Datasets:

---

mayk00

/

maykel_dataset

like 0

[]

[]

[ "شبکه های متن به گفتار ابتدا متن داده شده را به بازنمایی میانی" ]

[]

[]

[ "You can clone the `togethercomputer/RedPajama-Data-1T-Sample` repo and load the dataset with `load_dataset(\"path/to/cloned_repo\")` to use it offline.", "@mariosasko Thank you for your kind reply! I'll try it as a workaround.\r\nDoes that mean that currently it's not supported to simply load with a short name?", "It is, but manually downloading repo files to the cache can easily lead to failure (the HF cache is not meant to be modified by a user besides deleting the files 🙂), as in your case. Hence, the clone + `load_dataset(\"path/to/cloned_repo\")` workflow should be used instead." ]

[]

[]

[ "Unlike Pandas, Arrow is strict with types, so converting the problematic strings to ints (or ints to strings) to ensure all the values have the same type is the only fix. \r\n\r\nJSON support has been requested in Arrow [here](https://github.com/apache/arrow/issues/32538), but I don't expect this to be implemented soon. \r\n\r\nAlso, this type could be represented with the Arrow Union type. However, due to low usage, the Union type has limited support in the Arrow ecosystem (e.g., IIRC Parquet still does not support it). So, we should probably wait a bit more before adding support for it in `datasets`", "> Unlike Pandas, Arrow is strict with types, so converting the problematic strings to ints (or ints to strings) to ensure all the values have the same type is the only fix.\r\n> \r\n> JSON support has been requested in Arrow [here](https://github.com/apache/arrow/issues/32538), but I don't expect this to be implemented soon.\r\n> \r\n> Also, this type could be represented with the Arrow Union type. However, due to low usage, the Union type has limited support in the Arrow ecosystem (e.g., IIRC Parquet still does not support it). So, we should probably wait a bit more before adding support for it in `datasets`\r\n\r\nOk many thanks, I was able to mitigate the problem by manually checking and converting all problematic fields now." ]

[]

[]

[ "I've also encountered this issue recently and want to ask if this has been seen.\r\n\r\n@albertvillanova for visibility - I'm not sure who the right person is to tag, but I noticed you were active recently so perhaps you can direct this to the right person.\r\n\r\nThanks!" ]

[]

[]

[ "The expression \"train+test\" concatenates the splits.\r\n\r\nThe individual splits as separate datasets can be obtained as follows:\r\n```python\r\ntrain_ds, test_ds = load_dataset(\"<dataset_name>\", split=[\"train\", \"test\"])\r\ntrain_10pct_ds, test_10pct_ds = load_dataset(\"<dataset_name>\", split=[\"train[:10%]\", \"test[:%10]\"])\r\n```" ]

[]

[]

