Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "After looking into this, discovered that this is already supported if the `dataset_infos.json` file is configured correctly! Here is a working example: https://huggingface.co/datasets/abidlabs/test-audio-13\r\n\r\nThis should be probably be documented, though. " ]

[ "Thanks for reporting. I think this can be fixed by improving the `CachedDatasetModuleFactory` and making it look into the `parquet` cache directory (datasets from push_to_hub are loaded with the parquet dataset builder). I'll look into it", "Hi, I'm having the same issue. Is there any update on this?", "We haven't had a chance to fix this yet. If someone would like to give it a try I'd be happy to give some guidance", "@lhoestq Do you have an idea of what changes need to be made to `CachedDatasetModuleFactory`? I would be willing to take a crack at it. Currently unable to train with datasets I have `push_to_hub` on a cluster whose compute nodes are not connected to the internet.\r\n\r\nIt looks like it might be this line:\r\n\r\nhttps://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L994\r\n\r\nWhich wouldn't pick up the stuff saved under `\"datasets/allenai___parquet/*\"`. Additionally, the datasets saved under `\"datasets/allenai___parquet/*\"` appear to have hashes in their name, e.g. `\"datasets/allenai___parquet/my_dataset-def9ee5552a1043e\"`. This would not be detected by `CachedDatasetModuleFactory`, which currently looks for subdirectories\r\n\r\nhttps://github.com/huggingface/datasets/blob/0c1d099f87a883e52c42d3fd1f1052ad3967e647/src/datasets/load.py#L995-L999", "`importable_directory_path` is used to find a **dataset script** that was previously downloaded and cached from the Hub\r\n\r\nHowever in your case there's no dataset script on the Hub, only parquet files. So the logic must be extended for this case.\r\n\r\nIn particular I think you can add a new logic in the case where `hashes is None` (i.e. if there's no dataset script associated to the dataset in the cache).\r\n\r\nIn this case you can check directly in the in the datasets cache for a directory named `<namespace>__parquet` and a subdirectory named `<config_id>`. The config_id must match `{self.name.replace(\"/\", \"--\")}-*`. \r\n\r\nIn your case those two directories correspond to `allenai___parquet` and then `allenai--my_dataset-def9ee5552a1043e`\r\n\r\nThen you can find the most recent version of the dataset in subdirectories (e.g. sorting using the last modified time of the `dataset_info.json` file).\r\n\r\nFinally, we will need return the module that is used to load the dataset from the cache. It is the same module than the one that would have been normally used if you had an internet connection.\r\n\r\nAt that point you can ping me, because we will need to pass all this:\r\n- `module_path = _PACKAGED_DATASETS_MODULES[\"parquet\"][0]`\r\n- `hash` it corresponds the name of the directory that contains the .arrow file, inside `<namespace>__parquet/<config_id>`\r\n- ` builder_kwargs = {\"hash\": hash, \"repo_id\": self.name, \"config_id\": config_id}`\r\nand currently `config_id` is not a valid argument for a `DatasetBuilder`\r\n\r\nI think in the future we want to change this caching logic completely, since I don't find it super easy to play with.", "Hi! Is there a workaround for the time being?\r\nLike passing `data_dir` or something like that?\r\n\r\nI would like to use [this diffuser example](https://github.com/huggingface/diffusers/tree/main/examples/unconditional_image_generation) on my cluster whose nodes are not connected to the internet. I have downloaded the dataset online form the login node.", "Hi ! Yes you can save your dataset locally with `my_dataset.save_to_disk(\"path/to/local\")` and reload it later with `load_from_disk(\"path/to/local\")`\r\n\r\n(removing myself from assignees since I'm currently not working on this right now)", "Still not fixed? ......", "Any idea @lhoestq who to tag to fix this ? This is a very annoying bug, which is becoming more and more present since the push_to_hub API is getting used more ?", "Perhaps @mariosasko ? Thanks a lot for the great work on the lib !", "It should be easier to implement now that we improved the caching of datasets from `push_to_hub`: each dataset has its own directory in the cache.\r\n\r\nThe cache structure has been improved in https://github.com/huggingface/datasets/pull/5331. Now the cache structure is `\"{namespace__}<dataset_name>/<config_name>/<version>/<hash>/\"` which contains the arrow files `\"<dataset_name>-<split>.arrow\"` and `\"dataset_info.json\"`. \r\n\r\nThe idea is to extend `CachedDatasetModuleFactory` to also check if this directory exists in the cache (in addition to the already existing cache check) and return the requested dataset module. The module name can be found in the JSON file in the `builder_name` field.", "Any progress?", "I started a PR to draft the logic to reload datasets from the cache fi they were created with push_to_hub: https://github.com/huggingface/datasets/pull/6459\r\n\r\nFeel free to try it out", "It seems that this does not support dataset with uppercase name ", "Which version of `datasets` are you using ? This issue has been fixed with `datasets` 2.16", "I can confirm that this problem is still happening with `datasets` 2.17.0, installed from pip", "Can you share a code or a dataset that reproduces the issue ? It seems to work fine on my side.", "Yeah, \r\n```python\r\ndataset = load_dataset(\"roneneldan/TinyStories\")\r\n```\r\nI tried it with:\r\n```python\r\ndataset = load_dataset(\"roneneldan/tinystories\")\r\n```\r\nand it worked.\r\n\r\n> It seems that this does not support dataset with uppercase name\r\n\r\n@fecet was right, but if you just put the name lowercase, it works. " ]

