Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "I think `stream` could be misleading since the data is not being streamed from remote endpoints (one could think that's the case when they see `load_dataset` followed by `stream`). Hence, I prefer the second option.\r\n\r\nPS: When we resolve https://github.com/huggingface/datasets/issues/4542, we could add `as_tf_dataset` to the API for consistency and deprecate `to_tf_dataset`." ]

[ "Could you share the full stack trace please ?\r\n\r\n\r\nCan you also try running this code ? It can be useful to determine if the issue comes from `datasets` or `fsspec` (streaming) or `pyarrow` (parquet reading):\r\n```python\r\nds = load_dataset(\"parquet\", data_files=a_parquet_file_url, use_auth_token=True)\r\n```", "Here's the full trace\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/loubna_huggingface_co/load.py\", line 15, in <module>\r\n ds_all = load_dataset(\"bigcode/the-stack-dedup-pjj\", data_dir=\"data/java\",use_auth_token=True, split=\"train\", revision=\"v1.1.a1\")\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 1742, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 814, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 905, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 1502, in _prepare_split\r\n for key, table in logging.tqdm(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/std.py\", line 1195, in __iter__\r\n for obj in iterable:\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py\", line 67, in _generate_tables\r\n parquet_file = pq.ParquetFile(f)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/pyarrow/parquet/__init__.py\", line 286, in __init__\r\n self.reader.open(\r\n File \"pyarrow/_parquet.pyx\", line 1227, in pyarrow._parquet.ParquetReader.open\r\n File \"pyarrow/error.pxi\", line 100, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.\r\n```\r\n\r\nwhen running\r\n```python\r\nds = load_dataset(\"parquet\", data_files=\"https://huggingface.co/datasets/bigcode/the-stack-dedup-pjj/blob/v1.1.a1/data/java/data_0000.parquet\", use_auth_token=True)\r\n```\r\nI get 401 error, but that's the case for the python subset too which I can load properly\r\n```\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 1719, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 1497, in load_dataset_builder\r\n dataset_module = dataset_module_factory(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 1134, in dataset_module_factory\r\n return PackagedDatasetModuleFactory(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 707, in get_module\r\n data_files = DataFilesDict.from_local_or_remote(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/data_files.py\", line 795, in from_local_or_remote\r\n DataFilesList.from_local_or_remote(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/data_files.py\", line 764, in from_local_or_remote\r\n origin_metadata = _get_origin_metadata_locally_or_by_urls(data_files, use_auth_token=use_auth_token)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/data_files.py\", line 710, in _get_origin_metadata_locally_or_by_urls\r\n return thread_map(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/contrib/concurrent.py\", line 94, in thread_map\r\n return _executor_map(ThreadPoolExecutor, fn, *iterables, **tqdm_kwargs)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/contrib/concurrent.py\", line 76, in _executor_map\r\n return list(tqdm_class(ex.map(fn, *iterables, **map_args), **kwargs))\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/std.py\", line 1183, in __iter__\r\n for obj in iterable:\r\n File \"/opt/conda/envs/venv/lib/python3.9/concurrent/futures/_base.py\", line 609, in result_iterator\r\n yield fs.pop().result()\r\n File \"/opt/conda/envs/venv/lib/python3.9/concurrent/futures/_base.py\", line 446, in result\r\n return self.__get_result()\r\n File \"/opt/conda/envs/venv/lib/python3.9/concurrent/futures/_base.py\", line 391, in __get_result\r\n raise self._exception\r\n File \"/opt/conda/envs/venv/lib/python3.9/concurrent/futures/thread.py\", line 58, in run\r\n result = self.fn(*self.args, **self.kwargs)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/data_files.py\", line 701, in _get_single_origin_metadata_locally_or_by_urls\r\n return (request_etag(data_file, use_auth_token=use_auth_token),)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/utils/file_utils.py\", line 411, in request_etag\r\n response.raise_for_status()\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/requests/models.py\", line 960, in raise_for_status\r\n raise HTTPError(http_error_msg, response=self)\r\nrequests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/datasets/bigcode/the-stack-dedup-pjj/blob/v1.1.a1/data/python/data_0000.parquet```", "Can you check you used the right token ? You shouldn't get a 401 using your token", "I checked it’s the right token, when loading the full dataset I get the error after data extraction so I can access the files. \r\n```\r\nDownloading and preparing dataset parquet/bigcode--the-stack-dedup-pjj to /home/loubna_huggingface_co/.cache/huggingface/datasets/bigcode___parquet/bigcode--the-stack-dedup-pjj-872ffac7f4bb46ca/0.0.0/2a3b91fbd88a2c90d1dbbb32b460cf621d31bd5b05b934492fdef7d8d6f236ec...\r\nDownloading data files: 100%|██████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 22.38it/s]\r\nExtracting data files: 100%|███████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 49.91it/s]\r\nTraceback (most recent call last):\r\n File \"/home/loubna_huggingface_co/load_ds.py\", line 5, in <module>\r\n ds = load_dataset(\"bigcode/the-stack-dedup-pjj\", data_dir=\"data/java\", use_auth_token=True,split=\"train\", revision=\"v1.1.a1\")\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py\", line 1742, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 814, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 905, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py\", line 1502, in _prepare_split\r\n for key, table in logging.tqdm(\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/std.py\", line 1195, in __iter__\r\n for obj in iterable:\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py\", line 67, in _generate_tables\r\n parquet_file = pq.ParquetFile(f)\r\n File \"/opt/conda/envs/venv/lib/python3.9/site-packages/pyarrow/parquet/__init__.py\", line 286, in __init__\r\n self.reader.open(\r\n File \"pyarrow/_parquet.pyx\", line 1227, in pyarrow._parquet.ParquetReader.open\r\n File \"pyarrow/error.pxi\", line 100, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.\r\n```\r\nCould it be that I'm using a wrong url, I just copied it from the address bar", "The URL is wrong indeed, the right one is the one with \"resolve\" (the one you get when clicking on \"download\")- otherwise you try to download an html page ;)\r\n```\r\nhttps://huggingface.co/datasets/bigcode/the-stack-dedup-pjj/resolve/v1.1.a1/data/java/data_0000.parquet\r\n```", "Ah thanks! So I tried it with the first parquet file and it works, is there a way to know which parquet file was causing the issue since there are a lot of shards?", "I think you have to try them all :/\r\n\r\nAlternatively you can add a try/catch in `parquet.py` in `datasets` to raise the name of the file that fails at doing `parquet_file = pq.ParquetFile(f)` when you run your initial code\r\n```python\r\nload_dataset(\"bigcode/the-stack-dedup-pjj\", data_dir=\"data/java\", split=\"train\", revision=\"v1.1.a1\", use_auth_token=True)\r\n```\r\nbut it will still iterate on all the files until it fails", "Ok I will do that", "I did find the file, and I get the same error as before \r\n```\r\nDownloading data files: 100%|███████████████████| 1/1 [00:00<00:00, 8160.12it/s]\r\nExtracting data files: 100%|████████████████████| 1/1 [00:00<00:00, 1447.81it/s]\r\n \r\n---------------------------------------------------------------------------\r\nArrowInvalid Traceback (most recent call last)\r\nInput In [22], in <cell line: 7>()\r\n 4 data_features = (data[\"train\"].features)\r\n 6 url = \"/home/loubna_huggingface_co/.cache/huggingface/datasets/downloads/93431bc4380de07de8b0ab533666cb5a6120cbe266779e0a63c86bf7717475d7\"\r\n----> 7 data = load_dataset(\"parquet\", \r\n 8 data_files=url,\r\n 9 split=\"train\",\r\n 10 features=data_features,\r\n 11 use_auth_token=True)\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/datasets/load.py:1742, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, keep_in_memory, save_infos, revision, use_auth_token, task, streaming, **config_kwargs)\r\n 1739 try_from_hf_gcs = path not in _PACKAGED_DATASETS_MODULES\r\n 1741 # Download and prepare data\r\n-> 1742 builder_instance.download_and_prepare(\r\n 1743 download_config=download_config,\r\n 1744 download_mode=download_mode,\r\n 1745 ignore_verifications=ignore_verifications,\r\n 1746 try_from_hf_gcs=try_from_hf_gcs,\r\n 1747 use_auth_token=use_auth_token,\r\n 1748 )\r\n 1750 # Build dataset for splits\r\n 1751 keep_in_memory = (\r\n 1752 keep_in_memory if keep_in_memory is not None else is_small_dataset(builder_instance.info.dataset_size)\r\n 1753 )\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py:814, in DatasetBuilder.download_and_prepare(self, output_dir, download_config, download_mode, ignore_verifications, try_from_hf_gcs, dl_manager, base_path, use_auth_token, file_format, max_shard_size, storage_options, **download_and_prepare_kwargs)\r\n 808 if not downloaded_from_gcs:\r\n 809 prepare_split_kwargs = {\r\n 810 \"file_format\": file_format,\r\n 811 \"max_shard_size\": max_shard_size,\r\n 812 **download_and_prepare_kwargs,\r\n 813 }\r\n--> 814 self._download_and_prepare(\r\n 815 dl_manager=dl_manager,\r\n 816 verify_infos=verify_infos,\r\n 817 **prepare_split_kwargs,\r\n 818 **download_and_prepare_kwargs,\r\n 819 )\r\n 820 # Sync info\r\n 821 self.info.dataset_size = sum(split.num_bytes for split in self.info.splits.values())\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py:905, in DatasetBuilder._download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)\r\n 901 split_dict.add(split_generator.split_info)\r\n 903 try:\r\n 904 # Prepare split will record examples associated to the split\r\n--> 905 self._prepare_split(split_generator, **prepare_split_kwargs)\r\n 906 except OSError as e:\r\n 907 raise OSError(\r\n 908 \"Cannot find data file. \"\r\n 909 + (self.manual_download_instructions or \"\")\r\n 910 + \"\\nOriginal error:\\n\"\r\n 911 + str(e)\r\n 912 ) from None\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/datasets/builder.py:1502, in ArrowBasedBuilder._prepare_split(self, split_generator, file_format, max_shard_size)\r\n 1500 total_num_examples, total_num_bytes = 0, 0\r\n 1501 try:\r\n-> 1502 for key, table in logging.tqdm(\r\n 1503 generator,\r\n 1504 unit=\" tables\",\r\n 1505 leave=False,\r\n 1506 disable=not logging.is_progress_bar_enabled(),\r\n 1507 ):\r\n 1508 if max_shard_size is not None and writer._num_bytes > max_shard_size:\r\n 1509 num_examples, num_bytes = writer.finalize()\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/tqdm/std.py:1195, in tqdm.__iter__(self)\r\n 1192 time = self._time\r\n 1194 try:\r\n-> 1195 for obj in iterable:\r\n 1196 yield obj\r\n 1197 # Update and possibly print the progressbar.\r\n 1198 # Note: does not call self.update(1) for speed optimisation.\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py:67, in Parquet._generate_tables(self, files)\r\n 65 for file_idx, file in enumerate(itertools.chain.from_iterable(files)):\r\n 66 with open(file, \"rb\") as f:\r\n---> 67 parquet_file = pq.ParquetFile(f)\r\n 68 try:\r\n 69 for batch_idx, record_batch in enumerate(\r\n 70 parquet_file.iter_batches(batch_size=self.config.batch_size, columns=self.config.columns)\r\n 71 ):\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/pyarrow/parquet/__init__.py:286, in ParquetFile.__init__(self, source, metadata, common_metadata, read_dictionary, memory_map, buffer_size, pre_buffer, coerce_int96_timestamp_unit, decryption_properties, thrift_string_size_limit, thrift_container_size_limit)\r\n 280 def __init__(self, source, *, metadata=None, common_metadata=None,\r\n 281 read_dictionary=None, memory_map=False, buffer_size=0,\r\n 282 pre_buffer=False, coerce_int96_timestamp_unit=None,\r\n 283 decryption_properties=None, thrift_string_size_limit=None,\r\n 284 thrift_container_size_limit=None):\r\n 285 self.reader = ParquetReader()\r\n--> 286 self.reader.open(\r\n 287 source, use_memory_map=memory_map,\r\n 288 buffer_size=buffer_size, pre_buffer=pre_buffer,\r\n 289 read_dictionary=read_dictionary, metadata=metadata,\r\n 290 coerce_int96_timestamp_unit=coerce_int96_timestamp_unit,\r\n 291 decryption_properties=decryption_properties,\r\n 292 thrift_string_size_limit=thrift_string_size_limit,\r\n 293 thrift_container_size_limit=thrift_container_size_limit,\r\n 294 )\r\n 295 self.common_metadata = common_metadata\r\n 296 self._nested_paths_by_prefix = self._build_nested_paths()\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/pyarrow/_parquet.pyx:1227, in pyarrow._parquet.ParquetReader.open()\r\n\r\nFile /opt/conda/envs/venv/lib/python3.9/site-packages/pyarrow/error.pxi:100, in pyarrow.lib.check_status()\r\n\r\nArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.\r\n```", "Can you check the JSON file associated to `/home/loubna_huggingface_co/.cache/huggingface/datasets/downloads/93431bc4380de07de8b0ab533666cb5a6120cbe266779e0a63c86bf7717475d7` ? In the JSON file we can know from where it was downloaded\r\n\r\nYou can find it at `/home/loubna_huggingface_co/.cache/huggingface/datasets/downloads/93431bc4380de07de8b0ab533666cb5a6120cbe266779e0a63c86bf7717475d7.json`", "It's this file `https://huggingface.co/datasets/bigcode/the-stack-dedup-pjj/resolve/f48656daa9f3a3607dacf8b57a65810a6a7a7f73/data/java/data_0022.parquet` loading it gives the same error", "I'm able to load it properly using\r\n```python\r\nds = load_dataset(\"parquet\", data_files=a_parquet_file_url, use_auth_token=token)\r\n```\r\n\r\nMy guess is that your download was corrupted. Please delete `93431bc4380de07de8b0ab533666cb5a6120cbe266779e0a63c86bf7717475d7` and `93431bc4380de07de8b0ab533666cb5a6120cbe266779e0a63c86bf7717475d7.json` locally and try again", "That worked, thanks! But I thought if something went wrong with a download `datasets` creates new cache for all the files, that's not the case? (at some point I even changed dataset versions so it was still using that cache?)", "Cool !\r\n\r\n> But I thought if something went wrong with a download datasets creates new cache for all the files\r\n\r\nWe don't perform integrity verifications if we don't know in advance the hash of the file to download.\r\n\r\n> at some point I even changed dataset versions so it was still using that cache?\r\n\r\n`datasets` caches the files by URL and ETag. If the content of a file changes, then the ETag changes and so it redownloads the file", "I see, thank you!\r\n", "I experience the same error in v 2.12.0. But found out it was due to one column from polars was a categorical dtype (related to the error from #5706. Temporarily resolved it by casting the column to str instead." ]

[]

[ "Hi ! It looks like your \"isDif\" column is a Sequence of Value(\"string\"), not a Sequence of ClassLabel.\r\n\r\nYou can convert your Value(\"string\") feature type to a ClassLabel feature type this way:\r\n```python\r\nfrom datasets import ClassLabel, Sequence\r\n\r\n# provide the label_names yourself\r\nlabel_names = [...]\r\n# OR get them from the dataset\r\nlabel_names = sorted(set(label for labels in raw_datasets[\"train\"][\"isDif\"] for label in labels))\r\n\r\n# Cast to ClassLabel\r\nraw_datasets = raw_datasets.cast_column(\"isDif\", Sequence(ClassLabel(names=label_names)))\r\n```\r\n", "thank you \r\nit works 💯 " ]

[ "cc @albertvillanova the author would like to add this dataset to the hub: https://github.com/microsoft/table-transformer/issues/68#issuecomment-1319114621. Could you help him out?" ]

[ "Thanks for reporting, @adiprasad.\r\n\r\nWe are having a look at it.", "I have opened an issue in that dataset Community tab on the Hub: https://huggingface.co/datasets/consumer-finance-complaints/discussions/1\r\n\r\nPlease note that in the meantime, you can load the dataset by passing `ignore_verifications=True`:\r\n```python\r\n>>> ds = load_dataset(\"consumer-finance-complaints\", ignore_verifications=True)\r\n>>> ds\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['Date Received', 'Product', 'Sub Product', 'Issue', 'Sub Issue', 'Complaint Text', 'Company Public Response', 'Company', 'State', 'Zip Code', 'Tags', 'Consumer Consent Provided', 'Submitted via', 'Date Sent To Company', 'Company Response To Consumer', 'Timely Response', 'Consumer Disputed', 'Complaint ID'],\r\n num_rows: 3079747\r\n })\r\n})\r\n```", "PR fixing this issue: https://huggingface.co/datasets/consumer-finance-complaints/discussions/2" ]

[ "I notice a comment in the code says:\r\n`Having lists of different sizes makes sharding ambigious, raise an error in this case until we decide how to define sharding without ambiguity for users` \r\n \r\n ... which suggests this update was pushed knowing that it might break some things. But, it didn't seem to have a useful error message of an argument that could be passed to avoid the error.", "Sorry for the inconvenience, I opened a PR in your repo to fix this: https://huggingface.co/datasets/sil-ai/bloom-speech/discussions/2\r\n\r\nBasically we've always considered lists in `gen_kwargs` to be a shard list that we can split and pass into different workers to generate the dataset (e.g. if you pass `num_proc=` in `load_dataset()` to generate the dataset in parallel), but it was documented only recently", "@lhoestq Thanks for the help. It looks like that took care of it." ]

[ "The bulk edit is running...\r\n\r\nSee for example: \r\n- A single config: https://huggingface.co/datasets/acronym_identification/discussions/2\r\n- Multiple configs: https://huggingface.co/datasets/babi_qa/discussions/1", "TODO: Add \"dataset_info\" YAML metadata to:\r\n- [x] \"chr_en\" has no metadata JSON file, nor \"dataset_info\" YAML tag in its card\r\n - Fixing PR: https://huggingface.co/datasets/chr_en/discussions/1 \r\n- [x] \"conll2000\" has no metadata JSON file, but it has \"dataset_info\" YAML tag in its card\r\n- [x] \"crime_and_punish\" has no metadata JSON file, but it has \"dataset_info\" YAML tag in its card\r\n- [x] \"dart\" has no metadata JSON file, but it has \"dataset_info\" YAML tag in its card\r\n- [x] \"iwslt2017\" has no metadata JSON file, but it has \"dataset_info\" YAML tag in its card\r\n- [ ] \"mc4\" has no metadata JSON file, nor \"dataset_info\" YAML tag in its card\r\n- [ ] \"the_pile\" has no metadata JSON file, nor \"dataset_info\" YAML tag in its card\r\n- [ ] \"timit_asr\" has no metadata JSON file, nor \"dataset_info\" YAML tag in its card", "The bulk edit is finished." ]

