Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "@lhoestq @LysandreJik", "<s>Hi :) Can you share with us the code you used ?</s>\r\n\r\nEDIT: from https://github.com/huggingface/transformers/issues/11665#issuecomment-838348291 I understand you're using the run_clm.py script. Can you share your logs ?\r\n", "Also note that for the caching to work, you must reuse the exact same parameters as in the first run. Did you change any parameter ? The `preprocessing_num_workers` should also stay the same", "> Also note that for the caching to work, you must reuse the exact same parameters as in the first run. Did you change any parameter ? The `preprocessing_num_workers` should also stay the same\r\n\r\nI only changed the `preprocessing_num_workers` maybe it is the problem~ I will try again~" ]

[ "Hi ! We don't have `join`/`merge` on a certain column as in pandas.\r\nMaybe you can just use the [concatenate_datasets](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=concatenate#datasets.concatenate_datasets) function.\r\n", "Hi! You can use `datasets_sql` for that now. As of recently, PyArrow also supports querying tables via Substrait, so I think we can start adding these methods to the API soon." ]

[ "Hi! Columns are removed **after** applying the function and **before** updating the examples with the function's output (as per the docs [here](https://huggingface.co/docs/datasets/package_reference/main_classes#datasets.Dataset.map.remove_columns)). I agree the docs on this should be more clear." ]

[ "I've raised a PR for this. Should work with `dataset = load_dataset(\"web_of_science\", \"WOS11967\", ignore_verifications=True)`once it gets merged into the main branch. Thanks for reporting this! " ]

[]

[]

[ "Hi @lhoestq,\r\nShould we change `desc` in [pbar](https://github.com/huggingface/datasets/blob/81fcf88172ed5e3026ef68aed4c0ec6980372333/src/datasets/arrow_dataset.py#L1860) to something meaningful?", "I think the user could pass the `desc` parameter to `map` so that it can be displayed in the tqdm progress bar, as suggested by @cccntu.\r\n\r\nWhen there's no multiprocessing, the `desc` of the progress bar could be the `desc` passed by the user.\r\nIn multiprocessing, we were already using a `desc` equal to `\"#\" + str(rank)`.\r\nWe can change it to be `(desc or \"\") + \"#\" + str(rank)` instead.\r\n\r\nIn the end, since both `desc` and `rank` could be None, we can have:\r\n```python\r\npbar_desc = (desc or \"\") + \"#\" + str(rank) if rank is not None else desc\r\n```\r\n\r\nFinally let's remember that if we add `desc` as a new parameter to `map`, we should add it to the `ignore_kwargs` list of the `@fingerprint_transform` decorator of `Dataset._map_single` since we don't want this parameter to affect the fingerprint of the resulting dataset." ]

[ "cc @beurkinger but I think this has been fixed internally and will soon be updated right ?", "This issue has been fixed." ]

[ "Upgrading datasets to version 1.6 fixes the issue", "This bug was fixed in #1995. Upgrading the `datasets` should work! ", "Thanks @ekeleshian for having reported.\r\n\r\nI am closing this issue once that you updated `datasets`. Feel free to reopen it if the problem persists." ]

[ "I tried upgrading to `datasets==1.6.2` and downgrading to `1.6.0`. Both versions produce the same output.\r\n\r\nDowngrading to `1.5.0` works and produces the following output for me:\r\n\r\n```bash\r\nDownloading: 9.20kB [00:00, 3.94MB/s] \r\nDownloading: 5.99kB [00:00, 3.29MB/s] \r\nNo config specified, defaulting to: sst/default\r\nDownloading and preparing dataset sst/default (download: 6.83 MiB, generated: 3.73 MiB, post-processed: Unknown size, total: 10.56 MiB) to /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b...\r\n Dataset sst downloaded and prepared to /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b. Subsequent calls will reuse this data.\r\nexecuted [0, 1]\r\n#0: 0%| | 0/5 [00:00<?, ?ba/s]\r\n#1: 0%| | 0/5 [00:00<?, ?ba/s]\r\nexecuted [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]\r\nexecuted [4272, 4273, 4274, 4275, 4276, 4277, 4278, 4279, 4280, 4281]\r\nexecuted [1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009]\r\nexecuted [5272, 5273, 5274, 5275, 5276, 5277, 5278, 5279, 5280, 5281]\r\nexecuted [2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009]\r\nexecuted [6272, 6273, 6274, 6275, 6276, 6277, 6278, 6279, 6280, 6281]\r\nexecuted [3000, 3001, 3002, 3003, 3004, 3005, 3006, 3007, 3008, 3009]\r\nexecuted [7272, 7273, 7274, 7275, 7276, 7277, 7278, 7279, 7280, 7281]\r\nexecuted [4000, 4001, 4002, 4003, 4004, 4005, 4006, 4007, 4008, 4009]\r\n#0: 100%|██████████| 5/5 [00:00<00:00, 94.83ba/s]\r\nexecuted [8272, 8273, 8274, 8275, 8276, 8277, 8278, 8279, 8280, 8281]\r\n#1: 100%|██████████| 5/5 [00:00<00:00, 92.75ba/s]\r\nexecuted [0, 1]\r\n#0: 0%| | 0/1 [00:00<?, ?ba/s]\r\n#1: 0%| | 0/1 [00:00<?, ?ba/s]\r\nexecuted [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]\r\nexecuted [551, 552, 553, 554, 555, 556, 557, 558, 559, 560]\r\n#0: 100%|██████████| 1/1 [00:00<00:00, 118.81ba/s]\r\n#1: 100%|██████████| 1/1 [00:00<00:00, 123.06ba/s]\r\nexecuted [0, 1]\r\n#0: 0%| | 0/2 [00:00<?, ?ba/s]\r\n#1: 0%| | 0/2 [00:00<?, ?ba/s]\r\nexecuted [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]\r\nexecuted [1105, 1106, 1107, 1108, 1109, 1110, 1111, 1112, 1113, 1114]\r\nexecuted [1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009]\r\n#0: 100%|██████████| 2/2 [00:00<00:00, 119.42ba/s]\r\nexecuted [2105, 2106, 2107, 2108, 2109, 2110, 2111, 2112, 2113, 2114]\r\n#1: 100%|██████████| 2/2 [00:00<00:00, 123.33ba/s]\r\n\r\n\r\n\r\n ############################## \r\n\r\n\r\n\r\nexecuted [0, 1]\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-6079777aa097c8f8.arrow\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-2dc05c46f68eda6e.arrow\r\nexecuted [0, 1]\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-1ca347e7430b98f1.arrow\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-c0f1a73ce3ba40cd.arrow\r\nexecuted [0, 1]\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-832a1407bf1ac5b7.arrow\r\nLoading cached processed dataset at /home/johannes/.cache/huggingface/datasets/sst/default/1.0.0/a16a45566b63b2c3179e6c1d0f8edadde56e45570ee8cf99394fbb738491d34b/cache-036316a259b773c4.arrow\r\n- Datasets: 1.5.0\r\n- Python: 3.8.3 (default, May 19 2020, 18:47:26) \r\n[GCC 7.3.0]\r\n- Platform: Linux-5.4.0-72-generic-x86_64-with-glibc2.10\r\n```", "Hi,\r\n\r\nset `keep_in_memory` to False when loading a dataset (`sst = load_dataset(\"sst\", keep_in_memory=False)`) to prevent it from loading in-memory. Currently, in-memory datasets fail to find cached files due to this check (always False for them):\r\n\r\nhttps://github.com/huggingface/datasets/blob/241a0b4a3a868778ee91e767ad406f9da7610df2/src/datasets/arrow_dataset.py#L1718\r\n\r\n@albertvillanova It seems like this behavior was overlooked in #2182.\r\n\r\n", "Hi @villmow, thanks for reporting. \r\n\r\nAs @mariosasko has pointed out, we did not consider this case when introducing the feature of automatic in-memory for small datasets. This needs to be fixed.", "Hi ! Currently a dataset that is in memory doesn't know doesn't know in which directory it has to read/write cache files.\r\nOn the other hand, a dataset that loaded from the disk (via memory mapping) uses the directory from which the dataset is located to read/write cache files.\r\n\r\nBecause of that, currently in-memory datasets simply don't use caching.\r\n\r\nMaybe a Dataset object could have a `cache_dir` that is set to the directory where the arrow files are created during `load_dataset` ?", "Fixed once reverted the default in-memory feature:\r\nClosed by #2460 (to close issue #2458).", "Please @villmow, feel free to update to `Datasets` latest version (1.8)." ]

[ "Thanks for reporting, @sgraaf.\r\n\r\nI am going to have a look at it. \r\n\r\nI guess the expected codec is \"UTF-8\". Normally, when no explicitly codec is passed, Python uses one which is platform-dependent. For Linux machines, the default codec is `utf_8`, which is OK. However for Windows machine, the default codec is `cp1252`, which causes the problem.", "Awesome, thank you. 😃 ", "@sgraaf, I have just merged the fix in the master branch.\r\n\r\nYou can either:\r\n- install `datasets` from source code\r\n- wait until we make the next release of `datasets`\r\n- set the `utf-8` codec as your default instead of `cp1252`. This can be done by activating the Python [UTF-8 mode](https://www.python.org/dev/peps/pep-0540) either by passing the command-line option `-X utf8` or by setting the environment variable `PYTHONUTF8=1`." ]

[ "Hi @richardliaw, \r\n\r\nFirst, thanks for the compliments.\r\n\r\nIn relation with your request, currently, the dynamic modules path is obtained this way:\r\n```python\r\nfrom datasets.load import init_dynamic_modules, MODULE_NAME_FOR_DYNAMIC_MODULES\r\n\r\ndynamic_modules_path = init_dynamic_modules(MODULE_NAME_FOR_DYNAMIC_MODULES)\r\n```\r\n\r\nLet me know if it is OK for you this way. \r\n\r\nI could set `MODULE_NAME_FOR_DYNAMIC_MODULES` as default value, so that you could instead obtain the path with:\r\n```\r\ndynamic_modules_path = datasets.load.init_dynamic_modules()\r\n```", "Hi @albertvillanova, the default value proposal seems great :) Looking forward to this!", "I like the idea as well ! thanks @albertvillanova ", "Hi @richardliaw, the feature is on the master branch and will be included in the next release in a couple of weeks.", "awesome work @albertvillanova !" ]

[ "Fixed by #2317." ]

[ "Hi @gchhablani, \r\n\r\nThere are some 3rd-party dependencies that require to build code in C. In this case, it is the library `python-Levenshtein`.\r\n\r\nOn Windows, in order to be able to build C code, you need to install at least `Microsoft C++ Build Tools` version 14. You can find more info here: https://visualstudio.microsoft.com/visual-cpp-build-tools/", "Hi @albertvillanova \r\n\r\nSorry for such a trivial issue ;-; \r\n\r\nThanks a lot." ]

[ "I'm happy to take this on:) One question: The original unlabelled data is stored unsegmented (see e.g. https://github.com/facebookresearch/voxpopuli/blob/main/voxpopuli/get_unlabelled_data.py#L30), but segmenting the audio in the dataset would require a dependency on something like soundfile or torchaudio. An alternative could be to provide the segments start and end times as a Sequence and then it's up to the user to perform the segmentation on-the-fly if they wish?", "Hey @jfainberg,\r\n\r\nThis sounds great! I think adding a dependency would not be a big problem, however automatically segmenting the data probably means that it would take a very long time to do:\r\n\r\n```python\r\ndataset = load_dataset(\"voxpopuli\", \"french\")\r\n```\r\n\r\n=> so as a start I think your option 2 is the way to go!", "@polinaeterna VoxPopuli is available [here](https://huggingface.co/datasets/facebook/voxpopuli), so we can close this issue, right?\r\n", "@mariosasko yes, sure, closing it" ]

[]

[]

[ "Hi ! Have you tried other values for `preprocessing_num_workers` ? Is it always process 0 that is slower ?\r\nThere are no difference between process 0 and the others except that it processes the first shard of the dataset.", "Hi, I have found the reason of it. Before using the map function to tokenize the data, I concatenate the wikipedia and bookcorpus first, like this:\r\n```if args.dataset_name1 is not None:\r\n dataset1 = load_dataset(args.dataset_name1, args.dataset_config_name1, split=\"train\")\r\n dataset1 = dataset1.remove_columns('title')\r\n if args.dataset_name2 is not None:\r\n dataset2 = load_dataset(args.dataset_name2, args.dataset_config_name2,split=\"train\")\r\n assert dataset1.features.type == dataset2.features.type, str(dataset1.features.type)+';'+str(dataset2.features.type)\r\n datasets12 = concatenate_datasets([dataset1, dataset2], split='train')\r\n```\r\nWhen I just use one datasets, e.g. wikipedia, the problem seems no longer exist:\r\n\r\n\r\nBookcorpus has more row numbers than Wikipedia, however, it takes much more time to process each batch of wiki than that of bookcorpus. When we first concatenate two datasets and then use _map_ to process the concatenated datasets, e.g. `num_proc=5`, process 0 has to process all of the wikipedia data, causing the problem that #0 takes a longer time to finish the job. \r\n\r\nThe problem is caused by the different characteristic of different datasets. One solution might be using _map_ first to process two datasets seperately, then concatenate the tokenized and processed datasets before input to the `Dataloader`.\r\n\r\n", "That makes sense ! You can indeed use `map` on both datasets separately and then concatenate.\r\nAnother option is to concatenate, then shuffle, and then `map`." ]

[ "Hi,\r\n\r\nthis is not a standard CSV file (requires additional preprocessing) so I wouldn't label this as s bug. You could parse the examples with the regex module or the string API to extract the data, but the following approach is probably the easiest (once you load the data):\r\n```python\r\nimport ast\r\n# load the dataset and copy the features\r\ndef process(ex):\r\n return {\"tokens\": ast.literal_eval(ex[\"tokens\"]), \"labels\": ast.literal_eval(ex[\"labels\"])}\r\ndataset = dataset.map(process, features=new_features)\r\n```\r\n", "Hi,\r\n\r\nThanks for the reply.\r\nI have already used ```ast.literal_eval``` to evaluate the string into list, but I was getting another error:\r\n```\r\nArrowInvalid: Could not convert X with type str: tried to convert to int\r\n```\r\nWhy this happens ? Should labels be mapped to their ids and use int instead of str ?", "Yes, just map the labels to their ids." ]

[ "\r\nI received an answer for this question on the HuggingFace Datasets forum by @lhoestq\r\n\r\nHi !\r\n\r\nIf you want to tokenize line by line, you can use this:\r\n\r\n```\r\nmax_seq_length = 512\r\nnum_proc = 4\r\n\r\ndef tokenize_function(examples):\r\n# Remove empty lines\r\nexamples[\"text\"] = [line for line in examples[\"text\"] if len(line) > 0 and not line.isspace()]\r\nreturn tokenizer(\r\n examples[\"text\"],\r\n truncation=True,\r\n max_length=max_seq_length,\r\n)\r\n\r\ntokenized_dataset = dataset.map(\r\ntokenize_function,\r\nbatched=True,\r\nnum_proc=num_proc,\r\nremove_columns=[\"text\"],\r\n)\r\n```\r\n\r\nThough the TextDataset was doing a different processing by concatenating all the texts and building blocks of size 512. If you need this behavior, then you must apply an additional map function after the tokenization:\r\n\r\n```\r\n# Main data processing function that will concatenate all texts from\r\n# our dataset and generate chunks of max_seq_length.\r\ndef group_texts(examples):\r\n# Concatenate all texts.\r\nconcatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}\r\ntotal_length = len(concatenated_examples[list(examples.keys())[0]])\r\n# We drop the small remainder, we could add padding if the model supported it instead of this drop,\r\n# you can customize this part to your needs.\r\ntotal_length = (total_length // max_seq_length) * max_seq_length\r\n# Split by chunks of max_len.\r\nresult = {\r\n k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]\r\n for k, t in concatenated_examples.items()\r\n}\r\nreturn result\r\n\r\n# Note that with `batched=True`, this map processes 1,000 texts together,\r\n# so group_texts throws away a remainder for each of those groups of 1,000 texts.\r\n# You can adjust that batch_size here but a higher value might be slower to preprocess.\r\n\r\ntokenized_dataset = tokenized_dataset.map(\r\ngroup_texts,\r\nbatched=True,\r\nnum_proc=num_proc,\r\n)\r\n```\r\n\r\nThis code comes from the processing of the run_mlm.py example script of transformers\r\n\r\n", "Resolved" ]