[]

[ "Hi ! Thanks for your message :) This is a great idea indeed. We haven't started working on this yet though. For now I guess you can host your metric on the Hub (either with your model or your dataset) and use `hf_hub_download` to download it (docs [here](https://github.com/huggingface/huggingface_hub/blob/main/docs/hub/how-to-downstream.md#cached_download))", "This is a great solution in the meantime, thanks!", "Here's the code I used, in case it can be of help to someone else:\r\n```python\r\nimport os, shutil\r\nfrom huggingface_hub import hf_hub_download\r\ndef download_metric(repo_id, file_path):\r\n # repo_id: for models \"username/model_name\", for datasets \"datasets/username/model_name\"\r\n local_metric_path = hf_hub_download(repo_id=repo_id, filename=file_path)\r\n updated_local_metric_path = (os.path.dirname(local_metric_path) + os.path.basename(local_metric_path).replace(\".\", \"_\") + \".py\")\r\n shutil.copy(local_metric_path, updated_local_metric_path)\r\n return updated_local_metric_path\r\n\r\nmetric = load_metric(download_metric(REPO_ID, FILE_PATH))\r\n```", "Solved with https://github.com/huggingface/evaluate 🤗 ", "Yay!! cc @lvwerra @sashavor @douwekiela \r\n\r\nPlease share your feedback @eladsegal =)" ]

[ "This should also resolve: https://github.com/huggingface/datasets/issues/3185." ]

[]

[ "known issue due to https://github.com/huggingface/datasets/pull/2362 (and [internal](https://github.com/huggingface/moon-landing/issues/946)) , will be solved soon", "hmm actually i have no recollection of why I said that", "Because it has dots in some YAML keys, it can't be parsed and indexed by the back-end" ]

[]

[ "This issue is not reproducible." ]

[ "In the framework of BigScience:\r\n- bigscience-workshop/data_tooling#121\r\n\r\nI have created this dataset as a community dataset: https://huggingface.co/datasets/albertvillanova/pmc_open_access\r\n\r\nHowever, I was wondering that it may be more appropriate to move it under an org namespace: `pubmed_central` or `pmc`\r\nThis way, we could add other datasets I'm also working on: Author Manuscript Dataset, Historical OCR Dataset, LitArch Open Access Subset.\r\n\r\nWhat do you think? @lhoestq @mariosasko ", "Why not ! Having them under such namespaces would also help people searching for this kind of datasets.\r\nWe can also invite people from pubmed at one point", "DONE: https://huggingface.co/datasets/pmc/open_access" ]

[ "Thanks for reporting @VictorSanh.\r\n\r\nI'm looking at it... " ]

[ "Hi, which dataset is giving you an error?" ]

[ "Can you reopen with the correct dataset name (if relevant)?\r\n\r\nThanks", "The dataset viewer was down tonight. It works again." ]

[ "Hi ! According to the DPR paper, the wikipedia dump is the one from Dec. 20, 2018.\r\nEach instance contains a paragraph of at most 100 word, as well as the title of the wikipedia page it comes from and the DPR embedding (a 768-d vector).", "Closed by:\r\n- #3534" ]