[ "Hi! Instead of processing a single example at a time, you should use the batched `map` for the best performance (with `num_proc=1`) - the fast tokenizers can process a batch's samples in parallel in that scenario.\r\n\r\nE.g., the following code in Colab takes an hour to complete:\r\n```python\r\n# !pip install datasets transformers\r\nfrom datasets import load_dataset\r\nfrom transformers import AutoTokenizer\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")\r\ndataset = dataset.map(lambda ex: tokenizer(ex[\"text\"]), batched=True, remove_columns=[\"text\", \"meta\"])\r\n```", "Batched is far worse. A single batch of 1000 took hours and that was only 1%\r\n\r\n\r\nOn Thu, Oct 19, 2023, 2:26 PM Mario Šaško ***@***.***> wrote:\r\n\r\n> Hi! You should use the batched map for the best performance (with\r\n> num_proc=1) - the fast tokenizers can process a batch's samples in\r\n> parallel.\r\n>\r\n> E.g., the following code in Colab takes an hour to complete:\r\n>\r\n> # !pip install datasets transformersfrom datasets import load_datasetfrom transformers import AutoTokenizertokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"]), batched=True, remove_columns=[\"text\", \"meta\"])\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771503757>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/ABDD3ZJHPSRVDEXFNMXR2N3YAFWFZAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRGUYDGNZVG4>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "Can you please provide a self-contained reproducer?", "Which specific version of datasets are you using?\r\n\r\nWhat is the architecture of your colab setup? Ram? Cores? OS?\r\n\r\n\r\nOn Thu, Oct 19, 2023, 2:27 PM pensive introvert ***@***.***>\r\nwrote:\r\n\r\n> Batched is far worse. A single batch of 1000 took hours and that was only\r\n> 1%\r\n>\r\n>\r\n> On Thu, Oct 19, 2023, 2:26 PM Mario Šaško ***@***.***>\r\n> wrote:\r\n>\r\n>> Hi! You should use the batched map for the best performance (with\r\n>> num_proc=1) - the fast tokenizers can process a batch's samples in\r\n>> parallel.\r\n>>\r\n>> E.g., the following code in Colab takes an hour to complete:\r\n>>\r\n>> # !pip install datasets transformersfrom datasets import load_datasetfrom transformers import AutoTokenizertokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"]), batched=True, remove_columns=[\"text\", \"meta\"])\r\n>>\r\n>> —\r\n>> Reply to this email directly, view it on GitHub\r\n>> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771503757>,\r\n>> or unsubscribe\r\n>> <https://github.com/notifications/unsubscribe-auth/ABDD3ZJHPSRVDEXFNMXR2N3YAFWFZAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRGUYDGNZVG4>\r\n>> .\r\n>> You are receiving this because you authored the thread.Message ID:\r\n>> ***@***.***>\r\n>>\r\n>\r\n", "from functools import partial\r\nimport transformers\r\nfrom datasets import load_dataset, concatenate_datasets, load_from_disk\r\n\r\nmodel_name_or_path=\"/opt/data/data/daryl149/llama-2-7b-chat-hf\"\r\noutput_dir=\"/opt/data/data/LongLoRA/checkpoints\"\r\ncache_dir=\"/opt/data/data/LongLoRA/cache\"\r\nmodel_max_length=16384\r\n\r\nIGNORE_INDEX = -100\r\nDEFAULT_PAD_TOKEN = \"[PAD]\"\r\nDEFAULT_EOS_TOKEN = \"</s>\"\r\nDEFAULT_BOS_TOKEN = \"<s>\"\r\nDEFAULT_UNK_TOKEN = \"<unk>\"\r\n\r\n\r\ntokenizer = transformers.LlamaTokenizerFast.from_pretrained(\r\n model_name_or_path,\r\n cache_dir=cache_dir,\r\n model_max_length=model_max_length,\r\n padding_side=\"right\",\r\n use_fast=True,\r\n #use_fast=False\r\n)\r\n\r\nspecial_tokens_dict = dict()\r\nif tokenizer.pad_token is None:\r\n special_tokens_dict[\"pad_token\"] = DEFAULT_PAD_TOKEN\r\nif tokenizer.eos_token is None:\r\n special_tokens_dict[\"eos_token\"] = DEFAULT_EOS_TOKEN\r\nif tokenizer.bos_token is None:\r\n special_tokens_dict[\"bos_token\"] = DEFAULT_BOS_TOKEN\r\nif tokenizer.unk_token is None:\r\n special_tokens_dict[\"unk_token\"] = DEFAULT_UNK_TOKEN\r\n\r\ntokenizer.add_special_tokens(special_tokens_dict)\r\n\r\ndef tokenize_fn(tokenizer, example):\r\n context_length = tokenizer.model_max_length\r\n outputs = tokenizer(\r\n tokenizer.eos_token.join(example[\"text\"]),\r\n #truncation=False,\r\n truncation=True,\r\n return_tensors=\"pt\",\r\n #return_tensors=\"np\",\r\n pad_to_multiple_of=context_length,\r\n padding=True,\r\n )\r\n return {\"input_ids\": outputs[\"input_ids\"].view(-1, context_length)}\r\n\r\nfor idx in range(100):\r\n dataset = load_dataset(\"togethercomputer/RedPajama-Data-1T-Sample\",\r\ncache_dir=cache_dir, split=f'train[{idx}%:{idx+1}%]')\r\n dataset = dataset.map(partial(tokenize_fn, tokenizer), batched=False,\r\nnum_proc=16, remove_columns=[\"text\", \"meta\"])\r\n dataset.save_to_disk(training_args.cache_dir + f\"/training_data_{idx}\")\r\n\r\n\r\nOn Thu, Oct 19, 2023 at 2:30 PM Mario Šaško ***@***.***>\r\nwrote:\r\n\r\n> Can you please provide a self-contained reproducer?\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771509229>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/ABDD3ZNBZ3BE7Q4EQZZK6MLYAFWURAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRGUYDSMRSHE>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "I changed the tokenizer to one without \"Fast suffix, and something changed.\r\nThe fraction, although still slowed a lot at 80% was able to get over the\r\nfinish line of 100%\r\n\r\nI have to do more testng, see if the whole set can be processed\r\n\r\n\r\n\r\nOn Thu, Oct 19, 2023 at 3:03 PM pensive introvert <\r\n***@***.***> wrote:\r\n\r\n> from functools import partial\r\n> import transformers\r\n> from datasets import load_dataset, concatenate_datasets, load_from_disk\r\n>\r\n> model_name_or_path=\"/opt/data/data/daryl149/llama-2-7b-chat-hf\"\r\n> output_dir=\"/opt/data/data/LongLoRA/checkpoints\"\r\n> cache_dir=\"/opt/data/data/LongLoRA/cache\"\r\n> model_max_length=16384\r\n>\r\n> IGNORE_INDEX = -100\r\n> DEFAULT_PAD_TOKEN = \"[PAD]\"\r\n> DEFAULT_EOS_TOKEN = \"</s>\"\r\n> DEFAULT_BOS_TOKEN = \"<s>\"\r\n> DEFAULT_UNK_TOKEN = \"<unk>\"\r\n>\r\n>\r\n> tokenizer = transformers.LlamaTokenizerFast.from_pretrained(\r\n> model_name_or_path,\r\n> cache_dir=cache_dir,\r\n> model_max_length=model_max_length,\r\n> padding_side=\"right\",\r\n> use_fast=True,\r\n> #use_fast=False\r\n> )\r\n>\r\n> special_tokens_dict = dict()\r\n> if tokenizer.pad_token is None:\r\n> special_tokens_dict[\"pad_token\"] = DEFAULT_PAD_TOKEN\r\n> if tokenizer.eos_token is None:\r\n> special_tokens_dict[\"eos_token\"] = DEFAULT_EOS_TOKEN\r\n> if tokenizer.bos_token is None:\r\n> special_tokens_dict[\"bos_token\"] = DEFAULT_BOS_TOKEN\r\n> if tokenizer.unk_token is None:\r\n> special_tokens_dict[\"unk_token\"] = DEFAULT_UNK_TOKEN\r\n>\r\n> tokenizer.add_special_tokens(special_tokens_dict)\r\n>\r\n> def tokenize_fn(tokenizer, example):\r\n> context_length = tokenizer.model_max_length\r\n> outputs = tokenizer(\r\n> tokenizer.eos_token.join(example[\"text\"]),\r\n> #truncation=False,\r\n> truncation=True,\r\n> return_tensors=\"pt\",\r\n> #return_tensors=\"np\",\r\n> pad_to_multiple_of=context_length,\r\n> padding=True,\r\n> )\r\n> return {\"input_ids\": outputs[\"input_ids\"].view(-1, context_length)}\r\n>\r\n> for idx in range(100):\r\n> dataset = load_dataset(\"togethercomputer/RedPajama-Data-1T-Sample\",\r\n> cache_dir=cache_dir, split=f'train[{idx}%:{idx+1}%]')\r\n> dataset = dataset.map(partial(tokenize_fn, tokenizer), batched=False,\r\n> num_proc=16, remove_columns=[\"text\", \"meta\"])\r\n> dataset.save_to_disk(training_args.cache_dir + f\"/training_data_{idx}\")\r\n>\r\n>\r\n> On Thu, Oct 19, 2023 at 2:30 PM Mario Šaško ***@***.***>\r\n> wrote:\r\n>\r\n>> Can you please provide a self-contained reproducer?\r\n>>\r\n>> —\r\n>> Reply to this email directly, view it on GitHub\r\n>> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771509229>,\r\n>> or unsubscribe\r\n>> <https://github.com/notifications/unsubscribe-auth/ABDD3ZNBZ3BE7Q4EQZZK6MLYAFWURAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRGUYDSMRSHE>\r\n>> .\r\n>> You are receiving this because you authored the thread.Message ID:\r\n>> ***@***.***>\r\n>>\r\n>\r\n", "So, using LlamaTokenizerFast was the problem. Changing it to LlamaTokenizer\r\nfixed things,\r\n\r\nOn Thu, Oct 19, 2023 at 4:04 PM pensive introvert <\r\n***@***.***> wrote:\r\n\r\n> I changed the tokenizer to one without \"Fast suffix, and something\r\n> changed. The fraction, although still slowed a lot at 80% was able to get\r\n> over the finish line of 100%\r\n>\r\n> I have to do more testng, see if the whole set can be processed\r\n>\r\n>\r\n>\r\n> On Thu, Oct 19, 2023 at 3:03 PM pensive introvert <\r\n> ***@***.