[ "From the trace this seems like an error in the tokenizer library instead.\r\n\r\nDo you mind opening an issue at https://github.com/huggingface/tokenizers instead?", "Hi @tginart, thanks for reporting.\r\n\r\nI think this issue is already open at `tokenizers` library: https://github.com/huggingface/tokenizers/issues/685" ]

[ "Hi ! I think you might have to ask on the `transformers` repo on or the forum at https://discuss.huggingface.co/\r\n\r\nClosing since it's not related to this library" ]

[ "Therefore, when I try to concatenate larger datasets (5x 35GB data sets) I also get an out of memory error, since over 90GB of swap space was used at the time of the crash:\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nMemoryError Traceback (most recent call last)\r\n<ipython-input-6-9766d77530b9> in <module>\r\n 20 print(file_name)\r\n 21 cv_batch = load_from_disk(file_name)\r\n---> 22 cv_sampled_train = concatenate_datasets([cv_sampled_train, cv_batch])\r\n 23 \r\n 24 print(\"Saving to disk!\")\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\arrow_dataset.py in concatenate_datasets(dsets, info, split, axis)\r\n 2891 \r\n 2892 # Concatenate tables\r\n-> 2893 table = concat_tables([dset._data for dset in dsets if len(dset._data) > 0], axis=axis)\r\n 2894 table = update_metadata_with_features(table, None)\r\n 2895 \r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\table.py in concat_tables(tables, axis)\r\n 837 if len(tables) == 1:\r\n 838 return tables[0]\r\n--> 839 return ConcatenationTable.from_tables(tables, axis=axis)\r\n 840 \r\n 841 \r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\table.py in from_tables(cls, tables, axis)\r\n 697 return result\r\n 698 \r\n--> 699 blocks = to_blocks(tables[0])\r\n 700 for table in tables[1:]:\r\n 701 table_blocks = to_blocks(table)\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\table.py in to_blocks(table)\r\n 669 return [[InMemoryTable(table)]]\r\n 670 elif isinstance(table, ConcatenationTable):\r\n--> 671 return copy.deepcopy(table.blocks)\r\n 672 else:\r\n 673 return [[table]]\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 151 copier = getattr(x, \"__deepcopy__\", None)\r\n 152 if copier is not None:\r\n--> 153 y = copier(memo)\r\n 154 else:\r\n 155 reductor = dispatch_table.get(cls)\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\table.py in __deepcopy__(self, memo)\r\n 143 # by adding it to the memo, self.table won't be copied\r\n 144 memo[id(self.table)] = self.table\r\n--> 145 return _deepcopy(self, memo)\r\n 146 \r\n 147 def __getstate__(self):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\datasets\\table.py in _deepcopy(x, memo)\r\n 62 memo[id(x)] = result\r\n 63 for k, v in x.__dict__.items():\r\n---> 64 setattr(result, k, copy.deepcopy(v, memo))\r\n 65 return result\r\n 66 \r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 170 y = x\r\n 171 else:\r\n--> 172 y = _reconstruct(x, memo, *rv)\r\n 173 \r\n 174 # If is its own copy, don't memoize.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)\r\n 262 if deep and args:\r\n 263 args = (deepcopy(arg, memo) for arg in args)\r\n--> 264 y = func(*args)\r\n 265 if deep:\r\n 266 memo[id(x)] = y\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in <genexpr>(.0)\r\n 261 deep = memo is not None\r\n 262 if deep and args:\r\n--> 263 args = (deepcopy(arg, memo) for arg in args)\r\n 264 y = func(*args)\r\n 265 if deep:\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 170 y = x\r\n 171 else:\r\n--> 172 y = _reconstruct(x, memo, *rv)\r\n 173 \r\n 174 # If is its own copy, don't memoize.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)\r\n 262 if deep and args:\r\n 263 args = (deepcopy(arg, memo) for arg in args)\r\n--> 264 y = func(*args)\r\n 265 if deep:\r\n 266 memo[id(x)] = y\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in <genexpr>(.0)\r\n 261 deep = memo is not None\r\n 262 if deep and args:\r\n--> 263 args = (deepcopy(arg, memo) for arg in args)\r\n 264 y = func(*args)\r\n 265 if deep:\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_tuple(x, memo, deepcopy)\r\n 208 \r\n 209 def _deepcopy_tuple(x, memo, deepcopy=deepcopy):\r\n--> 210 y = [deepcopy(a, memo) for a in x]\r\n 211 # We're not going to put the tuple in the memo, but it's still important we\r\n 212 # check for it, in case the tuple contains recursive mutable structures.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in <listcomp>(.0)\r\n 208 \r\n 209 def _deepcopy_tuple(x, memo, deepcopy=deepcopy):\r\n--> 210 y = [deepcopy(a, memo) for a in x]\r\n 211 # We're not going to put the tuple in the memo, but it's still important we\r\n 212 # check for it, in case the tuple contains recursive mutable structures.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_tuple(x, memo, deepcopy)\r\n 208 \r\n 209 def _deepcopy_tuple(x, memo, deepcopy=deepcopy):\r\n--> 210 y = [deepcopy(a, memo) for a in x]\r\n 211 # We're not going to put the tuple in the memo, but it's still important we\r\n 212 # check for it, in case the tuple contains recursive mutable structures.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in <listcomp>(.0)\r\n 208 \r\n 209 def _deepcopy_tuple(x, memo, deepcopy=deepcopy):\r\n--> 210 y = [deepcopy(a, memo) for a in x]\r\n 211 # We're not going to put the tuple in the memo, but it's still important we\r\n 212 # check for it, in case the tuple contains recursive mutable structures.\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 144 copier = _deepcopy_dispatch.get(cls)\r\n 145 if copier is not None:\r\n--> 146 y = copier(x, memo)\r\n 147 else:\r\n 148 if issubclass(cls, type):\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in _deepcopy_list(x, memo, deepcopy)\r\n 203 append = y.append\r\n 204 for a in x:\r\n--> 205 append(deepcopy(a, memo))\r\n 206 return y\r\n 207 d[list] = _deepcopy_list\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\copy.py in deepcopy(x, memo, _nil)\r\n 159 reductor = getattr(x, \"__reduce_ex__\", None)\r\n 160 if reductor is not None:\r\n--> 161 rv = reductor(4)\r\n 162 else:\r\n 163 reductor = getattr(x, \"__reduce__\", None)\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\pyarrow\\io.pxi in pyarrow.lib.Buffer.__reduce_ex__()\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\pyarrow\\io.pxi in pyarrow.lib.Buffer.to_pybytes()\r\n\r\nMemoryError: \r\n\r\n```", "Hi ! this looks like an important issue. Let me try to reproduce this.\r\nCc @samsontmr this might be related to the memory issue you have in #2134 ", "@lhoestq Just went to open a similar issue.\r\n\r\nIt seems like deep copying (tested on master) the dataset object writes the table's record batches (`dset._data._batches`) into RAM.\r\n\r\nTo find the bug, I modified the `_deepcopy` function in `table.py` as follows:\r\n```python\r\ndef _deepcopy(x, memo: dict):\r\n \"\"\"deepcopy a regular class instance\"\"\"\r\n import psutil # pip install this package\r\n import time\r\n cls = x.__class__\r\n result = cls.__new__(cls)\r\n memo[id(x)] = result\r\n for k, v in x.__dict__.items():\r\n print(\"=\"* 50)\r\n print(\"Current memory:\", psutil.virtual_memory().percent)\r\n print(f\"Saving object {k} with value {v}\")\r\n setattr(result, k, copy.deepcopy(v, memo))\r\n time.sleep(5)\r\n print(\"Memory after copy:\", psutil.virtual_memory().percent)\r\n return result\r\n```\r\nTest script:\r\n```python\r\nimport copy\r\nfrom datasets import load_dataset\r\nbk = load_dataset(\"bookcorpus\", split=\"train\")\r\nbk_copy = copy.deepcopy(bk)\r\n```", "Thanks for the insights @mariosasko ! I'm working on a fix.\r\nSince this is a big issue I'll make a patch release as soon as this is fixed", "Hi @samsontmr @TaskManager91 the fix is on the master branch, feel free to install `datasets` from source and let us know if you still have issues", "We just released `datasets` 1.6.2 that includes the fix :)", "thanks it works like a charm! :)" ]

[ "Hi @puzzler10, \r\nThose examples where `gold_label` field was empty, -1 label was alloted to it. In order to remove it you can filter the samples from train/val/test splits. Here's how you can drop those rows from the dataset:\r\n`dataset = load_dataset(\"snli\")`\r\n`dataset_test_filter = dataset['test'].filter(lambda example: example['label'] != -1)`\r\n\r\nI agree it should have been mentioned in the documentation. I'll raise a PR regarding the same. Thanks for pointing out!" ]

[ "This has been fixed in this commit: https://github.com/huggingface/datasets/pull/2254/commits/88676c930216cd4cc31741b99827b477d2b46cb6\r\n\r\nIt was introduced in #2246 : using map with `input_columns` doesn't return the other columns anymore" ]

[ "See PR #2274 " ]

[ "Thanks for reporting ! We're looking into it", "I'm not able to reproduce this, do you think you can provide a code that creates a DatasetDict that has this issue when saving and reloading ?", "Hi, I just ran into a similar error. Here is the minimal code to reproduce:\r\n```python\r\nfrom datasets import load_dataset, DatasetDict\r\nds = load_dataset('super_glue', 'multirc')\r\n\r\nds.save_to_disk('tempds')\r\n\r\nds = DatasetDict.load_from_disk('tempds')\r\n\r\n```\r\n\r\n```bash\r\nReusing dataset super_glue (/home/idahl/.cache/huggingface/datasets/super_glue/multirc/1.0.2/2fb163bca9085c1deb906aff20f00c242227ff704a4e8c9cfdfe820be3abfc83)\r\nTraceback (most recent call last):\r\n File \"/home/idahl/eval-util-expl/multirc/tmp.py\", line 7, in <module>\r\n ds = DatasetDict.load_from_disk('tempds')\r\n File \"/home/idahl/miniconda3/envs/eval-util-expl/lib/python3.9/site-packages/datasets/dataset_dict.py\", line 710, in load_from_disk\r\n dataset_dict[k] = Dataset.load_from_disk(dataset_dict_split_path, fs, keep_in_memory=keep_in_memory)\r\n File \"/home/idahl/miniconda3/envs/eval-util-expl/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 687, in load_from_disk\r\n return Dataset(\r\n File \"/home/idahl/miniconda3/envs/eval-util-expl/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 274, in __init__\r\n raise ValueError(\r\nValueError: External features info don't match the dataset:\r\nGot\r\n{'answer': Value(dtype='string', id=None), 'idx': {'answer': Value(dtype='int32', id=None), 'paragraph': Value(dtype='int32', id=None), 'question': Value(dtype='int32', id=None)}, 'label': ClassLabel(num_classes=2, names=['False', 'True'], names_file=None, id=None), 'paragraph': Value(dtype='string', id=None), 'question': Value(dtype='string', id=None)}\r\nwith type\r\nstruct<answer: string, idx: struct<answer: int32, paragraph: int32, question: int32>, label: int64, paragraph: string, question: string>\r\n\r\nbut expected something like\r\n{'answer': Value(dtype='string', id=None), 'idx': {'paragraph': Value(dtype='int32', id=None), 'question': Value(dtype='int32', id=None), 'answer': Value(dtype='int32', id=None)}, 'label': Value(dtype='int64', id=None), 'paragraph': Value(dtype='string', id=None), 'question': Value(dtype='string', id=None)}\r\nwith type\r\nstruct<answer: string, idx: struct<paragraph: int32, question: int32, answer: int32>, label: int64, paragraph: string, question: string>\r\n\r\n```\r\n\r\nThe non-matching part seems to be\r\n`'label': ClassLabel(num_classes=2, names=['False', 'True'], names_file=None, id=None),`\r\nvs \r\n`'label': Value(dtype='int64', id=None),`\r\n\r\nAnd the order in the `<struct...` being different, which might cause the [features.type != inferred_features.type](https://github.com/huggingface/datasets/blob/master/src/datasets/arrow_dataset.py#L274) condition to become true and raise this ValueError.\r\n\r\n\r\nI am using datasets version 1.6.2.\r\n\r\nEdit: can confirm, this works without error in version 1.5.0", "My current workaround is to remove the idx feature:\r\n\r\n```\r\n\r\nfrom datasets import load_dataset, DatasetDict, Value\r\nds = load_dataset('super_glue', 'multirc')\r\nds = ds.remove_columns('idx')\r\n\r\nds.save_to_disk('tempds')\r\n\r\nds = DatasetDict.load_from_disk('tempds')\r\n\r\n```\r\n\r\nworks.", "It looks like this issue comes from the order of the fields in the 'idx' struct that is different for some reason.\r\nI'm looking into it. Note that as a workaround you can also flatten the nested features with `ds = ds.flatten()`", "I just pushed a fix on `master`. We'll do a new release soon !\r\n\r\nThanks for reporting" ]

[ "Thanks @bhavitvyamalik for the fix !\r\nThe fix will be available in the next release.\r\nIt's already available on the `master` branch. For now you can either install `datasets` from source or use `script_version=\"master\"` in `load_dataset` to use the fixed version of this dataset." ]

[ "Thanks for reporting ! We are working on this and we'll do a patch release very soon.", "We did a patch release to fix this issue.\r\nIt should be fixed in the new version 1.6.1\r\n\r\nThanks again for reporting and for the details :)" ]