[ "IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work), so that for datasets that have a working dataset preview, we can remove the dummy data IMO. On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.\r\n\r\nI don't really have an opinion regarding the JSON metadata as I don't know enough about it.\r\n\r\n", "I don't know all the details, but generally I'd be in favor of unifying the metadata formats into YAML inside .md (and so deprecating the dataset_infos.json) \r\n\r\n(Ultimately the CI can run on \"HuggingFace Actions\" instead of on GitHub)", "The dataset_infos.json file currently has these useful infos for each dataset configuration, that I think can be moved to the dataset tags:\r\n- Size of the dataset in MB: download size, arrow file size, and total size (sum of download + arrow)\r\n- Size of each split in MB and number of examples. Again this can be moved to the dataset tags\r\n- Feature type of each column\r\n- supported task templates (it defines what columns correspond to the features and labels for example)\r\n\r\nBut it also has this, which I'm not sure if it should be in the tags or not:\r\n- Checksums of the downloaded files for integrity verifications\r\n\r\nSo ultimately this file could probably be deprecated in favor of having the infos in the tags.\r\n\r\n> Also note that for generating both (dataset_infos.json file and dummy data), the entire dataset needs being downloaded. This can be an issue for huge datasets (like WIT, with 400 GB of data).\r\n\r\nTo get the exact number of examples and size in MB of the dataset, one needs to download and generate it completely. IMO these infos are very important when someone considers using a dataset. Though using streaming we could do some extrapolation to have approximate values instead.\r\n\r\nFor the integrity verifications we also need the number of examples and the checksums of the downloaded files, so it requires the dataset to be fully downloaded once. This can be optional though.\r\n\r\n> IMO, the data streaming test is good enough of a test that the dataset works correctly (assuming that we can more or less ensure that if streaming works then the non-streaming case will also work)\r\n\r\nI agree with this. Usually if a dataset works in streaming mode, then it works in non-streaming mode (the other way around is not true though).\r\n\r\n> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data.\r\n\r\nYes indeed, or at least make sure that it was tested on the true data.", "(note that if we wanted to display sizes, etc we could also pretty easily parse the `dataset_infos.json` on the hub side)", "I agree that we can move the relevant parts of `dataset_infos.json` to the YAML tags.\r\n\r\n> On the other hand, it seems like not all datasets have streaming enabled yet and for those datasets (if they are used a lot), I think it would be nice to continue testing some dummy data. <\r\n> > Yes indeed, or at least make sure that it was tested on the true data.\r\n\r\nI like the idea of testing streaming and falling back to the dummy data test if streaming does not work. Generating dummy data can be very tedious, so this would be a nice incentive for the contributors to make their datasets streamable. ", "CC: @severo ", "About dummy data, please see e.g. this PR: https://github.com/huggingface/datasets/pull/3692/commits/62368daac0672041524a471386d5e78005cf357a\r\n- I updated the previous dummy data: I just had to rename the file and its directory\r\n - the dummy data zip contains only a single file: `pubmed22n0001.xml.gz`\r\n\r\nThen I discover it fails: https://app.circleci.com/pipelines/github/huggingface/datasets/9800/workflows/173a4433-8feb-4fc6-ab9e-59762084e3e1/jobs/60437\r\n```\r\nNo such file or directory: '.../dummy_data/pubmed22n0002.xml.gz'\r\n```\r\n- it needs dummy data for all the 1114 files: \r\n `_URLs = [f\"ftp://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed22n{i:04d}.xml.gz\" for i in range(1, 1115)]`\r\n- this confirms me that it never passed the test: these dummy data files were not present before my PR\r\n- therefore, is it really useful the data test if we just ignore it when it does not pass?\r\n\r\nIn relation with JSON metadata, I'm generating the file for `pubmed` (see above) in a GCP instance: it's running for more than 3 hours and only 9 million examples generated so far (before my PR, it had 32 million, now it has more).", "I mention in https://github.com/huggingface/datasets-server/wiki/Preliminary-design that the future \"datasets server\" could be in charge of generating both the dummy data and the dataset-info.json file if required (or their equivalent).", "Hi ! I think dummy data generation is out of scope for the datasets server, since it's about generating the original data files.\r\n\r\nThat would be amazing to have it generate the dataset_infos.json though !", "From some offline discussion with @mariosasko and especially for vision datasets, we'll probably not require dummy data anymore and use streaming instead :) This will make adding a new dataset much easier.\r\nThis should also make sure that streaming works as expected directly in the CI, without having to check the dataset viewer once the PR is merged", "OK. I removed the \"dummy data\" item from the services of the dataset server", "It seems that migration from dataset-info.json to dataset card YAML has been acted.\r\n\r\nProbably it's a good idea, but I didn't find the pros and cons of this decision, so I put some I could think of:\r\n\r\npros:\r\n- only one file to parse, share, sync\r\n- it gives a hint to the users that if you write your dataset card, you should also specify the metadata\r\n\r\ncons:\r\n- the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums\r\n- YAML vs JSON: not sure which one is easier for users to fill and maintain\r\n- two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.\r\n- [low priority] besides the JSON file, we might want to support yaml or toml file if the user prefers (as [prettier](https://prettier.io/docs/en/configuration.html) and others do for their config files, for example). Inside the md, I understand that only YAML is allowed", "> the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums\r\n\r\nNote that we could simply not have the checksums in the YAML metadata at all, or maybe at one point have a pointer to another file instead.\r\n\r\nWe can also choose to hide (collapse) certain sections in the YAML by default when we open the dataset card editor.\r\n\r\n> two concepts are mixed in the same file (metadata and documentation). This means that if you're interested only in one of them, you still have to know how to parse the whole file.\r\n\r\nI think it's fine for now. Later if we really end up with too many YAML sections we can see if we need to tweak the API endpoints or the `datasets`/`huggingface_hub` tools\r\n\r\n> YAML vs JSON: not sure which one is easier for users to fill and maintain\r\n\r\nRegarding YAML vs JSON: I think YAML is easier to write by hand, and I also think that it's better for consistency - i.e. we're using more and more YAML to configure models/datasets/spaces", "I didn't know the decision was already taken. Good to know. 😅", "> the metadata header might be very long, before reaching the start of the README/dataset card. It might be surprising when you edit the file because the metadata is not shown on top when the dataset card is rendered (dataset page). It also somewhat prevents including large strings like the checksums\r\n\r\nWe can definitely work on this on the hub side to make the UX better", "Tensorflow Datasets catalog includes a community catalog where you can find and use HF datasets (see [here](https://www.tensorflow.org/datasets/community_catalog/huggingface)).\r\n\r\nFYI I noticed today that they are using the exported dataset_infos.json files from github to get the metadata (see their code [here](https://github.com/tensorflow/datasets/blob/a482f01c036a10496f5e22e69a2ef81b707cc418/tensorflow_datasets/scripts/documentation/build_community_catalog.py#L261))", "Metadata is now stored as YAML, and dummy data is deprecated, so I think we can close this issue." ]