[ "Also cc @mariosasko and @lhoestq ", "Cool ! Let us know if you have questions or if we can help :)\r\n\r\nI guess we'll also have to create the NYU CS Department on the Hub ?", "> I guess we'll also have to create the NYU CS Department on the Hub ?\r\n\r\nYes, you're right! Let me add it to my profile first, and then we can transfer. Meanwhile, if it's recommended to loop the dataset author in here, let me know. \r\n\r\nAlso, the NYU Depth dataset seems big. Any example scripts for creating image datasets that I could refer? ", "You can check the imagenet-1k one.\r\n\r\nPS: If the licenses allows it, it'b be nice to host the dataset as sharded TAR archives (like imagenet-1k) instead of the ZIP format they use:\r\n- it will make streaming much faster\r\n- ZIP compression is not well suited for images\r\n- it will allow parallel processing of the dataset (you can pass a subset of shards to each worker)\r\n\r\n> if it's recommended to loop the dataset author in here, let me know.\r\n\r\nIt's recommended indeed, you can send them an email once you have the dataset ready and invite them to the org on the Hub", "> You can check the imagenet-1k one.\r\n\r\nWhere can I find the script? Are you referring to https://huggingface.co/docs/datasets/image_process ? Or is there anything more specific? ", "You can find it here: https://huggingface.co/datasets/imagenet-1k/blob/main/imagenet-1k.py", "Update: started working on it here: https://huggingface.co/datasets/sayakpaul/nyu_depth_v2. \r\n\r\nI am facing an issue and I have detailed it here: https://huggingface.co/datasets/sayakpaul/nyu_depth_v2/discussions/1\r\n\r\nEdit: The issue is gone. \r\n\r\nHowever, since the dataset is distributed as a single TAR archive (following the [URL used in TensorFlow Datasets](https://github.com/tensorflow/datasets/tree/master/tensorflow_datasets/datasets/nyu_depth_v2/nyu_depth_v2_dataset_builder.py)) the loading is taking longer. How would suggest to shard the single TAR archive? \r\n\r\n@lhoestq \r\n\r\n", "A Colab Notebook demonstrating the dataset loading part: \r\n\r\nhttps://colab.research.google.com/gist/sayakpaul/aa0958c8d4ad8518d52a78f28044d871/scratchpad.ipynb\r\n\r\n@osanseviero @lhoestq \r\n\r\nI will work on a notebook to work with the dataset including data visualization.", "@osanseviero @lhoestq things seem to work fine with the current version of the dataset [here](https://huggingface.co/datasets/sayakpaul/nyu_depth_v2). Here's a notebook I developed to help with visualization: https://colab.research.google.com/drive/1K3ZU8XUPRDOYD38MQS9nreQXJYitlKSW?usp=sharing. \r\n\r\n@lhoestq I need your help with the following:\r\n\r\n> However, since the dataset is distributed as a single TAR archive (following the [URL used in TensorFlow Datasets](https://github.com/tensorflow/datasets/tree/master/tensorflow_datasets/datasets/nyu_depth_v2/nyu_depth_v2_dataset_builder.py)) the loading is taking longer. How would suggest to shard the single TAR archive?\r\n\r\n@osanseviero @lhoestq question for you:\r\n\r\nWhere should we host the dataset? I think hosting it under hf.co/datasets (that is HF is the org) is fine as we have ImageNet-1k hosted similarly. We could then reach out to Diana Wofk (author of [Fast Depth](https://github.com/dwofk/fast-depth) and the owner of the repo on which TFDS NYU Depth V2 is based) for a review. WDYT? ", "> However, since the dataset is distributed as a single TAR archive (following the [URL used in TensorFlow Datasets](https://github.com/tensorflow/datasets/tree/master/tensorflow_datasets/datasets/nyu_depth_v2/nyu_depth_v2_dataset_builder.py)) the loading is taking longer. How would suggest to shard the single TAR archive?\r\n\r\nFirst you can separate the train data and the validation data.\r\n\r\nThen since the dataset is quite big, you can even shard the train split and the validation split in multiple TAR archives. Something around 16 archives for train and 4 for validation would be fine for example.\r\n\r\nAlso no need to gzip the TAR archives, the images are already compressed in png or jpeg.", "> Then since the dataset is quite big, you can even shard the train split and the validation split in multiple TAR archives. Something around 16 archives for train and 4 for validation would be fine for example.\r\n\r\nYes, I got you. But this process seems to be manual and should be tailored for the given dataset. Do you have any script that you used to create the ImageNet-1k shards? \r\n\r\n> Also no need to gzip the TAR archives, the images are already compressed in png or jpeg.\r\n\r\nI was not going to do that. Not sure what brought it up. ", "> Yes, I got you. But this process seems to be manual and should be tailored for the given dataset. Do you have any script that you used to create the ImageNet-1k shards?\r\n\r\nI don't, but I agree it'd be nice to have a script for that !\r\n\r\n> I was not going to do that. Not sure what brought it up.\r\n\r\nThe original dataset is gzipped for some reason", "Oh, I am using this URL for the download: https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/datasets/nyu_depth_v2/nyu_depth_v2_dataset_builder.py#L24. ", "> Where should we host the dataset? I think hosting it under hf.co/datasets (that is HF is the org) is fine as we have ImageNet-1k hosted similarly.\r\n\r\nMaybe you can create an org for NYU Courant (this is the institute of the lab of the main author of the dataset if I'm not mistaken), and invite the authors to join.\r\n\r\nWe don't add datasets without namespace anymore", "Updates: https://huggingface.co/datasets/sayakpaul/nyu_depth_v2/discussions/5\r\n\r\nThe entire process (preparing multiple archives, preparing data loading script, etc.) was fun and engaging, thanks to the documentation. I believe we could work on a small blog post that would work as a reference for the future contributors following this path. What say? \r\n\r\nCc: @lhoestq @osanseviero ", "> I believe we could work on a small blog post that would work as a reference for the future contributors following this path. What say?\r\n\r\n@polinaeterna already mentioned it would be nice to present this process for audio (it's exactly the same), I believe it can be useful to many people", "Cool. Let's work on that after the NYU Depth Dataset is fully in on Hub (under the appropriate org). 🤗", "@lhoestq need to discuss something while I am adding the dataset card to https://huggingface.co/datasets/sayakpaul/nyu_depth_v2/. \r\n\r\nAs per [Papers With Code](https://paperswithcode.com/dataset/nyuv2), NYU Depth v2 is used for many different tasks:\r\n\r\n* Monocular depth estimation\r\n* Depth estimation \r\n* Semantic segmentation\r\n* Plane instance segmentation \r\n* ...\r\n\r\nSo, while writing the supported task part of the dataset card, should we focus on all these? IMO, we could focus on just depth estimation and semantic segmentation for now since we have supported models for these two. WDYT?\r\n\r\nAlso, I am getting: \r\n\r\n\r\n```\r\nremote: Your push was accepted, but with warnings:\r\nremote: - Warning: The task_ids \"depth-estimation\" is not in the official list: acceptability-classification, entity-linking-classification, fact-checking, intent-classification, multi-class-classification, multi-label-classification, multi-input-text-classification, natural-language-inference, semantic-similarity-classification, sentiment-classification, topic-classification, semantic-similarity-scoring, sentiment-scoring, sentiment-analysis, hate-speech-detection, text-scoring, named-entity-recognition, part-of-speech, parsing, lemmatization, word-sense-disambiguation, coreference-resolution, extractive-qa, open-domain-qa, closed-domain-qa, news-articles-summarization, news-articles-headline-generation, dialogue-generation, dialogue-modeling, language-modeling, text-simplification, explanation-generation, abstractive-qa, open-domain-abstractive-qa, closed-domain-qa, open-book-qa, closed-book-qa, slot-filling, masked-language-modeling, keyword-spotting, speaker-identification, audio-intent-classification, audio-emotion-recognition, audio-language-identification, multi-label-image-classification, multi-class-image-classification, face-detection, vehicle-detection, instance-segmentation, semantic-segmentation, panoptic-segmentation, image-captioning, grasping, task-planning, tabular-multi-class-classification, tabular-multi-label-classification, tabular-single-column-regression, rdf-to-text, multiple-choice-qa, multiple-choice-coreference-resolution, document-retrieval, utterance-retrieval, entity-linking-retrieval, fact-checking-retrieval, univariate-time-series-forecasting, multivariate-time-series-forecasting, visual-question-answering, document-question-answering\r\nremote: ----------------------------------------------------------\r\nremote: Please find the documentation at:\r\nremote: https://huggingface.co/docs/hub/model-cards#model-card-metadata\r\n```\r\n\r\nWhat should be the plan of action for this?\r\n\r\nCc: @osanseviero \r\n\r\n", "> What should be the plan of action for this?\r\n\r\nWhen you merged https://github.com/huggingface/hub-docs/pull/488, there is a JS Interfaces GitHub Actions workflow that runs https://github.com/huggingface/hub-docs/actions/workflows/js-interfaces-tests.yml. It has a step called [export-task scripts](https://github.com/huggingface/hub-docs/actions/runs/3622479064/jobs/6107238948) which exports an interface you can use in `dataset`. If you look at the logs, it prints out a map. This map can replace https://github.com/huggingface/datasets/blob/main/src/datasets/utils/resources/tasks.json (tasks.json was generated with this script), which should add depth estimation\r\n", "Thanks @osanseviero. \r\n\r\nhttps://github.com/huggingface/datasets/pull/5335", "Closing the issue as the dataset has been successfully added: https://huggingface.co/datasets/sayakpaul/nyu_depth_v2" ]

[ "After a discussion with @mishig25:\r\n- He said that this action should be triggered if we call our release branch according to the regex `v*-release`, as transformers does\r\n- I said that our procedure is different: our release branch is *temporary* and it is deleted just after the release PR is merged to main\r\n - Indeed the release tag is not yet created when we make the release PR (not event when this is merged to main), but when we make the Release itself.\r\n\r\nI was thinking that maybe we could change the triggering event: use `release` instead of `push`.\r\n\r\nWhat do you think, @huggingface/datasets?", "Why is it an issue if our branch is temporary ?", "He says not; but the branch has no tag yet; does the doc building require the tag? Or just the version number in `__init__.py` or setup.py?", "It uses `module.__version__` (i.e. the one defined in `__init__.py`) - no need to have a tag\r\n\r\nhttps://github.com/huggingface/doc-builder/blob/81575cf081964c30ea5fd39450f4820db963f18e/src/doc_builder/commands/build.py#L69", "Thanks, @lhoestq.\r\n\r\n@mishig25 has manually forced the generation of the docs, that are live for 2.7.0 version: https://huggingface.co/docs/datasets/v2.7.0/en/index ", "Cool ! this can be closed then ?", "I was waiting for #5250 to be merged to close this.", "just to confirm, is there anything I need to do from my side ? Or is everything good here ?" ]

[ "It seems all the tasks have been addressed, meaning this issue can be closed, no?" ]

[ "Hi @peregilk this bug is directly related to https://github.com/huggingface/datasets/issues/3888, and still not fixed... But I'll try to have a look!", "Thanks @alvarobartt. It is great if you are able to fix it, but when reading the explanation it seems like it is possible to work around it.\r\n\r\nWe also tried keeping the 'info.features' and then adding a modified version back after the remove/rename. Unforutunately that leads to a dataset that is not possible to iterate over.", "So if you iterate over the `IterableDataset` as `next(iter(ds))` and then run `rename_columns` when checking that data it will work, but in the end, it's just renaming the column one example/batch at a time, not renaming the column name for all the entries in the dataset, which is the ideal.", "@alvarobartt Thanks. My use case was that I wanted to do multiple things, ie removing all unnecessary columns, renaming some valid columns, and then using cast (in my case checking if the audio is not 16K and casting it). It is just convenient to look into the info.features between each of these operations. Alternatively, I will just plan ahead...;) To me it seems like all the operations are working.\r\n\r\nThanks for the advice. It was very useful.", "If we know the features before renaming, then we know the features after renaming, so we can pass the new features to the returned dataset in `rename_column` indeed ! If anyone is interested in contributing, feel free to open a PR and I'd be happy to help / give some pointers :)", "Sure @lhoestq thanks! I’ll try to work on that", "#self-assign" ]

[ "Hi ! What kind of private source ? We're exploring adding support for cloud storage and URIs like s3://, gs:// etc. with authentication in the download manager", "Hello! It's a google cloud storage, so gs://, but I'm using it with https.\r\nBeing able to provide a file system like [here](https://huggingface.co/docs/datasets/main/filesystems#load-serialized-datasets) would be even more practical indeed.\r\nI've found a quite complicated workaround which consists of monkey patching all of the functions in streaming_download_manager.py to use my own _get_authentication_headers_for_url_ . \r\n\r\nA support for this use case would be greatly appreciated!\r\n\r\nFor reference my _get_authentication_headers_for_url_ looks like this:\r\n```\r\nimport os\r\nfrom typing import Optional, Union\r\n\r\nfrom datasets import config\r\nfrom huggingface_hub import HfFolder\r\nfrom gcsfs.credentials import GoogleCredentials\r\n\r\nDEFAULT_PROJECT = os.environ.get(\"GCSFS_DEFAULT_PROJECT\", \"\")\r\naccess = \"full_control\"\r\ngcs_token = os.environ.get(\"GCS_TOKEN\")\r\n\r\n\r\ndef get_authentication_headers_for_url(url: str, use_auth_token: Optional[Union[str, bool]] = None) -> dict:\r\n \"\"\"Handle the HF authentication\"\"\"\r\n headers = {}\r\n if url.startswith(config.HF_ENDPOINT):\r\n if use_auth_token is False:\r\n token = None\r\n elif isinstance(use_auth_token, str):\r\n token = use_auth_token\r\n else:\r\n token = HfFolder.get_token()\r\n elif url.startswith(\"https://storage.googleapis.com\"):\r\n credentials = GoogleCredentials(DEFAULT_PROJECT, access, gcs_token)\r\n credentials.maybe_refresh()\r\n token = credentials.credentials.token\r\n else:\r\n token = None\r\n if token:\r\n headers[\"authorization\"] = f\"Bearer {token}\"\r\n return headers\r\n```", "I would be a big fan of this feature! @Hubert-Bonisseur if this doesn't become a supported feature, would you mind sharing your code? Thanks!", "> I would be a big fan of this feature! @Hubert-Bonisseur if this doesn't become a supported feature, would you mind sharing your code? Thanks!\r\n\r\nI published it here:\r\nhttps://github.com/Hubert-Bonisseur/private-dataset-hub\r\n\r\nI modified the names of a lot of functions for privacy and I don't have time to test it again so you may get import errors, but you have the code. The custom_load_dataset is the function you are interested in I think.\r\n\r\nIt relies a lot on patching, if you find a better way to do this, I'd be interested.", "Given the amount of patching it does, this is likely to break at one point. I'd encourage you to wait for a proper support in `datasets` directly if you can wait." ]

[ "Hi @capsabogdan! Unfortunately, it's hard to implement because quite often datasets data is being hosted in a single archive for all splits :( So we have to download the whole archive to split it into splits. This is the case for CommonVoice too. \r\n\r\nHowever, for cases when data is distributed in separate archives ащк different splits I suppose it can (and will) be implemented someday. \r\n\r\n\r\nBtw for quick check of the dataset you can use [streaming](https://huggingface.co/docs/datasets/stream):\r\n```python\r\ncv = load_dataset(\"mozilla-foundation/common_voice_11_0\", \"en\", split=\"test\", streaming=True)\r\ncv = iter(cv)\r\nprint(next(cv))\r\n\r\n>> {'client_id': 'a07b17f8234ded5e847443ea6f423cef745cbbc7537fb637d58326000aa751e829a21c4fd0a35fc17fb833aa7e95ebafce5efd19beeb8d843887b85e4eb35f5b',\r\n>> 'path': None,\r\n>> 'audio': {'path': 'cv-corpus-11.0-2022-09-21/en/clips/common_voice_en_100363.mp3',\r\n>> 'array': array([ 0.0000000e+00, 1.1748125e-14, 1.5450088e-14, ...,\r\n>> 1.3011958e-06, -6.3548953e-08, -9.9098514e-08], dtype=float32),\r\n>> ...}\r\n\r\n```", "thank you for the answer but am not sure if this will not be helpful, as we\nneed maybe just 10% of the datasets for some experiment\n\ncan we get just a portion of the dataset with stream?\n\n\nis there really no solution? :(\n\nAm Di., 15. Nov. 2022 um 16:55 Uhr schrieb Polina Kazakova <\n***@***.***>:\n\n> Hi @capsabogdan <https://github.com/capsabogdan>! Unfortunately, it's\n> hard to implement because quite often datasets data is being hosted in a\n> single archive for all splits :( So we have to download the whole archive\n> to split it into splits. This is the case for CommonVoice too.\n>\n> However, for cases when data is distributed in separate archives in\n> different splits I suppose it can be implemented someday.\n>\n> Btw for quick check of the dataset you can use streaming\n> <https://huggingface.co/docs/datasets/stream>:\n>\n> cv = load_dataset(\"mozilla-foundation/common_voice_11_0\", \"en\", split=\"test\", streaming=True)cv = iter(cv)print(next(cv))\n> >> {'client_id': 'a07b17f8234ded5e847443ea6f423cef745cbbc7537fb637d58326000aa751e829a21c4fd0a35fc17fb833aa7e95ebafce5efd19beeb8d843887b85e4eb35f5b',>> 'path': None,>> 'audio': {'path': 'cv-corpus-11.0-2022-09-21/en/clips/common_voice_en_100363.mp3',>> 'array': array([ 0.0000000e+00, 1.1748125e-14, 1.5450088e-14, ...,>> 1.3011958e-06, -6.3548953e-08, -9.9098514e-08], dtype=float32),>> ...}\n>\n> —\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/5243#issuecomment-1315512887>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ALSIFOC3JYRCTH54OBRUJULWIOW6PANCNFSM6AAAAAASAYO2LY>\n> .\n> You are receiving this because you were mentioned.Message ID:\n> ***@***.***>\n>\n", "maybe it would be nice if you guys ould do some sort of shard before\nloading the dataset, so users can download just chunks of data :)\n\nI think this would be very helpful\n\nAm Di., 15. Nov. 2022 um 19:24 Uhr schrieb Bogdan Capsa <\n***@***.***>:\n\n> thank you for the answer but am not sure if this will not be helpful, as\n> we need maybe just 10% of the datasets for some experiment\n>\n> can we get just a portion of the dataset with stream?\n>\n>\n> is there really no solution? :(\n>\n> Am Di., 15. Nov. 2022 um 16:55 Uhr schrieb Polina Kazakova <\n> ***@***.***>:\n>\n>> Hi @capsabogdan <https://github.com/capsabogdan>! Unfortunately, it's\n>> hard to implement because quite often datasets data is being hosted in a\n>> single archive for all splits :( So we have to download the whole archive\n>> to split it into splits. This is the case for CommonVoice too.\n>>\n>> However, for cases when data is distributed in separate archives in\n>> different splits I suppose it can be implemented someday.\n>>\n>> Btw for quick check of the dataset you can use streaming\n>> <https://huggingface.co/docs/datasets/stream>:\n>>\n>> cv = load_dataset(\"mozilla-foundation/common_voice_11_0\", \"en\", split=\"test\", streaming=True)cv = iter(cv)print(next(cv))\n>> >> {'client_id': 'a07b17f8234ded5e847443ea6f423cef745cbbc7537fb637d58326000aa751e829a21c4fd0a35fc17fb833aa7e95ebafce5efd19beeb8d843887b85e4eb35f5b',>> 'path': None,>> 'audio': {'path': 'cv-corpus-11.0-2022-09-21/en/clips/common_voice_en_100363.mp3',>> 'array': array([ 0.0000000e+00, 1.1748125e-14, 1.5450088e-14, ...,>> 1.3011958e-06, -6.3548953e-08, -9.9098514e-08], dtype=float32),>> ...}\n>>\n>> —\n>> Reply to this email directly, view it on GitHub\n>> <https://github.com/huggingface/datasets/issues/5243#issuecomment-1315512887>,\n>> or unsubscribe\n>> <https://github.com/notifications/unsubscribe-auth/ALSIFOC3JYRCTH54OBRUJULWIOW6PANCNFSM6AAAAAASAYO2LY>\n>> .\n>> You are receiving this because you were mentioned.Message ID:\n>> ***@***.***>\n>>\n>\n", "+1 on this feature request - I am running into the same problem, where I only need the test set for a dataset that has a huge training set", "Hey, I'm also interested in that as a feature. I'm having the same problem with Common Voice 13.0. The dataset is super big but I only want the test data to benchmark multilingual models, but I don't have much Terabytes to store all the dataset...", "Consider this approach: Download and save individual audio files by streaming each split, then compile a CSV file that contains the file names and corresponding text.\r\n\r\n```python3\r\nimport os\r\nimport shutil\r\nfrom pathlib import Path\r\n\r\nimport datasets\r\nimport pandas as pd\r\nimport soundfile\r\nfrom datasets import Dataset, concatenate_datasets, load_dataset\r\n\r\n\r\ndataset = load_dataset(\"librispeech_asr\", 'clean', split=\"train.100\", streaming=True)\r\ndataset = iter(dataset)\r\n\r\ndownload_path = os.path.join(os.getcwd(), 'librispeech', 'clips')\r\ncsv_name = os.path.join(os.getcwd(), 'librispeech', 'clean_train_100.csv')\r\n\r\nrows = []\r\nfor i, row in enumerate(dataset):\r\n print(i)\r\n path = os.path.join(download_path, row['audio']['path'])\r\n soundfile.write(path, row['audio']['array'], row['audio']['sampling_rate'])\r\n\r\n del row['audio']\r\n rows.append(row)\r\n\r\ndf = pd.DataFrame(rows)\r\ndf.to_csv(csv_name, index=False, header=True)\r\n```" ]

[ "cc @abhishekkrthakur @SBrandeis " ]

[ "Thanks for reporting, @vinaykakade.\r\n\r\nWe are discussing about making a release early this week.\r\n\r\nPlease note that in the meantime, in your specific case (as we also pointed out here: https://github.com/huggingface/datasets/issues/5162#issuecomment-1291720293), you can circumvent the issue by pinning `multiprocess` to 0.70.13 version (instead of using latest 0.70.14).\r\n\r\nDuplicate of:\r\n- https://github.com/huggingface/datasets/issues/5162", "You can also make `pip-compile` work by using the backtracking resolver (instead of the legacy one): https://pip-tools.readthedocs.io/en/latest/#a-note-on-resolvers\r\n```\r\npip-compile --resolver=backtracking requirements.in\r\n```\r\nThis resolver will automatically use `multiprocess` 0.70.13 version.\r\n" ]

[ "In case others find this, the problem was not with set_format, but my usages of `to_pandas()` and `from_pandas()` which I was using during dataset splitting; somewhere in the chain of converting to and from pandas the `Array2D/Array3D` types get converted to series of `Sequence()` types" ]