[ "Hi ! Sorry to hear that. This may come from another issue then.\r\n\r\nFirst can we check if this latency comes from the dataset itself ?\r\nYou can try to load your dataset and benchmark the speed of querying random examples inside it ?\r\n```python\r\nimport time\r\nimport numpy as np\r\n\r\nfrom datasets import load_from_disk\r\n\r\ndataset = load_from_disk(...) # or from load_dataset...\r\n\r\n_start = time.time()\r\nn = 100\r\nfor i in np.random.default_rng(42).integers(0, len(dataset), size=n):\r\n _ = dataset[i]\r\nprint(time.time() - _start)\r\n```\r\n\r\nIf we see a significant speed difference between your two datasets then it would mean that there's an issue somewhere", "Hi @lhoestq, here is the result. I additionally measured time to `load_from_disk`:\r\n* 60GB\r\n```\r\nloading took: 22.618776321411133\r\nramdom indexing 100 times took: 0.10214924812316895\r\n```\r\n\r\n* 600GB\r\n```\r\nloading took: 1176.1764674186707\r\nramdom indexing 100 times took: 2.853600025177002\r\n```\r\n\r\nHmm.. I double checked that it's version 1.6.0. The difference seems quite big, could it be related to the running environment? \r\n", "I'm surprised by the speed change. Can you give more details about your dataset ?\r\nThe speed depends on the number of batches in the arrow tables and the distribution of the lengths of the batches.\r\nYou can access the batches by doing `dataset.data.to_batches()` (use only for debugging) (it doesn't bring data in memory).\r\n\r\nAlso can you explain what parameters you used if you used `map` calls ?\r\nAlso if you have some code that reproduces the issue I'd be happy to investigate it.", "Also if you could give us more info about your env like your OS, version of pyarrow and if you're using an HDD or a SSD", "Here are some details of my 600GB dataset. This is a dataset AFTER the `map` function and once I load this dataset, I do not use `map` anymore in the training. Regarding the distribution of the lengths, it is almost uniform (90% is 512 tokens, and 10% is randomly shorter than that -- typical setting for language modeling).\r\n```\r\nlen(batches):\r\n492763\r\n\r\nbatches[0]: \r\npyarrow.RecordBatch\r\nattention_mask: list<item: uint8>\r\n child 0, item: uint8\r\ninput_ids: list<item: int16>\r\n child 0, item: int16\r\nspecial_tokens_mask: list<item: uint8>\r\n child 0, item: uint8\r\ntoken_type_ids: list<item: uint8>\r\n child 0, item: uint8\r\n```\r\n\r\nHere the some parameters to `map` function just in case it is relevant:\r\n```\r\nnum_proc=1 # as multi processing is slower in my case\r\nload_from_cache_file=False\r\n```\r\n", "Regarding the environment, I am running the code on a cloud server. Here are some info:\r\n```\r\nUbuntu 18.04.5 LTS # cat /etc/issue\r\npyarrow 3.0.0 # pip list | grep pyarrow\r\n```\r\nThe data is stored in SSD and it is mounted to the machine via Network File System.\r\n\r\nIf you could point me to some of the commands to check the details of the environment, I would be happy to provide relevant information @lhoestq !", "I am not sure how I could provide you with the reproducible code, since the problem only arises when the data is big. For the moment, I would share the part that I think is relevant. Feel free to ask me for more info.\r\n\r\n```python\r\nclass MyModel(pytorch_lightning.LightningModule)\r\n def setup(self, stage):\r\n self.dataset = datasets.load_from_disk(path)\r\n self.dataset.set_format(\"torch\")\r\n\r\n def train_dataloader(self):\r\n collate_fn = transformers.DataCollatorForLanguageModeling(\r\n tokenizer=transformers.ElectraTokenizerFast.from_pretrained(tok_path)\r\n )\r\n dataloader = torch.utils.DataLoader(\r\n self.dataset,\r\n batch_size=32,\r\n collate_fn=collate_fn,\r\n num_workers=8,\r\n pin_memory=True,\r\n )\r\n```", "Hi ! Sorry for the delay I haven't had a chance to take a look at this yet. Are you still experiencing this issue ?\r\nI'm asking because the latest patch release 1.6.2 fixed a few memory issues that could have lead to slow downs", "Hi! I just ran the same code with different datasets (one is 60 GB and another 600 GB), and the latter runs much slower. ETA differs by 10x.", "@lhoestq and @hwijeen\r\n\r\nDespite upgrading to datasets 1.6.2, still experiencing extremely slow (2h00) loading for a 300Gb local dataset shard size 1.1Gb on local HDD (40Mb/s read speed). This corresponds almost exactly to total data divided by reading speed implying that it reads the entire dataset at each load.\r\n\r\nStack details:\r\n=========\r\n\r\n> GCC version: Could not collect\r\n> Clang version: Could not collect\r\n> CMake version: Could not collect\r\n> \r\n> Python version: 3.7 (64-bit runtime)\r\n> Is CUDA available: True\r\n> CUDA runtime version: 10.2.89\r\n> GPU models and configuration: GPU 0: GeForce GTX 1050\r\n> Nvidia driver version: 457.63\r\n> cuDNN version: C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.2\\bin\\cudnn64_7.dll\r\n> HIP runtime version: N/A\r\n> MIOpen runtime version: N/A\r\n> \r\n> Versions of relevant libraries:\r\n> [pip3] datasets==1.6.2\r\n> [pip3] transformers==4.5.1\r\n> [pip3] numpy==1.19.1\r\n> [pip3] numpydoc==1.1.0\r\n> [pip3] pytorch-metric-learning==0.9.98\r\n> [pip3] torch==1.8.1\r\n> [pip3] torchaudio==0.8.1\r\n> [pip3] torchvision==0.2.2\r\n> [conda] blas 2.16 mkl conda-forge\r\n> [conda] cudatoolkit 10.2.89 hb195166_8 conda-forge\r\n> [conda] libblas 3.8.0 16_mkl conda-forge\r\n> [conda] libcblas 3.8.0 16_mkl conda-forge\r\n> [conda] liblapack 3.8.0 16_mkl conda-forge\r\n> [conda] liblapacke 3.8.0 16_mkl conda-forge\r\n> [conda] mkl 2020.1 216\r\n> [conda] numpy 1.19.1 py37hae9e721_0 conda-forge\r\n> [conda] numpydoc 1.1.0 py_1 conda-forge\r\n> [conda] pytorch 1.8.1 py3.7_cuda10.2_cudnn7_0 pytorch\r\n> [conda] pytorch-metric-learning 0.9.98 pyh39e3cac_0 metric-learning\r\n> [conda] torchaudio 0.8.1 py37 pytorch\r\n> [conda] torchvision 0.2.2 py_3 pytorch", "Hi @BenoitDalFerro how do your load your dataset ?", "Hi @lhoestq thanks for the quick turn-around, actually the plain vanilla way, without an particular knack or fashion, I tried to look into the documentation for some alternative but couldn't find any\r\n\r\n> dataset = load_from_disk(dataset_path=os.path.join(datasets_dir,dataset_dir))", "I’m facing the same issue when loading a 900GB dataset (stored via `save_to_disk`): `load_from_disk(path_to_dir)` takes 1.5 hours and htop consistently shows high IO rates > 120 M/s.", "@tsproisl same here, smells like ~~teen spirit~~ intended generator inadvertently ending up iterator\r\n\r\n@lhoestq perhaps solution to detect bug location in code is to track its signature via HD read usage monitoring, option is to add tracking decorator on top each function and sequentially close all hatches from top to bottom, suggest PySmart https://pypi.org/project/pySMART/ a Smartmontools implementation", "I wasn't able to reproduce this on a toy dataset of around 300GB:\r\n\r\n```python\r\nimport datasets as ds\r\n\r\ns = ds.load_dataset(\"squad\", split=\"train\")\r\ns4000 = ds.concatenate_datasets([s] * 4000)\r\nprint(ds.utils.size_str(s4000.data.nbytes)) # '295.48 GiB'\r\n\r\ns4000.save_to_disk(\"tmp/squad_4000\")\r\n```\r\n\r\n```python\r\nimport psutil\r\nimport time\r\nfrom datasets import load_from_disk\r\n\r\ndisk = \"disk0\" # You may have to change your disk here\r\niocnt1 = psutil.disk_io_counters(perdisk=True)[disk]\r\ntime1 = time.time()\r\n\r\ns4000_reloaded = load_from_disk(\"tmp/squad_4000\")\r\n\r\ntime2 = time.time()\r\niocnt2 = psutil.disk_io_counters(perdisk=True)[disk]\r\n\r\nprint(f\"Blocks read {iocnt2.read_count - iocnt1.read_count}\") # Blocks read 18\r\nprint(f\"Elapsed time: {time2 - time1:.02f}s\") # Elapsed time: 14.60s\r\n```\r\n\r\nCould you run this on your side and tell me if how much time it takes ? Please run this when your machine is idle so that other processes don't interfere.\r\n\r\nI got these results on my macbook pro on datasets 1.6.2", "@lhoestq thanks, test running as we speak, bear with me", "Just tried on google colab and got ~1min for a 15GB dataset (only 200 times SQuAD), while it should be instantaneous. The time is spent reading the Apache Arrow table from the memory mapped file. This might come a virtual disk management issue. I'm trying to see if I can still speed it up on colab.", "@lhoestq what is Google Colab's HD read speed, is it possible to introspect incl. make like SSD or HDD ?", "@lhoestq Thank you! The issue is getting more interesting. The second script is still running, but it's definitely taking much longer than 15 seconds.", "Okay, here’s the ouput:\r\nBlocks read 158396\r\nElapsed time: 529.10s\r\n\r\nAlso using datasets 1.6.2. Do you have any ideas, how to pinpoint the problem?", "@lhoestq, @tsproisl mmmh still writing on my side about 1h to go, thinking on it are your large datasets all monoblock unsharded ? mine is 335 times 1.18Gb shards.", "The 529.10s was a bit too optimistic. I cancelled the reading process once before running it completely, therefore the harddrive cache probably did its work.\r\n\r\nHere are three consecutive runs\r\nFirst run (freshly written to disk):\r\nBlocks read 309702\r\nElapsed time: 1267.74s\r\nSecond run (immediately after):\r\nBlocks read 113944\r\nElapsed time: 417.55s\r\nThird run (immediately after):\r\nBlocks read 42518\r\nElapsed time: 199.19s\r\n", "@lhoestq \r\nFirst test\r\n> elapsed time: 11219.05s\r\n\r\nSecond test running bear with me, for Windows users slight trick to modify original \"disk0\" string:\r\n\r\nFirst find physical unit relevant key in dictionnary\r\n```\r\nimport psutil\r\npsutil.disk_io_counters(perdisk=True)\r\n```\r\n\r\n> {'PhysicalDrive0': sdiskio(read_count=18453286, write_count=4075333, read_bytes=479546467840, write_bytes=161590275072, read_time=20659, write_time=2464),\r\n> 'PhysicalDrive1': sdiskio(read_count=1495778, write_count=388781, read_bytes=548628622336, write_bytes=318234849280, read_time=426066, write_time=19085)}\r\n\r\nIn my case it's _PhysicalDrive1_\r\n\r\nThen insert relevant key's string as _disk_ variable\r\n\r\n```\r\npsutil.disk_io_counters()\r\ndisk = 'PhysicalDrive1' # You may have to change your disk here\r\niocnt1 = psutil.disk_io_counters(perdisk=True)[disk]\r\ntime1 = time.time()\r\ns4000_reloaded = load_from_disk(\"your path here\")\r\ntime2 = time.time()\r\niocnt2 = psutil.disk_io_counters(perdisk=True)[disk]\r\nprint(f\"Blocks read {iocnt2.read_count - iocnt1.read_count}\") # Blocks read 18\r\nprint(f\"Elapsed time: {time2 - time1:.02f}s\") # Elapsed time: 14.60s\r\n```", "@lhoestq\r\nSecond test\r\n\r\n> Blocks read 1265609\r\n> Elapsed time: 11216.55s", "@lhoestq any luck ?", "Unfortunately no. Thanks for running the benchmark though, it shows that you machine does a lot of read operations. This is not expected: in other machines it does almost no read operations which enables a very fast loading.\r\n\r\nI did some tests on google colab and have the same issue. The first time the dataset arrow file is memory mapped takes always a lot of time (time seems linear with respect to the dataset size). Reloading the dataset is then instantaneous since the arrow file has already been memory mapped.\r\n\r\nI also tried using the Arrow IPC file format (see #1933) instead of the current streaming format that we use but it didn't help.\r\n\r\nMemory mapping is handled by the OS and depends on the disk you're using, so I'm not sure we can do much about it. I'll continue to investigate anyway, because I still don't know why in some cases it would go through the entire file (high `Blocks read ` as in your tests) and in other cases it would do almost no reading.", "@lhoestq thanks for the effort, let's stay in touch", "Just want to say that I am seeing the same issue. Dataset size if 268GB and it takes **3 hours** to load `load_from_disk`, using dataset version `1.9.0`. Filesystem underneath is `Lustre` ", "Hi @lhoestq, confirmed Windows issue, exact same code running on Linux OS total loading time about 3 minutes.", "Hmm that's different from what I got. I was on Ubuntu when reporting the initial issue." ]

[]

[ "Hi,\r\n\r\n1. try\r\n ```python\r\n dataset = load_dataset(\"text\", data_files={\"train\": [\"a1.txt\", \"b1.txt\"], \"test\": [\"c1.txt\"]})\r\n ```\r\n instead.\r\n\r\n Sadly, I can't reproduce the error on my machine. If the above code doesn't resolve the issue, try to update the library to the \r\n newest version (`pip install datasets --upgrade`).\r\n\r\n2. https://github.com/huggingface/transformers/blob/3ed5e97ba04ce9b24b4a7161ea74572598a4c480/examples/pytorch/language-modeling/run_mlm.py#L258-L259\r\nThis is the original code. You'll have to modify the example source to work with multiple train files. To make it easier, let's say \"|\" will act as a delimiter between files:\r\n ```python\r\n if data_args.train_file is not None:\r\n data_files[\"train\"] = data_args.train_file.split(\"|\") # + .split(\"|\")\r\n ```\r\n Then call the script as follows (**dataset_name must be None**):\r\n ```bash\r\n python run_mlm.py [... other args] --train_file a1.txt|b1.txt\r\n ```", "i meet the same error with datasets 1.11.0, is there any insight about this?" ]

[ "Hi @villmow, thanks for reporting.\r\n\r\nCould you please try with the Datasets version 1.6? We released it yesterday and it fixes some issues about the processing speed. You can see the fix implemented by @lhoestq here: #2122.\r\n\r\nOnce you update Datasets, please confirm if the problem persists.", "Hi @albertvillanova, thanks for the reply. I just tried the new version and the problem still persists. \r\n\r\nDo I need to rebuild the saved dataset (which I load from disk) with the 1.6.0 version of datasets? My script loads this dataset and creates new datasets from it. I tried it without rebuilding.\r\n\r\nSee this short video of what happens. It does not create all processes at the same time:\r\n\r\nhttps://user-images.githubusercontent.com/2743060/115720139-0da3a500-a37d-11eb-833a-9bbacc70868d.mp4\r\n\r\n", "There can be a bit of delay between the creations of the processes but this delay should be the same for both your `map` calls. We should look into this.\r\nAlso if you hav some code that reproduces this issue on google colab that'd be really useful !\r\n\r\nRegarding the speed differences:\r\nThis looks like a similar issue as https://github.com/huggingface/datasets/issues/1992 who is experiencing the same speed differences between processes.\r\nThis is a known bug that we are investigating. As of now I've never managed to reproduce it on my machine so it's pretty hard for me to find where this issue comes from.\r\n", "Upgrade to 1.6.1 solved my problem somehow. I did not change any of my code, but now it starts all processes around the same time.", "Nice ! I'm glad this works now.\r\nClosing for now, but feel free to re-open if you experience this issue again." ]

[ "This should be fixed now!\r\n\r\ncc @srush " ]

[ "Hi @odellus, thanks for reporting.\r\n\r\nThe `wikihow` dataset has 2 versions:\r\n- `all`: Consisting of the concatenation of all paragraphs as the articles and the bold lines as the reference summaries.\r\n- `sep`: Consisting of each paragraph and its summary.\r\n\r\nTherefore, in order to load it, you have to specify which version you would like, for example:\r\n```python\r\ndataset = load_dataset('wikihow', 'all')\r\n```\r\n\r\nPlease, tell me if this solves your problem.", "Good call out. I did try that and that's when it told me to download the\ndataset. Don't believe I have tried it with local files. Will try first\nthing in the morning and get back to you.\n\nOn Mon, Apr 19, 2021, 11:17 PM Albert Villanova del Moral <\n***@***.***> wrote:\n\n> Hi @odellus <https://github.com/odellus>, thanks for reporting.\n>\n> The wikihow dataset has 2 versions:\n>\n> - all: Consisting of the concatenation of all paragraphs as the\n> articles and the bold lines as the reference summaries.\n> - sep: Consisting of each paragraph and its summary.\n>\n> Therefore, in order to load it, you have to specify which version you\n> would like, for example:\n>\n> dataset = load_dataset('wikihow', 'all')\n>\n> Please, tell me if this solves your problem.\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2239#issuecomment-823004146>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABDYI3HVRTBI2QT3BOG262DTJUL57ANCNFSM43GV5BZQ>\n> .\n>\n", "Hi @odellus, yes you are right.\r\n\r\nDue to the server where the `wikihow` dataset is hosted, the dataset can't be downloaded automatically by `huggingface` and you have to download it manually as you did.\r\n\r\nNevertheless, you have to specify which dataset version you would like to load anyway:\r\n```python\r\ndataset = load_dataset('wikihow', 'all', data_dir='./wikihow')\r\n```\r\nor\r\n```python\r\ndataset = load_dataset('wikihow', 'sep', data_dir='./wikihow')\r\n```\r\nI find that the instructions given by `huggingface` are not clear enough: I am going to fix this.\r\nPlease tell me if this eventually works for you.", "That was it. Thank you Albert!" ]

[ "@albertvillanova I would like to take this up. It would be great if you could point me as to how the dataset size is calculated in HF. Thanks!" ]

[]

