score', 'is_ques_subjective', 'review_id', 'id', 'title', 'context', 'question', 'answers'],\r\n num_rows: 255\r\n })\r\n})\r\n```\r\n\r\nCould you please try again and see if the problem persists?\r\n\r\nIf that is the case, you can circumvent the issue by passing `ignore_verifications`:\r\n```python\r\nds = load_dataset(\"subjqa\", \"electronics\", ignore_verifications=True)", "Thanks checking!\r\n\r\nYou're totally right. I don't know what's changed, but I'm glad it's working now!\r\n\r\n" ]

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

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Hi @abhi-mosaic, thanks for reporting.\r\n\r\nI'm looking at it... ", "also hitting this issue", "Hey @albertvillanova, sorry to reopen this... I can confirm that on `master` branch the dataset is downloadable now but it is still broken in streaming mode:\r\n\r\n```python\r\n >>> import datasets\r\n >>> pubmed_train = datasets.load_dataset('pubmed', split='train', streaming=True)\r\n >>> next(iter(pubmed_train))\r\n```\r\n```\r\n No such file or directory: 'gzip://pubmed22n0001.xml::ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed22n0001.xml.gz'\r\n```\r\n", "Hi @abhi-mosaic, would you mind opening another issue for this new problem?\r\n\r\nFirst issue (already solved) was a ConnectionError due to the yearly update release of PubMed: we fixed it by updating the URLs from year 2021 to year 2022.\r\n\r\nHowever this is another problem: to make pubmed streamable. Please note that NOT all our datastes are streamable: we are making streamable more and more of them... but this is an on-going process...\r\n\r\nThanks.", "@albertvillanova \r\nWhen I tried below codes, I got the similar error\r\n\r\n```\r\n\r\ndataset=load_dataset(\"pubmed\",split=\"train\")\r\n\r\nCouldn't reach ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed21n0601.xml.gz\r\n```", "@y-rok you need to update `datasets`:\r\n```shell\r\npip install -U datasets\r\n```" ]

[]

[]

[ "@lhoestq I've created a notebook for you to see the difference: https://colab.research.google.com/drive/1by3EqnoKvC2p-yKW4lPDGOFOZHyGVyeQ?usp=sharing.\r\n\r\nFeel free to suggest better descriptions for the progress bars. \r\n\r\nIf everything looks good, think we can merge." ]

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

[]

[]

[]

[]

[ "The original openslr site mixed both names https://openslr.org/72/ :-)", "Yeah, I filed the issue to have it fixed there last year, but it looks like they missed a few." ]

[]

[]

[ "> all the tests pass, but I'm still not able to import the dataset\r\n\r\nSince it's not merged on `master` yet, you have to provide the path to your local `wiki_bio.py` to use it.\r\nIndeed the library downloads the dataset files from `master` if you have a dev installation of the library.\r\n\r\nI agree it would be nice to change that, and use the local dataset scripts from the `datasets` directory - it feels definitely more natural.", "Cool, thanks for your help and I agree!" ]

[]

[]

[ "Hi @thomasw21, I remember suggesting `imap_unordered` to @lhoestq at that time to speed up `to_json` further but after trying `pool_imap` on multiple datasets (>9GB) , memory utilisation was almost constant and we decided to go ahead with that only. \r\n\r\n1. Did you try this without `gzip`? Because `gzip` feature was introduced recently and I didn't check multi_proc thing with `gzip`. One thing I know is that `gzip` is slow in our implementation than `zip` (it's a WIP #3551) \r\n2. You can try reducing your batch size, this can also help in avoiding OOM errors!", "Thanks @bhavitvyamalik ! I see. I'm not sure this PR actually fixes things for me either (I ended up reducing the num_proc/batch_size to lower it). It does allow the process to run for longer, but I think the reason why it was waiting is that one of the process crashes .... Unfortunately I was working on a setup with a low RAM/cpu core ratio. I'm actually very surprised that it doesn't change memory utilization, otherwise I don't see the purpose of `imap_unordered` existing. I think it's main purpose are when you have high variance in samples (in terms of bytes), which causes unecessary accumulation in `imap`\r\n 1. Did not try without `gzip`\r\n 2. Yeah or `num_proc`", "Can you please try without `gzip` to see how it performs? If it works fine then we can improve `gzip` from our side (I'm already working on it)", "I'll be busy for next few weeks on another project, will do as soon as I have some bandwidth.\r\n" ]

[ { "id": 2067376369, "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request", "name": "dataset request", "color": "e99695", "default": false, "description": "Requesting to add a new dataset" }, { "id": 3608944167, "node_id": "LA_kwDODunzps7XHB4n", "url": "https://api.github.com/repos/huggingface/datasets/labels/multimodal", "name": "multimodal", "color": "19E633", "default": false, "description": "Multimodal datasets" } ]

[]

[]

[]

[]

[]

[]

[]

[ "Sure, let's include this in today's release.", "Cool ! The windows CI should be fixed on master now, feel free to merge :)" ]

[]

[]

[ "Works smoothly with the `transformers.Trainer` class now, thank you!" ]

[]

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

[]

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

[]