[ "Hi! This is probably due to the fact that `IterableDataset.map` sets `features` to `None` before mapping examples. We can fix the issue by passing the old `features` dict to the map generator and performing encoding/decoding there (before calling the map transform function)." ]

[ "Hi @ToddMorrill, thanks for reporting.\r\n\r\nThree weeks ago I contacted the team who created the Pile dataset to report this issue with their data host server: https://the-eye.eu\r\n\r\nThey told me that unfortunately, the-eye was heavily affected by the recent tornado catastrophe in the US. They hope to have their data back online asap.", "Hi @ToddMorrill, people from the Pile team have mirrored their data in a new host server: https://mystic.the-eye.eu\r\n\r\nSee:\r\n- #3627\r\n\r\nIt should work if you update your URL.\r\n\r\nWe should also update the URL in our course material.", "The old URL is still present in the HuggingFace course here: \r\nhttps://huggingface.co/course/chapter5/4?fw=pt\r\n\r\nI have created a PR for the Notebook here: https://github.com/huggingface/notebooks/pull/148\r\nNot sure if the HTML is in a public repo. I wasn't able to find it. ", "Fixed the other two URLs here: \r\nhttps://github.com/mwunderlich/notebooks/pull/1", "Both URLs are broken now\r\n`HTTPError: 404 Client Error: Not Found for URL: https://the-eye.eu/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst`\r\nAnd\r\n`ConnectTimeout: HTTPSConnectionPool(host='mystic.the-eye.eu', port=443): Max retries exceeded with url: /public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst (Caused by ConnectTimeoutError(, 'Connection to mystic.the-eye.eu timed out. (connect timeout=10.0)'))`\r\n\r\n\r\n", "I was able to find a torrent with \"The Pile\" dataset here: [The Pile An 800GB Dataset of Diverse Text for Language Modeling ](https://academictorrents.com/details/0d366035664fdf51cfbe9f733953ba325776e667)\r\n\r\nThe complete dataset is huge, so I would suggest you to download only the \"PUBMED_title_abstracts_2019_baseline.jsonl.zst\" file, which is about 7GB. You can do this by using a torrent client of your choice (I typically utilize Transmission, which is pre-installed in Ubuntu distributions).\r\n\r\n", "@albertvillanova another issue:\r\n```\r\n15 experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights()\r\n16 File \"/lfs/ampere1/0/brando9/beyond-scale-language-data-diversity/src/diversity/div_coeff.py\", line 474, in experiment_compute_diveristy_coeff_single_dataset_then_combined_datasets_with_domain_weights\r\n17 column_names = next(iter(dataset)).keys()\r\n18 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py\", line 1353, in __iter__\r\n19 for key, example in ex_iterable:\r\n20 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/iterable_dataset.py\", line 207, in __iter__\r\n21 yield from self.generate_examples_fn(**self.kwargs)\r\n22 File \"/lfs/ampere1/0/brando9/.cache/huggingface/modules/datasets_modules/datasets/EleutherAI--pile/ebea56d358e91cf4d37b0fde361d563bed1472fbd8221a21b38fc8bb4ba554fb/pile.py\", line 236, in _generate_examples\r\n23 with zstd.open(open(files[subset], \"rb\"), \"rt\", encoding=\"utf-8\") as f:\r\n24 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/streaming.py\", line 74, in wrapper\r\n25 return function(*args, download_config=download_config, **kwargs)\r\n26 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/datasets/download/streaming_download_manager.py\", line 496, in xopen\r\n27 file_obj = fsspec.open(file, mode=mode, *args, **kwargs).open()\r\n28 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py\", line 134, in open\r\n29 return self.__enter__()\r\n30 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/core.py\", line 102, in __enter__\r\n31 f = self.fs.open(self.path, mode=mode)\r\n32 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/spec.py\", line 1241, in open\r\n33 f = self._open(\r\n34 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py\", line 356, in _open\r\n35 size = size or self.info(path, **kwargs)[\"size\"]\r\n36 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py\", line 121, in wrapper\r\n37 return sync(self.loop, func, *args, **kwargs)\r\n38 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py\", line 106, in sync\r\n39 raise return_result\r\n40 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/asyn.py\", line 61, in _runner\r\n41 result[0] = await coro\r\n42 File \"/lfs/ampere1/0/brando9/miniconda/envs/beyond_scale/lib/python3.10/site-packages/fsspec/implementations/http.py\", line 430, in _info\r\n43 raise FileNotFoundError(url) from exc\r\n44 FileNotFoundError: https://the-eye.eu/public/AI/pile_preliminary_components/NIH_ExPORTER_awarded_grant_text.jsonl.zst\r\n```\r\n\r\nany suggestions?", "related: https://github.com/huggingface/datasets/issues/6144", "this seems to work but it's rather annoying.