***> wrote:\r\n>\r\n>> from functools import partial\r\n>> import transformers\r\n>> from datasets import load_dataset, concatenate_datasets, load_from_disk\r\n>>\r\n>> model_name_or_path=\"/opt/data/data/daryl149/llama-2-7b-chat-hf\"\r\n>> output_dir=\"/opt/data/data/LongLoRA/checkpoints\"\r\n>> cache_dir=\"/opt/data/data/LongLoRA/cache\"\r\n>> model_max_length=16384\r\n>>\r\n>> IGNORE_INDEX = -100\r\n>> DEFAULT_PAD_TOKEN = \"[PAD]\"\r\n>> DEFAULT_EOS_TOKEN = \"</s>\"\r\n>> DEFAULT_BOS_TOKEN = \"<s>\"\r\n>> DEFAULT_UNK_TOKEN = \"<unk>\"\r\n>>\r\n>>\r\n>> tokenizer = transformers.LlamaTokenizerFast.from_pretrained(\r\n>> model_name_or_path,\r\n>> cache_dir=cache_dir,\r\n>> model_max_length=model_max_length,\r\n>> padding_side=\"right\",\r\n>> use_fast=True,\r\n>> #use_fast=False\r\n>> )\r\n>>\r\n>> special_tokens_dict = dict()\r\n>> if tokenizer.pad_token is None:\r\n>> special_tokens_dict[\"pad_token\"] = DEFAULT_PAD_TOKEN\r\n>> if tokenizer.eos_token is None:\r\n>> special_tokens_dict[\"eos_token\"] = DEFAULT_EOS_TOKEN\r\n>> if tokenizer.bos_token is None:\r\n>> special_tokens_dict[\"bos_token\"] = DEFAULT_BOS_TOKEN\r\n>> if tokenizer.unk_token is None:\r\n>> special_tokens_dict[\"unk_token\"] = DEFAULT_UNK_TOKEN\r\n>>\r\n>> tokenizer.add_special_tokens(special_tokens_dict)\r\n>>\r\n>> def tokenize_fn(tokenizer, example):\r\n>> context_length = tokenizer.model_max_length\r\n>> outputs = tokenizer(\r\n>> tokenizer.eos_token.join(example[\"text\"]),\r\n>> #truncation=False,\r\n>> truncation=True,\r\n>> return_tensors=\"pt\",\r\n>> #return_tensors=\"np\",\r\n>> pad_to_multiple_of=context_length,\r\n>> padding=True,\r\n>> )\r\n>> return {\"input_ids\": outputs[\"input_ids\"].view(-1, context_length)}\r\n>>\r\n>> for idx in range(100):\r\n>> dataset = load_dataset(\"togethercomputer/RedPajama-Data-1T-Sample\",\r\n>> cache_dir=cache_dir, split=f'train[{idx}%:{idx+1}%]')\r\n>> dataset = dataset.map(partial(tokenize_fn, tokenizer), batched=False,\r\n>> num_proc=16, remove_columns=[\"text\", \"meta\"])\r\n>> dataset.save_to_disk(training_args.cache_dir +\r\n>> f\"/training_data_{idx}\")\r\n>>\r\n>>\r\n>> On Thu, Oct 19, 2023 at 2:30 PM Mario Šaško ***@***.***>\r\n>> wrote:\r\n>>\r\n>>> Can you please provide a self-contained reproducer?\r\n>>>\r\n>>> —\r\n>>> Reply to this email directly, view it on GitHub\r\n>>> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771509229>,\r\n>>> or unsubscribe\r\n>>> <https://github.com/notifications/unsubscribe-auth/ABDD3ZNBZ3BE7Q4EQZZK6MLYAFWURAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRGUYDSMRSHE>\r\n>>> .\r\n>>> You are receiving this because you authored the thread.Message ID:\r\n>>> ***@***.***>\r\n>>>\r\n>>\r\n", "Indeed, the tokenizer is super slow. Perhaps @ArthurZucker knows the reason why.\r\n\r\n([This](https://colab.research.google.com/drive/1VgeurX-4Fl2X6aBQTwh_X4kuQKZ6K9L1?usp=sharing) simplified Colab can be used to reproduce the behavior)", "same issue here\r\nsample to reproduce: https://github.com/philschmid/document-ai-transformers/blob/main/training/donut_sroie.ipynb\r\nwith following map line\r\nhttps://github.com/philschmid/document-ai-transformers/blob/main/training/donut_sroie.ipynb\r\n\r\nIf I directly iterate over the dataset and call the mapping method, it is very fast\r\n```py\r\nfor sample in dataset:\r\n def preprocess_documents_for_donut(sample):\r\n```\r\n\r\nif i removed `.convert('RGB')` It can run to completion without getting stuck. I suspect it has something to do with the Image.\r\n\r\nIf I use batch, it's even slower.", "@ewfian \r\n\r\n> If I directly iterate over the dataset and call the mapping method, it is very fast\r\n\r\n`Dataset.map` must also convert the images into bytes to write them to an Arrow file (the write itself takes some time, too). \r\n\r\nYou can make the `map` faster by manually converting the images into an \"arrow-compatible\" representation. Otherwise, the Pillow defaults are used when saving an image, which seems particularly slow for the notebook's case.\r\n\r\n```python\r\ndef preprocess_documents_for_donut(sample):\r\n text = json.loads(sample[\"text\"])\r\n d_doc = task_start_token + json2token(text) + eos_token\r\n image = sample[\"image\"].convert('RGB')\r\n # convert image to bytes\r\n buffer = io.BytesIO()\r\n image.save(buffer, format=\"PNG\", compress_level=1)\r\n return {\"image\": {\"bytes\": buffer.getvalue()}, \"text\": d_doc}\r\n\r\nproc_dataset = dataset.map(preprocess_documents_for_donut, writer_batch_size=50)\r\n```", "The problem I had was to do with map using fork and copying locks from the\r\nparent process in acquired state. I ended up changing the context to use\r\nforkserver instead.\r\n\r\n\r\nOn Wed, Nov 29, 2023, 10:04 PM Mario Šaško ***@***.***> wrote:\r\n\r\n> @ewfian <https://github.com/ewfian>\r\n>\r\n> If I directly iterate over the dataset and call the mapping method, it is\r\n> very fast\r\n>\r\n> Dataset.map must also convert the images into bytes to write them to an\r\n> Arrow file (the write itself takes some time, too).\r\n>\r\n> You can make the map faster by manually converting the images into an\r\n> \"arrow-compatible\" representation. Otherwise, the Pillow defaults are used\r\n> when saving an image, which seems particularly slow for the notebook's case.\r\n>\r\n> def preprocess_documents_for_donut(sample):\r\n> text = json.loads(sample[\"text\"])\r\n> d_doc = task_start_token + json2token(text) + eos_token\r\n> image = sample[\"image\"].convert('RGB')\r\n> # convert image to bytes\r\n> buffer = io.BytesIO()\r\n> image.save(buffer, format=\"PNG\", compress_level=1)\r\n> return {\"image\": {\"bytes\": buffer.getvalue()}, \"text\": d_doc}\r\n> proc_dataset = dataset.map(preprocess_documents_for_donut, writer_batch_size=50)\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1833033973>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/ABDD3ZKKEKJVWBFH7QHLRJ3YG7ZUJAVCNFSM6AAAAAA6HDKPSCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMZTGAZTGOJXGM>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "I face the same issue many times.\r\n\r\nNot only when using the transformers' tokenizer, but also when applying nltk's [pos_tag](https://www.nltk.org/api/nltk.tag.pos_tag.html) to the entire English Wikipedia. So I suspect the cause is not in the tokenizer but in the Dataset.map\r\n\r\nMy case:\r\nAt the beginning of the run, the speed was 600 samples/s, but it slowed down to 20 samples/s at around 90% (after 3 hours). I am concerned that the CPU usage was only about 5% at the end of the run, even though there was still lots of data left.", "It is the interaction of fork() inside the map and tokenizer mutexes/locks.\r\n\r\nYou have to set up your own process pool and use fork server instead of\r\nfork.\r\n\r\n\r\nOn Tue, Aug 6, 2024, 11:44 AM yuji96 ***@***.***> wrote:\r\n\r\n> I face the same issue many times.\r\n>\r\n> Not only when using the transformers' tokenizer, but also when applying\r\n> nltk's pos_tag <https://www.nltk.org/api/nltk.tag.pos_tag.html> to the\r\n> entire English Wikipedia. So I suspect the cause is not in the tokenizer\r\n> but in the Dataset.map\r\n>\r\n> My case:\r\n> At the beginning of the run, the speed was 600 samples/s, but it slowed\r\n> down to 20 samples/s at around 90% (after 3 hours). I am concerned that the\r\n> CPU usage was only about 5% at the end of the run, even though there was\r\n> still lots of data left.\r\n>\r\n> #6319 (comment)\r\n> <https://github.com/huggingface/datasets/issues/6319#issuecomment-1771629160>\r\n> It's very nice to hear that the run is complete, but the original issue\r\n> has not been solved, which is that it gets slower and slower. As it is now,\r\n> Dataset.map will not be able to handle the large datasets that are getting\r\n> larger day by day.\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/6319#issuecomment-2271603976>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/ABDD3ZLFHSIGNNXAEJJIXWLZQDVPTAVCNFSM6AAAAABMCTVK2SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZRGYYDGOJXGY>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "Thank you for your advice!\r\n\r\nI added `multiprocess.set_start_method(\"forkserver\")` but the result seemed to be the same. In my case, it may be due to the very simple fact that about 10% of the process, which includes long text, never ends. I'll try shard by data size.\r\n\r\n", "Would recommend using `LlamaTokenizerFast` not `LlamaTokenizer` ! " ]