[ "Hi ! At the moment you can use `to_pandas()` to get a pandas DataFrame that supports `group_by` operations (make sure your dataset fits in memory though)\r\n\r\nWe use Arrow as a back-end for `datasets` and it doesn't have native group by (see https://github.com/apache/arrow/issues/2189) unfortunately.\r\n\r\nI just drafted what it could look like to have `group_by` in `datasets`:\r\n```python\r\nfrom datasets import concatenate_datasets\r\n\r\ndef group_by(d, col, join): \r\n \"\"\"from: https://github.com/huggingface/datasets/issues/3644\"\"\"\r\n # Get the indices of each group\r\n groups = {key: [] for key in d.unique(col)} \r\n def create_groups_indices(key, i): \r\n groups[key].append(i) \r\n d.map(create_groups_indices, with_indices=True, input_columns=col) \r\n # Get one dataset object per group\r\n groups = {key: d.select(indices) for key, indices in groups.items()} \r\n # Apply join function\r\n groups = {\r\n key: dataset_group.map(join, batched=True, batch_size=len(dataset_group), remove_columns=d.column_names)\r\n for key, dataset_group in groups.items()\r\n } \r\n # Return concatenation of all the joined groups\r\n return concatenate_datasets(groups.values())\r\n```\r\n\r\nexample of usage:\r\n```python\r\n\r\ndef join(batch): \r\n # take the batch of all the examples of a group, and return a batch with one aggregated example\r\n # (we could aggregate examples into several rows instead of one, if you want)\r\n return {\"total\": [batch[\"i\"]]} \r\n\r\nd = Dataset.from_dict({\r\n \"i\": [i for i in range(50)],\r\n \"group_key\": [i % 4 for i in range(50)],\r\n})\r\nprint(group_by(d, \"group_key\", join))\r\n# total\r\n# 0 [0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48]\r\n# 1 [1, 5, 9, 13, 17, 21, 25, 29, 33, 37, 41, 45, 49]\r\n# 2 [2, 6, 10, 14, 18, 22, 26, 30, 34, 38, 42, 46]\r\n# 3 [3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47]\r\n```\r\n\r\nLet me know if that helps !\r\n\r\ncc @albertvillanova @mariosasko for visibility", "@lhoestq As of PyArrow 7.0.0, `pa.Table` has the [`group_by` method](https://arrow.apache.org/docs/python/generated/pyarrow.Table.html#pyarrow.Table.group_by), so we should also consider using that function for grouping. ", "Any update on this?", "You can use https://github.com/mariosasko/datasets_sql by @mariosasko to go group by operations using SQL queries", "Hi, I have a similar issue as OP but the suggested solutions do not work for my case. Basically, I process documents through a model to extract the last_hidden_state, using the \"map\" method on a Dataset object, but would like to average the result over a categorical column at the end (i.e. groupby this column).\r\n- A to_pandas() saturates the memory, although it gives me the desired result through a .groupby().apply(np.mean, axis=0) on a smaller use-case,\r\n- The solution posted on Feb 4 is much too slow,\r\n- datasets_sql seems to not like the fact that I'm averaging np.arrays.\r\nSo I'm kinda out of \"non brute force\" options... Any help appreciated", "> Hi, I have a similar issue as OP but the suggested solutions do not work for my case. Basically, I process documents through a model to extract the last_hidden_state, using the \"map\" method on a Dataset object, but would like to average the result over a categorical column at the end (i.e. groupby this column).\r\n \r\nIf you haven't yet, you could explore using [Polars](https://www.pola.rs/) for this. It's a new DataFrame library written in Rust with Python bindings. It is Pandas like it in many ways ,but does have some biggish differences in syntax/approach so it's definitely not a drop-in replacement. \r\n\r\nPolar's also uses Arrow as a backend but also supports out-of-memory operations; in this case, it's probably easiest to write out your dataset to parquet and then use the polar's `scan_parquet` method (this will lazily read from the parquet file). The thing you get back from that is a `LazyDataFrame` i.e. nothing is loaded into memory until you specify a query and call a `collect` method. \r\n\r\nExample below of doing a groupby on a dataset which definitely wouldn't fit into memory on my machine:\r\n\r\n```\r\nfrom datasets import load_dataset\r\nimport polars as pl\r\n\r\nds = load_dataset(\"blbooks\")\r\nds['train'].to_parquet(\"test.parquet\")\r\ndf = pl.scan_parquet(\"test.parquet\")\r\ndf.groupby('date').agg([pl.count()]).collect()\r\n```\r\n\r\n>datasets_sql seems to not like the fact that I'm averaging np.arrays.\r\n\r\nI am not certain how Polars will handle this either. It does have NumPy support (https://pola-rs.github.io/polars-book/user-guide/howcani/interop/numpy.html) but I assume Polars will need to have at least enough memory in each group you want to average over so you may still end up needing more memory depending on the size of your dataset/groups. \r\n\r\n\r\n", "Hi @davanstrien , thanks a lot, I didn't know about this library and the answer works! I need to try it on the full dataset now, but I'm hopeful. Here's what my code looks like:\r\n```\r\nlist_size = 768\r\ndf.groupby(\"date\").agg(\r\n pl.concat_list(\r\n [\r\n pl.col(\"hidden_state\")\r\n .arr.slice(n, 1)\r\n .arr.first()\r\n .mean()\r\n for n in range(0, list_size)\r\n ]\r\n ).collect()\r\n```\r\n\r\nFor some reasons, the following code was giving me a \"mean() got unexpected argument 'axis'\":\r\n```\r\ndf2 = df.groupby('date').agg(\r\n pl.col(\"hidden_state\").map(np.mean).alias(\"average_hidden_state\")\r\n).collect()\r\n\r\n```\r\n\r\nEDIT: The solution works on my large dataset, the memory does not crash and the time is reasonable, thanks a lot again!", "@jeremylhour glad this worked for you :) ", "I find this functionality missing in my workflow as well and the workarounds with SQL and Polars unsatisfying. Since PyArrow has exposed this functionality, I hope this soon makes it into a release. (:" ]