[ "I opened [this issue](https://github.com/python/cpython/issues/99401).\r\nPython's maintainers say that the issue is caused by [this change](https://docs.python.org/3.11/whatsnew/3.11.html#dataclasses).\r\nI believe adding a `__hash__` method to `datasets.utils.version.Version` should solve (at least partially) this issue.", "Has this been fixed? I am running into this issue now. \r\n\r\nIf this has been fixed, could have a new release with this?\r\n", "Hi, I am getting error while training \r\n\r\n(tensorflow) C:\\tensorflow\\models\\research\\object_detection>python train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config\r\nTraceback (most recent call last):\r\n File \"C:\\tensorflow\\models\\research\\object_detection\\train.py\", line 54, in <module>\r\n from object_detection.legacy import trainer\r\n File \"C:\\tensorflow\\models\\research\\object_detection\\legacy\\trainer.py\", line 27, in <module>\r\n from object_detection.builders import optimizer_builder\r\n File \"C:\\tensorflow\\models\\research\\object_detection\\builders\\optimizer_builder.py\", line 25, in <module>\r\n from official.modeling.optimization import ema_optimizer\r\n File \"C:\\tensorflow\\models\\official\\modeling\\optimization\\__init__.py\", line 19, in <module>\r\n from official.modeling.optimization.configs.optimization_config import *\r\n File \"C:\\tensorflow\\models\\official\\modeling\\optimization\\configs\\optimization_config.py\", line 31, in <module>\r\n @dataclasses.dataclass\r\n ^^^^^^^^^^^^^^^^^^^^^\r\n File \"C:\\Users\\x0133252\\AppData\\Local\\anaconda3\\envs\\tensorflow\\Lib\\dataclasses.py\", line 1223, in dataclass\r\n return wrap(cls)\r\n ^^^^^^^^^\r\n File \"C:\\Users\\x0133252\\AppData\\Local\\anaconda3\\envs\\tensorflow\\Lib\\dataclasses.py\", line 1213, in wrap\r\n return _process_class(cls, init, repr, eq, order, unsafe_hash,\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"C:\\Users\\x0133252\\AppData\\Local\\anaconda3\\envs\\tensorflow\\Lib\\dataclasses.py\", line 958, in _process_class\r\n cls_fields.append(_get_field(cls, name, type, kw_only))\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"C:\\Users\\x0133252\\AppData\\Local\\anaconda3\\envs\\tensorflow\\Lib\\dataclasses.py\", line 815, in _get_field\r\n raise ValueError(f'mutable default {type(f.default)} for field '\r\nValueError: mutable default <class 'official.modeling.optimization.configs.optimizer_config.SGDConfig'> for field sgd is not allowed: use default_factory", "@Jayanth1812 and anyone else receiving a similar issue, it most likely has to do with your Python version. Downgrading to Python 3.9 works for me, but doing a downgrade might impact a lot of things. So to be safe and what worked for me was creating a new conda environment and following the installations here: https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/install.html\r\n\r\nAnd for Tensorflow GPU compatibility, after installing TensorFlow follow the instructions in section 4 'GPU Setup' in this document: https://www.tensorflow.org/install/pip", "@Jayanth1812, you can see in your error stack trace, that the error is caused by the `tensorflow` library, not by the `datasets` library. See:\r\n```\r\nFile \"C:\\Users\\x0133252\\AppData\\Local\\anaconda3\\envs\\tensorflow\\Lib\\dataclasses.py\"\r\n```\r\n\r\nYou should open an issue in their repository instead: https://github.com/tensorflow/tensorflow " ]

[ "Hi! \r\n\r\nWe could address this by calling `.item()` on such tensors to extract the value, but this would lose us the type, which could lead to storing the generated dataset in a suboptimal format. Considering this, I think the only proper fix would be implementing support for 0-D tensors on Apache Arrow's side (Arrow is the underlying format we use to store datasets on disk/in memory). WDYT @lhoestq?", "I think we can just convert the item to a numpy typed scalar using `.numpy()` ?\r\n\r\nFor example this works:\r\n```python\r\nimport numpy as np\r\nimport pyarrow as pa\r\n\r\nassert pa.array([np.float64(1.0)]).type == pa.float64()\r\nassert pa.array([np.float32(1.0)]).type == pa.float32()\r\nassert pa.array([np.int32(1)]).type == pa.int32()\r\nassert pa.array([np.int64(1)]).type == pa.int64()\r\n```\r\n\r\nAnd therefore it would work the same as for PyTorch N-D Tensors: convert to Numpy Array to keep the type in `_cast_to_python_objects`, then convert to Arrow" ]

[ "`load_dataset` first checks for a local directory before checking for the Hub.\r\n\r\nTo make it explicit that it has to fetch the Hub, we could support the `hffs` syntax:\r\n```python\r\nload_dataset(\"hf://datasets/glue\")\r\n```\r\n\r\nwould that work for you ? Also cc @mariosasko who's leading the `hffs` project", "yeah, that would be a fine solution.", "This still has no proper solution in 2.11\r\n\r\nperhaps have a `download_config=\"force_remote\"` or just backtrack once you reach `EmptyDatasetError` locally and then try to load it from the hub (or a local cache, as that only gets checked if there is no local folder...?)" ]

[ "Fixed. Please close.", "how to fix?i need your help" ]

[ "Hi ! Datasets are memory mapped from your disk, i.e. they're not loaded in RAM. This is possible thanks to the Arrow data format.\r\n\r\nTherefore the column you remove is not in RAM, so removing it doesn't cause the RAM to decrease.", "Thanks for the explanation! @lhoestq \r\nI wonder since it is memory mapped, can we reduce or remove this memory map?", "Yes you can `del common_voice` for example or wait for it to be garbage collected" ]

[ "@NielsRogge @rwightman may have additional requirements regarding this feature.\r\n\r\nWhen adding a new (decodable) type, the hardest part is choosing the right decoding library. What I mean by \"right\" here is that it has all the features we need and is easy to install (with GPU support?).\r\n\r\nSome candidates/options:\r\n* [`decord`](https://github.com/dmlc/decord): no longer [maintained](https://github.com/dmlc/decord/issues/214), not trivial to install with GPU support\r\n* [`pyAV`](https://github.com/PyAV-Org/PyAV): used for CPU decoding in `torchvision`, GPU decoding not supported if I'm not mistaken, otherwise the best candidate probably\r\n* [`video_reader`](https://github.com/pytorch/vision/blob/de350bc01ad2193ea2888f0ce8a6a346d3cba5a9/torchvision/csrc/io/video_reader/video_reader.cpp): used for GPU decoding in `torchvision`, depends on `torch'\r\n* OpenCV: uses `ffmpeg` for video decoding under the hood\r\n* ...\r\n\r\nAnd the last resort is building our own library, which is the most flexible solution but also requires the most work.\r\n\r\nPS: I'm adding a link to an article that compares various video decoding libraries: https://towardsdatascience.com/lightning-fast-video-reading-in-python-c1438771c4e6", "@mariosasko is GPU decoding a hard requirement here? Do we really need it? (I don't know)\r\n\r\nSomething to consider with `decord` is that it doesn't (AFAIK) support writing videos, so you'd still need something else for that. also I've noticed [issues](https://github.com/dmlc/decord/issues/242) with decord's ability to decode stereo audio streams along side the video (which you don't run into with PyAV).\r\n\r\n---\r\n\r\nI think PyAV should be able to do the job just fine to start. If we write the video io utilities as their own functions, we can hot swap them later if we find/write a different solution that's faster/better.", "Video is still a bit of a mess, but I'd say pyAV is likely the best approach (or supporting all three via pytorchvideo, but that adds a middle man dependency).\r\n\r\nBeing able to decode on the GPU, into memory that could be passed off to a Tensor in whatever framework is being used would be the dream, I don't think there is any interop of that nature working right now. Number of decoder instances per GPU is limited so it's not clear if balancing load btw GPU decoders and CPUs would be needed in say large scale video training.\r\n\r\nAny of these solutions is less than ideal due to the nature of video, having a simple Python interface video / start -> end results in lots of extra memory (you need to decode whole range of the clips into a buffer before using anything). Any scalable video system would be streaming on the fly (issuing frames via callbacks as soon as the stream is far enough along to have re-ordered the frames and synced audio+video+other metadata (sensors, CC, etc).\r\n\r\n", "For standalone usage, decoding on GPU could be ideal but isn't async processing of inputs on CPUs while letting the accelerator busy for training the de-facto? Of course, I am aware of other advanced mechanisms such as CPU offloading, but I think my point is conveyed. ", "Here's a minimal implementation of the helper functions we'd need from PyAV, a lot of which I borrowed from `pytorchvideo`, stripping out the `torch` specific stuff:\r\n\r\n[](https://colab.research.google.com/gist/nateraw/c327cb6ff6b074e6ddc8068d19c0367d/pyav-io.ipynb)\r\n \r\nIt's not too much code...@mariosasko we could probably just maintain these helper fns within the `datasets` library, right? ", "Also wanted to note I added a PR for video classification in `transformers` here, which uses `decord`. It's still open...should we make a decision now to align the libraries we are using between `datasets` and `transformers`? (CC @Narsil )\r\n\r\nhttps://github.com/huggingface/transformers/pull/20151", "Fully agree on at least trying to unite things.\r\n\r\nMaking clear function boundaries to help us change dependency if needed seems like a good idea since there doesn't seem to be a clear winner.\r\n\r\nI also happen to like directly calling ffmpeg. For some reason it was a lot faster than pyav. " ]

[ "I just tried to do the same but changing the `.wav` files to `.mp3` files and that doesn't fix it.", "I don't know if anyone will ever read this but I've tried to upload the same dataset with google colab and the output seems more clarifying. I didn't specify the train/test split so the dataset wasn't fully uploaded (or that is what I understood, might be wrong!!).\r\n\r\nNow, including the `drop_metadata` flag I can load the dataset normally (at least with colab notebook):\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"audiofolder\", data_dir=\"../archive/Dataset\", , drop_metadata=True)\r\n```\r\n\r\nI'll close the issue.", "@uriii3 Hello, I understand correctly that you converted your wav files to mp3?", "Yes but it didn't matter. I don't remember which of them I ended up working with." ]

[ "cc @McPatate maybe you know what's happening ?", "Yes I think I know what is happening. We check in zips for pickles, and the UI must display the pickle jar when a scan has an associated list of imports, even when empty.\r\n~I'll fix ASAP !~", "> I'll fix ASAP !\r\n\r\nActually I'd rather leave it like that for now, as it indicates that we checked for pickles and nothing dangerous appeared :)", "Closing the issue with the typical \"feature not a bug\" " ]

[ "Did you run `huggingface-cli lfs-enable-largefiles` before committing or before adding ? Maybe you can try before adding\r\n\r\nAnyway I'd encourage you to split your data into several TAR archives if possible, this way the dataset can loaded faster using multiprocessing (by giving each process a subset of shards to process)", "@lhoestq \r\nThanks for the help!\r\n> Maybe you can try before adding\r\n\r\nIt did not help\r\n\r\nBut I totally got your point about split into multiple TAR archives. It really helped!" ]

[ "I think this behavior comes from PyArrow:\r\n```python\r\nimport pyarrow as pa\r\nt = pa.table({\"a\": [[0]]})\r\nt.to_pandas().a.values[0]\r\n# array([0])\r\n```\r\n\r\nI believe this has to do with zero-copy: you can get a pandas DataFrame without copying the buffers from arrow, and therefore end up with numpy arrays.", "That's interesting, I guess not much to do here then." ]

[ "Hi ! Interesting :) Can you provide concrete examples of cases where it can be useful ?", "Few example blogs and posts that might help on this - \r\n\r\n1. https://hevodata.com/learn/databricks-delta-tables/\r\n2. https://docs.databricks.com/delta/index.html\r\n\r\nBasically, we are looking at utility of Datasets library with Delta Lake Tables.\r\n", "`datasets` can already read/write from parquet from/to a cloud storage using fsspec, if I understand correctly it's should be possible to load parquet files as delat lake tables no ? :) Or is there someting missing ?", "@lhoestq Per my understanding, delta lake table is a bunch of paruqet files together with the meta to support ACID. For example file 1 contains v0.1 of record A while file 2 contains v0.2 of record A. I am assuming the Hugging face dataset would delegate the read/write delta table to 3rd party lib, maybe pyarrow. Correct me if I was wrong @reichenbch \r\n\r\nAnd I am assuming, people are asking the versioning of Hugging face datasets. But I am assuming Hugging face delegate this function to github and it is not the key requirement for Public Data set. It actually the key function of ML Ops, I am not sure whether hugging face would like expand to that area." ]

[]

[ "Hi ! I think there exist tools to dump and reload an index in your elastic search but I'm not super familiar with it.\r\n\r\nAnyway after reloading an index in elastic search you can call `ds.load_elasticsearch_index` which will connect the index to the dataset without re-indexing" ]

[ "@merveenoyan Do you want different files to be splits or configurations?\r\n\r\nFrom [what you specified in `Readme.md`](https://huggingface.co/datasets/inria-soda/tabular-benchmark/commit/fb4575853772c62a20203bdd6cc0202f5db4ce4e) I hypothesize that you want to have 4 **configs** corresponding to directories: `\"clf_cat\", \"clf_num\", \"reg_cat\", \"reg_num\"`. And inside each config you require to have as many splits as there are `csv` files\r\nso if you run \r\n```python\r\nload_dataset(\"inria-soda/tabular-benchmark\", \"clf_cat\", split=\"compass\")\r\n```\r\nyou will generate the data only from `compass.csv` file.\r\nIn this case, running `load_dataset(\"inria-soda/tabular-benchmark\", \"clf_cat\"`) without split parameter will return `DatasetDict` object with `\"KDDCup09_upselling\", \"cat_compass\", \"cat_covertype\", ... \"road_safety\"` keys (which values are splits - `Dataset` objects)\r\n\r\n**or**\r\ndo you want each file to be a separate config? Like:\r\n```python\r\nload_dataset(\"inria-soda/tabular-benchmark\", \"clf_cat_compass\") # returns DatasetDict with a single \"train\" split\r\n```\r\n**or**\r\nmaybe smth completely different? :smile: \r\n\r\nAnyway, now I have an impression that this is probably rather a matter of automatically inferring configs from repository structure rather than providing parameters in metadata yaml.\r\n", "@polinaeterna I want the latter where you can think of every CSV file as a config, like MNLI from GLUE.", "@merveenoyan @lhoestq I see two solutions to this case. \r\n1. Parse configurations automatically from directories names. That is, if you have data structure like:\r\n```\r\ntabular-benchmark\r\n └─clf_cat_compass\r\n └─compass.csv\r\n └─clf_cat_cat_covertype\r\n └─covertype.csv\r\n ...\r\n └─reg_cat_house_sales\r\n └─house_sales.csv\r\n```\r\nyou'll get \"clf_cat_compass\", \"clf_cat_cat_covertype\", ... \"reg_cat_house_sales\" configurations that would contain **only files from corresponding directories**. \r\n**\\+** this is a requested change and needed in general and would solve other problems, see https://github.com/huggingface/datasets/issues/4578, would also help with https://github.com/huggingface/datasets/pull/5213 which I'm working on currently\r\n**\\+** would allow users to do just `load_dataset(“inria-soda/tabular-benchmark”, “clf_cat_compass”)`, no `data_files` param required\r\n**\\-** in this specific case it would require restructuring of the data - putting each file in a directory named as a config name (to me personally it doesn't seem to be a big deal) \r\n\r\n2. More or less what we discussed before - add support for manually specifying parameters in the metadata. We can add new metadata yaml field (say, `\"custom_configs_info\"`), so that we can provide smth like:\r\n```yaml\r\n---\r\n...\r\ndataset_info:\r\n ... \r\ncustom_configs_info:\r\n- config_name: reg_cat_house_sales\r\n data_files:\r\n - reg_cat/house_sales.csv\r\n- config_name: clf_cat_compass\r\n data_files:\r\n - clf_cat/compass.csv\r\n...\r\n---\r\n```\r\n**\\+** Would be useful not only for tabular data and not only for `data_files` parameter - any packaged dataset’s viewer can be customized to use specific, non-default parameters. @merveenoyan do you maybe have any other examples/use cases in mind where you want to provide any specific parameters to the viewer? \r\n**\\-** I'm not sure here but assume that it might require changes in interaction with the viewer on the hub side - to parse these configurations, as they not default configurations (not in `BUILDER_CONFIGS` list). cc @severo But probably this can be solved on the `datasets` side too.\r\n\r\nOverall, I would start from implementing the first solution since it's related to what I'm doing now and is super useful for `datasets` in general. And then if we agree that having more flexibility in providing parameters to the viewer is required, I can implement the second one. Let me know what you think :) ", "> We can add new metadata yaml field (say, \"custom_configs_info\"), so that we can provide smth like:\r\n\r\nLove it ! Some other ideas to name the \"custom_configs_info\" field: \"configs\", \"parameters\", \"config_args\", \"configurations\"\r\n\r\n> it might require changes in interaction with the viewer on the hub side - to parse these configurations, as they not default configurations (not in BUILDER_CONFIGS list)\r\n\r\nIf we update the `get_dataset_config_names()` function in `datasets` in inspect.py we should be fine - that's what the viewer is using\r\n\r\n> Overall, I would start from implementing the first solution since it's related to what I'm doing now and is super useful for datasets in general. And then if we agree that having more flexibility in providing parameters to the viewer is required, I can implement the second one. Let me know what you think :)\r\n\r\nActually I feel like the second solution includes the first use case you mentioned. If you implement the second solution, then users would just have to add a few lines of YAML and their directories would be considered configurations no ? Maybe there's no need to implement two different logics to do the same thing", "is there any update on this? 🕵🏻", "@merveenoyan I haven't started working on this yet, working on adding configs to packaged datasets instead: https://github.com/huggingface/datasets/pull/5213 because this both would allow you to solve your issue and is a frequently requested feature.\r\n\r\nadding arbitrary parameters to yaml would be my next task i think!", "@merveenoyan ignore my comment above, I'm switching to this task now :D", "I want to be able to create folders in a model.", "Addressed in #5331 " ]

[ "Hi ! It looks like an issue with your python environment, can you make sure you're able to run GET requests to https://huggingface.co using `requests` in python ?", "Thanks for your reply. Does this mean that I have to use the `do_dataset `function and the `requests `function to download the dataset from the company's proxy environment?\r\n\r\n\r\n* Reference: \r\n```bash\r\n### How to load this dataset directly with the [datasets](https://github.com/huggingface/datasets) library\r\n\r\n\r\n* https://huggingface.co/datasets/moyix/debian_csrc\r\n\r\n* from datasets import load_dataset\r\ndataset = load_dataset(\"moyix/debian_csrc\")\r\n\r\n\r\n\r\n### Or just clone the dataset repo\r\n\r\n\r\ngit lfs install\r\ngit clone https://huggingface.co/datasets/moyix/debian_csrc\r\n# if you want to clone without large files – just their pointers\r\n# prepend your git clone with the following env var:\r\nGIT_LFS_SKIP_SMUDGE=1\r\n```", "You can use `requests` to see if downloading a file from the Hugging Face Hub works. If so, then `datasets` should work as well. If not, then you have to find another way using an internet connection that works", "I resolved this issue by applying to \"unblock websites\" at https://huggingface.com in a corporate network environment with a firewall. \r\n", "> Hi ! It looks like an issue with your python environment, can you make sure you're able to run GET requests to https://huggingface.co using `requests` in python ?\r\n\r\nyes，but still not work\r\n\r\n\r\n\r\n", "I read https://github.com/huggingface/datasets/blob/main/src/datasets/load.py, it fail when get the dataset metadata, so download_config has not worked.\r\n```python\r\n hf_api = HfApi(config.HF_ENDPOINT)\r\n try:\r\n dataset_info = hf_api.dataset_info(\r\n repo_id=path,\r\n revision=revision,\r\n token=download_config.token,\r\n timeout=100.0,\r\n )\r\n except Exception as e: # noqa catch any exception of hf_hub and consider that the dataset doesn't exist\r\n if isinstance(\r\n e,\r\n (\r\n OfflineModeIsEnabled,\r\n requests.exceptions.ConnectTimeout,\r\n requests.exceptions.ConnectionError,\r\n ),\r\n ):\r\n raise ConnectionError(f\"Couldn't reach '{path}' on the Hub ({type(e).__name__})\")\r\n```\r\nI configure the huggingface_hub api, use configure_http_backend\r\n```python\r\nfrom huggingface_hub import configure_http_backend\r\ndef backend_factory() -> requests.Session:\r\n session = requests.Session()\r\n session.proxies = proxy\r\n session.verify = False\r\n return session\r\n\r\nconfigure_http_backend(backend_factory=backend_factory)\r\n```\r\nIt works.", "Even tough it does not look like a certificate error in the error message, I had the same error and adding following lines to my code solved my problem.\r\n\r\nimport os\r\nos.environ['CURL_CA_BUNDLE'] = ''", "@kuikuikuizzZ Could you please explain where the configuration code is added?", "> Even tough it does not look like a certificate error in the error message, I had the same error and adding following lines to my code solved my problem.\r\n> \r\n> import os os.environ['CURL_CA_BUNDLE'] = ''\r\n\r\nWorked for as well!\r\nI faced the issue while submitting jobs through SLURM.", "> Even tough it does not look like a certificate error in the error message, I had the same error and adding following lines to my code solved my problem.\r\n> \r\n> import os os.environ['CURL_CA_BUNDLE'] = ''\r\n\r\ndoesn't work , what does this code mean?", "If you're working on a cluster, may be that they disabled remote connections for security purposes, you will have to download the files on your local machine and then transfer them to your cluster through scp or some other transfer protocol. I know you've probably resolved the issue, but that is for anyone in the future who might stumble across this thread and needs help because I struggled with that even after reading this thread.", "> Even tough it does not look like a certificate error in the error message, I had the same error and adding following lines to my code solved my problem.\r\n> \r\n> import os os.environ['CURL_CA_BUNDLE'] = ''\r\n\r\nIf this not work, try this:\r\n```bash\r\nexport http_proxy=\"http://127.0.0.1:10810\"\r\nexport https_proxy=\"http://127.0.0.1:10810\"\r\ngit config --global http.proxy http://127.0.0.1:10810\r\ngit config --global https.proxy http://127.0.0.1:10810\r\n\r\njupyter notebook\r\n```\r\n\r\nset your proxy env first, then start notebook **in this session**\r\n", "> If you're working on a cluster, may be that they disabled remote connections for security purposes, you will have to download the files on your local machine and then transfer them to your cluster through scp or some other transfer protocol. I know you've probably resolved the issue, but that is for anyone in the future who might stumble across this thread and needs help because I struggled with that even after reading this thread.\r\n\r\nThank you buddy!" ]