[]

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

[ "Thanks for reporting. We are investigating the issue.", "We have opened an issue in the corresponding Hub dataset: https://huggingface.co/datasets/sentiment140/discussions/3\r\n\r\nLet's continue the discussion there." ]

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

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

[]

[]

[]

[ "Thanks for reporting! We've spotted the bugs with the `array.values` handling and are fixing them in https://github.com/huggingface/datasets/pull/6283 (should be part of the next release).", "> Thanks for reporting! We've spotted the bugs with the `array.values` handling and are fixing them in #6283 (should be part of the next release).\r\n\r\ni encounter another exception while cast_column to type `Sequence(feature={\"points\": Array2D(shape=(-1, 2), dtype=\"int64\"), \"label\": ClassLabel(num_classes=num_classes, names=names)})`\r\n\r\nwhile my data like this: '{\"points\": [[0.6,0.6], [0.7,0.7], [0.8,0.8]], \"label\": \"A1\"}'\r\n\r\nhere is the backtrace info:\r\n\r\n```\r\n out = func(dataset, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 2110, in cast_column\r\n return self.cast(features)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 2055, in cast\r\n dataset = dataset.map(\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 592, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 557, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 3097, in map\r\n for rank, done, content in Dataset._map_single(**dataset_kwargs):\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 3474, in _map_single\r\n batch = apply_function_on_filtered_inputs(\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/arrow_dataset.py\", line 3353, in apply_function_on_filtered_inputs\r\n processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2329, in table_cast\r\n return cast_table_to_schema(table, schema)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2288, in cast_table_to_schema\r\n arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2288, in <listcomp>\r\n arrays = [cast_array_to_feature(table[name], feature) for name, feature in features.items()]\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1831, in wrapper\r\n return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1831, in <listcomp>\r\n return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2073, in cast_array_to_feature\r\n arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2073, in <listcomp>\r\n arrays = [_c(array.field(name), subfeature) for name, subfeature in feature.items()]\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1833, in wrapper\r\n return func(array, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2095, in cast_array_to_feature\r\n casted_values = _c(array.values, feature.feature)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1833, in wrapper\r\n return func(array, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 2144, in cast_array_to_feature\r\n return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1833, in wrapper\r\n return func(array, *args, **kwargs)\r\n File \"/home/protoss.gao/.local/lib/python3.9/site-packages/datasets/table.py\", line 1967, in array_cast\r\n return pa_type.wrap_array(array)\r\n File \"pyarrow/types.pxi\", line 1369, in pyarrow.lib.BaseExtensionType.wrap_array\r\nTypeError: Incompatible storage type for extension<arrow.py_extension_type<Array2DExtensionType>>: expected list<item: list<item: double>>, got list<item: double>\r\n```\r\nand i print(array) in datasets/table.py:1967 indeed get 2D list. is that same issue in #6283 ?\r\n\r\nbesides this, hugging face datasets seems don't naturally support multi-labels which means `Sequence(ClassLabel)` illegal if data is [\"label1\", \"label2\"]. so i have to define a class derived from `ClassLabel`, like this:\r\n\r\n```\r\nclass AisClassLabels(ClassLabel):\r\n def encode_example(self, example_data):\r\n if self.num_classes is None:\r\n raise ValueError(\r\n \"Trying to use ClassLabel feature with undefined number of class. \"\r\n \"Please set ClassLabel.names or num_classes.\"\r\n )\r\n if not isinstance(example_data, list):\r\n example_data = [example_data]\r\n\r\n for i in range(len(example_data)):\r\n if isinstance(example_data[i], str):\r\n example_data[i] = self.str2int(example_data[i])\r\n if not -1 <= example_data[i] < self.num_classes:\r\n raise ValueError(f\"Class label {example_data:d} greater than configured num_classes {self.num_classes}\")\r\n return example_data\r\n```\r\nand it works well in my case. but is there any recommend way to implement multi-labels?", "`Incompatible storage type for extension<arrow.py_extension_type<Array2DExtensionType>>: expected list<item: list<item: double>>, got list<item: double>`\r\nif i change `Array2D(shape=(-1, 2), dtype=\"int64\")` to `Sequence(Value(\"int64\"))` , every thing goes well. but my data is 2D int list", "i test Sequence(ClassLabel) is ok if one column is label list. but it is not ok in nested column such as `Sequence(feature= {\"points\": Sequence(Value(\"int32\")), \"label\": Sequence(ClassLabel(num_classes....)))`. in this case i need override ClassLabels. encode_example as i given above." ]

[]

[]

[ "This (Windows) issue was fixed in `fsspec` in https://github.com/fsspec/filesystem_spec/pull/1275. So, to avoid the error, update the `fsspec` installation with `pip install -U fsspec`.", "> This (Windows) issue was fixed in `fsspec` in [fsspec/filesystem_spec#1275](https://github.com/fsspec/filesystem_spec/pull/1275). So, to avoid the error, update the `fsspec` installation with `pip install -U fsspec`.\r\n\r\nafter I run `pip install -U fsspec`\r\n\r\nit occurs a new error:\r\n```\r\nERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflict\r\ns.\r\ndatasets 2.14.5 requires fsspec[http]<2023.9.0,>=2023.1.0, but you have fsspec 2023.9.2 which is incompatible.\r\n\r\n```", "The `fsspec<2023.9.0` upper bound will be removed in the next release. The `ResourceError` fix is also present in version 2023.6.0, so use that version in the meantime (`pip install fsspec==2023.6.0`).", "> The `fsspec<2023.9.0` upper bound will be removed in the next release. The `ResourceError` fix is also present in version 2023.6.0, so use that version in the meantime (`pip install fsspec==2023.6.0`).\r\n\r\nthanks for reply！" ]

[]

[]

[ "more information：\r\n``` \r\nFile \"text2vec\\__init__.py\", line 8, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"text2vec\\bertmatching_model.py\", line 19, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"text2vec\\bertmatching_dataset.py\", line 7, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"datasets\\__init__.py\", line 52, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"datasets\\inspect.py\", line 30, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"datasets\\load.py\", line 58, in <module>\r\nFile \"<frozen importlib._bootstrap>\", line 1027, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 688, in _load_unlocked\r\nFile \"PyInstaller\\loader\\pyimod02_importers.py\", line 499, in exec_module\r\nFile \"datasets\\packaged_modules\\__init__.py\", line 31, in <module>\r\nFile \"inspect.py\", line 1147, in getsource\r\nFile \"inspect.py\", line 1129, in getsourcelines\r\nFile \"inspect.py\", line 958, in findsource\r\nOSError: could not get source code\r\n```\r\n", "Can you share a reproducer? I haven't been able to reproduce the error myself.", "> '\r\n\r\nthanks,I solve it.it's about pyinstaller.", "1", "> > '\r\n> \r\n> thanks,I solve it.it's about pyinstaller.\r\n\r\nI encountered the same error, how to solve it?" ]

[]

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

[ "Thanks for reporting, @finiteautomata.\r\n\r\nWe are investigating it. ", "There is a bug in `datasets`. You can see our proposed fix:\r\n- #6309 " ]

[]

[]

[ "You can find the reasoning behind this naming scheme [here](https://github.com/huggingface/transformers/pull/16343#discussion_r931182168).\r\n\r\nThis point has been raised several times, so I'd be okay with starting with `00001-` (also to be consistent with the `transformers` sharding), but I'm not sure @lhoestq agrees.", "We start at 0 in `datasets` for consistency with Apache Spark, Apache Beam, Dask and others.\r\n\r\nAlso note `transformers` isn't a good reference on this topic. I talked with the maintainers when they added shards but it was already released this way. Though we found that there is a backward-compatible way in `transformers` to start at 0, but no request from `transformers` users to changes this AFAIK.", "not sure it would be a good idea to break the consistency now, IMO", "Makes sense to start at 0 for plenty of good reasons so I'm on board.\r\n\r\nWhat about the second part `-of-0000X`? With single commit PR #6269 just getting merged, there was a note about issues with 100+ file edits https://github.com/huggingface/datasets/pull/6269#issuecomment-1755428581.\r\n\r\nThat would be my last remaining concern in the context of the `push_to_hub(..., append=True)` work to be done, where appending a single file to the full dataset will require renaming every other existing file in the dataset. If it doesn't seem like a big issue for this work then all the better 👍" ]

[]

[]