[ "Hi ! Indeed there's no verification on the uniqueness nor the types of the keys.\r\nDo you already have some ideas of what you would like to implement and how ?", "Hey @lhoestq, thank you so much for the opportunity.\r\nAlthough I haven't had much experience with the HF Datasets code, after a careful look at how the `ArrowWriter` functions, I think we can implement this as follows:\r\n\r\n1. First, we would have to update the `ArrowWriter.write()` function here:\r\nhttps://github.com/huggingface/datasets/blob/fcd3c3c8e3b1d9a2f3686a496082e21f06591380/src/datasets/arrow_writer.py#L296\r\nso that it accepts an additional argument `key` which would be appended along with the example here after hashing.\r\n\r\n2. Then, we would need to create a `Hasher` class which will take the key as its input and return a hash for it (We might need to use some hash salt which can be passed to the ArrowWriter.writer() with value equal to the `split_name` for differentiating between same keys of different splits)\r\n\r\n We can use the `hashlib.md5` function for hashing which will conert each key to its byte code before hashing (depending on the data type of the key) **Thus, the `key` type will be verified here**.\r\n\r\n3. Now, we would have to edit this\r\nhttps://github.com/huggingface/datasets/blob/fcd3c3c8e3b1d9a2f3686a496082e21f06591380/src/datasets/arrow_writer.py#L257\r\n so that it iterates over each `(hash, example)` pair (sorted according to hash). We can then simply **check whether each hash is different from the previous hash** (since they will be sorted)\r\n\r\nHowever, since I'm not very familiar with how the data is being written on disk in the form of a table, I might need some guidance for Step 3. \r\nPlease let me know your thought on this. Thanks!", "Interesting !\r\nWe keep the dataset sorted in the order examples are generated by the builder (we expect the dataset builders to generate examples in deterministic order). Therefore I don't think we should shuffle the examples with the hashing. Let me know what you think.\r\nOther that that, I really like the idea of checking for keys duplicates in `write_examples_on_file` :)\r\n\r\nThis looks like a great plan ! Feel free to open a PR and ping me if you have questions or if I can help\r\n", "@lhoestq I'm glad you liked the idea!\r\nI think that since the keys will be unique and deterministic in the nature themselves, so even if we shuffle the examples according to the hash, a deterministic order would still be maintained (as the keys will always have the same hash, whenever the dataset is generated). \r\nAnd since, we are not dealing with time series data (which would require the data to be in original order), I don't think the order of examples would matter much, as long as the order is deterministic and constant for all users.\r\n\r\nI think that this is also what was originally envisioned as mentioned in the documentation here:\r\nhttps://github.com/huggingface/datasets/blob/6775661b19d2ec339784f3d84553a3996a1d86c3/src/datasets/builder.py#L973\r\n\r\nAlso, if we avoid this, we would need to keep track of all the hashed keys in some place and compare each individual key with all others. This can cause some major overhead as each dataset consists of tens of thousands of examples.\r\nLet me know your thoughts in it! I would be opening a PR soon :)", "When users load their own data, they expect the order to stay the same. I think that shuffling the data can make things inconvenient.\r\n\r\n> I think that this is also what was originally envisioned as mentioned in the documentation here:\r\n\r\nThis part was originally developed by tensorflow datasets, and tensorflow datasets indeed does the shuffling. However in this library this is probably not what we want in the general case. But if @albertvillanova and @thomwolf you have opinions on this please let us know.\r\n\r\n> Also, if we avoid this, we would need to keep track of all the hashed keys in some place and compare each individual key with all others. This can cause some major overhead as each dataset consists of tens of thousands of examples.\r\n\r\nMaybe we cam simply keep track of the hashes of of each batch being written ? The size of the batch when the data are save in arrow is 10 000 examples. This would only ensure that we don't have duplicates in each batch, but there might still be duplicates across batches. For 10 000 examples the hashes can just be stored as a python `set`.\r\n\r\nOtherwise if we want full deduplication, we need an extra tool that allows to temporarily save and query hashes that may need to use disk space rather than memory.", "Yes I think we want to keep the original order by default and only shuffle when the user ask for it (for instance by calling `dataset.shuffle()`). That’s how I had it in mind originally.", "Hey @lhoestq, I just had a more in-depth look at the original TFDS code about why the keys and hash were used in the first place.\r\n\r\nIn my opinion, the only use that the `hash(key)` serves is that it allows us to shuffle the examples in a deterministic order (as each example will always yield the same key and thus, the same hash on every system) so that the same dataset is generated for each user, irrespective of the order the examples are yielded by the dataset builder on different user systems.\r\n\r\nOtherwise, if we are not shuffling, then while yielding and writing the data, after getting the key and hashing it for an example, I can't quite see the use of the hash or the key. The hash will simply be generated for each example but not actually used anywhere?\r\n\r\n@lhoestq @thomwolf It would be great if you could explain a bit more about the usage of keys. Thanks!\r\n", "In `datasets` the keys are currently ignored.\r\nFor shuffling we don't use the keys. Instead we shuffle an array of indices. Since both the original order of the dataset and the indices shuffling are deterministic, then `dataset.shuffle` is deterministic as well.\r\nWe can use it to:\r\n1. detect duplicates\r\n2. verify that the generation order is indeed deterministic\r\n3. maybe more ?", "Thanks a lot @lhoestq. I think I understand what we need to do now. The keys can indeed be used for detecting duplicates in generated examples as well as ensuring the order.\r\n\r\n> Maybe we cam simply keep track of the hashes of of each batch being written ? The size of the batch when the data are save in arrow is 10 000 examples. This would only ensure that we don't have duplicates in each batch,\r\n\r\nI think that checking for duplicates in every batch independently would be sufficient as the probability of collisions using something like `MD5` is very low. I would be opening a draft PR soon. It would be great to have your guidance. Thanks!" ]

[ "Hi ! Sure sounds good. Also if you find other datasets that use tuples instead of str/int, you can also fix them !\r\nthanks :)", "@lhoestq I have sent a PR for fixing the issue. Would be great if you could have a look! Thanks!" ]

[ "I found the problem. I called `set_format` on some columns before. This makes it crash. Here is a complete example to reproduce:\r\n\r\n```python\r\nfrom datasets import load_dataset\r\nsst = load_dataset(\"sst\")\r\nsst.set_format(\"torch\", columns=[\"label\"], output_all_columns=True)\r\nds = sst[\"train\"]\r\n\r\n# crashes\r\nds.map(\r\n lambda x: {\"a\": list(range(20))},\r\n remove_columns=ds.column_names,\r\n load_from_cache_file=False,\r\n num_proc=1,\r\n batched=True,\r\n)\r\n```", "Thanks for reporting and for providing this code to reproduce the issue, this is really helpful !", "I merged a fix, it should work on `master` now :)\r\nWe'll do a new release soon !" ]

[]

[ "Hi,\r\n\r\ncurrently the datasets API doesn't have a dedicated function to remove duplicate rows, but since the LAMA dataset is not too big (it fits in RAM), we can leverage pandas to help us remove duplicates:\r\n```python\r\n>>> from datasets import load_dataset, Dataset\r\n>>> dataset = load_dataset('lama', split='train')\r\n>>> dataset = Dataset.from_pandas(dataset.to_pandas().drop_duplicates(subset=...)) # specify a subset of the columns to consider in a list or use all of the columns if None\r\n```\r\n\r\nNote that the same can be achieved with the `Dataset.filter` method but this would requrie some extra work (filter function, speed?).", "Oh, seems like my question wasn't specified well. I'm _not_ asking how to remove duplicates, but whether duplicates should be removed if I want to do the evaluation on the LAMA dataset as it was proposed in the original paper/repository? In other words, will I get the same result if evaluate on the de-duplicated dataset loaded from HF's `datasets` as the results I'd get if I use the original data format and data processing script in https://github.com/facebookresearch/LAMA? ", "So it looks like the person who added LAMA to the library chose to have one item per piece of evidence rather than one per relation - and in this case, there are duplicate pieces of evidence for the target relation\r\n\r\nIf I understand correctly, to reproduce reported results, you would have to aggregate predictions for the several pieces of evidence provided for each relation (each unique `uuid`), but the original authors will know better \r\n\r\ncc @fabiopetroni " ]

[ "Hi @nsaphra, thanks for reporting.\r\n\r\nThis issue was fixed in `datasets` version 1.3.0. Could you please update `datasets` and tell me if the problem persists?\r\n```shell\r\npip install -U datasets\r\n```", "There might be a bug in the conda version of `datasets` 1.2.1 where the datasets/metric scripts are downloaded from `master` instead of the `1.2.1` repo.\r\n\r\nYou can try setting the env var `HF_SCRIPTS_VERSION=\"1.2.1\"` as a workaround. Let me know if that helps.", "I just faced the same issue. I was using 1.2.1 from conda and received the same AttributeError complaining about 'add_start_docstrings'. Uninstalling the conda installed datasets and then installing the latest datasets (version 1.5.0) using pip install solved the issue for me. I don't like mixing up conda and pip installs in the same environments but this will have to do for now, until 1.5.0 is made available through conda.", "Yep, seems to have fixed things! The conda package could really do with an update. Thanks!" ]

[ "Hi ! Apparently the data are not available from this url anymore. We'll replace it with the new url when it's available", "I saw this on their website when we request to download the dataset:\r\n\r\n\r\nCan we still request them link for the dataset and make a PR? @lhoestq @yjernite ", "I've contacted Martin (first author of the fquad paper) regarding a possible new url. Hopefully we can get one soon !", "They now made a website to force people who want to use the dataset for commercial purposes to seek a commercial license from them ...", "The script has been adopted to support manual download from the website, so I'm closing this issue." ]

[ "Hi,\r\n\r\nI've already opened a PR with the fix. If you are in a hurry, just build the project from source and run:\r\n```bash\r\ndatasets-cli test datasets/lc_quad --save_infos --all_configs --ignore_verifications\r\n```\r\n\r\n", "Ah sorry, I tried searching but couldn't find any related PR. \r\n\r\nThank you! " ]

[ "Hi ! Yes this is an issue with `datasets<=1.5.0`\r\nThis issue has been fixed by #2122 , we'll do a new release soon :)\r\nFor now you can test it on the `master` branch.", "Hi, thank you for your answer. I did not realize that my issue stems from the same problem. " ]

[ "Hi ! The ClassLabel feature type encodes the labels as integers.\r\nThe integer corresponds to the index of the label name in the `names` list of the ClassLabel.\r\nHere that means that the labels are 'entailment' (0), 'neutral' (1), 'contradiction' (2).\r\n\r\nYou can get the label names back by using `a.features['label'].int2str(i)`.\r\n", "Hi! You can also easily reorder the label with the [`Dataset.align_labels_with_mapping`](https://huggingface.co/docs/datasets/master/en/process#align) method." ]

[ "Hi,\r\n\r\nthe output of the tokenizers is treated specially in the lib to optimize the dataset size (see the code [here](https://github.com/huggingface/datasets/blob/master/src/datasets/arrow_writer.py#L138-L141)). It looks like that one of the values in a dictionary returned by the tokenizer is out of the assumed range.\r\nCan you please provide a minimal reproducible example for more help?", "Hi @yana-xuyan, thanks for reporting.\r\n\r\nAs clearly @mariosasko explained, `datasets` performs some optimizations in order to reduce the size of the dataset cache files. And one of them is storing the field `special_tokens_mask` as `int8`, which means that this field can only contain integers between `-128` to `127`. As your message error states, one of the values of this field is `50259`, and therefore it cannot be stored as an `int8`.\r\n\r\nMaybe we could implement a way to disable this optimization and allow using any integer value; although the size of the cache files would be much larger.", "I'm facing same issue @mariosasko @albertvillanova \r\n\r\n```\r\nArrowInvalid: Integer value 50260 not in range: -128 to 127\r\n```\r\n\r\nTo reproduce:\r\n```python\r\nSPECIAL_TOKENS = ['<bos>','<eos>','<speaker1>','<speaker2>','<pad>']\r\nATTR_TO_SPECIAL_TOKEN = {\r\n 'bos_token': '<bos>', \r\n 'eos_token': '<eos>', \r\n 'pad_token': '<pad>',\r\n 'additional_special_tokens': ['<speaker1>', '<speaker2>']\r\n }\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"gpt2\", use_fast=False)\r\nnum_added_tokens =tokenizer.add_special_tokens(ATTR_TO_SPECIAL_TOKEN)\r\nvocab_size = len(self.tokenizer.encoder) + num_added_tokens\r\nvocab =tokenizer.get_vocab()\r\n\r\npad_index = tokenizer.pad_token_id\r\neos_index = tokenizer.eos_token_id\r\nbos_index = tokenizer.bos_token_id\r\nspeaker1_index = vocab[\"<speaker1>\"]\r\nspeaker2_index = vocab[\"<speaker2>\"]\r\n```\r\n\r\n```python\r\ntokenizer.decode(['50260'])\r\n'<speaker1>'\r\n```", "@mariosasko \r\nI am hitting this bug in the Bert tokenizer too. I see that @albertvillanova labeled this as a bug back in April. Has there been a fix released yet?\r\nWhat I did for now is to just disable the optimization in the HF library. @yana-xuyan and @thomas-happify, is that what you did and did that work for you?\r\n\r\n", "Hi @gregg-ADP, \r\n\r\nThis is still a bug.\r\n\r\nAs @albertvillanova has suggested, maybe it's indeed worth adding a variable to `config.py` to have a way to disable this behavior.\r\n\r\nIn the meantime, this forced optimization can be disabled by specifying `features` (of the returned examples) in the `map` call:\r\n```python\r\nfrom datasets import *\r\n... # dataset init\r\nds.map(process_example, features=Features({\"special_tokens_mask\": Sequence(Value(\"int32\")), ... rest of the features}) \r\n```\r\n\r\ncc @lhoestq so he is also aware of this issue", "Thanks for the quick reply @mariosasko. What I did was to changed the optimizer to use int32 instead of int8. \r\nWhat you're suggesting specifies the type for each feature explicitly without changing the HF code. This is definitely a better option. However, we are hitting a new error later:\r\n```\r\n File \"/Users/ccccc/PycharmProjects/aaaa-ml/venv-source/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 1051, in _call_impl\r\n return forward_call(*input, **kwargs)\r\nTypeError: forward() got an unexpected keyword argument 'pos'\r\n\r\n```\r\nWhere 'pos' is the name of a new feature we added. Do you agree that your way of fixing the optimizer issue will not fix our new issue? If not, I will continue with this optimizer fix until we resolve our other issue.\r\n", "Hi @gwc4github,\r\n\r\nthe fix was merged a few minutes ago, and it doesn't require any changes on the user side (e.g. no need for specifying `features`). If you find time, feel free to install `datasets` from master with:\r\n```\r\npip install git+https://github.com/huggingface/datasets.git\r\n```\r\nand let us know if it works for your use case! " ]

[ "Hi ! This might be related to #2153 \r\n\r\nYou're right the ArrowWriter should be initialized with `features=self.info.features` ! Good catch\r\nI'm opening a PR to fix this and also to figure out how it was not caught in the tests\r\n\r\nEDIT: opened #2201", "> Hi ! This might be related to #2153\r\n> \r\n> You're right the ArrowWriter should be initialized with `features=self.info.features` ! Good catch\r\n> I'm opening a PR to fix this and also to figure out how it was not caught in the tests\r\n> \r\n> EDIT: opened #2201\r\n\r\nGlad to hear that! Thank you for your fix, I'm new to huggingface, it's a fantastic project 😁" ]

[ "Hi ! Files that starts with `cache-*` are cached computation files, i.e. they are the cached results of map/filter/cast/etc. operations. For example if you used `map` on your dataset to transform it, then the resulting dataset is going to be stored and cached in a `cache-*` file. These files are used to avoid having to load the dataset in RAM, even after many transforms", "Thanks @lhoestq! Hmm.. that's strange because I specifically turned off auto caching, and saved mapped result, using `save_to_disk`, to another location. At this location, the following file is created:`355G\tcache-ed205e500a7dc44c.arrow`\r\n\r\nTo my observation, both `load_dataset` and `map` creates `cache-*` files, and I wonder what the `cache-*` file from `load_dataset` is for (as I believe the same information is stored in `json-train.arrow`.", "This is a wrong report -- `cache-*` files are created only my `map`, not by `load_dataset`. " ]

[ "Thanks for reporting @samsontmr.\r\n\r\nIt seems a backward compatibility issue...", "Thanks @samsontmr this should be fixed on master now\r\n\r\nFeel free to reopen if you're still having issues" ]