[ "Actually upon closer inspection, it is documented in the code that this behavior is intentional, so I'll close this." ]

[ "#self-assign", "@lhoestq can you close this issue as part of the recent #5205 merge? Thanks 🤗 ", "Thank you :)" ]

[ "Fixed by: https://huggingface.co/datasets/paws/discussions/1" ]

[ "Thanks for catching this, I can go through the docs and replace the links to their corresponding datasets on the Hub!" ]

[ "Also unrelated but still: https://huggingface.co/docs/datasets/image_dataset#generate-the-dataset\r\n```If your loading script passed the test, you should now have a dataset_infos.json file in your dataset folder.```\r\nIt's not the case anymore as it's now in the readme.md, it was confusing to me", "And here is my data loader script: https://huggingface.co/datasets/corentinm7/MyoQuant-SDH-Data/blob/main/SDH_16k.py\r\nI have one file archive to download that contains the images for all splits and one `metadata.jsonl` to download that contains the informations about what image goes into what split.", "Hi @lambda-science! It seems that your repo is recognized as a packaged module [ImageFolder](https://huggingface.co/docs/datasets/main/en/image_dataset#imagefolder), not as a dataset with the custom loading script, because loader looks for a script that has the same name as the dataset repo. So please try to rename your script to `MyoQuant-SDH-Data.py`, this should help.", "> Hi @lambda-science! It seems that your repo is recognized as a packaged module [ImageFolder](https://huggingface.co/docs/datasets/main/en/image_dataset#imagefolder), not as a dataset with the custom loading script, because loader looks for a script that has the same name as the dataset repo. So please try to rename your script to `MyoQuant-SDH-Data.py`, this should help.\r\n\r\nHi !\r\n\r\nThank you for your answer. That was... embarrassingly easy, sorry for this issue, everything is fixed now ! \r\n\r\nHave a nice day ! :)", "@lambda-science that's not embarrassing at all! it's actually not clear from the documentation that the script should have the same name, so thank you for the issue, we'll add this information to the docs :) " ]

[ "Hi! Yes, this is expected behavior - we do this as a security measure to not leak local paths (this info would be useless on other users' machines anyways) and only push audio bytes. \r\n" ]

[ "I have to admit I'm not a fan of this idea, as this would result in a non-consistent behavior between tabular and non-tabular datasets, which is confusing if done without the context you provided. Instead, we could consider returning a `Dataset` object rather than `DatasetDict` if there is only one split in the generated dataset. But then again, I think this lib is a bit too old to make such changes. @lhoestq @albertvillanova WDYT?\r\n\r\n", "We can brainstorm here to see how we could make it happen ? And then depending on the options we see if it's a change we can do.\r\n\r\nI'm starting with a first reasoning\r\n\r\nCurrently not passing `split=` in `load_dataset` means \"return a dict with each split\".\r\n\r\nNow what would happen if a dataset has no split ? Ideally it should return one Dataset. And passing `split=` would have no sense. So depending on the dataset content, not passing `split=` should return a dict or a Dataset. In particular, those two cases should work:\r\n```python\r\n# case 1: dataset without split\r\nds = load_dataset(\"dataset_without_split\")\r\nds[0], ds[\"column_name\"], list(ds) # we want this\r\n\r\n# case 2: dataset with splits\r\nds = load_dataset(\"dataset_with_splits\")\r\nds[\"train\"] # this works and can't be changed\r\nds = load_dataset(\"dataset_with_splits\", split=\"train\")\r\nds[0], ds[\"column_name\"], list(ds) # this works and can't be changed\r\n```\r\n\r\nI can see several ideas:\r\n1. allowing `load_dataset` to return a different object based on the dataset content - either a Dataset or a DatasetDict\r\n - we can update `get_dataset_split_names` to return None or a list if users want to know in advance what object will be returned. They can also use `isinstance` _a posteriori_\r\n - but in this case we expect users to be careful when loading datasets and always to extra steps to check if they got a Dataset or DatasetDict\r\n2. merge Dataset and DatasetDict objects\r\n - they already share many functions: map, filter, push_to_hub etc.\r\n - we can define `ds[0]` to be the first item of the first split, and consider that the uses accesses rows from the full table of all the splits concatenated\r\n - however there is a collision when doing `ds[\"column_name\"]` or `ds[\"train\"]` that we need to address: the first returns a list, while the other returns a Dataset.\r\n\r\nWhat are your opinions on those two ideas ? Do you have other ideas in mind ?", "I like the first idea more (concatenating splits doesn't seem useful, no?). This is a significant breaking change, so I think we should do a poll (or something similar) to gather more info on the actual \"expected behavior\" and wait for Datasets 3.0 if we decide to implement it.\r\n\r\nPS: @thomwolf also suggested the same thing a while ago (https://github.com/huggingface/datasets/issues/743#issuecomment-746074641).", "I think it's an interesting improvement to the user experience for a case that comes often (no split) so I would definitively support it.\r\n\r\nI would be more in favor of option 2 rather than returning various types of objects from load_dataset and handling carefully the possible collisions indeed", "Related: if a dataset only has one split, we don't show the splits select control in the dataset viewer on the Hub, eg. compare https://huggingface.co/datasets/hf-internal-testing/fixtures_image_utils/viewer/image/test with https://huggingface.co/datasets/glue/viewer/mnli/test.\r\n\r\nSee https://github.com/huggingface/moon-landing/pull/3858 for more details (internal)", "I feel like the second idea is a bit more overkill. \r\n@severo I would say it's a bit irrelevant to the problem we have but is a separate problem @polinaeterna is solving at the moment. 😅 (also discussed on slack)", "OK, sorry for polluting the thread. The relation I saw with the dataset viewer is that from a UX point of view, we hide the concepts of split and configuration whenever possible -> this issue feels like doing the same in the datasets library.", "I would agree that returning different types based on the content of the dataset might be confusing.\r\n\r\nWe can do something similar to what `fetch_*` or `load_*` from `sklearn.datasets` do, which is to have an arg which changes the type of the returned type. For instance, `load_iris` would return a dict, but `load_iris(..., return_X_y=True)` would return a tuple.\r\n\r\nHere we can have a similar arg such as `return_X` which would then only return a single `DataSet` or an array.", "> I feel like the second idea is a bit more overkill.\r\n\r\nOverkill in what sense ?\r\n\r\n> Here we can have a similar arg such as return_X which would then only return a single DataSet or an array.\r\n\r\nRight now one can already pass `split=\"all\"` to get one `Dataset` object with all the data in it (unsplit). We could also have something like `return_all=True` so make the API clearer.\r\n\r\n> I would be more in favor of option 2 rather than returning various types of objects from load_dataset and handling carefully the possible collisions indeed\r\n\r\nI think it would be ok to handle the collision by allowing both `ds[\"train\"]` and `ds[\"column_name\"]` (and maybe adding something like `ds.splits` for those who want to iterate over the splits or add new ones)", "Would it make sense to remove the notion of \"split\" in `load_dataset`? I feel a lof of it comes from the want to have some sort of group of more or less similar dataset. \"train\"/\"test\"/\"validation\" are the traditional ones, but there are some datasets that have much more splits.\r\n\r\nWould it make sense to force `load_dataset` to only load a single `Dataset` object, and fail if it doesn't point to one. And have another method that's like `load_dataset_group_info` that can return a very arbitrary info class (Dict, List whatever), but you need to pass individual infos to `load_dataset` to run anything? Typically I don't think `DatasetDict.map` is really that helpful, but that's my personal opinion. This would help make things more readable (typically knowing if an object is a `Dataset` or a `DatasetDict`)", "> Would it make sense to remove the notion of \"split\" in load_dataset?\r\n\r\nI think we need to keep it - though in practice people can name the splits whatever they want anyway.\r\n\r\n> Would it make sense to force load_dataset to only load a single Dataset object, and fail if it doesn't point to one.\r\n\r\nWe need to keep backward compatibility ideally - in particular the load_dataset + ds[\"train\"] one", "> I think we need to keep it - though in practice people can name the splits whatever they want anyway.\r\n\r\nIt was my understanding that the whole issue was that `load_dataset` returned multiple types of objects.\r\n\r\n> We need to keep backward compatibility ideally - in particular the load_dataset + ds[\"train\"] one\r\n\r\nYeah sorry I meant ideally. One can always start developing `load_dataset_v2` can deprecate the first one and remove it in the longer term.", "> It was my understanding that the whole issue was that load_dataset returned multiple types of objects.\r\n\r\nYes indeed, but we still want to keep a way to load the train/val/test/whatever splits alone ;)", "@thomasw21's solution is good but it will break backwards compatibility. 😅", "Started to experiment with merging Dataset and DatasetDict. My plan is to define the splits of a Dataset in Dataset.info.splits (already exists, but never used). A Dataset would then be the concatenation of its splits if they exist.\r\n\r\nNot sure yet this is the way to go. My plan is to play with it and see and share it with you, so we can see if it makes sense from a UX point of view.", "So just to make sure that I understand the current direction, people will have to be extra careful when handling splits right?\r\nImagine \"potato\" a dataset containing train/validation split:\r\n```\r\nload_dataset(\"potato\") # returns the concatenation of all the splits\r\n```\r\nPreviously the design would force you to choose a split (it would raise otherwise), or manually concat them if you really wanted to play with concatenated splits. Now it would potentially run without raising for a bit of time until you figure out that you've been training on both train and validation split.\r\n\r\nWould it make sense to use a dataset specific default instead of using the concatenation, typically \"potato\" dataset's default would be train?\r\n```\r\nload_dataset(\"potato\") # returns \"train\" split\r\nload_dataset(\"potato\", split=\"train\") # returns \"train\" split\r\nload_dataset(\"potato\", split=\"validation\") # returns \"validation\" split\r\nconcatenate_datasets([load_dataset(\"potato\", split=\"train\"), load_dataset(\"potato\", split=\"validation\")]) # returns concatenation\r\n```", "> load_dataset(\"potato\") # returns \"train\" split\r\n\r\nTo avoid a breaking change we need to be able to do `load_dataset(\"potato\")[\"validation\"]` as well.\r\n\r\nIn that case I'd wonder where the validation split comes from, since the rows of the dataset wouldn't contain the validation split according to your example. That's why I'm more in favor of concatenating.\r\n\r\nA dataset is one table, that optionally has some split info about subsets (e.g. for training an evaluation)\r\n\r\nThis also allows anyone to re-split the dataset the way they want if they're not happy with the default:\r\n\r\n```python\r\nds = load_dataset(\"potato\").train_test_split(test_size=0.2)\r\ntrain_ds = ds[\"train\"]\r\ntest_ds = ds[\"test\"]\r\n```", "Just thinking about this, we could just have `to_dataframe()` as `load_dataset(\"blah\").to_dataframe()` to get the whole dataset, and not change anything else.", "I have a first implementation of option 2 (merging Dataset and DatasetDict) in this PR: https://github.com/huggingface/datasets/pull/5301/\r\n\r\nFeel free to play with it if you're interested, and let me know what you think. In this PR, a dataset is one table that optionally has some split info about subsets.", "@adrinjalali we already have [to_pandas](https://huggingface.co/docs/datasets/package_reference/main_classes#datasets.Dataset.to_pandas) AFAIK that essentially does the same thing (for a dataset, not for a dataset dict), I was wondering if it makes sense to have this as I don't know portion of people who load non-tabular datasets into dataframes. @lhoestq I saw your PR and it will break a lot of things imo, WDYT of this option? ", "> we already have [to_pandas](https://huggingface.co/docs/datasets/package_reference/main_classes#datasets.Dataset.to_pandas) AFAIK that essentially does the same thing (for a dataset, not for a dataset dict)\r\n\r\nyes correct :)\r\n\r\n> I saw your PR and it will break a lot of things imo\r\n\r\nDo you have concrete examples you can share ?\r\n\r\n> WDYT of this option?\r\n\r\nThe to_dataframe option ? I think it not enough, since you'd still get a `DatasetDict({\"train\": Dataset()})` if you load a dataset with no splits (e.g. one CSV), and this doesn't really make sense.\r\n\r\nNote that in the PR I opened you can do\r\n```python\r\nds = load_dataset(\"dataset_with_just_one_csv\") # Dataset type\r\ndf = load_dataset(\"dataset_with_just_one_csv\").to_pandas() # DataFrame type\r\n```", "@lhoestq no I think @adrinjalali and I meant when user calls `to_dataframe` if there's only train split in `DatasetDict` we could directly load that into dataframe. This might cause a confusion given there's to_pandas but I think it's more intuitive and least breaking change. (given people -who use `datasets` for tabular workflows- will eventually call `to_pandas` anyway) ", "So in that case it would be fine to still end up with a dataset dict with a \"train\" split ?", "yeah what I mean is this:\r\n\r\n```py\r\ndataset = load_dataset(\"blah\")\r\n\r\n# deal with a split of the dataset\r\ntrain = dataset[\"train\"]\r\ntrain_df = dataset[\"train\"].to_dataframe()\r\n\r\n# deal with the whole dataset\r\ndataset_df = dataset.to_dataframe()\r\n```\r\n\r\nSo we do two things to improve tabular experience:\r\n- allow datasets to have a single split\r\n- add `to_dataframe` to the root dict level so that users can simply call `df = load_dataset(\"blah\").to_dataframe()` and have it in their `pandas.DataFrame` object.", "Ok ! Note that we already have `Dataset.to_pandas()` so for consistency I'd call it `DatasetDict.to_pandas()` as well, does it sound good to you ? This is something we can add pretty easily", "yeah that sounds perfect @lhoestq !", "> So just to make sure that I understand the current direction, people will have to be extra careful when handling splits right?\r\n\r\nWe can raise an error if someone does `load_dataset(...)[0]` if the dataset is made of several splits, and return the first example if there's one or zero splits (i.e. when it's not ambiguous). Had this idea from the dicussions in #5312 WDYT @thomasw21 ?", "> We can raise an error if someone does load_dataset(...)[0] if the dataset is made of several splits,\r\n\r\nBut then how is that different to have the distinction between DatasetDict and Dataset then? Is it just that \"default behaviour when there are no splits or single split, it returns directly the split when there's no ambiguity\".\r\n\r\nAlso I was wondering how the concatenation could have heavy impacts when running mapping functions/filtering in batch? Typically can batch be somehow mixed?", "> But then how is that different to have the distinction between DatasetDict and Dataset then?\r\n\r\nBecause it doesn't make sense to be able to do `example = ds[0]` or `examples = list(ds)` on a class named `DatasetDict` of type `Dict[str, Dataset]`.\r\n\r\n> Also I was wondering how the concatenation could have heavy impacts when running mapping functions/filtering in batch? Typically can batch be somehow mixed?\r\n\r\nNo, we run each function on each split separated", "> Because it doesn't make sense to be able to do example = ds[0] or examples = list(ds) on a class named DatasetDict of type Dict[str, Dataset].\r\n\r\nHum but you're still going to raise an exception in both those cases with your current change no? (actually list(ds) would return the name of the splits no?)\r\n\r\n> No, we run each function on each split separated\r\n\r\nNice!" ]

[ "Hi! The first `Dataset.from_sql` call also outputs the \"ImportError: Using URI string without sqlalchemy installed.\" message, but you also get \"During handling of the above exception another exception occurred: ...\" after which the ValueError is printed. I agree that this behavior makes it easy to miss the original error. \r\n\r\nI think we can improve this by not throwing the writer's ValueError if the error from a dataset script is already being handled to make debugging easier. @lhoestq @albertvillanova wdyt?", "Yup ! Alternatively the error can be raised in sql.py before generating the examples ? In `_info` for example", "yea @lhoestq that would probably be good. The 2nd error is useless if the 1st error is the real reason it failed. " ]

[]

[]

[ "Yea this would be great! We would need an object detection tutorial notebook too if it doesn't already exist there. ", "There is one: https://huggingface.co/docs/datasets/object_detection.\r\n\r\nI will start the work. " ]

[ "Sure this sounds great! Would this be pure torchvision, albumentations, or something else?", "I am considering `torchvision` and `albumentations`. Also [works with TensorFlow](https://github.com/deep-diver/segformer-tf-transformers/blob/main/notebooks/TFSegFormer_Finetune.ipynb). \r\n\r\nI am assigning the issue to myself then. " ]

[ "The beam utilities allow to prepare a dataset as parquet in your cloud storage. From my perspective this CLI is not super easy to use, but we've been working on a new python API to prepare a dataset in your cloud storage:\r\n```python\r\nfrom datasets import load_dataset_builder\r\n\r\nbuilder = load_dataset_builder(\"c4\", \"en\")\r\nbuilder.download_and_prepapre(\"s3://my-bucket/c4\", file_format=\"parquet\")\r\n```\r\n\r\nAnd to use Beam you can do:\r\n```python\r\nbeam_runner = ... # one of \"SparkRunner\", \"DataFlowRunner\", \"DirectRunner\", etc.\r\nbeam_options = ...\r\n\r\nbuilder.download_and_prepapre(\r\n \"s3://my-bucket/c4\",\r\n file_format=\"parquet\",\r\n beam_runner=beam_runner,\r\n beam_options=beam_options\r\n)\r\n```\r\n\r\nThough Beam can be used ONLY if there is a dataset script based on the `BeamBasedBuilder` right now - it doesn't work on an arbitrary dataset (see [wikipedia.py](https://huggingface.co/datasets/wikipedia/blob/main/wikipedia.py) for example).", "Thanks! \r\n\r\nWould be nice to have something similar for creating large image datasets. " ]

[ "Cc: @lhoestq ", "You can end up with a list instead of a tensor if all the tensors inside the list can't be stacked together - can you make sure all your inputs are tensors with the same shape ?", "Is there an easy way to ensure it?", "You can make sure your `tokenize` function always return tensors of the same shape", "I modified my `tokenize()` function to be like so:\r\n\r\n```py\r\ndef tokenize(batch):\r\n return tokenizer(batch[\"text\"], padding=\"longest\")\r\n```\r\n\r\nso that the padding always happens w.r.t to the length of the longest sequence in a batch. The issue still persists. Is there any other way? ", "tbh I though your first implementation was fine\r\n```python\r\ndef tokenize(batch):\r\n return tokenizer(batch[\"text\"], padding=True, truncation=True)\r\n```\r\n\r\nMaybe you can try to see what the erroring data looks like by adding a try/except in `get_test_accuracy` ?", "This is what I got. \r\n\r\nFor the non-erroring data, it looks like (without the labels):\r\n\r\n```\r\ntensor([[ 101, 10047, 3110, ..., 0, 0, 0],\r\n [ 101, 1045, 2514, ..., 0, 0, 0],\r\n [ 101, 1045, 2514, ..., 0, 0, 0],\r\n ...,\r\n [ 101, 1045, 2005, ..., 0, 0, 0],\r\n [ 101, 1045, 2572, ..., 0, 0, 0],\r\n [ 101, 10047, 7481, ..., 0, 0, 0]]) 128\r\ntensor([[1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n ...,\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0]]) 128\r\n```\r\n\r\nFor the erroring part:\r\n\r\n```\r\n[tensor([ 101, 1045, 2064, 2102, 2393, 3110, 2066, 2242, 6355, 3047, 2004, 2574,\r\n 2004, 1996, 8629, 2357, 2125, 4299, 1045, 2071, 2424, 2009, 2006, 7858,\r\n 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0]), tensor([ 101, 10047, 5458, 1997, 3110, 11654, 1998, 11055, 102, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0]), tensor([ 101, 1045, 2074, 2064, 2102, 6073, 1996, 3110, 2008, 2026,\r\n 14982, 2000, 5587, 2203, 16650, 29563, 2030, 2569, 4506, 2052,\r\n 2191, 1037, 2738, 11552, 2208, 17044, 14540, 2100, 3375, 102,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0]),\r\n...\r\n\r\n[tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]), tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),\r\n...\r\n```\r\n\r\nI also tried investigating the shapes of the individual entries within a `batch` without the labels:\r\n\r\n```py\r\ndef get_test_accuracy(model):\r\n def fn(batch): \r\n try:\r\n inputs = {k:v.to(device) for k,v in batch.items() \r\n if k in tokenizer.model_input_names}\r\n with torch.no_grad():\r\n output = model(**inputs)\r\n pred_label = torch.argmax(output.logits, axis=-1)\r\n return {\"predicted_label\": pred_label.cpu().numpy()}\r\n except:\r\n for k in batch:\r\n if k != \"label\":\r\n for i in range(len(batch[k])):\r\n print(batch[k][i].shape)\r\n return fn\r\n```\r\n\r\nThey are:\r\n\r\n```\r\n...\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([66])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\ntorch.Size([69])\r\n```\r\n\r\nThere are differing shapes. I understand if I set `batch_size=None` in `emotions_encoded = emotions.map(tokenize, batched=True)` the problem should be fixed as the whole dataset would be treated as a single batch. But is there a way to do that in batches? ", "If you use the same batch_size for your two maps, you should get the exact same batches - therefore all containing the same shapes", "Oh I see. Thanks. Closing this issue. " ]