\r\n\r\nSummary of how to make it work:\r\n1. get urls to parquet files into a list\r\n2. load list to load_dataset via `load_dataset('parquet', data_files=urls)` (note api names to hf are really confusing sometimes)\r\n3. then it should work, print a batch of text.\r\n\r\npresudo code\r\n```python\r\nurls_hacker_news = [\r\n \"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00000-of-00004.parquet\",\r\n \"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00001-of-00004.parquet\",\r\n \"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00002-of-00004.parquet\",\r\n \"https://huggingface.co/datasets/EleutherAI/pile/resolve/refs%2Fconvert%2Fparquet/hacker_news/pile-train-00003-of-00004.parquet\"\r\n]\r\n\r\n...\r\n\r\n\r\n # streaming = False\r\n from diversity.pile_subset_urls import urls_hacker_news\r\n path, name, data_files = 'parquet', 'hacker_news', urls_hacker_news\r\n # not changing\r\n batch_size = 512\r\n today = datetime.datetime.now().strftime('%Y-m%m-d%d-t%Hh_%Mm_%Ss')\r\n run_name = f'{path} div_coeff_{num_batches=} ({today=} ({name=}) {data_mixture_name=} {probabilities=})'\r\n print(f'{run_name=}')\r\n\r\n # - Init wandb\r\n debug: bool = mode == 'dryrun'\r\n run = wandb.init(mode=mode, project=\"beyond-scale\", name=run_name, save_code=True)\r\n wandb.config.update({\"num_batches\": num_batches, \"path\": path, \"name\": name, \"today\": today, 'probabilities': probabilities, 'batch_size': batch_size, 'debug': debug, 'data_mixture_name': data_mixture_name, 'streaming': streaming, 'data_files': data_files})\r\n # run.notify_on_failure() # https://community.wandb.ai/t/how-do-i-set-the-wandb-alert-programatically-for-my-current-run/4891\r\n print(f'{debug=}')\r\n print(f'{wandb.config=}')\r\n\r\n # -- Get probe network\r\n from datasets import load_dataset\r\n import torch\r\n from transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n\r\n tokenizer = GPT2Tokenizer.from_pretrained(\"gpt2\")\r\n if tokenizer.pad_token_id is None:\r\n tokenizer.pad_token = tokenizer.eos_token\r\n probe_network = GPT2LMHeadModel.from_pretrained(\"gpt2\")\r\n device = torch.device(f\"cuda:{0}\" if torch.cuda.is_available() else \"cpu\")\r\n probe_network = probe_network.to(device)\r\n\r\n # -- Get data set\r\n def my_load_dataset(path, name):\r\n print(f'{path=} {name=} {streaming=}')\r\n if path == 'json' or path == 'bin' or path == 'csv':\r\n print(f'{data_files_prefix+name=}')\r\n return load_dataset(path, data_files=data_files_prefix+name, streaming=streaming, split=\"train\").with_format(\"torch\")\r\n elif path == 'parquet':\r\n print(f'{data_files=}')\r\n return load_dataset(path, data_files=data_files, streaming=streaming, split=\"train\").with_format(\"torch\")\r\n else:\r\n return load_dataset(path, name, streaming=streaming, split=\"train\").with_format(\"torch\")\r\n # - get data set for real now\r\n if isinstance(path, str):\r\n dataset = my_load_dataset(path, name)\r\n else:\r\n print('-- interleaving datasets')\r\n datasets = [my_load_dataset(path, name).with_format(\"torch\") for path, name in zip(path, name)]\r\n [print(f'{dataset.description=}') for dataset in datasets]\r\n dataset = interleave_datasets(datasets, probabilities)\r\n print(f'{dataset=}')\r\n batch = dataset.take(batch_size)\r\n print(f'{next(iter(batch))=}')\r\n column_names = next(iter(batch)).keys()\r\n print(f'{column_names=}')\r\n\r\n # - Prepare functions to tokenize batch\r\n def preprocess(examples):\r\n return tokenizer(examples[\"text\"], padding=\"max_length\", max_length=128, truncation=True, return_tensors=\"pt\")\r\n remove_columns = column_names # remove all keys that are not tensors to avoid bugs in collate function in task2vec's pytorch data loader\r\n def map(batch):\r\n return batch.map(preprocess, batched=True, remove_columns=remove_columns)\r\n tokenized_batch = map(batch)\r\n print(f'{next(iter(tokenized_batch))=}')\r\n```\r\n\r\nhttps://stackoverflow.com/questions/76891189/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-th/76902681#76902681\r\n\r\nhttps://discuss.huggingface.co/t/how-to-download-data-from-hugging-face-that-is-visible-on-the-data-viewer-but-the-files-are-not-available/50555/5?u=severo", "If some people stumble upon this thread and still have this problem, i reuploaded the dataset to HF [here](https://huggingface.co/datasets/casinca/PUBMED_title_abstracts_2019_baseline)\r\n\r\nIts the exact same dataset you just have to change the url from the course, for example:\r\n\r\n```python\r\nfrom datasets import load_dataset, DownloadConfig\r\n\r\ndata_files = \"https://huggingface.co/datasets/casinca/PUBMED_title_abstracts_2019_baseline/resolve/main/PUBMED_title_abstracts_2019_baseline.jsonl.zst\"\r\npubmed_dataset = load_dataset(\r\n \"json\",\r\n data_files=data_files,\r\n split=\"train\",\r\n download_config=DownloadConfig(delete_extracted=True), # optional argument\r\n)\r\n```" ]