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

[ "Hi ! Yes we are working on making `filter` significantly faster. You can look at related PRs here: #2060 #2178 \r\n\r\nI think you can expect to have the fast version of `filter` available next week.\r\n\r\nWe'll make it only select one column, and we'll also make the overall filtering operation way faster by avoiding many arrow<->python conversions especially during writing.\r\n\r\nI'll let you know how it goes !", "@lhoestq Thanks for the response— it's great to hear that we'll be getting a much faster `filter` method soon. However, my use case does also involve using `map` over a single column in order to pre-compute roughly uniformly sized batches, and right now that is also very slow. Is there any plan to make `map` faster for single column operations?\r\n\r\nIf that's not a priority for the maintainers right now, I could try my hand at adding the feature, but I can't guarantee I would do a good job given my lack of familiarity with pyarrow.", "Currently the optimal setup for single-column computations is probably to do something like\r\n```python\r\nresult = dataset.map(f, input_columns=\"my_col\", remove_columns=dataset.column_names)\r\n```\r\nThis has two advantages:\r\n- input_columns=\"my_col\" allows to only read the column \"my_col\"\r\n- remove_columns=dataset.column_names makes `map` only keep the output of your function `f`, and it drops the other columns of the dataset instead of keeping them.\r\n\r\nLet me know if it improves speed on your side.\r\n\r\nYou can also get more speed by using `batched=True` and setting `num_proc=` for multiprocessing", "Hi @lhoestq ,\r\n\r\nI'm hijacking this issue, because I'm currently trying to do the approach you recommend:\r\n\r\n> Currently the optimal setup for single-column computations is probably to do something like\r\n> \r\n> ```python\r\n> result = dataset.map(f, input_columns=\"my_col\", remove_columns=dataset.column_names)\r\n> ```\r\n\r\nHere is my code: (see edit, in which I added a simplified version\r\n\r\n```\r\nThis is the error:\r\n```bash\r\npyarrow.lib.ArrowInvalid: Column 1 named tokens expected length 8964 but got length 1000\r\n```\r\nI wonder why this error occurs, when I delete every column? Can you give me a hint?\r\n\r\n### Edit:\r\nI preprocessed my dataset before (using map with the features argument) and saved it to disk. May this be part of the error? I can iterate over the\r\ncomplete dataset and print every sample before calling map. There seems to be no other problem with the dataset.\r\n\r\nI tried to simplify the code that crashes:\r\n\r\n```python\r\n# works\r\nlog.debug(dataset.column_names)\r\nlog.debug(dataset)\r\nfor i, sample in enumerate(dataset):\r\n log.debug(i, sample)\r\n\r\n# crashes\r\ncounted_dataset = dataset.map(\r\n lambda x: {\"a\": list(range(20))},\r\n input_columns=column,\r\n remove_columns=dataset.column_names,\r\n load_from_cache_file=False,\r\n num_proc=num_workers,\r\n batched=True,\r\n)\r\n```\r\n\r\n```\r\npyarrow.lib.ArrowInvalid: Column 1 named tokens expected length 20 but got length 1000\r\n```\r\n\r\nEdit2: \r\n\r\nMay this be a problem with a schema I set when preprocessing the dataset before? I tried to add the `features` argument to the function and then I get a new error:\r\n\r\n```python\r\n# crashes\r\ncounted_dataset = dataset.map(\r\n lambda x: {\"a\": list(range(20))},\r\n input_columns=column,\r\n remove_columns=dataset.column_names,\r\n load_from_cache_file=False,\r\n num_proc=num_workers,\r\n batched=True,\r\n features=datasets.Features(\r\n {\r\n \"a\": datasets.Sequence(datasets.Value(\"int32\"))\r\n }\r\n )\r\n)\r\n```\r\n\r\n```\r\n File \"env/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 1704, in _map_single\r\n writer.write_batch(batch)\r\n File \"env/lib/python3.8/site-packages/datasets/arrow_writer.py\", line 312, in write_batch\r\n col_type = schema.field(col).type if schema is not None else None\r\n File \"pyarrow/types.pxi\", line 1341, in pyarrow.lib.Schema.field\r\nKeyError: 'Column tokens does not exist in schema'\r\n```", "Hi ! Can you open a separate issue for that ?\r\nAlso if you could provide a google colab or a sample code to reproduce this issue that would be helpful.\r\nOn my side I was not able to reproduce this error.", "@lhoestq Sorry I'm just responding now. I'm currently using your recommendation for the map on a single column, and I've gotten it to be fast enough to sort of work for my use case by just setting `num_proc=10`, although it's still quite slow. It's clear that it is still loading the entirety of each row into memory and then discarding everything except the selected column, instead of exploiting the columnar data format to only load the selected column.\r\n\r\nMy code is like this:\r\n```\r\n self.dataset = self.dataset.sort('num_tokens')\r\n batch_dataset = self.dataset.map(\r\n\tcompute_uniform_sized_batches,\r\n\tbatched=True, batch_size=10_000, num_proc=10, input_columns=['num_tokens'],\r\n\tremove_columns=get_columns_all_equal(self.dataset),\r\n\twith_indices=True,\r\n\tfn_kwargs=dict(max_size=tokens_per_batch)\r\n)\r\nself.batches = {\r\n\tname: list(zip(split['start'], split['length']))\r\n\tfor name, split in batch_dataset.items()\r\n}\r\n```\r\nI find that the processes with higher IDs take significantly longer to complete, presumably because the dataset is sorted by article length and they're loading the entire article text into memory, instead of just the 'num_tokens' column.\r\n\r\nI should note that my batching procedure would work best if I just used `batch_size=None` and loaded the whole column into memory at once, but I found that this was intolerably slow and gave me no progress information, so I'm using the less than ideal `batch_size=10_000`.", "Hi @norabelrose ! I'm glad you managed to make this work on your side.\r\nRegarding memory usage, you can try to drop the columns that you don't want to use for your `map` for now.\r\n\r\nIn the future we'll try to find a way to not load unnecessary columns in memory in `map`. Currently the way it works is that it gets the batch as a python dict, then it updates it using the output of your mapping function, and finally it removes columns from `remove_columns`. Therefore for a moment some columns are loaded in memory even if you remove them or don't use them for your mapping function.\r\n\r\nIt would be nice to have a way to optimize memory for cases such as yours !", "@lhoestq After looking through the source code, it looks like the following solution has at least some chance of working:\r\n- refactor `Dataset.map()` so that the `input_columns` parameter is implemented by using the `self.formatted_as()` context manager with `columns=input_columns`\r\n- change `Dataset._getitem()` so that it passes `self._data.drop(drop_columns)` to the `query_table()` function whenever `format_columns` is non-None and `output_all_columns` is False, instead of `self._data` itself", "Looks like a great direction :)\r\nNote that `query_table` doesn't bring data into memory. Only `format_table` does.\r\nAlso the dataset may already have a format with `columns=` already defined so we would need to define the formatted `input_dataset` like:\r\n```python\r\n# before the `map` main for loop\r\ninput_columns = input_columns if input_columns is not None else self.column_names\r\nif not self._output_all_columns:\r\n columns = [col for col in input_columns if self._format_columns is None or col in self._format_columns]\r\n input_dataset = self.with_format(\r\n type=self._format_type,\r\n columns=columns\r\n )\r\nelse:\r\n # in this case we could find a way to filter both format_columns and unformatted columns eventually\r\n input_dataset = self\r\n# then input_dataset can be used in the main for loop of `map`\r\n```\r\n\r\nEDIT: oh and regarding streaming format versus file format for arrow, we plan to start using the file format #1933 at one point (though I'm not sure if it would improve performance)", "Good to know about `query_table` not bringing anything into memory. I was under the impression that it did because a while back I looked at my `map` operation in pdb and it looked like it was spending forever in line 93 of formatting.py, `return pa.concat_tables(....)`, although that was before the `fast_slice` interpolation search was implemented, so it may have had more to do with the slow ChunkedArray slice implementation than anything else.\r\n\r\nIf `query_table` is I/O free then the fix may be as simple as just adding this to line 1779 of arrow_dataset.py:\r\n```python\r\n# Only load the columns we actually need\r\nif input_columns:\r\n stack.enter_context(self.formatted_as(\r\n self._format_type,\r\n columns=input_columns,\r\n output_all_columns=False,\r\n **self._format_kwargs\r\n ))\r\n```\r\nIt's not clear to me why the `[col for col in input_columns if self._format_columns is None or col in self._format_columns]` check would be necessary— it seems like either `input_columns` should simply temporarily override the `_format_columns` within the `map` operation, or we should throw an error if there are any conflicts. Currently it doesn't look like this case is checked for at all within `map`, but maybe I'm just missing it.", "`query_table` simply slices/concatenates parts of the table. The actual data inside the table is not brought in memory.\r\nAlso I'm more in favor of declaring `input_dataset = self.with_format(...)` since `formatted_as` may update the dataset fingerprint of `self`, which is not expected when someone runs `map`.\r\n\r\n> It's not clear to me why the [col for col in input_columns if self._format_columns is None or col in self._format_columns] check would be necessary— it seems like either input_columns should simply temporarily override the _format_columns within the map operation, or we should throw an error if there are any conflicts. Currently it doesn't look like this case is checked for at all within map, but maybe I'm just missing it.\r\n\r\nActually yes we can just use input_columns. And we do need to add a check to make sure there are not conflicts or this could lead to confusing errors.", "That sounds good to me! I just submitted a PR (#2246) implementing your approach. I also changed how `_query_table` handles Iterable keys since it still seemed like `pa.concat_tables` was taking a long time to create the table for each batch. Now my whole `map()` operation takes 1 min 46 seconds where it used to take somewhere on the order of 10 minutes." ]

[ "Hi @anassalamah,\r\n\r\nCould you please try with this:\r\n```python\r\ntrain_ds = load_dataset(\"news_commentary\", lang1=\"ar\", lang2=\"en\", split='train[:98%]')\r\nval_ds = load_dataset(\"news_commentary\", lang1=\"ar\", lang2=\"en\", split='train[98%:]')\r\n```", "Hello @albertvillanova, \r\n\r\nThanks for the suggestion. I didn't know you could do that. however, it didn't resolve the issue\r\n\r\n\r\n" ]

[ "Hi ! We refactored save_to_disk in #2025 so this doesn't happen.\r\nFeel free to try it on master for now\r\nWe'll do a new release soon" ]

[ "Hi ! Thanks for reporting\r\nIf I recall correctly this has been recently fixed #1995\r\nCan you try to upgrade your local version of `datasets` ?\r\n```\r\npip install --upgrade datasets\r\n```", "Hi Ihoestq,\r\n\r\nThank you. It works after upgrading the datasets\r\n" ]

[ "An educated guess: does this refer to the fact that depending on the custom column names in the dataset files (csv in this case), there is a dataset loader being created? and this dataset loader - using the \"custom data configuration\" is used among all jobs running using this particular csv files? (thinking out loud here...)\r\n\r\nIf this is the case, it may be ok for my use case (have to think about it more), still a bit surprising given that datasets caching is disabled (or so I hope) by the lines I pasted above. ", "Hi ! Currently disabling the caching means that all the dataset transform like `map`, `filter` etc. ignore the cache: it doesn't write nor read processed cache files.\r\nHowever `load_dataset` reuses datasets that have already been prepared: it does reload prepared dataset files.\r\n\r\nIndeed from the documentation:\r\n> datasets.set_caching_enabled(boolean: bool)\r\n\r\n> When applying transforms on a dataset, the data are stored in cache files. The caching mechanism allows to reload an existing cache file if it’s already been computed.\r\n> Reloading a dataset is possible since the cache files are named using the dataset fingerprint, which is updated after each transform.\r\n> If disabled, the library will no longer reload cached datasets files when applying transforms to the datasets. More precisely, if the caching is disabled:\r\n> - cache files are always recreated\r\n> - cache files are written to a temporary directory that is deleted when session closes\r\n> - cache files are named using a random hash instead of the dataset fingerprint - use datasets.Dataset.save_to_disk() to save a transformed dataset or it will be deleted when session closes\r\n> - caching doesn’t affect datasets.load_dataset(). If you want to regenerate a dataset from scratch you should use the download_mode parameter in datasets.load_dataset().", "Thank you for the clarification. \r\n\r\nThis is a bit confusing. On one hand, it says that cache files are always recreated and written to a temporary directory that is removed; on the other hand the last bullet point makes me think that since the default according to the docs for `download_mode (Optional datasets.GenerateMode) – select the download/generate mode - Default to REUSE_DATASET_IF_EXISTS` => it almost sounds that it could reload prepared dataset files. Where are these files stored? I guess not in the temporary directory that is removed... \r\n\r\nI find this type of api design error-prone. When I see as a programmer `datasets.set_caching_enabled(False)` I expect no reuse of anything in the cache. ", "It would be nice if the documentation elaborated on all the possible values for `download_mode` and/or a link to `datasets.GenerateMode`. \r\nThis info here:\r\n```\r\n \"\"\"`Enum` for how to treat pre-existing downloads and data.\r\n The default mode is `REUSE_DATASET_IF_EXISTS`, which will reuse both\r\n raw downloads and the prepared dataset if they exist.\r\n The generations modes:\r\n | | Downloads | Dataset |\r\n | -----------------------------------|-----------|---------|\r\n | `REUSE_DATASET_IF_EXISTS` (default)| Reuse | Reuse |\r\n | `REUSE_CACHE_IF_EXISTS` | Reuse | Fresh |\r\n | `FORCE_REDOWNLOAD` | Fresh | Fresh |\r\n```", "I have another question. Assuming that I understood correctly and there is reuse of datasets files when caching is disabled (!), I'm guessing there is a directory that is created based on some information on the dataset file. I'm interested in the situation where I'm loading a (custom) dataset from local disk. What information is used to create the directory/filenames where the files are stored?\r\n\r\nI'm concerned about the following scenario: if I have a file, let's say `train.csv` at path `the_path`, run once, the dataset is prepared, some models are run, etc. Now let's say there is an issue and I recreate `train.csv` at the same path `the_path`. Is there enough information in the temporary name/hash to *not* reload the *old* prepared dataset (e.g., timestamp of the file)? Or is it going to reload the *old* prepared file? ", "Thanks for the feedback, we'll work in improving this aspect of the documentation.\r\n\r\n> Where are these files stored? I guess not in the temporary directory that is removed...\r\n\r\nWe're using the Arrow file format to load datasets. Therefore each time you load a dataset, it is prepared as an arrow file on your disk. By default the file is located in the ~/.cache/huggingface/datasets/<dataset_name>/<config_id>/<version> directory.\r\n\r\n> What information is used to create the directory/filenames where the files are stored?\r\n\r\nThe config_id contains a hash that takes into account:\r\n- the dataset loader used and its source code (e.g. the \"csv\" loader)\r\n- the arguments passed to the loader (e.g. the csv delimiter)\r\n- metadata of the local data files if any (e.g. their timestamps)\r\n\r\n> I'm concerned about the following scenario: if I have a file, let's say train.csv at path the_path, run once, the dataset is prepared, some models are run, etc. Now let's say there is an issue and I recreate train.csv at the same path the_path. Is there enough information in the temporary name/hash to not reload the old prepared dataset (e.g., timestamp of the file)? Or is it going to reload the old prepared file?\r\n\r\nYes the timestamp of the local csv file is taken into account. If you edit your csv file, the config_id will change and loading the dataset will create a new arrow file.", "Thank you for all your clarifications, really helpful! \r\n\r\nIf you have the bandwidth, please do revisit the api wrt cache disabling. Anywhere in the computer stack (hardware included) where you disable the cache, one assumes there is no caching that happens. ", "That makes total sense indeed !\r\nI think we can do the change", "I have another question about caching, this time in the case where FORCE_REDOWNLOAD is used to load the dataset, the datasets cache is one directory as defined by HF_HOME and there are multiple concurrent jobs running in a cluster using the same local dataset (i.e., same local files in the cluster). Does anything in the naming convention and/or file access/locking that you're using prevent race conditions between the concurrent jobs on the caching of the local dataset they all use?\r\n\r\nI noticed some errors (can provide more details if helpful) in load_dataset/prepare_split that lead to my question above. \r\n\r\nLet me know if my question is clear, I can elaborate more if needed @lhoestq Thank you!", "I got another error that convinces me there is a race condition (one of the test files had zero samples at prediction time). I think it comes down to the fact that the `config_id` above (used in the naming for the cache) has no information on who's touching the data. If I have 2 concurrent jobs, both loading the same dataset and forcing redownload, they may step on each other foot/caching of the dataset. ", "We're using a locking mechanism to prevent two processes from writing at the same time. The locking is based on the `filelock` module.\r\nAlso directories that are being written use a suffix \".incomplete\" so that reading is not possible on a dataset being written.\r\n\r\nDo you think you could provide a simple code to reproduce the race condition you experienced ?", "I can provide details about the code I'm running (it's really-really close to some official samples from the huggingface transformers examples, I can point to the exact sample file, I kept a record of that). I can also describe in which conditions this race occurs (I'm convinced it has to do with forcing the redownloading of the dataset, I've been running hundreds of experiments before and didn't have a problem before I forced the redownload). I also can provide samples of the different stack errors I get and some details about the level of concurrency of jobs I was running. I can also try to imagine how the race manifests (I'm fairly sure that it's a combo of one job cleaning up and another job being in the middle of the run).\r\n\r\nHowever, I have to cleanup all this to make sure I'm no spilling any info I shouldn't be spilling. I'll try to do it by the end of the week, if you think all this is helpful. \r\n\r\nFor now, I have a workaround. Don't use forcing redownloading. And to be ultra careful (although I don't think this is a problem), I run a series of jobs that will prepare the datasets and I know there is no concurrency wrt the dataset. Once that's done (and I believe even having multiple jobs loading the datasets at the same time doesn't create problems, as long as REUSE_DATASET_IF_EXISTS is the policy for loading the dataset, so the filelock mechanism you're using is working in that scenario), the prepared datasets will be reused, no race possible in any way. \r\n\r\nThanks for all the details you provided, it helped me understand the underlying implementation and coming up with workarounds when I ran into issues. ", "Hi! I have the same challenge with caching, where the **.cache** folder is required even though it isn't possible for me.\r\n\r\nI'd like to run transformers in Snowflake, using Snowpark for Python, this would mean I could provide configurable transformers in real-time for business users without having data leave an environment (for security reasons). With no need for data transfer,n the compute is faster. It is a large use case - is it possible to entirely disable caching in certain scenarios?\r\n@lhoestq ?\r\n", "You can try to change the location of the cache folder using the `HF_CACHE_HOME` environment variable, and set a location where you have read/write access.", "Thanks @lhoestq \r\n\r\nI wanted to do that, however, snowflake does not allow it to write at all. I'm asking around to see if they can help me out with that issue 😅" ]