[ "In the dumps page of the wiki (https://dumps.wikimedia.org/zhwiki/), I found the following dumps:\r\n```\r\nIndex of /zhwiki/\r\n[../](https://dumps.wikimedia.org/)\r\n[20220701/](https://dumps.wikimedia.org/zhwiki/20220701/) 21-Aug-2022 01:48 -\r\n[20220720/](https://dumps.wikimedia.org/zhwiki/20220720/) 02-Sep-2022 01:48 -\r\n[20220801/](https://dumps.wikimedia.org/zhwiki/20220801/) 21-Sep-2022 01:44 -\r\n[20220820/](https://dumps.wikimedia.org/zhwiki/20220820/) 01-Oct-2022 09:39 -\r\n[20220901/](https://dumps.wikimedia.org/zhwiki/20220901/) 20-Oct-2022 09:44 -\r\n[20220920/](https://dumps.wikimedia.org/zhwiki/20220920/) 23-Sep-2022 12:06 -\r\n[20221001/](https://dumps.wikimedia.org/zhwiki/20221001/) 04-Oct-2022 15:10 -\r\n[20221020/](https://dumps.wikimedia.org/zhwiki/20221020/) 01-Nov-2022 03:15 -\r\n[latest/](https://dumps.wikimedia.org/zhwiki/latest/) 01-Nov-2022 03:15 -\r\n```\r\n\r\nMaybe the older dumps are not available which caused the downloading failure? \r\n\r\nHowever, when I changed to the newer version:\r\n```\r\ndata = load_dataset('wikipedia', '20220701.zh', beam_runner='DirectRunner')\r\n```\r\n\r\nit shows:\r\n```\r\nValueError: BuilderConfig 20220701.zh not found. Available: ['20220301.aa', '20220301.ab', '20220301.ace', '20220301.ady', '20220301.af', '20220301.ak', '20220301.als', '20220301.am', '20220301.an', '20220301.ang', '20220301.ar', '20220301.arc', '20220301.arz', '20220301.as', '20220301.ast', '20220301.atj', '20220301.av', '20220301.ay', '20220301.az', '20220301.azb', '20220301.ba', '20220301.bar', '20220301.bat-smg', '20220301.bcl', '20220301.be', '20220301.be-x-old', '20220301.bg', '20220301.bh', '20220301.bi', '20220301.bjn', '20220301.bm', '20220301.bn', '20220301.bo', '20220301.bpy', '20220301.br', '20220301.bs', '20220301.bug', '20220301.bxr', '20220301.ca', '20220301.cbk-zam', '20220301.cdo', '20220301.ce', '20220301.ceb', '20220301.ch', '20220301.cho', '20220301.chr', '20220301.chy', '20220301.ckb', '20220301.co', '20220301.cr', '20220301.crh', '20220301.cs', '20220301.csb', '20220301.cu', '20220301.cv', '20220301.cy', '20220301.da', '20220301.de', '20220301.din', '20220301.diq', '20220301.dsb', '20220301.dty', '20220301.dv', '20220301.dz', '20220301.ee', '20220301.el', '20220301.eml', '20220301.en', '20220301.eo', '20220301.es', '20220301.et', '20220301.eu', '20220301.ext', '20220301.fa', '20220301.ff', '20220301.fi', '20220301.fiu-vro', '20220301.fj', '20220301.fo', '20220301.fr', '20220301.frp', '20220301.frr', '20220301.fur', '20220301.fy', '20220301.ga', '20220301.gag', '20220301.gan', '20220301.gd', '20220301.gl', '20220301.glk', '20220301.gn', '20220301.gom', '20220301.gor', '20220301.got', '20220301.gu', '20220301.gv', '20220301.ha', '20220301.hak', '20220301.haw', '20220301.he', '20220301.hi', '20220301.hif', '20220301.ho', '20220301.hr', '20220301.hsb', '20220301.ht', '20220301.hu', '20220301.hy', '20220301.ia', '20220301.id', '20220301.ie', '20220301.ig', '20220301.ii', '20220301.ik', '20220301.ilo', '20220301.inh', '20220301.io', '20220301.is', '20220301.it', '20220301.iu', '20220301.ja', '20220301.jam', '20220301.jbo', '20220301.jv', '20220301.ka', '20220301.kaa', '20220301.kab', '20220301.kbd', '20220301.kbp', '20220301.kg', '20220301.ki', '20220301.kj', '20220301.kk', '20220301.kl', '20220301.km', '20220301.kn', '20220301.ko', '20220301.koi', '20220301.krc', '20220301.ks', '20220301.ksh', '20220301.ku', '20220301.kv', '20220301.kw', '20220301.ky', '20220301.la', '20220301.lad', '20220301.lb', '20220301.lbe', '20220301.lez', '20220301.lfn', '20220301.lg', '20220301.li', '20220301.lij', '20220301.lmo', '20220301.ln', '20220301.lo', '20220301.lrc', '20220301.lt', '20220301.ltg', '20220301.lv', '20220301.mai', '20220301.map-bms', '20220301.mdf', '20220301.mg', '20220301.mh', '20220301.mhr', '20220301.mi', '20220301.min', '20220301.mk', '20220301.ml', '20220301.mn', '20220301.mr', '20220301.mrj', '20220301.ms', '20220301.mt', '20220301.mus', '20220301.mwl', '20220301.my', '20220301.myv', '20220301.mzn', '20220301.na', '20220301.nah', '20220301.nap', '20220301.nds', '20220301.nds-nl', '20220301.ne', '20220301.new', '20220301.ng', '20220301.nl', '20220301.nn', '20220301.no', '20220301.nov', '20220301.nrm', '20220301.nso', '20220301.nv', '20220301.ny', '20220301.oc', '20220301.olo', '20220301.om', '20220301.or', '20220301.os', '20220301.pa', '20220301.pag', '20220301.pam', '20220301.pap', '20220301.pcd', '20220301.pdc', '20220301.pfl', '20220301.pi', '20220301.pih', '20220301.pl', '20220301.pms', '20220301.pnb', '20220301.pnt', '20220301.ps', '20220301.pt', '20220301.qu', '20220301.rm', '20220301.rmy', '20220301.rn', '20220301.ro', '20220301.roa-rup', '20220301.roa-tara', '20220301.ru', '20220301.rue', '20220301.rw', '20220301.sa', '20220301.sah', '20220301.sat', '20220301.sc', '20220301.scn', '20220301.sco', '20220301.sd', '20220301.se', '20220301.sg', '20220301.sh', '20220301.si', '20220301.simple', '20220301.sk', '20220301.sl', '20220301.sm', '20220301.sn', '20220301.so', '20220301.sq', '20220301.sr', '20220301.srn', '20220301.ss', '20220301.st', '20220301.stq', '20220301.su', '20220301.sv', '20220301.sw', '20220301.szl', '20220301.ta', '20220301.tcy', '20220301.te', '20220301.tet', '20220301.tg', '20220301.th', '20220301.ti', '20220301.tk', '20220301.tl', '20220301.tn', '20220301.to', '20220301.tpi', '20220301.tr', '20220301.ts', '20220301.tt', '20220301.tum', '20220301.tw', '20220301.ty', '20220301.tyv', '20220301.udm', '20220301.ug', '20220301.uk', '20220301.ur', '20220301.uz', '20220301.ve', '20220301.vec', '20220301.vep', '20220301.vi', '20220301.vls', '20220301.vo', '20220301.wa', '20220301.war', '20220301.wo', '20220301.wuu', '20220301.xal', '20220301.xh', '20220301.xmf', '20220301.yi', '20220301.yo', '20220301.za', '20220301.zea', '20220301.zh', '20220301.zh-classical', '20220301.zh-min-nan', '20220301.zh-yue', '20220301.zu']\r\n```\r\n\r\nSo I guess adding the latest dumps versions to the `BuilderConfig` may solve the problem? But how to add it?", "Hi, @beyondguo, thanks for reporting.\r\n\r\nYou have all the information in the dataset card: https://huggingface.co/datasets/wikipedia\r\n\r\n> Then, you can load any subset of Wikipedia per language and per date this way:\r\n> ```python\r\n> from datasets import load_dataset\r\n> \r\n> load_dataset(\"wikipedia\", language=\"sw\", date=\"20220120\", beam_runner=...) \r\n> ```\r\n> where you can pass as beam_runner any Apache Beam supported runner for (distributed) data processing (see [here](https://beam.apache.org/documentation/runners/capability-matrix/)). Pass \"DirectRunner\" to run it on your machine.\r\n> \r\n> You can find the full list of languages and dates [here](https://dumps.wikimedia.org/backup-index.html).\r\n\r\nNote that you have to pass the language and date as keyword arguments, and the available dates depend on the language and can be found on Wikimedia website.", "Also:\r\n> Some subsets of Wikipedia have already been processed by HuggingFace, and you can load them just with:\r\n> ```python\r\n> load_dataset(\"wikipedia\", \"20220301.en\")\r\n> ```\r\n> The list of pre-processed subsets is:\r\n> - \"20220301.de\"\r\n> - \"20220301.en\"\r\n> - \"20220301.fr\"\r\n> - \"20220301.frr\"\r\n> - \"20220301.it\"\r\n> - \"20220301.simple\"" ]

[ "It looks like an issue with `gcsfs`, are you able to instantiate a `GCSFileSystem` manually ?", "closing since it was probably due to gcsfs" ]

[]

[]

[]

[]

[ "Thanks for reporting, @Rijgersberg.\r\n\r\nWe were waiting for the release of `dill` 0.3.6, that happened 2 days ago (24 Oct 2022): https://github.com/uqfoundation/dill/releases/tag/dill-0.3.6\r\n- See comment: https://github.com/huggingface/datasets/pull/4397#discussion_r880629543\r\n\r\nAlso `multiprocess` 0.70.14 was released 2 days ago: https://github.com/uqfoundation/multiprocess/releases/tag/multiprocess-0.70.14\r\n\r\nWe are addressing this issue to align dependencies.", "In your specific setup, I guess the compatible configuration is with `multiprocess` 0.70.13 (instead of 0.70.14).", "@Rijgersberg this issue is fixed. It will be available in our next `datasets` release.", "Thanks!", "> @Rijgersberg this issue is fixed. It will be available in our next `datasets` release.\n\nAny chance you have a eta? ", "@StefanSamba we are disussing about making a release early this week.", "@Rijgersberg, please also that you can make `pip-compile` work by using the backtracking resolver (instead of the legacy one): https://pip-tools.readthedocs.io/en/latest/#a-note-on-resolvers\r\n```\r\npip-compile --resolver=backtracking requirements.in\r\n```\r\nThis resolver will automatically use `multiprocess` 0.70.13 version. " ]

[ "Addressed in https://github.com/huggingface/datasets/pull/5191 (torch.Tensor objects now produce deterministic hashes)" ]

[ "Also cc @anton-l ", "BTW the exact same holds true for the audio folder", "I'm fine with adding a new column with the file name personally. Not sure how breaking this is though", "@patrickvonplaten do you mean just filename or full relative path inside the repo?\r\nI think it shouldn't be breaking, at least I cannot come up with any case where it is. Maybe @mariosasko can?\r\n\r\nalso I think that the problem here and in general is that Image/AudioFolder has default configuration which implies automatic label creation if there is not metadata file. It can be changed when you load the dataset with `load_dataset` but not on it's Hub page. \r\n\r\n", "> also I think that the problem here and in general Image/AudioFolder has default configuration which implies automatic label creation if there is not metadata file\r\n\r\nYea I agree it's often the wrong default. We can also imagine adding the builder's parameters as YAML in the repo.", "@lhoestq yes I also got the idea of some YAML config! not sure of what priority it is though.", "but it would actually also solve this issue: https://github.com/huggingface/datasets/issues/5153", "I meant just the file name (no path) that would already be super helpful IMO :-) (maybe dir+filename if there are dirs in the folder)", "@patrickvonplaten one more time, to be sure I understand you.\r\nFor example, we have data structure like this:\r\n```\r\n├─ data/\r\n│ └─ subdir/\r\n│ └── cats/\r\n│ ├── 0.jpg\r\n│ ├── 1.jpg\r\n│ └── 2.jpg\r\n│ └── dogs/\r\n│ ├── 0.jpg\r\n│ ├── 1.jpg\r\n│ └── 2.jpg\r\n└── another_subdir/\r\n ├── 10.jpg\r\n ├── 11.jpg\r\n └── 12.jpg\r\n```\r\nIs it okay to provide `\"data/subdir/cats/0.jpg\"`, `\"data/subdir/dogs/0.jpg\"`, `\"data/another_subdir/10.jpg\"`?\r\nI think providing just filenames might be confusing if they are not unique, as in this example. ", "Yes I think the relative path as you proposed makes a lot of sense :-) " ]

[ "There are currently 402 datasets with deprecated \"languages\" or \"licenses\".", "hey @albertvillanova ,i would love to work on this issue if you like.", "Hi @ayushthe1, thanks for your offer.\r\n\r\nBut as you can see, I self-assigned this issue.\r\n\r\nI have already fixed 200 out of the 402 datasets. My script is still running and fixing the rest.\r\n\r\nFor example: https://huggingface.co/datasets/fhamborg/news_sentiment_newsmtsc/discussions/2/files", "Thanks for your time. Will try next time. 😇", "@ayushthe1 feel free to take one of the non-assigned open issues: https://github.com/huggingface/datasets/issues", "This is done." ]

[ "Hi ! Maybe it's possible to have a consistent hash for a function defined in `__main__` and a function define in a notebook.\r\n\r\nHowever for functions imported from another location, pickle uses the location to identify the code, so in that case we can't do much I believe.\r\n\r\nWould it be ok for you if we only try to do this for functions in `__main__` / jupyter ?\r\n\r\nIf you'd like to contribute, you can read this part of the code and let me know if you have questions:\r\n\r\nhttps://github.com/huggingface/datasets/blob/7feeb5648a63b6135a8259dedc3b1e19185ee4c7/src/datasets/utils/py_utils.py#L617-L643\r\n\r\nI think the key here would be to also ignore the \"co_filename\" of functions defined in `__main__`", "Seems like a good solution, I will start a PR and see if I understood the changes needed. Thanks!" ]

[ "Hi ! From the `adlfs` docs, there are two filesystems you can use:\r\n> To use the Gen1 filesystem:\r\n> - known_implementations[‘adl’] = {‘class’: ‘adlfs.AzureDatalakeFileSystem’}\r\n> \r\n> To use the Gen2 filesystem:\r\n> - known_implementations[‘abfs’] = {‘class’: ‘adlfs.AzureBlobFileSystem’}\r\n\r\nIf I'm not mistaken you're using the second one - so you should use `abfs://` instead of `adl://`, and also run this at the beginning of your script:\r\n```python\r\nfrom fsspec.registry import known_implementations\r\nknown_implementations['abfs'] = {'class': 'adlfs.AzureDatalakeFileSystem'}\r\n```\r\n\r\n", "Thank you @lhoestq . Great call.\r\nUsing the default class from `known_implementations` dict solved my problem\r\n```\r\nknown_implementations[‘abfs’] = {‘class’: ‘adlfs.AzureBlobFileSystem’}\r\n```\r\nI'm closing this issue.", "> Thank you @lhoestq . Great call. Using the default class from `known_implementations` dict solved my problem\r\n> \r\n> ```\r\n> known_implementations[‘abfs’] = {‘class’: ‘adlfs.AzureBlobFileSystem’}\r\n> ```\r\n> \r\n> I'm closing this issue.\r\n\r\nHi so here `Saving serialized datasets\r\n\r\nAfter you have processed your dataset, you can save it to your cloud storage with [Dataset.save_to_disk()](https://huggingface.co/docs/datasets/v2.17.0/en/package_reference/main_classes#datasets.Dataset.save_to_disk):` what is the encoded dataset I have failed to save it ", "Uploading failed ? Did you get an error message ?" ]

[ "Makes sense! For the last structure, we could count the path segments (delimited by \"/\" for URLs and `os.sep` for local paths) to ensure all inferred labels are on the same level. Otherwise, I think it's safe to assume they are meaningless and ignore them.\r\n" ]

[]

[ "also asked in https://discuss.huggingface.co/t/create-multiple-dataset-configs-with-push-to-hub-method/25480" ]

[ "Hi! I can't reproduce the error following these steps. Can you please provide a reproducible example?", "I faced the same issue:\r\n\r\n### Repro\r\n```\r\n!pip install datasets==2.6.1\r\nimport datasets as Dataset\r\ndataset = Dataset.from_pandas(dataframe)\r\ndataset.save_to_disk(local)\r\n\r\n!pip install datasets==2.5.2\r\nimport datasets as Dataset\r\ndataset = Dataset.load_from_disk(local)\r\n```\r\n\r\n", "@Lokiiiiii And what are the contents of the \"dataframe\" in your example?", "I bumped into the issue too. @Lokiiiiii thanks for steps. I \"solved\" if for now by `pip install datasets>=2.6.1` everywhere.", "Hi all, \r\nI experienced the same issue. \r\nPlease note that the pull request is related to the IMDB example provided in the doc, and is a fix for that, in that context, to make sure that people can follow the doc example and have a working system. \r\nIt does not provide a fix for Datasets itself. ", "im getting the same error.\r\n- using the base AWS HF container that uses a datasets <2.\r\n- updating the AWS HF container to use dataset 2.4\r\n", "Same here, running on our SageMaker pipelines. It's only happening for some but not all of our saved Datasets.", "I am also receiving this error on Sagemaker but not locally, I have noticed that this occurs when the `.dataset/` folder does not contain a single file like:\r\n\r\n`dataset.arrow`\r\n\r\nbut instead contains multiple files like:\r\n\r\n`data-00000-of-00002.arrow`\r\n`data-00001-of-00002.arrow`\r\n\r\nI think that it may have something to do with this recent PR that updated the behaviour of `dataset.save_to_disk` by introducing sharding: https://github.com/huggingface/datasets/pull/5268\r\n\r\nFor now I can get around this by forcing datasets==2.8.0 on machine that creates dataset and in the huggingface instance for training (by running this at the start of training script `os.system(\"pip install datasets==2.8.0\")`)\r\n\r\nTo ensure the dataset is a single shard when saving the dataset locally:\r\n\r\n```python3\r\ndataset.flatten_indices().save_to_disk('path/to/dataset', num_shards=1)\r\n```\r\n\r\n and then manually changing the name afterwards from `path/to/dataset/data-00000-of-00001.arrow` to `path/to/dataset/dataset.arrow` and updating the `path/to/dataset/state.json` to reflect this name change. i.e. by changing `state.json` to this:\r\n\r\n```javascript\r\n{\r\n \"_data_files\": [\r\n {\r\n \"filename\": \"dataset.arrow\"\r\n }\r\n ],\r\n \"_fingerprint\": \"420086f0636f8727\",\r\n \"_format_columns\": null,\r\n \"_format_kwargs\": {},\r\n \"_format_type\": null,\r\n \"_output_all_columns\": false,\r\n \"_split\": null\r\n}\r\n```", "Does anyone know if this has been resolved?", "I have the same issue in datasets version 2.3.2" ]

[ "Hi, @santule.\r\n\r\nWe have transferred all dataset scripts from GitHub to the Hugging Face Hub: https://huggingface.co/datasets\r\n- Concretely, you have \"rvl_cdip\" here: https://huggingface.co/datasets/rvl_cdip\r\n\r\nTo be able to load them, you should update your `datasets` library:\r\n```\r\npip install -U datasets\r\n```", "thank you, it worked" ]