[]

[ "Hi ! Data streaming uses `fsspec` to read the data files progressively. IIRC the block size for buffering is 5MiB by default. So every time you finish iterating over a block, it downloads the next one. You can still try to increase the `fsspec` block size for buffering if it can help. To do so you just need to increase `fsspec.spec.AbstractBufferedFile.DEFAULT_BLOCK_SIZE `\r\n\r\nCurrently this is unfortunately done in a single thread, so it blocks the processing to download and uncompress the next block. At one point it would be nice to be able to do that in parallel !", "Hi! Thanks for the help, I will try it :)" ]

[ "Same error occures when using max samples with https://github.com/huggingface/transformers/blob/master/examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py", "I'm seeing this too, when using preprocessing_num_workers with \r\nhttps://github.com/huggingface/transformers/blob/master/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py" ]

[]

[]

[ "Hi ! `datasets` currently supports reading local files or files over HTTP. We may add support for other filesystems (cloud storages, hdfs...) at one point though :)" ]

[ "I think the test passes because it simply ignore what's after `gzip://`.\r\n\r\nThe returned urlpath is expected to look like `gzip://filename::url`, and the filename is currently considered to be what's after the final `/`, hence the result.\r\n\r\nWe can decide to change this and simply have `gzip://::url`, this way we don't need to guess the filename, what do you think ?" ]