[ "Hi, one workaround would be to save the mapped(tokenized in your case) file using `save_to_disk`, and having each process load this file using `load_from_disk`. This is what I am doing, and in this case, I turn off the ability to automatically load from the cache.\r\n\r\nAlso, multiprocessing the map function seems to be slower at the moment (#1992), hope this helps you.", "Thanks @hwijeen for the workaround, feels a bit prototypical but it works! (it seems files are written twice then though)\r\n\r\n(I haven't observed slowness using multiprocessed map function but I could be wrong)", "To my understanding, files are written twice anyhow(one after load_dataset, another aftet map). It's just that you now have it at the location where you can see, whereas it was secretlely saved at caching folder(.cache/huggingface/datasets by default)! Correct me if I'm wrong!", "Slowness in multiprocessing has been observed in certain environments but not others. We're investigating ;)", "So to answer my initial question, I was just doing something stupid as I was not re-giving the `preprocessing_num_workers` arguments when launching the distributed training (and it was then set to `None`). I initially thought the hash was computed only with the `tokenize_function` but it's all arguments. Thanks @lhoestq for clarifying!", "This cache process isn't really consistent. I just changed `per_device_train_batch_size` of training script and now it rebuilding the dataset cache!!!! Why?", "Hi ! A `map` function is recomputed if the code changes or if any of the variables it uses changes. Can you check that your function doesn't use `per_device_train_batch_size` or any variable that contains `per_device_train_batch_size` ?", "My code is actually a transformer's example for training t5, I modified a bit:\r\n\r\nhttps://github.com/puraminy/transformers/blob/4b40877132eedb566043f83de8f1d29a84d71430/examples/flax/language-modeling/run_t5_mlm_flax.py#L614\r\n\r\nNo, it doesn't use `per_device_train_batch_size`. I remember it worked for several times and then for no reason or various reasons like the above it started to build the cache again, as if it had an expiration date (maybe), or maybe I had changed the code! \r\n\r\nSo, to get rid of these problems I saved cache with a name (was forced to not use multiple_processes, because otherwise it generates multiple files) and then I load it from this cache file. " ]

[ "Hi ! Can you try to increase the block size ? For example\r\n```python\r\nblock_size_10MB = 10<<20\r\nload_dataset(\"json\", ..., block_size=block_size_10MB)\r\n```\r\nThe block size corresponds to how much bytes to process at a time from the input stream.\r\nThis will determine multi-threading granularity as well as the size of individual chunks in the dataset.\r\n\r\nYou can also try with bigger block sizes if needed", "Hi @lhoestq! Thank you for your prompt reply.\r\nI have experimented with (10<<20, 10<<28, 10<<30, 10<<33, 10<<34), since my machine has 192G of memory, but it's either the above-mentioned error or processed killed because of OOM.\r\n\r\nCould you give me a bit of background on why block size needs to be exactly calibrated?\r\nTo my understanding, small block sized should run just fine despite its slowness..\r\n\r\n\r\n", "We're using the JSON loader of pyarrow. It parses the file chunk by chunk to load the dataset.\r\nThis issue happens when there's no delimiter in one chunk of data. For json line, the delimiter is the end of line.\r\nSo with a big value for chunk_size this should have worked unless you have one extremely long line in your file.\r\n\r\nAlso what version of pyarrow are you using ?\r\n\r\nFInally I wonder if it could be an issue on pyarrow's side when using big json files. (I haven't tested big json files like yours)", "I'm using `pyarrow==3.0.0` with `datasets==1.5.0`.\r\n\r\nYour point totally makes sense. I will check if my jsonl file contains an extremely long file and let you know. \r\n\r\nHere are some different error messages that I got when tweaking `block_size`. I also suspect that this is related to the pyarrow... but I guess it would be wonderful if datasesets could give a clear guide on how to play with large datasets! (I am suddenly experiencing various issue when working with large datasets.. e.g. #1992 )\r\n```python\r\n return paj.ReadOptions(use_threads=self.use_threads, block_size=self.block_size)\r\n File \"pyarrow/_json.pyx\", line 56, in pyarrow._json.ReadOptions.__init__\r\n File \"pyarrow/_json.pyx\", line 81, in pyarrow._json.ReadOptions.block_size.__set__\r\nOverflowError: value too large to convert to int32_t\r\n```\r\n\r\n```python\r\n\r\nline 83, in _generate_tables\r\n parse_options=self.config.pa_parse_options,\r\n File \"pyarrow/_json.pyx\", line 247, in pyarrow._json.read_json\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 84, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Exceeded maximum rows\r\n```", "I am getting the same error. When I tweak the block_size, I also find:\r\n`OverflowError: value too large to convert to int32_t`\r\nand \r\n`pyarrow.lib.ArrowInvalid: Exceeded maximum rows`\r\n", "I made more tests. I used a smaller dataset and I was getting the same error, which means that it was not necessarily linked to the dataset size. To make both my smaller and larger datasets work, I got rid of lists with the json file. I had the following data format:\r\n```python\r\n[\r\n {'key': \"a\", 'value': ['one', 'two', 'three']},\r\n {'key': \"b\", 'value': ['four', 'five', 'six']}\r\n]\r\n```\r\nI changed to:\r\n\r\n```python\r\n {'key': \"a\", 'value': 'one\\ntwo\\nthree'},\r\n {'key': \"b\", 'value': 'four\\nfive\\nsix']}\r\n```\r\nand that worked!\r\n\r\nI used the following to reformat my json file:\r\n```python\r\nwith open(file_name, \"w\", encoding=\"utf-8\") as f:\r\n for item in list_:\r\n f.write(json.dumps(item) + \"\\n\")\r\n```\r\nThis works with `block_size_10MB = 10 << 20` or without specifying `block_size`.", "Thanks @hwijeen for reporting and thanks @jpilaul for pointing this out.\r\n\r\nIndeed, those are different JSON-like formats:\r\n- the first one is the **standard JSON** format: all the file content is JSON-valid, thus all content is either a JSON object (between curly brackets `{...}`) or a JSON array (between square brackets `[...]`)\r\n- the second one is called **JSON Lines**: the entire file content is not JSON-valid, but only every line (newline-delimited) is JSON-valid\r\n\r\nCurrently PyArrow only supports **JSON Lines** format: \r\n- https://arrow.apache.org/docs/python/generated/pyarrow.json.read_json.html\r\n > Currently only the line-delimited JSON format is supported.\r\n- https://arrow.apache.org/docs/python/json.html\r\n > Arrow supports reading columnar data from line-delimited JSON files.", "Thanks @albertvillanova for your explanation, it is helpful to know (maybe add to docs?)!\r\nHowever, the problem I described above happened when I was dealing with jsonl files 😿\r\nAlthough I did not thoroughly inspect, I suspect the cause was the one extremely long document in my case.", "I see... I guess there is another problem going one then, related to the size." ]

[]

[ "Hi @nelson-liu!\r\nHere is what I do to convert a string to class label:\r\n\r\n```python\r\nfrom datasets import load_dataset, features\r\n\r\n\r\ndset = load_dataset(...)\r\ncol_name = \"the string column name\"\r\n\r\nclass_names = dset.unique(col_name)\r\nclass_feature = features.ClassLabel(names=sorted(class_names))\r\ndset = dset.map(lambda str_value: {col_name: class_feature.str2int(str_value)}, input_columns=col_name)\r\n\r\ndset = dset.cast(features.Features({\r\n ...\r\n col_name: class_feature\r\n})\r\n```\r\n", "Hi! You can use `Dataset.class_encode_column` for this. And in the next release of `datasets` (this feature is only available on `master`), you'll also be able to use `cast` to do the conversion. \r\n\r\nAn example of conversion via `cast`: \r\n```python\r\nfrom datasets import Dataset, Features, ClassLabel\r\nd = Dataset.from_dict({\"a\": [\"no\", \"yes\", \"no\"]})\r\nd = d.cast(Features({\"a\": ClassLabel(names=[\"yes\", \"no\"])}))\r\n```" ]

[ "Actually, I found the answer [here](https://github.com/facebookresearch/faiss/wiki/FAQ#what-does-it-mean-when-a-search-returns--1-ids). \r\n\r\nSo we have to do some modifications to the code for instances where the index doesn't retrieve any IDs.", "@lhoestq @patrickvonplaten \r\n\r\nI also found another short bug in the retrieval part. Especially, when retrieving documents. If Faiss returns the -1 as the index, the retriever will always use the last element in the dataset.\r\n\r\nplease check [def get_doc_dicts function](https://github.com/huggingface/transformers/blob/master/src/transformers/models/rag/retrieval_rag.py#L222)\r\n\r\n\r\nDoes the use of the HNSW guarantee to retrieve valid indexes always? \r\n\r\n", "Hi !\r\nNo it happens sometimes to return -1, especially if your dataset is small.\r\nIf your dataset is big enough it shouldn't happen in my experience.\r\n\r\nIdeally we should ignore all the -1 that are returned. It should be possible to change that in RAG's code ", "I also checked with some indexes it returns more -1s. Specially with IVF\nwhen nprobr is very low. It doesn't happen when using HNSW though. But at\nthe moment if it happens, dataset will always return the last element.\nMaybe we should change it to repeat the most last valid retrieved doc id.\nWhat do you think?\n\nOn Wed, Apr 7, 2021, 21:09 Quentin Lhoest ***@***.***> wrote:\n\n> Hi !\n> No it happens sometimes to return -1, especially if your dataset is small.\n> If your dataset is big enough it shouldn't happen.\n>\n> Ideally we should ignore all the -1 that are returned. It should be\n> possible to change that in RAG's code\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2175#issuecomment-814746509>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGTENOTLBEZTXEO2RS3THQOMPANCNFSM42PRVYDA>\n> .\n>\n", "That would be an easy way to workaround this issue. Feel free to open a PR on `transformers` and ping me ! :)", "Sure. Will push everything together with RAG end to end. :) thanks a lot.\n\nOn Wed, Apr 7, 2021, 21:16 Quentin Lhoest ***@***.***> wrote:\n\n> That would be an easy way to workaround this issue. Feel free to open a PR\n> on transformers and ping me ! :)\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2175#issuecomment-814752589>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGWLROCGARKN7WOJYSTTHQPH5ANCNFSM42PRVYDA>\n> .\n>\n" ]

[ "It seems that this can be fixed from user's end by including a `date` argument, like this:\r\n\r\n`dataset = datasets.load_dataset('wikipedia', '20200501.en', date='20210420')`\r\n\r\nYou can get available dates from [here](https://dumps.wikimedia.org/enwiki/).\r\n\r\nThis is not a proper fix however as all the files will still have '20200501' in their file names." ]

[]

[ "Hi @vyraun ! The test references for CommonGen are not publicly available: you can reach out to the original dataset authors if you would like to ask for them, but we will not be releasing them as part of GEM (March 31st was the release date for the test set inputs, references are incidentally released for some of the test sets but shouldn't really be used for benchmark submissions)\r\n\r\ncc @sebastiangehrmann", "Oh okay, thanks @yjernite ! " ]

[ "Hi,\r\n\r\na HF dataset can be converted to a Torch Dataset with a simple wrapper as follows:\r\n```python\r\nfrom torch.utils.data import Dataset\r\n \r\nclass HFDataset(Dataset):\r\n def __init__(self, dset):\r\n self.dset = dset\r\n\r\n def __getitem__(self, idx):\r\n return self.dset[idx]\r\n\r\n def __len__(self):\r\n return len(self.dset)\r\n\r\ntrain_ds = HFDataset(train_ds)\r\n```\r\n@lhoestq Since the Arrow Dataset already provides `__getitem__` and `__len__`, I think we could use the [virtual subclass](https://docs.python.org/3/library/abc.html#abc.ABCMeta.register) mechanism from the `abc` module to elegantly solve this issue. This mechanism would allow the Arrow Dataset to be used in place of the Torch Dataset because the `isinstance(instance of Arrow Dataset, TorchDataset)` check would return True (DeepSpeed has this check [here](https://github.com/microsoft/DeepSpeed/blob/ab5534fc4c0f8ca21ada321f9730d723aa31288b/deepspeed/runtime/engine.py#L823)).\r\n\r\nAnd it requires a minimal change in the `arrow_dataset.py` file:\r\n```python\r\nif config.TORCH_AVAILABLE:\r\n from torch.utils.data import Dataset as TorchDataset\r\n TorchDataset.register(Dataset)\r\n```", "Interesting ! Thanks for sharing this @mariosasko . I like the idea\r\nThis looks like something we should add IMO", "@mariosasko \r\nThx for your code!\r\nIt perfectly works with a small modification for HF NLP dataset:\r\n```\r\noriginal_ds = nlp.load_dataset('scientific_papers', 'arxiv')\r\ntrain_ds = HFDataset(train_ds['train']) # needs splitting\r\n```", "@lhoestq Sadly, from Python 3.7 onwards `torch.utils.data.Dataset` doesn't support the virtual subclass mechanism due to `typing.Generic` type no longer having `abc.ABCMeta` as its metaclass.\r\n\r\nWith that in mind, another option is to remove a direct type check (`isinstance(dataset, torch.utils.data.Dataset)`) in `deepspeed.initalize` and to rewrite the checks in a manner similar to `torch.utils.data.DataLoader` ([link](https://github.com/pytorch/pytorch/blob/b80c6f863f2327c712c478f67c248b94d66b65ac/torch/utils/data/dataloader.py#L197-L239)). This is exactly why the `DataLoader` works with arbitrary objects that provide `__getitem__` and `__len__` (and in our case, the `ArrowDataset`). By doing so, their code wouldn't be any stricter in comparison to the `DataLoader`.\r\n\r\nSo if you agree, I can open an issue in their repo and fix this if they like the idea.", "That makes sense ! Feel free to open an issue on their repo and discuss this idea", "@y-rokutan Hi, now if you install `deepspeed` from master (this feature will be available in the next official release), the code should work without subclassing. Let us know if you still have any issues.", "Worth mentioning that any function that expects a `torch..Dataset` (like `torch..DataLoader`) will fail a mypy-esque typecheck if a `datasets.Dataset` is passed, even though it implements the interface correctly (I think). The virtual subclass idea was a good one- I wonder if there's another workaround given the Generic issue. What we're really talking about is something similar to the structural subtyping semantics that `typing.Protocol` defines. If `torch..DataLoader` accepted anything that supports `__getitem__` and `__len__` methods this would be much easier. Not sure if there's a way to do this without the wrapper from the perspective of `datasets`." ]