[ "`writer._num_bytes` is updated every `writer_batch_size`-th call to the `write` method (default `writer_batch_size` is 1000 (examples)). You should be able to see the update by passing a smaller `writer_batch_size` to the `load_dataset_builder`.\r\n\r\nWe could improve this by supporting the string `writer_batch_size` version as we do with `max_shard_size`, and capping `writer_batch_size` to `max_shard_size` in scenarios where the default `writer_batch_size` > `max_shard_size`. ", "Thanks, reducing `writer_batch_size` solved my problem :)" ]

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

[]

[]

[]

[]

[ "It looks to be the same issue as the one reported in https://discuss.huggingface.co/t/indexerror-invalid-key-16-is-out-of-bounds-for-size-0.\r\n\r\nCan you check the length of `train_dataset` before the `train_sampler = self._get_train_sampler()` (and after `_remove_unused_columns`) line?" ]

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

[]

[]

[]

[]

[ "Hi! Can you provide a code that reproduces the issue?\r\n\r\nAlso, which version of `datasets` are you using? You can check this by running `python -c \"import datasets; print(datasets.__version__)\"` inside the env. We added support for \"float images\" in `datasets 2.9`." ]

[]

[]

[ "Thanks for reporting! I've opened a PR with a fix" ]

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

[]

[ "Yea I think waiting for #6269 would be best, or branching from it. For reference, this [PR](https://github.com/LAION-AI/Discord-Scrapers/pull/2) is progressing pretty well which will do similar using the hf hub for our LAION dataset bot https://github.com/LAION-AI/Discord-Scrapers/pull/2. ", "Is there any update on this?", "Is there any update on this?", "No update so far on this feature request but for broader context, this announce will help with incremental datasets https://huggingface.co/blog/xethub-joins-hf :)" ]

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

[]

[ "A duplicate of https://github.com/huggingface/datasets/issues/4796.\r\n\r\nWe could get this for free by implementing the `Image` feature as an extension type, as shown in [this](https://colab.research.google.com/drive/1Uzm_tXVpGTwbzleDConWcNjacwO1yxE4?usp=sharing) Colab (example with UUIDs).\r\n", "+1 to this\r\nCalling this line with a df that contains a PIL image (as they are returned from load_dataset)\r\n`ds = Dataset.from_pandas(df)`\r\nResults in this error:\r\n`ArrowInvalid: ('Could not convert <PIL.PngImagePlugin.PngImageFile image mode=RGB size=1024x1024 at 0x2B41F2D70> with type PngImageFile: did not recognize Python value type when inferring an Arrow data type', 'Conversion failed for column image with type object')`" ]

[]

[]

[ "There is no \"text\" column in the `amazon_reviews_multi`, hence the `KeyError`. You can get the column names by running `dataset.column_names`." ]

[]

[]

[ "You should be able to load the images by modifying the `load_dataset` call like this:\r\n```python\r\ndataset = load_dataset(\"imagefolder\", data_dir=\"/content/datasets/PotholeDetectionYOLOv8-1\")\r\n```\r\n\r\nThe `imagefolder` builder expects the image files to be in `path/label/image_file` (e.g. .`.../train/dog/image_1.jpg`), so the solution for the labels in your case is to create metadata files (one for each split; as explained [here](https://huggingface.co/docs/datasets/image_dataset#imagefolder)) that map the images to their labels.", "> You should be able to load the images by modifying the `load_dataset` call like this:\r\n> \r\n> ```python\r\n> dataset = load_dataset(\"imagefolder\", data_dir=\"/content/datasets/PotholeDetectionYOLOv8-1\")\r\n> ```\r\n> \r\n> The `imagefolder` builder expects the image files to be in `path/label/image_file` (e.g. .`.../train/dog/image_1.jpg`), so the solution for the labels in your case is to create metadata files (one for each split; as explained [here](https://huggingface.co/docs/datasets/image_dataset#imagefolder)) that map the images to their labels.\r\n\r\nI tried like this but only uploads images and not labels, Andyrasika/potholes-dataset", "As explained in my previous comment, you need to define metadata files to load the labels or update the paths to be in the format `train/label/image` (`train- image /n -labels` is not supported by the loader).", "I downloaded my file after annotating using roboflow . It gives train-\r\nimages, labels , test- images, labels , valid- images, labels . I hope it\r\ngives you an idea of the dataset . Please advise on this dataset\r\n\r\nOn Tue, Oct 10, 2023 at 18:12 Mario Šaško ***@***.***> wrote:\r\n\r\n> As explained in my previous comment, you need to define metadata files to\r\n> load the labels or update the paths to be in the format train/label/image\r\n> (train- image /n -labels is not supported by the loader).\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/datasets/issues/6285#issuecomment-1755335215>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AE4LJNN56FWWTSBYTSTUWHLX6U7CVAVCNFSM6AAAAAA5YHCSTGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONJVGMZTKMRRGU>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n" ]

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

[]

[ "This dataset is already available on the Hub: https://huggingface.co/datasets/facebook/belebele.\r\n" ]

[]

[]

[ "Thanks for reporting! I've opened a PR with a fix.", "Thanks for the quick response @mariosasko! I just installed your branch via `poetry add 'git+https://github.com/huggingface/datasets#fix-array_values'` and I can confirm it works on the example provided.\r\n\r\nFollow up question for you, should `None`s be supported in these types of features as they are in others?\r\n\r\nFor example, the following script:\r\n\r\n```\r\nfrom datasets import Features, Value, Sequence, ClassLabel, Dataset\r\n\r\ndataset_features = Features({\r\n 'text': Value('string'),\r\n 'embedding': Sequence(Value('double'), length=2),\r\n 'categories': Sequence(ClassLabel(names=sorted([\r\n 'one',\r\n 'two',\r\n 'three'\r\n ]))),\r\n})\r\n\r\ndataset = Dataset.from_dict(\r\n {\r\n 'text': ['A'] * 10000,\r\n \"embedding\": [None] * 10000, # THIS LINE CHANGED\r\n 'categories': [[0]] * 10000,\r\n },\r\n features=dataset_features\r\n)\r\n\r\ndef test_mapper(r):\r\n r['text'] = list(map(lambda t: t + ' b', r['text']))\r\n return r\r\n\r\n\r\ndataset = dataset.map(test_mapper, batched=True, batch_size=10, features=dataset_features, num_proc=2)\r\n```\r\n\r\nfails with\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/multiprocess/pool.py\", line 125, in worker\r\n result = (True, func(*args, **kwds))\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/utils/py_utils.py\", line 1354, in _write_generator_to_queue\r\n for i, result in enumerate(func(**kwargs)):\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/arrow_dataset.py\", line 3493, in _map_single\r\n writer.write_batch(batch)\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/arrow_writer.py\", line 549, in write_batch\r\n array = cast_array_to_feature(col_values, col_type) if col_type is not None else col_values\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/table.py\", line 1831, in wrapper\r\n return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/table.py\", line 1831, in <listcomp>\r\n return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])\r\n File \"/home/jmif/.virtualenvs/llm-training/lib/python3.10/site-packages/datasets/table.py\", line 2160, in cast_array_to_feature\r\n raise TypeError(f\"Couldn't cast array of type\\n{array.type}\\nto\\n{feature}\")\r\nTypeError: Couldn't cast array of type\r\nfixed_size_list<item: double>[2]\r\nto\r\nSequence(feature=Value(dtype='float64', id=None), length=2, id=None)\r\n```\r\n\r\nIdeally we can have empty embedding columns as well!", "This part of PyArrow is buggy and inconsistent regarding features implemented across the types, so the only option is to operate on the Arrow buffer level to fix issues such as the above one.", "Ok - can you take the POC I did [here](https://github.com/huggingface/datasets/commit/15443098e9ce053943172f7ec6fce3769d7dff6e)? Happy to turn this into an actual PR but would appreciate feedback on the implementation before I take another pass!" ]