[ "Hi ! In the `transformers` issue the object to not hash is a `Pool` - I think you can instantiate it inside your function instead of passing it as a parameter. It's good practice that your function and all its fn_kwargs are picklable, in case you want to parallelize `map` using `num_proc>1`\r\n\r\nFor the other case `def fn(example, verbose=False):` however, I agree it would be nice to let the user specify that \"verbose\" needs to be ignored.\r\n\r\nDo you think providing a decorator could help ? Maybe\r\n```python\r\[email protected](ignore_kwargs=[\"verbose\"])\r\ndef func(example, verbose=False):\r\n ...\r\n```", "Hi @lhoestq! Thanks for your response.\r\n\r\nA `Pool` shouldn't be instantiated within the function, because there's a huge overhead in doing so. The main idea is that the same `Pool` should be used across all function calls. Parallel `map` is not helpful/desired in that specific scenario, because the heavy parallel computation is done by another lib (`pyctcdecode`, called within `transformer`'s model inference code).\r\n\r\nBut yes, it makes sense to be able to leverage parallel processing by just doing `num_proc>1` when possible.\r\n\r\nYour decorator suggestions seems like a pretty clean API to me. I didn't find a `datasets.hashing` module though. Would it be created for this specific purpose? Any downsides in just using `datasets.fingerprint`?\r\n\r\nAnd would `datasets.hashing.register` just add some metadata to `func` in your approach (so it could be inspected from `fingerprint_transform`)?\r\n\r\nAnd looking to the `datasets.Dataset` API, `.filter` would also benefited from this.", "> Would it be created for this specific purpose? Any downsides in just using datasets.fingerprint?\r\n\r\nThis can also go in datasets.fingerprint indeed - but maybe datasets.hashing tells more about what the register function does (i.e. register this function to have a custom hashing) ?\r\n\r\n> And would datasets.hashing.register just add some metadata to func in your approach (so it could be inspected from fingerprint_transform)?\r\n\r\nYup that's the idea :)\r\n\r\n> And looking to the datasets.Dataset API, .filter would also benefited from this.\r\n\r\nIndeed !\r\n\r\n-----\r\n\r\nIf you would like to contribute this you can assign yourself to this issue by posting #self-assign\r\nAnd of course if you have questions or if I can help, feel free to ping me !", "> This can also go in datasets.fingerprint indeed - but maybe datasets.hashing tells more about what the register function does (i.e. register this function to have a custom hashing) ?\r\n\r\nSure, it makes sense.\r\n\r\n---\r\n\r\nI don't plan to work on it right now, so I'll let it unassigned in case somebody wants to join. I'll get back at it as soon as possible though.\r\n" ]

[ "Thanks for reporting ! The issue doesn't come from shuffling, but from `beans` row order not being deterministic:\r\n\r\nhttps://huggingface.co/datasets/beans/blob/main/beans.py uses `dl_manager.iter_files` on ZIP archives and the file order doesn't seen to be deterministic and changes across machines", "Thank you for noticing indeed!", "This is still a bug, so I'd keep this one open if you don't mind ;)", "Besides the linked PR, to make the loading process fully deterministic, I believe we should also sort the data files [here](https://github.com/huggingface/datasets/blob/df4bdd365f2abb695f113cbf8856a925bc70901b/src/datasets/data_files.py#L276) and [here](https://github.com/huggingface/datasets/blob/df4bdd365f2abb695f113cbf8856a925bc70901b/src/datasets/data_files.py#L485) (e.g. fsspec's `LocalFileSystem.glob` relies on `os.scandir`, which yields the contents in arbitrary order). My concern is the overhead of these sorts... Maybe we could introduce a new flag to `load_dataset` similar to TFDS' [`shuffle_files`](https://www.tensorflow.org/datasets/determinism#determinism_when_reading) or sort only if the number of data files is small?", "We already return the result sorted at the end of `_resolve_single_pattern_locally` and `_resolve_single_pattern_in_dataset_repository` if I'm not mistaken", "@lhoestq Oh, you are right. Feel free to ignore my comment.", "I think the corresponding PR is ready to be merged :hugs: ", "@albertvillanova Thanks for the fix!" ]

[ "Thanks for reporting, @zhaowei-wang-nlp.\r\n\r\nYou are right, the documentation is confusing on the behavior of `remove_columns`. We should better explain it. ", "This is a duplicate of https://github.com/huggingface/datasets/issues/2343.", "I'm closing this issue because as @mariosasko pointed out, it is a duplicate of:\r\n- #2343" ]

[ "Hey ! Actually it works, just pass the right URL ;)\r\nThe URL must be the one with “/resolve/”\r\n\r\ne.g. https://huggingface.co/datasets/imagenet-1k/resolve/main/data/test_images.tar.gz\r\n\r\nYou can even pass a relative path to the dl_manager instead, like `dl_manager.download(\"data/test_images.tar.gz\")`", "Amazing it works, thanks!" ]

[ "I removed all the invalid task_ids in datasts without namespace, based on the <s>(internal)</s> types.ts", "(Types.ts is not internal it's public)", "I have opened PRs to fix the task_ids in all datasets within a namespace as well.\r\n\r\nWorking on task_categories...", "For future reference: this fix had some complications\r\n\r\nWhen trying to open a PR to fix the task tags, an exception was thrown if:\r\n- the metadata contained \"languages\" or \"licenses\" (instead of \"language\" or \"license\")\r\n- the metadata contained a non-valid language: `en-US` (instead of `en`), `no` (instead of `'no'`),...\r\n- the metadata contained a non-valid license\r\n- either `task_categories` or `task_ids` was not an array (a dict for each config)\r\n- the metadata contained non-valid tag names\r\n\r\nErrors:\r\n```\r\nValueError: - Error: \"languages\" is deprecated. Use \"language\" instead.\r\n```\r\n```\r\nValueError: - Error: \"licenses\" is deprecated. Use \"license\" instead.\r\n```\r\n```\r\nValueError: - Error: \"language[17]\" must only contain lowercase characters\r\n```\r\n```\r\nValueError: - Error: \"language[0]\" with value \"cz, de, it\" is not valid. It must be an ISO 639-1, 639-2 or 639-3 code (two/three letters), or a special value like \"code\", \"multilingual\". If you want to use BCP-47 identifiers, you can specify them in language_bcp47.\r\n```\r\n```\r\nValueError: - Error: \"task_ids\" must be an array\r\n```", "All Hub datasets are done.", "great job! did you have feedback from Hub users/i.E. repo authors?", "Yes, @julien-c. These are some of the feedbacks:\r\n- Most people just thank for the fix: [cahya/librivox-indonesia](https://huggingface.co/datasets/cahya/librivox-indonesia/discussions/1#6357cd8a292a050ebd705f84), [TurkuNLP/xlsum-fi](https://huggingface.co/datasets/TurkuNLP/xlsum-fi/discussions/1#6357828aa1f8ad1c31bcbe46), [coastalcph/fairlex](https://huggingface.co/datasets/coastalcph/fairlex/discussions/4#6351a527a8e595171ab1aef2)\r\n- Why are we changing their task names? [joelito/lextreme](https://huggingface.co/datasets/joelito/lextreme/discussions/1#6351b576fe367c0d9b12041b)\r\n - I take note of this for the next bulk operation; besides the PR title, we should also add a description to explain the reason for the change and also maybe putting a link to some pertinent GH Issue page\r\n- Some of them ask where to find the list of the supported task values is: [dennlinger/klexikon](https://huggingface.co/datasets/dennlinger/klexikon/discussions/3#6356b3ea80f8cb3ab777ac5c), [lmqg/qg_jaquad](https://huggingface.co/datasets/lmqg/qg_jaquad/discussions/1#635262467e4cc3135fd09f58)\r\n - Currently, the list is here: https://github.com/huggingface/hub-docs/blob/main/js/src/lib/interfaces/Types.ts#L85\r\n - Maybe we could made them more easily accessible\r\n- Some people do not agree about current \"hierarchy\":\r\n - text-scoring: [emrecan/nli_tr_for_simcse](https://huggingface.co/datasets/emrecan/nli_tr_for_simcse/discussions/1#6357c1b128792d8cdd51e9f9) (but referring to [emrecan/nli_tr_for_simcse](https://huggingface.co/datasets/emrecan/nli_tr_for_simcse/discussions/2/files))\r\n - Before \"text-scoring\" was a task_category, with task_ids [\"semantic-similarity-scoring\", \"sentiment-scoring\"]\r\n - Now all three are task_ids [\"text-scoring\", \"semantic-similarity-scoring\", \"sentiment-scoring\"] under the task_category \"text-classification\"\r\n - People complain that their scoring tasks are not classification task\r\n - binary-classification: why don't we have binary-classification? We have multi-class-classification, multi-label-classification and sentiment-classification, but not binary-classification\r\n - symbolic-regression: [yoshitomo-matsubara/srsd-feynman_hard](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_hard/discussions/2#63614194c12a09b8a31457cc), [yoshitomo-matsubara/srsd-feynman_medium](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_medium/discussions/2#6361418aeee0d27f04379e43), [yoshitomo-matsubara/srsd-feynman_easy](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_easy/discussions/2#6361416e00905b1ffb8d0112)\r\n - Why don't we have symbolic-regression task?\r\n\r\nNOTE: I'm editing this comment to add more feedback", "As someone with feedback on the updates (which I highly appreciate seeing included here :D), a few comments from a \"user perspective\": \r\n\r\n* I think the general confusion for me was also surrounding the hierarchy; it doesn't really become super clear (even when using the tagger space) that one is a subset of the other, especially since it seems to be still possible to include fine-grained tasks without the \"parent category\"?\r\n* The datasets explorer still shows tags that are no longer valid (e.g., super specific ones such as `summarization-other-paper-abstract-generation`, but also ones that should be `task_categories`, such as `summarization`). I'm assuming this will be fixed soon, but until then it can confuse people who don't understand why they suddenly can't use seemingly still valid tags anymore.\r\n* As I mentioned to @albertvillanova, having a dedicated page in the docs with explanations (especially wrt the difference between `task_categories` and `task_ids`) would be super helpful. However, I think it would have been sufficient to just include some description in the dataset PRs where you can link to the Github/other discussion on the topic :) That way, I can check myself what changes are expected to happen.\r\n\r\nThanks again for the streamlining process, I personally learned a fair bit about the tagging structure in the meantime!\r\nBest,\r\nDennis", "Thanks to you both for your feedback! super useful! cc'ing @osanseviero too 🙂\r\n\r\n> The datasets explorer still shows tags that are no longer valid\r\n\r\nwait which explorer is that? is it https://huggingface.co/datasets/viewer/ ?\r\n", "Sorry, this one: https://huggingface.co/datasets \r\nAnd then selecting the \"Fine-Grained Tasks\".", "good feedback! we'll improve this", "Super useful feedback, thanks a lot!", "- Some people do not agree about current \"hierarchy\":\r\n - symbolic-regression: [yoshitomo-matsubara/srsd-feynman_hard](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_hard/discussions/2#63614194c12a09b8a31457cc), [yoshitomo-matsubara/srsd-feynman_medium](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_medium/discussions/2#6361418aeee0d27f04379e43), [yoshitomo-matsubara/srsd-feynman_easy](https://huggingface.co/datasets/yoshitomo-matsubara/srsd-feynman_easy/discussions/2#6361416e00905b1ffb8d0112)\r\n - Why don't we have symbolic-regression task?", "@albertvillanova \r\nThank you for sharing our voice here!\r\n\r\nYes, we want `symbolic-regression` to be listed as a task. This task has been attracting attention from the machine learning/deep learning community, and unfortunately existing symbolic regression datasets are de-centralized in the community (hosted at individual platforms like author website, github, etc).\r\nIt would be great for the community if Hugging Face can support the task." ]

[]

[ "hey ,i would like to work on this issue . Please assign it to me.", "hey @mariosasko , i made a pr for this issue. Could you please review it.\r\nAlso i found multiple `OSError` in `extract.py` file which i thought could be replaced too but wasn't sure about them.\r\nPlease do tell if that also needs to be done." ]

[ "Hi! The Torch ops in your snippet are not equivalent to the NumPy ones, hence the difference. You can get the same behavior by replacing the line `feature = torch.mean(feature, dim=1)` with `feature = feature.squeeze().mean(1)` .", "> Hi! The Torch ops in your snippet are not equivalent to the NumPy ones, hence the difference. You can get the same behavior by replacing the line `feature = torch.mean(feature, dim=1)` with `feature = feature.squeeze().mean(1)` .\r\n\r\nThank you. " ]

[ "I can take this! #self-assign", "#self-assign", "@lazarust i'm already working on this issue :smile: ", "#self-assign", "hey @mariosasko , i made a pr for this issue. Could you please review it." ]

[ "any updates on this? It happens to me on [OpenWikiText-20%](https://huggingface.co/datasets/Bingsu/openwebtext_20p) dataset, but not on [OpenWebText-10k](https://huggingface.co/datasets/stas/openwebtext-10k). This is really strange because I don't change anything else in my running script.\r\n\r\ntransformers version 4.18.0.dev0\r\ndatasets version 1.18.0" ]

[ "Hi! Unfortunately, I can't reproduce this issue locally (in Python 3.7/3.10) or in Colab. I would assume this is due to a bug we fixed in the latest release, but your version is up-to-date, so I'm not sure if there is something we can do to help...", "Hi, 方子东. I tried running the code with exact the same configuration (both datasets 2.5.2 and 2.6.1, python, pyarrow, pandas), but on Linux. The results seem to be the expected `{<pyarrow.Int64Scalar: 4>, <pyarrow.Int64Scalar: 2>, <pyarrow.Int64Scalar: 3>, <pyarrow.Int64Scalar: 0>, <pyarrow.Int64Scalar: 1>}`.\r\nI don't have a Mac device. I can't verify whether this is a M1 chip-specific problem.", "I've just tested the code on my M1 Mac, and it behaves as expected.", "> Hi! Unfortunately, I can't reproduce this issue locally (in Python 3.7/3.10) or in Colab. I would assume this is due to a bug we fixed in the latest release, but your version is up-to-date, so I'm not sure if there is something we can do to help...\r\n\r\nThank you for your attention and feel sorry to take your time. Since this is a bug of old version, I think mybe my problem is because `cast` operation directaly used cached data generated by older verion of `datasets`. I tried to deleted the cached data and I got expected result.\r\n" ]

[ "@lhoestq I tested the script without accelerator, and I confirm this is due to datasets part as this gets similar results without accelerator.", "Hi ! You said it works on 1 GPU but doesn't wortk without accelerator - what's the difference between running on 1 GPU and running without accelerator in your case ?", "Hi @lhoestq \r\nthanks for coming back to me. Sorry for the confusion I made. I meant this works fine on 1 GPU, but on multi-gpu it is freezing. \"accelerator\" is not an issue as if you adapt the code without accelerator this still gets the same issue.\r\nIn order to test it. Please run \"accelerate config\", then use the setup for multi-gpu in one node.\r\nAfter that run \"accelerate launch code.py\" and then you would see the freezing occurs.", "Hi @lhoestq \r\ncould you have the chance to reproduce the error by running the minimal example shared?\r\nthanks", "I think you need to do `train_dataset = train_dataset.with_format(\"torch\")` to work with the DataLoader in a multiprocessing setup :)\r\n\r\nThe hang is probably caused by our streamign lib `fsspec` which doesn't work in multiprocessing out of the box - but we made it work with the PyTorch DataLoader when the dataset format is set to \"torch\"", "Hi @lhoestq \r\nthanks for the response. I added the line suggested right before calling `with accelerator.main_process_first():` in the code above and I confirm this also freezes. to reproduce it please run \"accelerate launch code.py\". I was wondering if you could have more suggestions for me? I do not have an idea how to fix this or debug this freezing. many thanks.", "Maybe the `fsspec` stuff need to be clearer even before - can you try to run this function at the very beginning of your script ?\r\n```python\r\nimport fsspec\r\n\r\ndef _set_fsspec_for_multiprocess() -> None:\r\n \"\"\"\r\n Clear reference to the loop and thread.\r\n This is necessary otherwise HTTPFileSystem hangs in the ML training loop.\r\n Only required for fsspec >= 0.9.0\r\n See https://github.com/fsspec/gcsfs/issues/379\r\n \"\"\"\r\n fsspec.asyn.iothread[0] = None\r\n fsspec.asyn.loop[0] = None\r\n\r\n_set_fsspec_for_multiprocess()\r\n```", "Hi @lhoestq \r\nthank you. I tried it, I am getting `AttributeError: module 'fsspec' has no attribute 'asyn'`. which version of fsspect do you use?\r\nI am using \r\n```fsspec 2022.8.2 pypi_0 pypi```\r\nthank you.", "Hi @lhoestq \r\nI solved `fsspec` error with this hack for now https://discuss.huggingface.co/t/attributeerror-module-fsspec-has-no-attribute-asyn/19255 but this is still freezing, I greatly appreciate if you could run this script on your side. Many thanks.\r\n\r\n```\r\nimport fsspec\r\n\r\ndef _set_fsspec_for_multiprocess() -> None:\r\n \"\"\"\r\n Clear reference to the loop and thread.\r\n This is necessary otherwise HTTPFileSystem hangs in the ML training loop.\r\n Only required for fsspec >= 0.9.0\r\n See https://github.com/fsspec/gcsfs/issues/379\r\n \"\"\"\r\n fsspec.asyn.iothread[0] = None\r\n fsspec.asyn.loop[0] = None\r\n\r\n\r\n_set_fsspec_for_multiprocess()\r\n\r\nfrom accelerate import Accelerator\r\nfrom accelerate.logging import get_logger\r\nfrom datasets import load_dataset\r\nfrom torch.utils.data.dataloader import DataLoader\r\nimport torch\r\nfrom datasets import load_dataset\r\nfrom transformers import AutoTokenizer\r\nimport torch\r\nfrom accelerate.logging import get_logger\r\nfrom torch.utils.data import IterableDataset\r\nfrom torch.utils.data.datapipes.iter.combinatorics import ShufflerIterDataPipe\r\n\r\n\r\nlogger = get_logger(__name__)\r\n\r\n\r\nclass ConstantLengthDataset(IterableDataset):\r\n \"\"\"\r\n Iterable dataset that returns constant length chunks of tokens from stream of text files.\r\n Args:\r\n tokenizer (Tokenizer): The processor used for proccessing the data.\r\n dataset (dataset.Dataset): Dataset with text files.\r\n infinite (bool): If True the iterator is reset after dataset reaches end else stops.\r\n max_seq_length (int): Length of token sequences to return.\r\n num_of_sequences (int): Number of token sequences to keep in buffer.\r\n chars_per_token (int): Number of characters per token used to estimate number of tokens in text buffer.\r\n \"\"\"\r\n\r\n def __init__(\r\n self,\r\n tokenizer,\r\n dataset,\r\n infinite=False,\r\n max_seq_length=1024,\r\n num_of_sequences=1024,\r\n chars_per_token=3.6,\r\n ):\r\n self.tokenizer = tokenizer\r\n # self.concat_token_id = tokenizer.bos_token_id\r\n self.dataset = dataset\r\n self.max_seq_length = max_seq_length\r\n self.epoch = 0\r\n self.infinite = infinite\r\n self.current_size = 0\r\n self.max_buffer_size = max_seq_length * chars_per_token * num_of_sequences\r\n self.content_field = \"text\"\r\n\r\n def __iter__(self):\r\n iterator = iter(self.dataset)\r\n more_examples = True\r\n while more_examples:\r\n buffer, buffer_len = [], 0\r\n while True:\r\n if buffer_len >= self.max_buffer_size:\r\n break\r\n try:\r\n buffer.append(next(iterator)[self.content_field])\r\n buffer_len += len(buffer[-1])\r\n except StopIteration:\r\n if self.infinite:\r\n iterator = iter(self.dataset)\r\n self.epoch += 1\r\n logger.info(f\"Dataset epoch: {self.epoch}\")\r\n else:\r\n more_examples = False\r\n break\r\n tokenized_inputs = self.tokenizer(buffer, truncation=False)[\"input_ids\"]\r\n all_token_ids = []\r\n for tokenized_input in tokenized_inputs:\r\n all_token_ids.extend(tokenized_input)\r\n for i in range(0, len(all_token_ids), self.max_seq_length):\r\n input_ids = all_token_ids[i : i + self.max_seq_length]\r\n if len(input_ids) == self.max_seq_length:\r\n self.current_size += 1\r\n yield torch.tensor(input_ids)\r\n\r\n def shuffle(self, buffer_size=1000):\r\n return ShufflerIterDataPipe(self, buffer_size=buffer_size)\r\n\r\n\r\ndef create_dataloaders(tokenizer, accelerator):\r\n ds_kwargs = {\"streaming\": True}\r\n # In distributed training, the load_dataset function gaurantees that only one process\r\n # can concurrently download the dataset.\r\n datasets = load_dataset(\r\n \"c4\",\r\n \"en\",\r\n cache_dir=\"cache_dir\",\r\n **ds_kwargs,\r\n )\r\n train_data, valid_data = datasets[\"train\"], datasets[\"validation\"]\r\n with accelerator.main_process_first():\r\n train_data = train_data.shuffle(buffer_size=10000, seed=None)\r\n train_dataset = ConstantLengthDataset(\r\n tokenizer,\r\n train_data,\r\n infinite=True,\r\n max_seq_length=256,\r\n )\r\n valid_dataset = ConstantLengthDataset(\r\n tokenizer,\r\n valid_data,\r\n infinite=False,\r\n max_seq_length=256,\r\n )\r\n train_dataset = train_dataset.shuffle(buffer_size=10000)\r\n train_dataloader = DataLoader(train_dataset, batch_size=160, shuffle=True)\r\n eval_dataloader = DataLoader(valid_dataset, batch_size=160)\r\n return train_dataloader, eval_dataloader\r\n\r\n\r\ndef main():\r\n # Accelerator.\r\n logging_dir = \"data_save_dir/log\"\r\n accelerator = Accelerator(\r\n gradient_accumulation_steps=1,\r\n mixed_precision=\"bf16\",\r\n log_with=\"tensorboard\",\r\n logging_dir=logging_dir,\r\n )\r\n # We need to initialize the trackers we use, and also store our configuration.\r\n # The trackers initializes automatically on the main process.\r\n if accelerator.is_main_process:\r\n accelerator.init_trackers(\"test\")\r\n tokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\r\n\r\n # Load datasets and create dataloaders.\r\n train_dataloader, _ = create_dataloaders(tokenizer, accelerator)\r\n\r\n train_dataloader = accelerator.prepare(train_dataloader)\r\n for step, batch in enumerate(train_dataloader, start=1):\r\n print(step)\r\n accelerator.end_training()\r\n\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n```", "Are you using `Pytorch 1.11`? Otherwise the script freezes because of the shuffling in this line: \r\n```\r\n return ShufflerIterDataPipe(self, buffer_size=buffer_size)\r\n```\r\n`ShufflerIterDataPipe` behavior must have changed for newer Pytorch versions. But this doesn't change whether you're using streaming or not in `datasets`, so probably not the same issue, but something to try.", "> Are you using `Pytorch 1.11`? Otherwise the script freezes because of the shuffling in this line:\r\n> \r\n> ```\r\n> return ShufflerIterDataPipe(self, buffer_size=buffer_size)\r\n> ```\r\n> \r\n> `ShufflerIterDataPipe` behavior must have changed for newer Pytorch versions. But this doesn't change whether you're using streaming or not in `datasets`, so probably not the same issue, but something to try.\r\n\r\nI met the same issue for pytorch 1.12 and 1.13, is there a way to work around for this function for newer pytorch versions?" ]