[ "This looks like an issue with the cc100 dataset itself but not sure\r\nDid you try loading cc100 on your machine ?", "Hi\nloading works fine, but the viewer only is broken\nthanks\n\nOn Wed, Apr 7, 2021 at 12:17 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> This looks like an issue with the cc100 dataset itself but not sure\n> Did you try loading cc100 on your machine ?\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2162#issuecomment-814793809>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMRUO33JSOYGT6RETWLTHQWNLANCNFSM42IUOR6Q>\n> .\n>\n", "Hi! This visualization tool is deprecated now. The viewer at https://huggingface.co/datasets/cc100 works fine, so I'm closing this issue." ]

[ "Not yet but it’s on the short/mid-term roadmap (requested by many indeed).", "oh, great, really awesome feature to have, thank you very much for the great, fabulous work", "We'll work on dataset streaming soon. This should allow you to only load the examples you need ;)", "thanks a lot Quentin, this would be really really a great feature to have\n\nOn Wed, Apr 7, 2021 at 12:14 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> We'll work on dataset streaming soon. This should allow you to only load\n> the examples you need ;)\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2161#issuecomment-814791922>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMROD62QAKIJMAKWISTTHQWBVANCNFSM42IUI5JQ>\n> .\n>\n", "Is streaming completed? On the 1.8.0 docs it is mentioned (https://huggingface.co/docs/datasets/dataset_streaming.html), but when following the example I get the following error:\r\n\r\n```\r\n>>> dataset2 = load_dataset(\"amazon_us_reviews\", \"Pet_Products_v1_00\", split='train', streaming=True)\r\n\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n<ipython-input-21-1eedab26cff1> in <module>()\r\n----> 1 en_dataset = load_dataset('oscar', \"unshuffled_deduplicated_en\", split='train', streaming=True)\r\n\r\n3 frames\r\n/usr/local/lib/python3.7/dist-packages/datasets/builder.py in _create_builder_config(self, name, custom_features, **config_kwargs)\r\n 339 if value is not None:\r\n 340 if not hasattr(builder_config, key):\r\n--> 341 raise ValueError(f\"BuilderConfig {builder_config} doesn't have a '{key}' key.\")\r\n 342 setattr(builder_config, key, value)\r\n 343 \r\n\r\nValueError: BuilderConfig OscarConfig(name='unshuffled_deduplicated_en', version=1.0.0, data_dir=None, data_files=None, description='Unshuffled and deduplicated, English OSCAR dataset') doesn't have a 'streaming' key.\r\n```\r\n\r\nUPDATE: Managed to get streaming working by building from source and installing the additional `datasets[streaming]` package:\r\n\r\n```\r\n!pip install git+https://github.com/huggingface/datasets.git\r\n!pip install datasets[streaming]\r\n```", "Hi ! Streaming is available on `master` only right now. We'll make a new release 1.9.0 on Monday :)" ]

[ "Hi.\r\nI cannot always reproduce this issue, and on later runs I did not see it so far. Sometimes also I set 8 processes but I see less being showed, is this normal, here only 5 are shown for 8 being set, thanks\r\n\r\n```\r\n#3: 11%|███████████████▊ | 172/1583 [00:46<06:21, 3.70ba/s]\r\n#4: 9%|█████████████▏ | 143/1583 [00:46<07:46, 3.09ba/s]\r\n#7: 6%|█████████ | 98/1583 [00:45<11:34, 2.14ba/s]\r\n#5: 8%|███████████▍ | 124/1583 [00:46<09:03, 2.68ba/s]\r\n#6: 7%|██████████▏ \r\n```", "closing since I cannot reproduce it again, thanks " ]

[ "closing since I think this is cc100, just the name has been changed. thanks " ]

[ "Thanks for reporting !\r\nThe viewer doesn't have all the dependencies of the datasets. We may add openpyxl to be able to show this dataset properly", "This viewer tool is deprecated now and the new viewer at https://huggingface.co/datasets/fake_news_english works fine, so I'm closing this issue" ]

[ "Hi ! Thanks for reporting. I opened a PR to fix this issue: #2201", "Nice question which helped me a lot! I have wasted a lot of time to the `DatasetDict` creation from a csv file. Hope the document of this module add some simple examples.", "Hi :) We're indeed working on tutorials that we will add to the docs !" ]

[ "Good catch ! Thanks for reporting\r\n\r\nI just opened #2180 to fix this", "Fixed in #2180" ]

[ "Hi @marrodion. \r\n\r\nThanks for pointing this out. It would be great to incorporate this metric-specific enhancement.\r\n\r\nAnother possibility would be to require the user to input the scheme as a string `mode=\"strict\", scheme=\"IOB2\"` and then dynamically import the corresponding module using Python `importlib`:\r\n```python\r\nif scheme:\r\n scheme = importlib.import_module(f\"seqeval.scheme.{scheme}\")\r\n```\r\n\r\nFeel free to create a Pull Request to make this contribution." ]

[ "Hi ! In the arrow file we store all the integers as uint8.\r\nSo your arrow file should weigh around `height x width x n_channels x n_images` bytes.\r\n\r\nWhat feature type do your TFDS dataset have ?\r\n\r\nIf it uses a `tfds.features.Image` type, then what is stored is the encoded data (as png or jpg for example). Since these encodings are made for compression, the resulting tfrecord is smaller that the arrow file.\r\n\r\nWe are working on adding a similar feature in `datasets`: the ability to store the encoded data instead of the raw integers for images, but also for audio data. This way, arrow files will have similar sizes as tfrecords for images.", "Thanks for the prompt response. You're right about the encoding, I have the `tfds.features.Image` feature type you mentioned.\r\nHowever, as described in the `dataset_info.json`, my dataset is made of 1479 (224x224x3) images. 1479 x 224 x 224 x 3 = 222630912 bytes which is far from the actual size 520803408 bytes. \r\n\r\nAnyway I look forward to the Image feature type in `datasets`. ", "@lhoestq I changed the data structure so I have a 2D Array feature type instead of a 3D Array by grouping the two last dimensions ( a 224x672 2D Array instead of a 224x224x3 3D Array). The file size is now 223973964 bytes, nearly half the previous size! Which is around of what I would expect.\r\nI found similar behavior in existing `datasets` collection, when comparing black and white vs color image, for example MNIST vs CIFAR. ", "Interesting !\r\nThis may be because of the offsets that are stored with the array data.\r\n\r\nCurrently the offsets are stored even if the `shape` of the arrays is fixed. This was needed because of some issues with pyarrow a few months ago. I think these issues have been addressed now, so we can probably try to remove them to make the file lighter.\r\n\r\nIdeally in your case the floats data should be 220 MB for both Array2D and Array3D", "Yeah for sure, can you be a bit more specific about where the offset is stored in the code base ? And any reference to pyarrow issues if you have some. I would be very interested in contributing to `datasets` by trying to fix this issue. ", "Pyarrow has two types of lists: variable length lists and fixed size lists.\r\nCurrently we store the ArrayXD data as variable length lists. They take more disk space because they must store both actual data and offsets.\r\nIn the `datasets` code this is done here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/dbac87c8a083f806467f5afc4ec9b401a7e4c15c/src/datasets/features.py#L346-L352\r\n\r\nTo use a fixed length list, one should use the `list_size` argument of `pyarrow.list_()`.\r\nI believe this would work directly modulo some changes in the numpy conversion here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/dbac87c8a083f806467f5afc4ec9b401a7e4c15c/src/datasets/features.py#L381-L395" ]

[ "That's how I loaded the dataset\r\n```python\r\nfrom datasets import load_dataset\r\nds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache')\r\n```", "Hi ! It looks like the arrow file in the folder\r\n`/usr/local/workspace/NAS_NLP/cache/wikipedia/20200501.en/1.0.0/50aa706aa417bb77d910ad61211cc672c0ef3e0f224225a5e0a18277ade8b931` is corrupted.\r\n\r\nCan you take a look and check that it's 18.3GB ?\r\n\r\nIf not, then maybe you need to redownload it:\r\n```python\r\nfrom datasets import load_dataset\r\nds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache', download_mode=\"force_redownload\")\r\n```", "> Hi ! It looks like the arrow file in the folder\r\n> `/usr/local/workspace/NAS_NLP/cache/wikipedia/20200501.en/1.0.0/50aa706aa417bb77d910ad61211cc672c0ef3e0f224225a5e0a18277ade8b931` is corrupted.\r\n> \r\n> Can you take a look and check that it's 18.3GB ?\r\n> \r\n> If not, then maybe you need to redownload it:\r\n> \r\n> ```python\r\n> from datasets import load_dataset\r\n> ds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache', download_mode=\"force_redownload\")\r\n> ```\r\n\r\nHi Ihoestq, thanks for the reply! Actually i think my issue is i couldn't download the dataset beyond 10.7G. It feels like the whole dataset is split into different volumes and after the first one was downloaded it crashed before proceeding to the next one. I did try 'force_redownload' mode but still got the same issue.", "I just tried on my side and got no issues.\r\nWhen downloading the dataset again, did it crash at 10.7GB as well ?", "> I just tried on my side and got no issues.\r\n> When downloading the dataset again, did it crash at 10.7GB as well ?\r\n\r\nYes i have tried it multiple times on different machines. I am wondering if you could share the screenshot of your dependency versions and i will try to make them the same as yours?", "I tried using `datasets` from `master` on macos with python 3.7.2\r\nI also have `requests==2.23.0` and `tqdm==4.45.0`." ]

[ "Hi !\r\nI think this has been fixed recently on `master`.\r\nCan you try again by installing `datasets` from `master` ?\r\n```\r\npip install git+https://github.com/huggingface/datasets.git\r\n```", "Hi!\r\n\r\nUsing that version of the code solves the issue. Thanks!" ]

[ "Hi ! Indeed only the languages of the `translate-train` data are included...\r\nI can't find a link to download the english train set on https://github.com/facebookresearch/MLQA though, do you know where we can download it ?", "Hi @lhoestq \r\nthank you very much for coming back to me, now I see, you are right, in the link you sent I see split of {split}-context-{context_language}-question-{question_language}.json with context_language=question_language=en, TFDS most probably has extracted english ones from these files as en language files, but translate-train/test do not have en indeed. thanks a lot for the great explanations", "I close the ticket, since I do not see any en existing, they have trained on \"SQuAD V1.1\" instead. Thanks. " ]

[ "Hi !\r\nIndeed `save_to_disk` doesn't call pickle anymore. Though the `OverflowError` can still appear for in-memory datasets bigger than 4GB. This happens when doing this for example:\r\n```python\r\nimport pyarrow as pa\r\nimport pickle\r\n\r\narr = pa.array([0] * ((4 * 8 << 30) // 64))\r\ntable = pa.Table.from_arrays([a], names=[\"foo\"])\r\npickle.dumps(table) # fails with an OverflowError\r\npickle.dumps(table, 4) # works !\r\n```\r\nWe'll do the change to use `protocol=4`.\r\n\r\nMoreover I've also seen other users complain about this error\r\n```\r\nstruct.error: 'I' format requires 0 <= number <= 4294967295\r\n```\r\n\r\nIt looks like something related to the 4GB limit as well but I'm not able to reproduce on my side.\r\nDo you think you can provide a script that reproduces the issue ?\r\nHow big is your dataset ? (number of bytes, number of rows)\r\n\r\n", "Hi!\r\nSo I've managed to created a minimum working (well technically crashing) example for the multiprocessing case, I create a huge list of zeros, like in your example, and then I try to .map(None, num_proc=2) over it, which then crashes, here's the code:\r\n\r\n```python\r\nfrom datasets import Dataset\r\n\r\nif __name__ == '__main__':\r\n ton_of_zeroes = [0] * ((12 * 8 << 30) // 64)\r\n large_dataset = Dataset.from_dict({'col': ton_of_zeroes})\r\n print(\"Start\")\r\n large_dataset.map(function=None, num_proc=2)\r\n print(\"Done - should not print\")\r\n```\r\n\r\nThe amount of zeros could probably be reduced, I haven't tried to minimize it to find the breaking point, I just increased it from your code (which by quick glance I assumed tried to allocate over 4 GiB)\r\n\r\nRunning this results in the following traceback:\r\n\r\n```\r\nParameter 'indices'=[ 0 1 2 ... 805306365 805306366 805306367] of the transform datasets.arrow_dataset.Dataset.select couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.\r\nTraceback (most recent call last):\r\n File \"./crash_multiproc_pickle.py\", line 7, in <module>\r\n large_dataset.map(function=None, num_proc=2)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1485, in map\r\n transformed_shards = [r.get() for r in results]\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1485, in <listcomp>\r\n transformed_shards = [r.get() for r in results]\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/pool.py\", line 657, in get\r\n raise self._value\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/pool.py\", line 431, in _handle_tasks\r\n put(task)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/connection.py\", line 209, in send\r\n self._send_bytes(_ForkingPickler.dumps(obj))\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/reduction.py\", line 54, in dumps\r\n cls(buf, protocol, *args, **kwds).dump(obj)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 454, in dump\r\n StockPickler.dump(self, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 437, in dump\r\n self.save(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 789, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 941, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 859, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 885, in _batch_setitems\r\n save(v)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 662, in save_reduce\r\n save(state)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 941, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 859, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 885, in _batch_setitems\r\n save(v)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 846, in _batch_appends\r\n save(tmp[0])\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 846, in _batch_appends\r\n save(tmp[0])\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 789, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 843, in _batch_appends\r\n save(x)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 732, in save_bytes\r\n self._write_large_bytes(BINBYTES + pack(\"<I\", n), obj)\r\nstruct.error: 'I' format requires 0 <= number <= 4294967295\r\n```\r\n\r\nMy datasets usually have hundreds of thousands to low millions of rows, with each row containing a list of 10 strings and list of vectors of different length (the strings tokenized), which in the worst case have 10\\*512\\*8 = 40960 bytes (but usually it is much smaller, as the vectors tend to be shorter. I need these groups of text lines to create training data for the Inverse Cloze Task.\r\n\r\nAnyway I don't think my particular dataset is relevant, as the tiny script I created also manages to crash.\r\nBut I think the issue is the same as the save_to_disk, from the traceback it seems that in multiprocessing, it tries to use dill to return the result of the map workers, which tries to pickle the data and can't do it, probably because it's again using the older pickle protocol. That's my guess anyway.", "I just merged a fix #2150 that allows to pickle tables bigger than 4GiB\r\nFeel free to try it on the `master` branch !", "awesome! I started getting this error as well when I tried to tokenize with a longer sequence length", "@prokopCerny does this fix work for you? I found that with the latest master, my container with 500GB RAM starts crashing when I try to map a large dataset using `num_proc`.\r\n\r\n@lhoestq would it be possible to implement some logic to keep the individual cache files small (say below 100mb)? I find this helps with loading large datasets, but the \"hack\" I was using (increasing `num_proc` to a large number) doesn't work anymore with the latest master; my container crashes even with `num_proc=200` now", "Closing since the original issue was fixed in #2150 \r\nFeel free to reopen if you are still experiencing it.\r\nFor the other problems, please open separate issues" ]