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

[]

[ "This is exactly what I was looking for. It would also be very useful for me :-)", "This issue is really smashing the selling point of HF datasets... The only workaround I've found so far is to create a customized IterableDataloader which improves the loading speed to some extent.\r\n\r\nFor example I've a HF dataset `dt_train` with `len(dt_train) == 1M`. Using plain DataLoader is extremely slow:\r\n```\r\n%%time\r\ndl_train = DataLoader(dt_train, batch_size=128, shuffle = True)\r\nfor batch in dl_train:\r\n pass\r\n``` \r\n\r\n```\r\nCPU times: user 24min 35s, sys: 704 ms, total: 24min 36s\r\nWall time: 24min 37s\r\n```\r\nAnd DataLoader works even worse with HF's iterable_dataset:\r\n```\r\n%%time\r\ndt_train_ = dt_train.with_format(None).to_iterable_dataset(num_shards=64).shuffle(buffer_size=10_000)\r\ndl_train = DataLoader(dt_train_, batch_size=128)\r\nfor batch in dl_train:\r\n pass\r\n```\r\n```\r\nCPU times: user 1h 6min 2s, sys: 4.28 s, total: 1h 6min 6s\r\nWall time: 1h 7min 53s\r\n```\r\nWorkaround by running a customized wrapper:\r\n```\r\n%%time\r\nfrom torch.utils.data import DataLoader, IterableDataset\r\n\r\nclass Dataset2Iterable(IterableDataset):\r\n \"\"\"\r\n Wrapper to use a HF dataset as pytorch IterableDataset to speed up data loading.\r\n \"\"\"\r\n def __init__(self, dataset, batch_size=1, shuffle=True):\r\n super(Dataset2Iterable).__init__()\r\n self.dataset = dataset\r\n self.batch_size = batch_size\r\n self.shuffle = shuffle\r\n\r\n def __iter__(self):\r\n if self.shuffle: self.dataset.shuffle()\r\n return self.dataset.iter(batch_size=self.batch_size)\r\n\r\ndl_train = DataLoader(Dataset2Iterable(dt_train, batch_size = 128), batch_size=1, num_workers=0)\r\nfor n in range(2):\r\n for batch in dl_train:\r\n pass\r\n```\r\nThe speed still is slower than using tensorflow's loader but improved a lot than previous code:\r\n```\r\nCPU times: user 4min 18s, sys: 0 ns, total: 4min 18s\r\nWall time: 4min 20s\r\n```\r\nNote that the way I implemented `Dataset2Iterable` will only work with `num_workers=0`.", "I can confirm that @zhh210's solution works with `num_workers=0`. However, for my use case, this was still slower than tokenizing on the fly through a collator and leveraging multiple workers in the dataloder.\r\n\r\n@lhoestq I think this is an important use case (e.g., streaming from a large dataset, online or stored on disk). What do you think might be the best solution to move forward?", "I guess it can be implemented using a batched`.map()` under the hood that returns a single item containing the input batch.\r\n\r\nIn the meantime you can use this:\r\n\r\n```python\r\ndef batch(unbatched: dict[str, list]) -> dict[str, list]:\r\n return {k: [v] for k, v in unbatched}\r\n\r\nbatched_dataset = dataset.map(batch, batched=True, batch_size=batch_size)\r\n```\r\n\r\nThough it would be great to have a `.batch()` method indeed, I'd be happy to help with anyone wants to open a PR", "If no one else is planning to work on this, I can take it on. I'll wait until next week, and if no one has started a PR by then, I'll go ahead and open one.", "It looks like the implementation of IterableDataset is still using a hardcoded batch size of 1. For example in line 2063 in [`/datsets/src/datasets/iterable_dataset.py`](https://github.com/huggingface/datasets/blob/3.0.2/src/datasets/iterable_dataset.py#L2063). Iterating over IterableDataset with large batch sizes therefore remains slow, even when using `batch()`. I guess then the data are not being read from one contiguous chunk of memory. Instead every example is retrieved one by one, leading to long dataloading times. As a minimal example: Load c4 dataset and iterate over it with a large batch size.\r\n```python\r\nimport datasets\r\nfrom timeit import default_timer as timer\r\nc4 = datasets.load_dataset(\"allenai/c4\", \"en\", streaming=True, split=\"train\")\r\nc4_batched = c4.batch(512**2) # use large batch size\r\niterator = iter(c4_batched)\r\nfor i in range(5):\r\n start_time=timer()\r\n next(iterator) # get next batch\r\n end_time = timer()\r\n print(f\"time for one batch: {end_time-start_time}\")\r\n```\r\nThis results in the following output for me:\r\ntime for one batch: 12.615376660600305\r\ntime for one batch: 13.011422813870013\r\ntime for one batch: 14.157325950451195\r\ntime for one batch: 14.225894245319068\r\ntime for one batch: 13.898222777992487\r\n\r\nBecause I want to use my IterableDataset with the pytorch dataloader I rewrote the `__iter_pytorch__` and the `__iter__` functions like so and am getting much faster dataloading times. I marked the lines I changed with \"# changed here\":\r\n\r\n```python\r\nfrom datasets.iterable_dataset import _convert_to_arrow\r\nfrom datasets.formatting import TensorFormatter, get_formatter\r\nfrom datasets.features.features import cast_to_python_objects\r\nimport sys\r\nimport fsspec.asyn\r\nfrom itertools import islice\r\nfrom datasets.utils.logging import get_logger\r\nfrom datasets.iterable_dataset import _examples_to_batch, _apply_feature_types_on_batch, _apply_feature_types_on_example\r\n\r\nlogger = get_logger(__name__)\r\n\r\ndef __iter__(self):\r\n if \"torch\" in sys.modules:\r\n import torch.utils.data\r\n\r\n worker_info = torch.utils.data.get_worker_info()\r\n if isinstance(self, torch.utils.data.IterableDataset) and worker_info is not None:\r\n # We're a torch.utils.data.IterableDataset in a PyTorch worker process\r\n yield from self._iter_pytorch()\r\n return\r\n\r\n ex_iterable = self._prepare_ex_iterable_for_iteration(batch_size=self.batch_size, drop_last_batch=self.drop_last_batch) # changed here\r\n if self._formatting:\r\n formatter = get_formatter(self._formatting.format_type, features=self.features)\r\n format_dict = (\r\n formatter.recursive_tensorize if isinstance(formatter, TensorFormatter) else cast_to_python_objects\r\n )\r\n else:\r\n format_dict = None\r\n\r\n if self._formatting and (ex_iterable.iter_arrow or self._formatting.format_type == \"arrow\"):\r\n if ex_iterable.iter_arrow:\r\n iterator = ex_iterable.iter_arrow()\r\n else:\r\n iterator = _convert_to_arrow(ex_iterable, batch_size=self.batch_size) # changed here\r\n for key, pa_table in iterator:\r\n yield formatter.format_row(pa_table)\r\n return\r\n\r\n for key, example in ex_iterable:\r\n if self.features and not ex_iterable.is_typed:\r\n # `IterableDataset` automatically fills missing columns with None.\r\n # This is done with `_apply_feature_types_on_example`.