[ "You can add columns that you wish to set into `torch` format using `dataset.set_format(\"torch\", ['id', 'abc'])` so that input batch of the transform only contains those columns", "Sorry, I miss `output_all_columns` args and thought after `dataset.set_format(\"torch\", columns=columns)` I can only get specific columns I assigned." ]

[ "Hi! \r\n\r\nYes, this makes sense for numeric values, but first I have to finish https://github.com/huggingface/datasets/pull/3336 because currently ArrayXD only allows the first dimension to be dynamic." ]

[ "Thanks for reporting @SaulLu.\r\n\r\nAt first sight I think the problem is caused because `pandas` only takes into account the `compression` parameter if called with a non-null file path or buffer. And in our implementation, we call pandas `to_json` with `None` `path_or_buf`.\r\n\r\nWe should fix this:\r\n- either handling directly the `compression` parameter ourselves\r\n- or refactoring to pass non-null path or buffer to pandas\r\n\r\nCC: @lhoestq", "I was thinking if we can handle the `compression` parameter by ourselves? Compression types will be similar to what `pandas` offer. Initially, we can try this with 2-3 compression types and see how good/bad it is? Let me know if it sounds good, I can raise a PR for this next week", "Hi ! Thanks for your help @bhavitvyamalik :)\r\nMaybe let's start with `gzip` ? I think it's the most common use case, then if we're fine with it we can add other compression methods" ]

[ "You're right, we have an issue today with the datasets preview. We're investigating.", "It should be fixed now. Thanks for reporting.", "Down again. ", "Fixed for good." ]

[ "Hi @puzzler10, thanks for reporting.\r\n\r\nPlease note this dataset is not hosted on Hugging Face Hub. See: \r\nhttps://github.com/huggingface/datasets/blob/c8f914473b041833fd47178fa4373cdcb56ac522/datasets/rotten_tomatoes/rotten_tomatoes.py#L42\r\n\r\nIf there are issues with the source data of a dataset, you should contact the data owners/creators instead. In the homepage associated with this dataset (http://www.cs.cornell.edu/people/pabo/movie-review-data/), you can find the authors of the dataset and how to contact them:\r\n> If you have any questions or comments regarding this site, please send email to Bo Pang or Lillian Lee.\r\n\r\nP.S.: Please also note that the example you gave of non-English review is in Portuguese (not Spanish). ;)", "Maybe best to just put a quick sentence in the dataset description that highlights this? " ]

[ "As discussed, I remember I set `decoded=False` here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the \"audio\" field (for Audio feature) was accessed.", "> I set decoded=False here to avoid decoding just by iterating over examples of dataset. We wanted to decode only if the \"audio\" field (for Audio feature) was accessed\r\n\r\nhttps://github.com/huggingface/datasets/pull/3430 will add more control to decoding, so I think it's OK to enable decoding in `__iter__` for now. After we merge the linked PR, the user can easily disable it again.", "@mariosasko I wonder why there is no issue in `Audio` feature with decoding disabled in `__iter__`, whereas there is in `Image` feature.\r\n\r\nEnabling decoding in `__iter__` will make fail Audio regressions tests: https://github.com/huggingface/datasets/runs/4608657230?check_suite_focus=true\r\n```\r\n=========================== short test summary info ============================\r\nFAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_not_decoded\r\nFAILED tests/features/test_audio.py::test_dataset_with_audio_feature_map_is_decoded\r\n========================= 2 failed, 15 passed in 8.37s =========================", "Please also note that the regression tests were implemented in accordance with the specifications:\r\n- when doing a `map` (wich calls `__iter__`) of a function that doesn't access the audio field, the decoding should be disabled; this is why the decoding is disabled in `__iter__` (and only enabled in `__getitem__`).", "> I wonder why there is no issue in Audio feature with decoding disabled in __iter__, whereas there is in Image feature.\r\n\r\n@albertvillanova Not sure if I understand this part. Currently, both the Image and the Audio feature don't decode data in `__iter__`, so their behavior is aligned there.\r\n", "Therefore, this is not an issue, neither for Audio nor Image feature.\r\n\r\nCould you please elaborate more on the expected use case? @lhoestq @NielsRogge \r\n\r\nThe expected use cases (in accordance with the specs: see #2324):\r\n- decoding should be enabled when accessing a specific item (`__getitem__`)\r\n- decoding should be disabled while iterating (`__iter__`) to allow preprocessing of non-audio/image features (like label or text, for example) using `.map`\r\n- decoding should be enabled in a `.map` only if the `.map` function accesses the audio/image feature (implemented using `LazyDict`)", "For me it's not an issue, actually. I just (mistakenly) tried to iterate over a PyTorch Dataset instead of a PyTorch DataLoader, \r\n\r\ni.e. I did this:\r\n\r\n`batch = next(iter(train_ds)) `\r\n\r\nwhereas I actually wanted to do\r\n\r\n`batch = next(iter(train_dataloader))`\r\n\r\nand then it turned out that in the first case, the image was a string of bytes rather than a Pillow image, hence Quentin opened an issue.", "Thanks @NielsRogge for the context.\r\n\r\nSo IMO everything is working as expected.\r\n\r\nI'm closing this issue. Feel free to reopen it again if further changes of the specs should be addressed.", "Thanks for the details :)\r\n\r\nI still think that it's unexpected to get different results when doing\r\n```python\r\nfor i in range(len(dataset)):\r\n sample = dataset[i]\r\n```\r\nand\r\n```python\r\nfor sample in dataset:\r\n pass\r\n```\r\neven though I understand that if you don't need to decode the data, then decoding image or audio data when iterating is a waste of time and resources.\r\n\r\nBut in this case users can still drop the column that need decoding to get the full speed back no ?" ]