[ "If you print those questions, you get readable texts:\r\n```python\r\n>>> questions = [\r\n... \"\\u0645\\u062a\\u0649 \\u0628\\u062f\\u0627\\u062a \\u0627\\u0644\\u0645\\u062c\\u0644\\u0629 \\u0627\\u0644\\u0645\\u062f\\u0631\\u0633\\u064a\\u0629 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645 \\u0628\\u0627\\u0644\\u0646\\u0634\\u0631?\",\r\n... \"\\u0643\\u0645 \\u0645\\u0631\\u0629 \\u064a\\u062a\\u0645 \\u0646\\u0634\\u0631\\u0647\\u0627 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0645\\u0627 \\u0647\\u064a \\u0627\\u0644\\u0648\\u0631\\u0642\\u0629 \\u0627\\u0644\\u064a\\u0648\\u0645\\u064a\\u0629 \\u0644\\u0644\\u0637\\u0644\\u0627\\u0628 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0643\\u0645 \\u0639\\u062f\\u062f \\u0627\\u0644\\u0627\\u0648\\u0631\\u0627\\u0642 \\u0627\\u0644\\u0627\\u062e\\u0628\\u0627\\u0631\\u064a\\u0629 \\u0644\\u0644\\u0637\\u0644\\u0627\\u0628 \\u0627\\u0644\\u062a\\u064a \\u0648\\u062c\\u062f\\u062a \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0641\\u064a \\u0627\\u064a \\u0633\\u0646\\u0629 \\u0628\\u062f\\u0627\\u062a \\u0648\\u0631\\u0642\\u0629 \\u0627\\u0644\\u0637\\u0627\\u0644\\u0628 \\u0627\\u0644\\u062d\\u0633 \\u0627\\u0644\\u0633\\u0644\\u064a\\u0645 \\u0628\\u0627\\u0644\\u0646\\u0634\\u0631 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\"\r\n... ]\r\n>>> print(questions)\r\n['متى بدات المجلة المدرسية في نوتردام بالنشر?', 'كم مرة يتم نشرها في نوتردام?', 'ما هي الورقة اليومية للطلاب في نوتردام?', 'كم عدد الاوراق الاخبارية للطلاب التي وجدت في نوتردام?', 'في اي سنة بدات ورقة الطالب الحس السليم بالنشر في نوتردام?']\r\n```\r\nI don't think we can change this", "Hi @dorost1234.\r\n\r\nIn Python 3, strings are sequences of Unicode _code points_. Unicode is a specification that maps all characters (and emoji symbols) with its unique representation in terms of code points. That is what you see: Unicode code points (represented by a \\u escaped sequence of 16-bit hex values).\r\n\r\nCharacters are usually represented (on screen and papers) with a graphical element called _glyph_. That is what you would like to see: glyphs. But Python does not care about glyphs: that is the job of the GUI or the terminal; glyphs are what you get with the `print` function (if your terminal is properly configured to display those glyphs).\r\n\r\nYou have more detailed information about Unicode in the Python documentation: https://docs.python.org/3/howto/unicode.html", "thank you so much for the insightful comments. " ]

[ "You can filter the languages this way:\r\n```python\r\ntydiqa_en = tydiqa_dataset.filter(lambda x: x[\"language\"] == \"english\")\r\n```\r\n\r\nOtherwise maybe we can have one configuration per language ?\r\nWhat do you think of this for example ?\r\n\r\n```python\r\nload_dataset(\"tydiqa\", \"primary_task.en\")\r\n```", "Hi\nthank you very much for the great response, this will be really wonderful\nto have one configuration per language, as one need the dataset in majority\nof case per language for cross-lingual evaluations.\nThis becomes also then more close to TFDS format, which is separated per\nlanguage https://www.tensorflow.org/datasets/catalog/tydi_qa which will be\nreally awesome to have.\nthanks\n\nOn Mon, Mar 29, 2021 at 6:17 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> You can filter the languages this way:\n>\n> tydiqa_en = tydiqa_dataset.filter(lambda x: x[\"language\"] == \"english\")\n>\n> Otherwise maybe we can have one configuration per language ?\n> What do you think of this for example ?\n>\n> load_dataset(\"tydiqa\", \"primary_task.en\")\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2132#issuecomment-809516799>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMXPW2PWSQ2RHG73O7TTGCY4LANCNFSM4Z7ER7IA>\n> .\n>\n", "@lhoestq I greatly appreciate any updates on this. thanks a lot" ]

[ "Hi ! Thanks for reporting\r\nI was able to reproduce this issue. This was caused by missing split infos if a worker reloads the cache of the other worker.\r\n\r\nI just opened https://github.com/huggingface/datasets/pull/2137 to fix this issue", "The PR got merged :)\r\nFeel free to try it out on the `master` branch", "Sorry for the late reply. \r\nNow everything just works well XD" ]

[ "Here please find TFDS format of this dataset: https://www.tensorflow.org/datasets/catalog/wikiann\r\nwhere there is a span column, this is really necessary to be able to use the data, and I appreciate your help @lhoestq ", "Hi !\r\nApparently you can get the spans from the NER tags using `tags_to_spans` defined here:\r\n\r\nhttps://github.com/tensorflow/datasets/blob/c7096bd38e86ed240b8b2c11ecab9893715a7d55/tensorflow_datasets/text/wikiann/wikiann.py#L81-L126\r\n\r\nIt would be nice to include the `spans` field in this dataset as in TFDS. This could be a good first issue for new contributors !\r\n\r\nThe objective is to use `tags_to_spans` in the `_generate_examples` method [here](https://github.com/huggingface/nlp/blob/c98e4b8f23e3770c401c6d9326e243e1ffd599ec/datasets/wikiann/wikiann.py#L292-L316) to create he `spans` for each example.", "Hi @lhoestq \r\nthank you very much for the help, it would be very nice to have it included, here is the full code, one need to also convert tags to string first:\r\n\r\n```\r\nimport datasets \r\nfrom datasets import load_dataset\r\n\r\ndef tags_to_spans(tags):\r\n \"\"\"Convert tags to spans.\"\"\"\r\n spans = set()\r\n span_start = 0\r\n span_end = 0\r\n active_conll_tag = None\r\n for index, string_tag in enumerate(tags):\r\n # Actual BIO tag.\r\n bio_tag = string_tag[0]\r\n assert bio_tag in [\"B\", \"I\", \"O\"], \"Invalid Tag\"\r\n conll_tag = string_tag[2:]\r\n if bio_tag == \"O\":\r\n # The span has ended.\r\n if active_conll_tag:\r\n spans.add((active_conll_tag, (span_start, span_end)))\r\n active_conll_tag = None\r\n # We don't care about tags we are\r\n # told to ignore, so we do nothing.\r\n continue\r\n elif bio_tag == \"B\":\r\n # We are entering a new span; reset indices and active tag to new span.\r\n if active_conll_tag:\r\n spans.add((active_conll_tag, (span_start, span_end)))\r\n active_conll_tag = conll_tag\r\n span_start = index\r\n span_end = index\r\n elif bio_tag == \"I\" and conll_tag == active_conll_tag:\r\n # We're inside a span.\r\n span_end += 1\r\n else:\r\n # This is the case the bio label is an \"I\", but either:\r\n # 1) the span hasn't started - i.e. an ill formed span.\r\n # 2) We have IOB1 tagging scheme.\r\n # We'll process the previous span if it exists, but also include this\r\n # span. This is important, because otherwise, a model may get a perfect\r\n # F1 score whilst still including false positive ill-formed spans.\r\n if active_conll_tag:\r\n spans.add((active_conll_tag, (span_start, span_end)))\r\n active_conll_tag = conll_tag\r\n span_start = index\r\n span_end = index\r\n # Last token might have been a part of a valid span.\r\n if active_conll_tag:\r\n spans.add((active_conll_tag, (span_start, span_end)))\r\n # Return sorted list of spans\r\n return sorted(list(spans), key=lambda x: x[1][0])\r\n\r\ndataset = load_dataset('wikiann', 'en', split=\"train\")\r\nner_tags = {\r\n 0:\"O\",\r\n 1:\"B-PER\",\r\n 2:\"I-PER\",\r\n 3:\"B-ORG\",\r\n 4:\"I-ORG\",\r\n 5:\"B-LOC\",\r\n 6:\"I-LOC\"\r\n}\r\n\r\ndef get_spans(tokens, tags):\r\n \"\"\"Convert tags to textspans.\"\"\"\r\n spans = tags_to_spans(tags)\r\n text_spans = [\r\n x[0] + \": \" + \" \".join([tokens[i]\r\n for i in range(x[1][0], x[1][1] + 1)])\r\n for x in spans\r\n ]\r\n if not text_spans:\r\n text_spans = [\"None\"]\r\n return text_spans\r\n\r\n\r\nfor i, d in enumerate(dataset):\r\n tokens = d['tokens']\r\n tags = d['ner_tags']\r\n tags = [ner_tags[i] for i in tags]\r\n spans = get_spans(tokens, tags)\r\n print(\"spans \", spans)\r\n print(d)\r\n if i > 10:\r\n break; \r\n```\r\nI am not sure how to contribute to the repository and how things work, could you let me know how one can access the datasets to be able to contribute to the repository? Maybe I could do it then\r\nthanks \r\n", "Cool ! Let me give you some context:\r\n\r\n#### Contribution guide\r\n\r\nYou can find the contribution guide here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md\r\n\r\nIt explains how to set up your dev environment in a few steps.\r\n\r\n#### Dataset loading\r\n\r\nEach Dataset is defined by a Table that have many rows (one row = one example) and columns (one column = one feature).\r\nTo change how a dataset is constructed, you have to modify its dataset script that you can find here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/master/datasets/wikiann/wikiann.py\r\n\r\nIt includes everything needed to load the WikiANN dataset.\r\nYou can load locally a modified version of `wikiann.py` with `load_dataset(\"path/to/wikiann.py\")`.\r\n\r\n#### Define a new column\r\n\r\nEach column has a name and a type. You can see how the features of WikiANN are defined here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/c98e4b8f23e3770c401c6d9326e243e1ffd599ec/datasets/wikiann/wikiann.py#L245-L263\r\n\r\nIdeally we would have one additional feature \"spans\":\r\n```python\r\n \"spans\": datasets.Sequence(datasets.Value(\"string\")),\r\n```\r\n\r\n#### Compute the content of each row\r\n\r\nTo build the WikiANN rows, the _generate_examples method from [here](https://github.com/huggingface/nlp/blob/c98e4b8f23e3770c401c6d9326e243e1ffd599ec/datasets/wikiann/wikiann.py#L292-L316) is used. This function `yield` one python dictionary for each example:\r\n```python\r\nyield guid_index, {\"tokens\": tokens, \"ner_tags\": ner_tags, \"langs\": langs}\r\n```\r\n\r\nThe objective would be to return instead something like\r\n```python\r\nspans = spans = get_spans(tokens, tags)\r\nyield guid_index, {\"tokens\": tokens, \"ner_tags\": ner_tags, \"langs\": langs, \"spans\": spans}\r\n```\r\n\r\nLet me know if you have questions !", "The PR was merged. Issue should be closed.\r\n\r\nCC: @lhoestq " ]

[ "Hi !\r\nWe're not using `TextDatasetForNextSentencePrediction` in `datasets`.\r\nAlthough you can probably use the `TextDatasetForNextSentencePrediction.create_examples_from_document` on a dataset to prepare it for next sentence prediction.", "Thanks.\r\n\r\nDo you mean that `TextDatasetForNextSentencePrediction.create_exapmles_from_document` can be applied to dataset object other than `TextDatasetForNextSentencePrediction` e.g. a `Dataset` object which is loaded by `datasets.load_dataset`?", "It would probably require a bit of tweaking, but you can apply it to a dataset, yes.\r\nThis should give you a new dataset with sentence pairs you can train a model on.\r\n\r\nYou can find the documentation about dataset processing here:\r\nhttps://huggingface.co/docs/datasets/processing.html#processing-data-with-map", "Thank you for detail information.\r\n\r\nI'll try to apply `create_examples_from_document` to `Dataset` object.\r\n" ]

[ "Hi\r\nGood catch ! Thanks for reporting\r\n\r\nIf you are interested in contributing, feel free to open a PR to fix this :) " ]

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

[ "I haven't played with it (yet) but it sounds really cool !\r\n" ]

[ "Hi,\r\n\r\nsadly I can't replicate the problem on my Windows machine. Try to update the library to the newest version with:\r\n```bash\r\npip install git+https://github.com/huggingface/datasets\r\n``` ", "Thanks for the answer! I updated the library but unfortunately it didn't solve the problem.", "Is there an error message ?\r\nWhat stacktrace do you get if you interrupt the execution of the program while downloading ?", "Sorry for the long time since my last comment, I tried again and don't seem to have the problem anymore, thanks for your support!", "Great ! I'm closing the issue then. Feel free to re-open if you experience this issue again" ]

[ "Thanks for reporting :) We're looking into it", "Back up. " ]

[ "@Frankie123421 what was the resolution to this?", "> @Frankie123421 what was the resolution to this?\r\n\r\nuse glue_metric.py instead of glue.py in load_metric", "thank you!" ]

[ "Hi,\r\n\r\nthe `_data` attribute is missing due to `MyDataset.__init__` not calling the parent `__init__`. However, I don't think it's a good idea to subclass the `datasets.Dataset` class (e.g. it's kind of dangerous to override `datasets.Dataset.__getitem__`). Instead, it's better to follow the \"association over inheritance\" approach with a simple wrapper class that delegates calls to a wrapped `Dataset` (map, etc.). Btw, the library offers the `datasets.Dataset.from_pandas` class method to directly create a `datasets.Dataset` from the dataframe." ]

[]

[]

[ "Hi ! Thanks for reporting\r\n\r\nBy looking at the raw `news-commentary-v14.en-kk.tsv` file, it looks like there are at least 17 lines with this issue.\r\nMoreover these issues are not always the same:\r\n- L97 is only `kk` text and must be appended at the end of the `kk` text of the **next** line\r\n- L2897 is only `kk` text and must be appended at the end of the `kk` text of the **previous** line\r\n- L1247 and L1248 are only `kk` texts and must be inserted at the **beginning** of the `kk` text of the next line\r\n- (and there are many others)\r\n\r\nIt would be nice to have a corrected version of this file ! The file is available in the `wmt/news-commentary` repository on the Datasets Hub here:\r\nhttps://huggingface.co/datasets/wmt/news-commentary/tree/main/v14/training\r\n\r\nThen maybe we can notify the WMT authors and host the corrected version somewhere" ]

[ "Hello @kyleclo! Currently, we are getting the data from your bucket, so if you remove it the HF script won't work anymore :) \r\n\r\nUntil you solve things on your end, @lhoestq suggested we just return a warning message when people try to load that dataset from HF. What would you like it to say?", "Hi @kyleclo, as of today, you have not removed your bucket data yet, and therefore HuggingFace can download it from there.\r\n\r\nIs it OK? Are you planning to eventually delete it? Thank you.", "Hi! Sorry I missed @yjernite 's previous message, thanks for responding! \r\n\r\nIs there an option where we can keep our data in our bucket, but the HF script no longer pulls data from it? " ]

[ "Hi ! `wiki_movies` was added in `datasets==1.2.0`. However it looks like you have `datasets==1.1.2`.\r\n\r\nTo use `wiki_movies`, please update `datasets` with\r\n```\r\npip install --upgrade datasets\r\n```", "Thanks a lot! That solved it and I was able to upload a model trained on it as well :)" ]

[ "Thanks for reporting :)\r\nMaybe we can concatenate fields only if they are different.\r\n\r\nCurrently this is done here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/349ac4398a3bcae6356f14c5754483383a60e8a4/src/datasets/info.py#L180-L196\r\n\r\nThis can be a good first contribution to the library.\r\nPlease comment if you'd like to improve this and open a PR :)" ]

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

[ "Hi ! This is 1.1 as specified by the download urls here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/349ac4398a3bcae6356f14c5754483383a60e8a4/datasets/squad/squad.py#L50-L55", "Got it. Thank you~" ]

[ "Hello. I've been looking for information about German Conll2003 and found your question. Official site (https://www.clips.uantwerpen.be/conll2003/ner/) mentions that organizers provide only annotation. German texts (ECI Multilingual Text Corpus) are not freely available and can be ordered from the Linguistic Data Consortium.\r\n\r\nBut maybe something has changed since 2003.", "You can find the reason for not including the German data here: https://github.com/huggingface/datasets/issues/4230." ]