\r\n example = _apply_feature_types_on_example(\r\n example, self.features, token_per_repo_id=self._token_per_repo_id\r\n )\r\n yield format_dict(example) if format_dict else example\r\n\r\n\r\n\r\ndef _iter_pytorch(self):\r\n ex_iterable = self._prepare_ex_iterable_for_iteration(batch_size=self.batch_size, drop_last_batch=self.drop_last_batch) # changed here\r\n # Fix for fsspec when using multiprocess to avoid hanging in the ML training loop. (only required for fsspec >= 0.9.0)\r\n # See https://github.com/fsspec/gcsfs/issues/379\r\n fsspec.asyn.reset_lock()\r\n # check if there aren't too many workers\r\n import torch.utils.data\r\n\r\n worker_info = torch.utils.data.get_worker_info()\r\n if self._is_main_process() and ex_iterable.n_shards < worker_info.num_workers:\r\n logger.warning(\r\n f\"Too many dataloader workers: {worker_info.num_workers} (max is dataset.n_shards={ex_iterable.n_shards}). \"\r\n f\"Stopping {worker_info.num_workers - ex_iterable.n_shards} dataloader workers.\"\r\n )\r\n logger.info(\r\n f\"To parallelize data loading, we give each process some shards (or data sources) to process. \"\r\n f\"Therefore it's unnecessary to have a number of workers greater than dataset.n_shards={ex_iterable.n_shards}. \"\r\n f\"To enable more parallelism, please split the dataset in more files than {ex_iterable.n_shards}.\"\r\n )\r\n # split workload\r\n _log_prefix = f\"node#{self._distributed.rank} \" if self._distributed else \"\"\r\n shards_indices = ex_iterable.split_shard_indices_by_worker(worker_info.id, worker_info.num_workers)\r\n if shards_indices:\r\n logger.debug(\r\n f\"{_log_prefix}dataloader worker#{worker_info.id}, ': Starting to iterate over {len(shards_indices)}/{ex_iterable.n_shards} shards.\"\r\n )\r\n ex_iterable = ex_iterable.shard_data_sources(worker_id=worker_info.id, num_workers=worker_info.num_workers)\r\n self._state_dict = ex_iterable._init_state_dict()\r\n if self._starting_state_dict:\r\n ex_iterable.load_state_dict(self._starting_state_dict)\r\n\r\n if self._formatting:\r\n formatter = get_formatter(self._formatting.format_type, features=self.features)\r\n format_dict = (\r\n formatter.recursive_tensorize if isinstance(formatter, TensorFormatter) else cast_to_python_objects\r\n )\r\n else:\r\n format_dict = None\r\n\r\n if self._formatting and (ex_iterable.iter_arrow or self._formatting == \"arrow\"):\r\n if ex_iterable.iter_arrow:\r\n iterator = ex_iterable.iter_arrow()\r\n else:\r\n iterator = _convert_to_arrow(ex_iterable, batch_size=self.batch_size) # changed here\r\n if self.batch_size > 1: # changed here until end of file\r\n for key, pa_table in iterator:\r\n yield formatter.format_batch(pa_table)\r\n return\r\n else:\r\n for key, pa_table in iterator:\r\n yield formatter.format_row(pa_table)\r\n return\r\n\r\n iterator = iter(ex_iterable)\r\n if self.batch_size > 1:\r\n for key, example in iterator:\r\n # If batched, first build the batch\r\n examples = [example] + [example for key, example in islice(iterator, self.batch_size - 1)]\r\n if self.drop_last_batch and len(examples) < self.batch_size: # ignore last batch\r\n return\r\n batch = _examples_to_batch(examples)\r\n if self.features and not ex_iterable.is_typed:\r\n # `IterableDataset` automatically fills missing columns with None.\r\n # This is done with `_apply_feature_types_on_batch`.\r\n batch = _apply_feature_types_on_batch(batch, self.features, token_per_repo_id=self._token_per_repo_id)\r\n yield format_dict(batch) if format_dict else batch\r\n else:\r\n for key, example in ex_iterable:\r\n if self.features and not ex_iterable.is_typed:\r\n # `IterableDataset` automatically fills missing columns with None.\r\n # This is done with `_apply_feature_types_on_example`.\r\n example = _apply_feature_types_on_example(\r\n example, self.features, token_per_repo_id=self._token_per_repo_id\r\n )\r\n yield format_dict(example) if format_dict else example\r\n logger.debug(\r\n f\"{_log_prefix}dataloader worker#{worker_info.id}, ': Finished iterating over {len(shards_indices)}/{ex_iterable.n_shards} shards.\"\r\n )\r\n else:\r\n logger.debug(\r\n f\"{_log_prefix}dataloader worker#{worker_info.id}, ': Stopping... Number of dataset shards < num_workers ({ex_iterable.n_shards}<{worker_info.num_workers}).\"\r\n )\r\n```\r\n\r\nFor anyone wanting to try it you can patch it into datasets by overwriting the function via `setattr(datasets.IterableDataset, '_iter_pytorch', _iter_pytorch)`\r\n\r\nI don't really know what most of the rest of the code is doing so no idea if this is a valid fix or not, but it seems to work for me.\r\nExample of running the fix:\r\n```python\r\nfrom torch.utils.data.dataloader import DataLoader\r\nc4.batch_size = 512**2 # set batch size here\r\ndataloader = Dataloader(c4, batch_size=None) # use custom batching from IterableDataset\r\niterator = iter(dataloader)\r\nfor i in range(5):\r\n start_time=timer()\r\n next(iterator) #get the batch\r\n end_time = timer()\r\n print(f\"time for one batch: {end_time-start_time}\")\r\n```\r\nI now get \r\ntime for one batch: 0.6047679269686341\r\ntime for one batch: 0.000248616561293602\r\ntime for one batch: 0.00017435848712921143\r\ntime for one batch: 0.00015910807996988297\r\ntime for one batch: 0.00015317369252443314\r\n\r\nI love the datasets library and it would be great if iterating with large batch sizes would be supported directly, either with a similar fix to mine or in some other way :)", "Hi @taczin , thanks for reporting !\r\n\r\nIndeed the `IterableDataset.batch()` implementation is quite naive is manipulates python objects:\r\n\r\nhttps://github.com/huggingface/datasets/blob/d37ed46ebf45981131bd3678173dbb4b7e2b2f1a/src/datasets/iterable_dataset.py#L3026-L3029\r\n\r\nHowever it can be much faster if it can be applied on the Arrow data, maybe using something like this (untested)\r\n\r\n```python\r\ndef batch_fn(unbatched): \r\n return {k: [v] for k, v in unbatched.items()} \r\n\r\ndef batch_fn_arrow(unbatched_pa_table): \r\n offsets = pa.array([0, len(unbatched_pa_table)])\r\n return pa.Table.from_arrays([\r\n pa.ListArray.from_arrays(offsets, unbatched_pa_table[k])\r\n for k in unbatched_pa_table.column_names\r\n ], unbatched_pa_table.column_names)\r\n\r\nif self._ex_iterabe.iter_arrow:\r\n return self.with_format(\"arrow\").map(\r\n batch_fn_arrow, batched=True, batch_size=batch_size, drop_last_batch=drop_last_batch\r\n ).with_format(self._formatting.format_type if self._formatting else None)\r\nelse:\r\n return self.map(batch_fn, batched=True, batch_size=batch_size, drop_last_batch=drop_last_batch)\r\n```", "Hi @lhoestq , thanks for your answer. I was wondering: is there a reason why the internal call to `ex_iterable = self._prepare_ex_iterable_for_iteration()` in the IterableDataset code does not pass the batch size even though it could? If not passed the default of 1 is used, leading to the observed slow loading.", "After calling `.batch()`, `_prepare_ex_iterable_for_iteration` should use batch_size=1 since now each row in the dataset is actually a batch of the original dataset." ]