[ "Thanks for reporting, @david1542." ]

[ "Hi @echatzikyriakidis, thanks for submitting the issue.\r\nWhich shell are you using exactly? I tried to run the command you sent, but I don't see colors at all 🧐\r\n\r\nI tried from bash and zsh as well.", "Hi @david1542 ,\r\n\r\nI use Google Colab.\r\n", "Got it. I [created a PR](https://github.com/huggingface/datasets/pull/5120) that fixes this issue. Turns out that the wrapping logic for the inner loop was slightly incorrect.", "Thank you!", "Hello @mariosasko \r\n\r\nI am still facing this issue. Was this problem fixed?\r\n\r\n\r\n\r\nI cleared the hugging face cache before running, and no error message was given. Let me know if you need a minimal repro of my code." ]

[ "Hi Hubert! Could you please probably create a publicly available `gs://` dataset link? I think this would be easier for others to directly start to debug.", "What seems to work is to change the line to:\r\n```\r\nfs.download(src_dataset_path, dataset_path.parent.as_posix(), recursive=True)\r\n```" ]

[ "The issue is here:\r\nhttps://github.com/huggingface/datasets/blob/3ad9644b9a2e4558dd1d0f1e43c67658674e6228/src/datasets/arrow_dataset.py#L2964", "@PartiallyTyped, @Muennighoff: the issue is fixed.\r\n\r\nWe are planning to make a patch release today.", "Thanks a lot for the swift response! For a brief moment yesterday I thought I had gone insane 🤣On 14 Oct 2022, at 15:44, Albert Villanova del Moral ***@***.***> wrote:\n@PartiallyTyped, @Muennighoff: the issue is fixed.\nWe are planning to make a patch release today.\n\n—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>" ]

[ "Same bug exists with `num_proc=1` on colab. `3.7.14 (default, Sep 8 2022, 00:06:44) [GCC 7.5.0]` ", "Thanks for reporting, @loubnabnl and for the additional information, @PartiallyTyped.\r\n\r\nHowever, I'm not able to reproduce this issue, neither locally nor on Colab:\r\n```\r\nDataset({\r\n features: ['repo_name', 'path', 'copies', 'size', 'content', 'license', 'hash', 'line_mean', 'line_max', 'alpha_frac', 'autogenerated'],\r\n num_rows: 10\r\n})\r\nDataset({\r\n features: ['repo_name', 'path', 'copies', 'size', 'content', 'license', 'hash', 'line_mean', 'line_max', 'alpha_frac', 'autogenerated'],\r\n num_rows: 10\r\n})\r\n```\r\nCC: @huggingface/datasets can anybody reproduce this?", "This is the minimum reproducible example. I ran this on the premium instances of colab.\r\n\r\n```\r\n# !pip install datasets\r\nimport datasets\r\nfrom datasets import load_dataset\r\nds = load_dataset(\"copenlu/answerable_tydiqa\").filter(\"english\".__eq__, input_columns=\"language\")\r\nassert all(map(\"english\".__eq__, ds[\"train\"][\"language\"]))\r\n```\r\n\r\nIn my case, the number of samples is correct, however, the samples selected when indexing are wrong.\r\n\r\n```python\r\nDatasetDict({\r\n validation: Dataset({\r\n features: ['question_text', 'document_title', 'language', 'annotations', 'document_plaintext', 'document_url'],\r\n num_rows: 990\r\n })\r\n train: Dataset({\r\n features: ['question_text', 'document_title', 'language', 'annotations', 'document_plaintext', 'document_url'],\r\n num_rows: 7389\r\n })\r\n})\r\n```\r\n\r\nThe number of rows is indeed correct, and i have checked it with a version that works.", "I can reproduce the issue on my mac too \r\n```\r\n- `datasets` version: 2.6.0\r\n- Platform: macOS-12.2.1-arm64-arm-64bit\r\n- Python version: 3.9.13\r\n- PyArrow version: 9.0.0\r\n- Pandas version: 1.4.3\r\n```\r\nBut not on Colab with python 3.7, maybe related to python version? (didn't manage to install python 3.9)\r\n```\r\n- `datasets` version: 2.6.0\r\n- Platform: Linux-5.10.133+-x86_64-with-Ubuntu-18.04-bionic\r\n- Python version: 3.7.14\r\n- PyArrow version: 9.0.0\r\n- Pandas version: 1.3.5\r\n```", "I have the same issue, here's a simple notebook to reproduce: https://colab.research.google.com/drive/1Lvo9fg5DSpGUUgXW5JAutZ0bFsR-WV--?usp=sharing\r\n\r\n\r\n\r\n", "I think there are 2 different issues here:\r\n- the one reported by @loubnabnl is related to multiprocessing in map and then filter; we should reproduce it first: I have tried with Python version 3.9.7 and I can't reproduce it either; maybe it is related to the version of PyArrow? To be checked.\r\n- the issue reported by @PartiallyTyped is related just to \"filter\" (without multiprocessing) and I can reproduce it.", "Could you create another issue for the @PartiallyTyped one please ?\r\n\r\nRegarding the OP issue, I also tried on colab or locally on py3.7 or py3.10 but didn't reproduce", "I have created another issue for the one reported by @PartiallyTyped: \r\n- #5112 ", "I managed to reproduce your issue @loubnabnl on colab by upgrading pyarrow to 9.0.0 instead of 6.0.1", "I managed to have a _super_ minimal reproducible example:\r\n```python\r\n\r\nfrom datasets import Dataset, concatenate_datasets\r\n\r\nds = concatenate_datasets([Dataset.from_dict({\"a\": [i]}) for i in range(10)])\r\nds2 = ds.map(lambda _: {}, batched=True)\r\nassert list(ds2) == list(ds)\r\n```\r\n(filter uses a batched `map` under the hood)", "> the one reported by @loubnabnl is related to multiprocessing in map and then filter; we should reproduce it first: I have tried with Python version 3.9.7 and I can't reproduce it either; maybe it is related to the version of PyArrow? To be checked.\r\n\r\nSo finally it was related to PyArrow version! :+1: ", "Doing a patch release asap :)", "Did the patch release yesterday, lmk if you still have issues", "It works now, thanks!\r\n" ]

[ "The hash used for caching is computed by pickling recursively the function passed to `map`. Maybe some objects don't have the same hash across sessions. In particular you can check the hash of your model using\r\n```python\r\nfrom datasets.fingerprint import Hasher\r\nobj = AutoModel.from_config(config=config, add_pooling_layer=False)\r\nprint(Hasher.hash(obj))\r\n```\r\n\r\nYou can find mode info here: https://huggingface.co/docs/datasets/about_cache\r\n\r\nYou can also provide your own unique hash in `map` if you want, with the `new_fingerprint` argument", "Indeed, the hash is changing. The `dumps` function serialize the model object in different ways because the model object is not deterministic\r\n```python\r\nfrom datasets.utils.py_utils import dumps\r\nobj1 = AutoModel.from_config(config=config, add_pooling_layer=False)\r\nobj2 = AutoModel.from_config(config=config, add_pooling_layer=False)\r\n\r\ndumps(bert) == dumps(bert2). # False\r\n```\r\n\r\n> You can find mode info here: https://huggingface.co/docs/datasets/about_cache\r\n> \r\n> You can also provide your own unique hash in map if you want, with the new_fingerprint argument\r\n\r\n\r\nThanks, the doc is so helpful. Indeed, we can fix the hash and get cache hit using `new_fingerprint`. Closing the issue." ]

[ "cc @lhoestq ", "Thanks for reporting, @cakiki.\r\n\r\nI would say the deletion of the dir is an expected behavior though...", "`dask.to_parquet` has an \"overwrite\" parameter and default is `False`, we could also have something similar", "Thank you both for your feedback!\r\n\r\n@albertvillanova I think I might have have the wrong mental model of what the function was meant to do. I thought it would be an API similar to the pandas `to_XX` write methods (Like the one @lhoestq mentions) so I just assumed it would download the dataframe to whichever folder I specififed (`\"./\"` in my case) so I could load it into a dask dataframe. I absolutely did not expect it to delete everything in my local directory, including the script where I called it from :smile: \r\n\r\nI think Quentin's proposed solution sounds like a reasonable feature!", "actually there's already a `download_mode` parameter that defaults to `REUSE_DATASET_IF_EXISTS` - so I guess it's just a matter of not deleting files unrelated to the dataset, and to overwrite existing dataset files if the download mode is `REUSE_CACHE_IF_EXISTS` or `FORCE_REDOWNLOAD`" ]

[ "Hi, thanks for reporting! The last line should be `dataset = Dataset.from_generator(my_gen)`.", "Can I work on this one?" ]

[]

[ "`datasets` doesn't seem to urlencode the directory names here\r\n\r\nhttps://github.com/huggingface/datasets/blob/7feeb5648a63b6135a8259dedc3b1e19185ee4c7/src/datasets/utils/file_utils.py#L109-L111\r\n\r\nfor example we should have\r\n```python\r\nfrom datasets.utils.file_utils import hf_hub_url\r\n\r\nurl = hf_hub_url(\"loubnabnl/bigcode_csharp\", \"data/c#/data_0003.jsonl\")\r\nprint(url)\r\n# Currently returns\r\n# https://huggingface.co/datasets/loubnabnl/bigcode_csharp/resolve/main/data/c#/data_0003.jsonl\r\n# while it should be \r\n# https://huggingface.co/datasets/loubnabnl/bigcode_csharp/resolve/main/data/c%23/data_0003.jsonl\r\n```", "I'll work on this :)", "@loubnabnl The dataset you linked in the description of the bug does not work and returns a 404. Where can I find the dataset to reproduce the bug?", "I think you can create a dataset repository on the Hub with a dummy file containing a `#`", "Ah sorry it was private I just made it public, I can also help with this if needed", "@lhoestq Should I url encode also repo_id and revision parameters? I'm not sure what are the valid characters there.\r\n\r\nPersonally, I would be cautious and only url encode the path parameter.", "These are possible solutions (assuming `from urllib.parse import quote`):\r\n\r\n1) url encode only the path parameter:\r\n```\r\n# src/datasets/utils/file_utils.py\r\ndef hf_hub_url(repo_id: str, path: str, revision: Optional[str] = None) -> str:\r\n revision = revision or config.HUB_DEFAULT_VERSION\r\n return config.HUB_DATASETS_URL.format(repo_id=repo_id, path=quote(path), revision=revision)\r\n```\r\n2) url encode all parameters:\r\n```\r\n# src/datasets/utils/file_utils.py\r\ndef hf_hub_url(repo_id: str, path: str, revision: Optional[str] = None) -> str:\r\n revision = revision or config.HUB_DEFAULT_VERSION\r\n return config.HUB_DATASETS_URL.format(repo_id=quote(repo_id), path=quote(path), revision=quote(revision))\r\n```\r\n3) url encode the whole url:\r\n```\r\n# src/datasets/config.py\r\nHUB_DATASETS_PATH = \"/datasets/{repo_id}/resolve/{revision}/{path}\"\r\nHUB_DATASETS_URL = HF_ENDPOINT + HUB_DATASETS_PATH\r\n```\r\n```\r\n# src/datasets/utils/file_utils.py\r\ndef hf_hub_url(repo_id: str, path: str, revision: Optional[str] = None) -> str:\r\n revision = revision or config.HUB_DEFAULT_VERSION\r\n return config.HF_ENDPOINT + quote(config.HUB_DATASETS_PATH.format(repo_id=repo_id, path=path, revision=revision))\r\n```", "repo_id can only contain alphanumeric characters and _- so it doesn't need to be encoded.\r\n\r\nHowever I agree it's a good idea to also apply `quote` to the revision as well as in 2. !", "Should be fixed by https://github.com/huggingface/datasets/issues/5099 - we'll do a release later today" ]

[ "It can be solved temporarily by remove `resolve` in \r\nhttps://github.com/huggingface/datasets/blob/bef23be3d9543b1ca2da87ab2f05070201044ddc/src/datasets/data_files.py#L278", "Hi, thanks for reporting and suggesting a fix! We still need to account for `.`/`..` in the file path, so a more robust fix would be `Path(os.path.abspath(filepath))`.", "> Hi, thanks for reporting and suggesting a fix! We still need to account for `.`/`..` in the file path, so a more robust fix would be `Path(os.path.abspath(filepath))`.\r\n\r\nThanks for your reply!" ]

[ "Thanks for reporting, @catalys1.\r\n\r\nThis seems a duplicate of:\r\n- #3310 \r\n\r\nThe source of the problem is in PyArrow:\r\n- [ARROW-15141: [C++] Fatal error condition occurred in aws_thread_launch](https://issues.apache.org/jira/browse/ARROW-15141)\r\n- [ARROW-17501: [C++] Fatal error condition occurred in aws_thread_launch](https://issues.apache.org/jira/browse/ARROW-17501)\r\n\r\nThe bug in their dependency is still unresolved:\r\n- https://github.com/aws/aws-sdk-cpp/issues/1809\r\n\r\nApparently, the `aws-sdk-cpp` PyArrow dependency needs to be pinned at version `1.8.186` if using conda. Have you updated it after installing PyArrow?\r\n```shell\r\nconda list aws-sdk-cpp\r\n```\r\n\r\nMaybe you should try to downgrade it to that version:\r\n```shell\r\nconda install -c conda-forge aws-sdk-cpp=1.8.186\r\n```" ]

[ "The transfer of the dummy dataset to the dummy org works as expected:\r\n```python\r\nIn [1]: from datasets import load_dataset; ds = load_dataset(\"dummy_canonical_dataset\", download_mode=\"force_redownload\"); ds\r\nDownloading builder script: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2.98k/2.98k [00:00<00:00, 2.01MB/s]\r\nDownloading and preparing dataset dummy_canonical_dataset/default (download: 411 bytes, generated: 385 bytes, post-processed: Unknown size, total: 796 bytes) to .../.cache/huggingface/datasets/dummy_canonical_dataset/default/1.0.0/100870c358637e269fee140585e61e1472d5075a9bf6f866719934c725e55fb4...\r\nDownloading data: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 411/411 [00:00<00:00, 293kB/s]\r\nDataset dummy_canonical_dataset downloaded and prepared to .../.cache/huggingface/datasets/dummy_canonical_dataset/default/1.0.0/100870c358637e269fee140585e61e1472d5075a9bf6f866719934c725e55fb4. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 304.16it/s]\r\nOut[1]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['langs', 'ner_tags', 'tokens'],\r\n num_rows: 3\r\n })\r\n})\r\n\r\nIn [2]: from datasets import load_dataset; ds = load_dataset(\"dummy-canonical-org/dummy_canonical_dataset\"); ds\r\nDownloading builder script: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2.98k/2.98k [00:00<00:00, 1.57MB/s]\r\nDownloading and preparing dataset dummy_canonical_dataset/default to .../.cache/huggingface/datasets/dummy-canonical-org___dummy_canonical_dataset/default/1.0.0/100870c358637e269fee140585e61e1472d5075a9bf6f866719934c725e55fb4...\r\nDataset dummy_canonical_dataset downloaded and prepared to .../.cache/huggingface/datasets/dummy-canonical-org___dummy_canonical_dataset/default/1.0.0/100870c358637e269fee140585e61e1472d5075a9bf6f866719934c725e55fb4. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 362.48it/s]\r\nOut[2]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['langs', 'ner_tags', 'tokens'],\r\n num_rows: 3\r\n })\r\n})\r\n```", "Cool ! 🚀 ", "Maybe we should be a bit more proactive with these transfers. There are only ≈70 canonical models, so reaching that number with datasets would be great, too. It's not easy considering the current number of ≈750 canonical datasets, but doable.\r\n\r\nFor instance, it shouldn't be too hard to transfer these datasets (partial list; all of them have more than > 1k downloads):\r\n\r\n<details>\r\n\r\n<summary> Datasets to transfer </summary>\r\n\r\n```\r\nquickdraw -> google\r\nopenai_humaneval -> openai\r\nc4 -> allenai/c4 (the canonical version reads data from the org version)\r\nmbpp -> google (ask jaaustin (author) where to transfer the dataset)\r\ncompetition_math -> hendrycks (author)\r\ngsm8k -> openai\r\nai2_arc -> allenai\r\nimdb -> stanfordai\r\ngreek_legal_code -> chrispap (author)\r\nspider -> Yale-LILY\r\nsquad and squad_v2 -> rajpurkarlab (or rajpurkar, a member of the org and one of the authors)\r\ncppe-5 -> rishitdagli\r\nnews_commentary -> Helsinki-NLP\r\njfleg -> keisks (author)\r\npubmed_qa -> qiaojin (author)\r\nmedmcqa -> infinitylogesh (author)\r\ncifar10 and cifar100 -> UniversityofToronto\r\ncc100 -> gwenzek (author)\r\nasset -> facebook\r\nblbooks -> BritishLibraryLabs\r\ncapes -> FLSRDS (maybe the author?)\r\ncc_news -> fhamborg (author)\r\nclue -> CLUE benchmark\r\ncoqa -> stanfordnlp\r\nlambada -> germank (author)\r\nlibrispeech_asr -> openslr\r\ndrop -> allenai\r\nduorc -> salesforce (ask amritasaha87 (author) where to transfer)\r\nglue -> nyu-mll ?\r\ngo_emotions -> google\r\ncommonsense_qa -> tau\r\ndbpedia_14 -> JensLehmann (author?)\r\ndiscofuse -> google\r\nmc4 -> allenai/c4\r\nopenbookqa -> allenai\r\nropes -> allene\r\ntrivia_qa -> mandarjoshi (author)\r\nwikiann -> afshinrahimi (author)\r\nxtreme -> google\r\nxscr -> INK-USC\r\nyelp_review_full -> Yelp\r\ntruthful_qa -> jacobhilton22 (author)\r\nbigbench -> google\r\nxnli -> facebook\r\nsciq -> allenai\r\nsst2 -> stanfordnlp\r\nblimp -> alexwarstadt (author)\r\ntweet_eval -> cardiffnlp\r\nbeans -> AI-Lab-Makerere\r\nlex_glue -> coastalcph\r\namericas_nli -> abteen (author)\r\nopus_euconst -> tiedeman (author)\r\nmedical_questions_pairs -> curaihealth\r\nweb_questions -> joberant (author)\r\nanli -> facebook\r\nrace -> CarnegieMellonCS\r\nklue -> klue\r\nwino_bias -> uclanlp\r\nwiki_qa -> microsoft\r\nxcopa -> cambridgeltl\r\nindic_glue -> ai4bharat\r\nboolq -> google\r\nadversarial_qa -> mbartolo (author)\r\nnq_open -> google\r\nsnli -> stanfordnlp\r\nstsb_multi_mt -> PhilipMay (author)\r\nmulti_nli -> sleepinyourhat (author)\r\npaws -> google\r\npaws-x -> google\r\nms_marco - microsoft\r\nxquad -> deepmind\r\nnarrativeqa -> deepmind\r\nkilt_tasks -> facebook\r\nhate_speech_offensive -> tdavidson (author)\r\nwiki40b -> google\r\ncovost2 -> facebook\r\ncommon_gen -> INKLAB\r\nmulti_eurlex -> kiddothe2b (author)\r\nexams -> mhardalov (author)\r\ntiny_shakespeare -> karpathy (author)\r\nblbooksgenre -> BritishLibraryLabs ?\r\nfood101 -> ethz ?\r\nscitail -> allenai\r\nbillsum -> FiscalNote\r\nimppres -> facebook\r\nquartz -> allenai\r\nqasc -> allenai\r\nquail -> textmachinelab\r\nwiki_lingua -> esdurmus\r\ncos_e -> salesforce ?\r\ncivil_comments -> google ? (create a “jigsaw” org) \r\nxquad_r -> google\r\nwikitext-> metamind (or salesforce)\r\n\r\n// deprecate c4 and mc4 in favor of allenai/c4 (add a dataset script to the org version to make it easier to use?)\r\n```\r\n</details>\r\n\r\nAlso, a space that allows users to claim the existing canonical datasets (for themselves or their organizations) could be nice.\r\n\r\nWDYT?", "Next week I can take care of some of them :) In most cases we just need to send an email to ask them if they're ok with it.\r\nLet's coordinate on slack ?", "Yup, sounds good to me!", "I can also continuing working on this if we agree this has become a priority now.", "cool stuff! \r\n\r\nthis morning on my side i moved huggingface.co/ctrl (a not very used model) to its rightful entity", "As a previous step before transferring the datasets, we decided we should convert them to Parquet, so that the viewer does not stop working (the viewer does not support datasets with scripts). \r\n\r\nDatasets converted to Parquet:\r\n- [x] adversarial_qa\r\n- [x] ai2_arc\r\n- [x] americas_nli\r\n- [x] anli\r\n- [x] asset\r\n- [x] beans\r\n- [ ] bigbench\r\n- [x] billsum\r\n- [ ] blbooks: it was already transferred to: TheBritishLibrary/blbooks\r\n- [ ] blbooksgenre: it was already transferred to: TheBritishLibrary/blbooksgenre\r\n- [x] blimp\r\n- [x] boolq\r\n- [ ] c4\r\n- [x] capes\r\n- [ ] cc100\r\n- [x] cc_news\r\n- [x] cifar10\r\n- [x] cifar100\r\n- [x] civil_comments\r\n- [x] clue\r\n- [x] common_gen\r\n- [x] commonsense_qa\r\n- [ ] competition_math: it was already transferred to: hendrycks/competition_math\r\n- [x] coqa\r\n- [x] cos_e\r\n- [ ] covost2: it requires manual download\r\n- [x] cppe-5\r\n- [x] dbpedia_14\r\n- [x] discofuse\r\n- [x] drop\r\n- [x] duorc\r\n- [x] exams\r\n- [x] food101\r\n- [x] glue\r\n- [x] go_emotions\r\n- [x] greek_legal_code\r\n- [x] gsm8k\r\n- [x] hate_speech_offensive\r\n- [x] imdb\r\n- [x] imppres\r\n- [x] indic_glue\r\n- [x] jfleg\r\n- [x] kilt_tasks\r\n- [x] klue\r\n- [x] lambada\r\n- [x] lex_glue\r\n- [ ] librispeech_asr\r\n- [x] mbpp\r\n- [ ] mc4\r\n- [x] medical_questions_pairs\r\n- [x] medmcqa\r\n- [x] ms_marco\r\n- [ ] multi_eurlex\r\n- [x] multi_nli\r\n- [ ] narrativeqa\r\n- [ ] news_commentary\r\n- [x] nq_open\r\n- [x] openai_humaneval\r\n- [x] openbookqa\r\n- [ ] opus_euconst\r\n- [x] paws\r\n- [x] paws-x\r\n- [x] pubmed_qa\r\n- [x] qasc\r\n- [x] quail\r\n- [x] quartz\r\n- [ ] quickdraw\r\n- [x] race\r\n- [x] ropes\r\n- [x] sciq\r\n- [x] scitail\r\n- [ ] snli\r\n- [x] spider\r\n- [x] squad\r\n- [x] squad_v2\r\n- [x] sst2\r\n- [x] stsb_multi_mt\r\n- [x] tiny_shakespeare\r\n- [x] trivia_qa\r\n- [x] truthful_qa\r\n- [x] tweet_eval\r\n- [x] web_questions\r\n- [ ] wiki40b\r\n- [x] wiki_lingua\r\n- [x] wiki_qa\r\n- [ ] wikiann\r\n- [x] wikitext\r\n- [x] wino_bias\r\n- [x] xcopa\r\n- [x] xcsr\r\n- [x] xnli\r\n- [x] xquad\r\n- [x] xquad_r\r\n- [ ] xtreme\r\n- [x] yelp_review_full\r\n", "For `c4` and `mc4` I was thinking of adding the corresponding configs to `allenai/c4` and redirect `c4` and `mc4` to `allenai/c4`. I'll open a PR on `allenai/c4` if it's good for you", "@davanstrien and @lhoestq, I have shared with you this spreadsheet: https://docs.google.com/spreadsheets/d/1GvNTd1UxmtTvEFOK-Eq6E3Str4FUWQuWZsEN0WVFirs/edit?usp=sharing\r\n\r\nThis way we can take datasets by batches to contact the authors and transfer to the organizations.", "We have already transferred all canonical datasets under organization/user namespaces." ]