[]

[]

[ "`evaluate.load(\"paws-x\", \"es\")` throws the error because there is no such metric in the `evaluate` lib.\r\n\r\nSo, this is unrelated to our lib." ]

[]

[]

[ "Since you are using Windows, maybe moving the `map` call inside `if __name__ == \"__main__\"` can fix the issue:\r\n```python\r\nif __name__ == \"__main__\":\r\n common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names[\"train\"], num_proc=4)\r\n```\r\n\r\nOtherwise, the only solution is to set `num_proc=1`.", "> Since you are using Windows, maybe moving the `map` call inside `if __name__ == \"__main__\"` can fix the issue:\r\n> \r\n> ```python\r\n> if __name__ == \"__main__\":\r\n> common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names[\"train\"], num_proc=4)\r\n> ```\r\n> \r\n> Otherwise, the only solution is to set `num_proc=1`.\r\n\r\nThank you very much for the response, i eventually tried setting `num_proc=1` and now the jupyter notebook kernel keers dying after running the command, what do you think the issue could be, could it be that my system is not capable of running the command \"i'm using a Lenovo Thinkpad T440 with no GPU\"", "Firstly, you didn't define feature_extractor variable. Secondly, it is large nlp model. Hence you should use proper gpu, otherwise your machine's cpu will be overclock and you can do nothing." ]

[]

[]

[ "Hi! The process of contributing a dataset is explained here: https://huggingface.co/docs/datasets/upload_dataset. Also, check https://huggingface.co/docs/datasets/image_dataset for a more detailed explanation of how to share an image dataset." ]

[]

[]

[ "Please tell me if the above info is not enough for solving the problem. I will then make my dataset public temporarily so that you can really reproduce the bug. ", "Hi! \r\nCould you share how to solve this problem? \r\nI faced this same error. " ]

[]

[]

[ "This has already been reported in the HF Course repo (https://github.com/huggingface/course/issues/623).", "@lhoestq @albertvillanova @lewtun I don't think we are allowed to host these data files on the Hub (due to DMCA), which means the only option is to use a different dataset in the course (and to re-record the video 🙂), no?", "Keeping the video is maybe fine, we can add a note on youtube to suggest to load a dataset with a different name. Maybe C4 ? And update the code snippets on the website ?", "Maybe you want to try it with the PUBMED dataset that I reproduced based on the The [PubMed Abstract GitHub Site](http://github.com/thoppe/The-Pile-PubMed) and uploaded on the HuggingFace:\r\n\r\n```\r\nfrom datasets import load_dataset\r\npubmed_dataset = load_dataset(\"hwang2006/PUBMED_title_abstracts_2020_baseline\")\r\npubmed_dataset\r\n\r\n#Downloading data: 100%\r\n#7.98G/7.98G [11:47<00:00, 9.68MB/s]\r\n#Generating train split: 17722096/0 [00:36<00:00, 505376.37 examples/s]\r\n\r\n#DatasetDict({\r\n# train: Dataset({\r\n# features: ['meta', 'text'],\r\n# num_rows: 17722096\r\n# })\r\n#})\r\n```", "孔令涛说感谢感谢" ]

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

[]

[ "Also reported in https://github.com/huggingface/datasets/issues/6259", "I think it's best to drop duplicates with a `set` (as a temporary fix) and improve the patterns when/if https://github.com/fsspec/filesystem_spec/pull/1382 gets merged. @lhoestq Do you have some other ideas?", "Alternatively we could just use this no ?\r\n\r\n```python\r\nif config.FSSPEC_VERSION < version.parse(\"2023.9.0\"):\r\n KEYWORDS_IN_PATH_NAME_BASE_PATTERNS = [\r\n \"{keyword}[{sep}/]**\",\r\n \"**[{sep}]{keyword}[{sep}/]**\",\r\n \"**/{keyword}[{sep}/]**\",\r\n ]\r\nelse:\r\n KEYWORDS_IN_PATH_NAME_BASE_PATTERNS = [\r\n \"{keyword}[{sep}/]**\",\r\n \"**/*[{sep}]{keyword}[{sep}/]**\",\r\n \"**/*/{keyword}[{sep}/]**\",\r\n ]\r\n```\r\n\r\nThis way no need to implement sets, which would require a bit of work since we've always considered a list of pattern to be resolved as the concatenated list of resolved files for each pattern (including duplicates)\r\n", "Arf `\"**/*/{keyword}[{sep}/]**\"` does return `data/keyword.txt` in latest `fsspec` but not in `glob.glob`\r\n\r\nEDIT: actually forgot to set `recursive=True`", "Actually `glob.glob` does return it with `recursive=True` ! my bad", "Pff just tested and my idea sucks, pattern 1 and 3 obviously give duplicates ", "> I think it's best to drop duplicates with a set (as a temporary fix)\r\n\r\nI started https://github.com/huggingface/datasets/pull/6278 to use DataFilesSet objects instead of DataFilesList" ]

[]

[]

[]

[]

[]

[ "`gen_kwargs` should be a `dict`, as stated in the docstring, but you are passing a `list`.\r\n\r\nSo, to fix the error, replace the list of dicts with a dict of lists (and slightly modify the generator function):\r\n```python\r\nfrom pathlib import Path\r\nimport datasets\r\n\r\ndef process_yaml(files):\r\n for f in files:\r\n # process\r\n yield dict(...)\r\n\r\n\r\nif __name__ == '__main__':\r\n import sys\r\n dir = Path(sys.argv[0]).parent\r\n ds = datasets.Dataset.from_generator(process_yaml, gen_kwargs={'files': [f for f in dir.glob('*.yml')]})\r\n ds.to_json('training.jsonl')\r\n```", "That runs, and because my dataset is small, it's what I did to get past the problem.\r\nHowever, it does not produce a sharded dataset. From the doc string I expect there ought to be a way to call from_generator such that num_shards in the resulting data set is equal to the number of items in the list.\r\nThe part of the doc string that your suggestion is not responsive to is:\r\n` You can define a sharded dataset by passing the list of shards in *g\r\nen_kwargs*.\r\n`\r\n\r\nWhat your suggestion does is calls the generator once, with the list argument, and produces a single shard dataset.\r\n", "The sharding mentioned here refers to using this function with `num_proc` (multiprocessing splits the `kwargs` into shards and passes them to the generator function)\r\n\r\n> That runs, and because my dataset is small, it's what I did to get past the problem.\r\n\r\n`from_generator` generates a memory-mapped dataset (can be larger than RAM), so the dataset size should not be an issue unless the generator function's implementation does not properly free the memory.\r\n", "It sounds like you are saying that num_proc affects the form of gen_kwargs.\r\nAre you saying that for non-zero num_proc gen_kwargs should be a list whose length is the same as num_proc?\r\nOr are you saying that for non-zero num_proc, gen_kwargs should be a dict whose elements are lists the length of num_proc?\r\n", "I ran some tests. So, it looks like with num_proc greater than 1, gen_kwargs is expected to be a dict of lists. It calls the generator also with a dict of lists, but the lists are split.\r\nI.E. if my original has `gen_kwargs=dict(a=[0,1,2])`, then my generator might get called with `gen_kwalrgs=dict([0])`.\r\nThat all makes sense, but I definitely think there is room for improvement in the doc string here.\r\nIn order to suggest improvements to the doc string, I need to look at how the gen_kwargs are split, and figure out if:\r\n* num_proc needs to exactly equal the length of the lists\r\n* num_proc needs to evenly divide the length of the lists\r\n* Or there's no required relationship.\r\nI'll look into that and then propose an improved doc string if no one else gets to it first.", "Okay, that was fun; I took a dive through the dataset code and feel like I have a much better understanding.\r\nHere is my understanding of the behavior:\r\n* max_proc is an upper limit on the number of shards that `from_generator` produces\r\n* If `max_proc` is greater than 1, then all lists in *gen_kwargs* must be the same length\r\n* If the lists in *gen_kwargs* are shorter than *num_proc* elements, *num_proc* will be reduced and a warning produced. Put another way, `min(list_length, num_shards)` shards will be produced\r\n* The members of the lists in *gen_kwargs* will be partitioned among the created jobs.\r\nTo validate the above, take a look at\r\n`_number_of_shards_in_gen_kwargs` and `_distribute_shards` and `_split_gen_kwargs` in utils/sharding.py.\r\nI've also chased down starting at *from_generator* all the way through to GeneratorBuilder and the calls to the functions in sharding.py.\r\nTomorrow I'll take a look at the contributing guidelines and see what's involved in putting together a PR to improve the doc string." ]