[ "Hi ! Could it be an Out of Memory issue that could have killed one of the processes ? can you check your memory ?", "Hi! I don't think it is a memory issue. I'm monitoring the main and spawn python processes and threads with `htop` and the memory does not peak. Besides, the example I've posted above should not be that demanding in terms of memory, right? (I have 32GB of RAM). ", "Indeed it should be fine. I couldn't reproduce the error though - I ran your script on my side and it works fine. What version of pytorch are you using ?", "Interesting.. I'm using `torch 1.12.1`", "I also tried on colab and it works fine 🤔 \r\nMaybe something is wrong with your installation of pytorch ?", "Oh actually I just saw that you're using python 3.9\r\n\r\nThis could be related to https://github.com/huggingface/datasets/issues/4113\r\n\r\nWe'll fix that as soon as we can, in the meantime you can try to use use single process, or use an older version of python maybe ?", "I tried with python 3.7 and the issue persists. In collab, which also uses 3.7 I don't get the issue, so yes I guess is something on mu side... will post it here if I manage to fix it", "Hi! Which version of transformers are you using? I test the code on Colab (so python 3.7) with transformers 4.23.1, torch 1.12.1 and pyarrow 9.0.0 (also 6.x), it worked without stuck.", "Hi, I have the same problem in use **datasets.IterableDatasetDict.map()**\r\nmy pytorch is 2.0.0a0+gitc263bd4\r\nmy python is 3.8.16(default, Jun 12 2023, 17:37:21)\r\nwork on aarch64 in 16 node, each node with 4*nVidia-A100-40G\r\nevery node have 4 process execute code as ↓\r\n\r\n```\r\nfrom datasets import load_dataset, interleave_datasets, IterableDatasetDict, concatenate_datasets\r\n```\r\n...\r\n```\r\n model_args.cache_dir = '/home/scx/.cache'\r\n for dataset_name in data_args.datasets_name:\r\n train_datasets.append(\r\n load_dataset(\r\n dataset_name,\r\n cache_dir=model_args.cache_dir,\r\n use_auth_token=True if model_args.use_auth_token else None,\r\n streaming=data_args.streaming,\r\n split='train'\r\n ).select_columns('text')\r\n )\r\n valid_datasets.append(\r\n load_dataset(\r\n dataset_name,\r\n cache_dir=model_args.cache_dir,\r\n use_auth_token=True if model_args.use_auth_token else None,\r\n streaming=data_args.streaming,\r\n split='validation'\r\n ).select_columns('text')\r\n )\r\n train_dataset = interleave_datasets(train_datasets,\r\n probabilities=data_args.datasets_probabilities, \r\n seed=training_args.seed,\r\n stopping_strategy='all_exhausted')\r\n raw_datasets = IterableDatasetDict({'train': train_dataset, 'validation': valid_dataset})\r\n```\r\n...\r\n\r\n```\r\n tokenized_datasets = None\r\n with training_args.main_process_first(desc=\"dataset map tokenization\"):\r\n if not data_args.streaming:\r\n tokenized_datasets = raw_datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n num_proc=data_args.preprocessing_num_workers,\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n desc=\"Running tokenizer on dataset\",\r\n remove_columns=column_names,\r\n )\r\n else:\r\n #TODO 20230722\r\n logger.info('{}: {}'.format(__file__, 'tokenized_datasets = raw_datasets.map('))\r\n logger.info('len raw_datasets: {}'.format(len(raw_datasets.items())))\r\n logger.info('raw_datasets:{}'.format(raw_datasets.items()))\r\n tokenized_datasets = raw_datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n batch_size=1000,\r\n remove_columns=column_names\r\n )\r\n logger.info('map ok!')\r\n logger.info('show train: {}'.format(next(iter(tokenized_datasets['train']))))\r\n logger.info('ok')\r\n # ### RAW CODE ###\r\n # tokenized_datasets = raw_datasets.map(\r\n # tokenize_function,\r\n # batched=True,\r\n # batch_size=1000,\r\n # remove_columns=column_names\r\n # )\r\n #TODO 20230722\r\n logger.info(\"Finish tokenization\")\r\n```\r\nthe output of my code is\r\n```\r\n07/22/2023 21:57:09 - INFO - __main__ - /demo/run_blue_space.py: tokenized_datasets = raw_datasets.map(\r\n07/22/2023 21:57:09 - INFO - __main__ - len raw_datasets: 2\r\n07/22/2023 21:57:09 - INFO - __main__ - raw_datasets:dict_items([('train', <datasets.iterable_dataset.IterableDataset object at 0x4005ee301190>), ('validation', <datasets.iterable_dataset.IterableDataset object at 0x4005ee5427f0>)])\r\n07/22/2023 21:57:09 - INFO - __main__ - map ok!\r\n07/22/2023 22:01:07 - INFO - __main__ - show train: {'input_ids': [14608, 26797, 31891, 34260, 12227, 33207, 5, 5, 31632, 26797, 31891, 34260, 12227, 33207, 7398, 28561, 31236, 31177, 31253, 33558, 31556, 31377, 72, 20732, 32383, 32295, 14027, 31178, 53, 61, 53, 55, 31189, 31146, 31321, 31235, 53, 61, 56, 58, 31189, 31145, 72, 53, 61, 58, 54, 31189, 54, 31245, 53, 60, 31224, 31896, 31178, 28561, 29331, 20732, 31888, 32637, 4426, 2824, 72, 53, 61, 60, 55, 31189, 53, 54, 31245, 53, 31224, 31896, 31178, 28561, 29331, 26137, 20732, 4426, 2824, 73, 54, 52, 52, 52, 31189, 61, 31245, 59, 31224, 31896, 31178, 29331, 28561, 20732, 4426, 2824, 73, 5], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}\r\n07/22/2023 22:01:07 - INFO - __main__ - ok\r\n```\r\n\r\n", "@bio-punk `IterableDatasetDict.map` does not support multiprocessing (only `DatasetDict.map` and `Dataset.map` do), so please open a new issue as this doesn't seem to be related to the original issue. ", "Closing as this issue doesn't seem to be related to `datasets`." ]

[ "Hi, thanks for reporting! This line should be replaced with \r\n```python\r\ndataset = dataset.map(lambda examples: tokenizer(examples[\"text\"], return_tensors=\"np\"), batched=True)\r\n```\r\nfor it to work (the `return_tensors` part inside the `tokenizer` call).", "Can I work on this?", "Fixed in https://github.com/huggingface/datasets/pull/5095" ]

[ "Nice!!" ]

[]

[ "Hi @junwang-wish, thanks for reporting.\r\n\r\nUnfortunately, I'm not able to reproduce the bug. Which version of `datasets` are you using? Does the problem persist if you update `datasets`?\r\n```shell\r\npip install -U datasets\r\n``` ", "Thanks @albertvillanova I updated `datasets` from `2.5.1` to `2.5.2` and tested copying the `json.gz` to a different directory and my mind was blown:\r\n\r\n```python\r\nfpath = '/data/junwang/.cache/general/57b6f2314cbe0bc45dda5b78f0871df2/test.json.gz'\r\nds_panda = DatasetDict(\r\n test=Dataset.from_pandas(\r\n pd.read_json(fpath, lines=True)\r\n )\r\n)\r\nds_direct = load_dataset(\r\n 'json', data_files={\r\n 'test': fpath\r\n }, features=Features(\r\n text_input=Value(dtype=\"string\", id=None),\r\n text_output=Value(dtype=\"string\", id=None)\r\n )\r\n)\r\nlen(ds_panda['test']), len(ds_direct['test'])\r\n```\r\nproduces \r\n```python\r\nUsing custom data configuration default-0e6cf24134163e8b\r\nFound cached dataset json (/data/junwang/.cache/huggingface/datasets/json/default-0e6cf24134163e8b/0.0.0/e6070c77f18f01a5ad4551a8b7edfba20b8438b7cad4d94e6ad9378022ce4aab)\r\n(1, 0)\r\n```\r\nbut then I ran below command to see if the same file in a different directory leads to same discrepancy\r\n```shell\r\ncp /data/junwang/.cache/general/57b6f2314cbe0bc45dda5b78f0871df2/test.json.gz tmp_test.json.gz\r\n```\r\nand so I ran\r\n```python\r\nfpath = 'tmp_test.json.gz'\r\nds_panda = DatasetDict(\r\n test=Dataset.from_pandas(\r\n pd.read_json(fpath, lines=True)\r\n )\r\n)\r\nds_direct = load_dataset(\r\n 'json', data_files={\r\n 'test': fpath\r\n }, features=Features(\r\n text_input=Value(dtype=\"string\", id=None),\r\n text_output=Value(dtype=\"string\", id=None)\r\n )\r\n)\r\nlen(ds_panda['test']), len(ds_direct['test'])\r\n```\r\nand behold, I get \r\n```python\r\nUsing custom data configuration default-f679b32ab0008520\r\nDownloading and preparing dataset json/default to /data/junwang/.cache/huggingface/datasets/json/default-f679b32ab0008520/0.0.0/e6070c77f18f01a5ad4551a8b7edfba20b8438b7cad4d94e6ad9378022ce4aab...\r\nDataset json downloaded and prepared to /data/junwang/.cache/huggingface/datasets/json/default-f679b32ab0008520/0.0.0/e6070c77f18f01a5ad4551a8b7edfba20b8438b7cad4d94e6ad9378022ce4aab. Subsequent calls will reuse this data.\r\n(1, 1)\r\n```\r\nThey match now !\r\n\r\nThis problem happens regardless of the shell I use (VScode jupyter extension or plain old Python REPL). \r\n\r\nI attached the `json.gz` here for reference: [test.json.gz](https://github.com/huggingface/datasets/files/9734843/test.json.gz)\r\n\r\n" ]

[ "FYI @lewtun ", "Hi @km5ar, thanks for reporting.\r\n\r\nThis should be fixed in the notebook:\r\n- the filename `datasets-issues-with-hf-doc-builder.jsonl` no longer exists on the repo; instead, current filename is `datasets-issues-with-comments.jsonl`\r\n- see: https://huggingface.co/datasets/lewtun/github-issues/tree/main\r\n\r\nAnyway, depending on your version of `datasets`, you can now use:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nissues_dataset = load_dataset(\"lewtun/github-issues\")\r\nissues_dataset\r\n```\r\ninstead of:\r\n```python\r\nfrom huggingface_hub import hf_hub_url\r\n\r\ndata_files = hf_hub_url(\r\n repo_id=\"lewtun/github-issues\",\r\n filename=\"datasets-issues-with-hf-doc-builder.jsonl\",\r\n repo_type=\"dataset\",\r\n)\r\nfrom datasets import load_dataset\r\n\r\nissues_dataset = load_dataset(\"json\", data_files=data_files, split=\"train\")\r\nissues_dataset\r\n```\r\n\r\nOutput:\r\n```python\r\nIn [25]: ds = load_dataset(\"lewtun/github-issues\")\r\nDownloading: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10.5k/10.5k [00:00<00:00, 5.75MB/s]\r\nUsing custom data configuration lewtun--github-issues-cff5093ecc410ea2\r\nDownloading and preparing dataset json/lewtun--github-issues to .../.cache/huggingface/datasets/lewtun___json/lewtun--github-issues-cff5093ecc410ea2/0.0.0/e6070c77f18f01a5ad4551a8b7edfba20b8438b7cad4d94e6ad9378022ce4aab...\r\nDownloading data: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 12.2M/12.2M [00:00<00:00, 26.5MB/s]\r\nDownloading data files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.70s/it]\r\nExtracting data files: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1589.96it/s]\r\nDataset json downloaded and prepared to .../.cache/huggingface/datasets/lewtun___json/lewtun--github-issues-cff5093ecc410ea2/0.0.0/e6070c77f18f01a5ad4551a8b7edfba20b8438b7cad4d94e6ad9378022ce4aab. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 133.95it/s]\r\n\r\nIn [26]: ds\r\nOut[26]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['url', 'repository_url', 'labels_url', 'comments_url', 'events_url', 'html_url', 'id', 'node_id', 'number', 'title', 'user', 'labels', 'state', 'locked', 'assignee', 'assignees', 'milestone', 'comments', 'created_at', 'updated_at', 'closed_at', 'author_association', 'active_lock_reason', 'pull_request', 'body', 'timeline_url', 'performed_via_github_app', 'is_pull_request'],\r\n num_rows: 3019\r\n })\r\n})\r\n```", "Thanks for reporting @km5ar and thank you @albertvillanova for the quick solution! I'll post a fix on the source too" ]

[ "~~It seems like #5043 fix (merged recently) is the root cause of such behaviour. When we empty indices mapping (because the dataset length equals to zero), we can no longer get column item like: `ds_filter_2['sentence']` which uses\r\n`ds_filter_1._indices.column(0)`~~\r\n\r\n**UPDATE:**\r\nEmpty datasets are returned without going through partial function on `map` method, which will not work to get indices for `filter`: we need to run `get_indices_from_mask_function` partial function on the dataset to get output = `{\"indices\": []}`. But this is complicated since functions used in args, in particular `get_indices_from_mask_function`, do not support empty datasets.\r\nWe can just handle empty datasets aside on filter method.", "#self-assign", "Thank you for solving this amazingly quickly!" ]

[ "hii @lhoestq, can you assign this issue to me? Though i am new to open source still I would love to put my best foot forward. I can see there isn't anyone right now assigned to this issue.", "Hi @zutarich ! This issue was fixed by #5852 - sorry I forgot to close it\r\n\r\nFeel free to look for other issues and ping me or @mariosasko if you have questions :)\r\nAlso let us know if we can help find an issue that can correspond to what you're looking for" ]

[ "tagging @nreimers ", "The dataset is sadly not really compatible to be loaded with `load_dataset`. So far it is better to git clone it and to use the files directly.\r\n\r\nA data loading script would be needed to be added to this dataset. But this was too much overhead / not really intuitive how to create it.", "Since the dataset is a bunch of TSVs we should not need a dataset script I think.\r\n\r\nBy default it tries to load all the TSVs at once, which fails here because they don't all have the same columns (pd.read_csv uses the first line as header by default). But those files have no header ! So, to properly load any TSV file in this repo, one has to pass `names=[...]` for pd.read_csv to know which column names to use.\r\n\r\nTo fix this situation, we can either do\r\n1. replace the TSVs by TSV with column names\r\n2. OR specify the pd.read_csv kwargs as YAML in the dataset card - and `datasets` would use that by default\r\n\r\nWDTY ?", "There are more issues in the dataset.\r\nTo load OpenSubtitles I have to provide this (see `skiprows`):\r\n\r\n```python\r\ndf_os = pd.read_csv(\r\n \"./parallel-sentences/OpenSubtitles/OpenSubtitles-en-de-train.tsv.gz\", \r\n sep=\"\\t\", \r\n quoting=csv.QUOTE_NONE,\r\n header=None,\r\n names=[\"en\", \"de\"],\r\n skiprows=[540344, 9151700, 10040173, 10040199, 11314673, 11338258, 11869223, 12159297, 12251078, 12303334],\r\n)\r\n```", "What's wrong with those lines exactly ?\r\nMaybe passing `error_bad_lines=False` (and maybe `warn_bad_lines=True`) can be helpful", "> What's wrong with those lines exactly ? \r\n\r\nStuff like this: `ParserError: Error tokenizing data. C error: Expected 2 fields in line 540345, saw 3`\r\n\r\n", "> Maybe passing error_bad_lines=False (and maybe warn_bad_lines=True) can be helpful\r\n\r\nYes. That would hide the issue but not solve it.", "@nreimers WDYT about the two options mentioned above ?" ]

[ "There is some discussion in https://github.com/huggingface/huggingface_hub/pull/1088 if it can help :)" ]

[ "@mariosasko I've raised a PR #5076 against this issue. Please help to review. Thanks." ]

[ "Hi, can I pick up this issue?", "#self-assign", "Looks like the top-level `datasource` directory was removed when https://github.com/huggingface/datasets/pull/4974 was merged, so there are 3 source files to fix." ]