[ "Hi @talha1503, thanks for reporting.\r\n\r\nIt seems there is an issue with one of the data files hosted at Google Drive:\r\n```\r\nGoogle Drive - Quota exceeded\r\n\r\nSorry, you can't view or download this file at this time.\r\n\r\nToo many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.\r\n```\r\n\r\nAs you probably know, Hugging Face does not host the data, and in this case the data owner decided to host their data at Google Drive, which has quota limits.\r\n\r\nIs there anything we could do, @lhoestq @mariosasko?", "This looks related to https://github.com/huggingface/datasets/issues/996", "It seems that [this](https://huggingface.co/datasets/ccdv/cnn_dailymail) copy of the dataset has fixed the problem", "thank you @AyhamAlom ...\r\nit resolved the error" ]

[ "Hi ! Can you try setting `datasets.config.MAX_TABLE_NBYTES_FOR_PICKLING` to a smaller value than `4 << 30` (4GiB), for example `500 << 20` (500MiB) ? It should reduce the maximum size of the arrow table being pickled during multiprocessing.\r\n\r\nIf it fixes the issue, we can consider lowering the default value for everyone.", "@lhoestq I tried that just now but didn't seem to help." ]

[]

[ "I think this is a duplicate of [#3190](https://github.com/huggingface/datasets/issues/3190)?", "Upgrading the datasets version as per #3190 fixes this bug. \r\nI'm Marking as closed." ]

[ "This dataset has files in ASF/AMC format. [ The skeleton file is the ASF file (Acclaim Skeleton File). The motion file is the AMC file (Acclaim Motion Capture data). ] \r\n\r\nSome questions : \r\n1. How do we go about representing these features using datasets.Features and generate examples ?\r\n2. The dataset download link for ASF/AMC files does not have metadata information, for eg : category and subcategory information. We will need to crawl the website for this information. The authors mention \"Please don't crawl this database for all motions.\" Can we mail the authors for this information ?\r\nThe dataset structure is as follows : \r\n```\r\nsubjects\r\n\t- 01\r\n\t\t- 01_01.amc\r\n\t\t- 01_02.amc\r\n\t\t.\r\n\t\t.\r\n\t\t.\r\n\t\t- 01.asf\r\n\t- 02\r\n\t\t- 02_01.amc\r\n\t\t- 02_02.amc\r\n\t\t.\r\n\t\t.\r\n\t\t.\r\n\t\t- 02.asf\r\n```\r\nThere is no metadata regarding the category, sub-category and motion description.\r\n\r\nNeed your inputs. @mariosasko / @lhoestq \r\nThank you.\r\n", "Hi @dnaveenr! Thanks for working on this!\r\n\r\n1. We can use the `Sequence(Value(\"string\"))` feature type for the subject's AMC files and `Value(\"string\")` for the subject's ASF file (`Value(\"string\")` represents the file paths) + the types for categories/subcategories and descriptions.\r\n2. We can use this URL to download the motion descriptions: http://mocap.cs.cmu.edu/search.php?subjectnumber=<subject_number>&motion=%%%&maincat=%&subcat=%&subtext=yes where `subject_number` is the number between 1 and 144. And to get categories/subcategories, feel free to contact the authors (they state in the FAQ they are happy to help) and ask them if they can provide the mapping from categories/subcategories to the AMC files to avoid crawling. You can also mention that your goal is to make their dataset more accessible by adding its loading script to the Hub.\r\n\r\nThe AMC files are also available in the tvd, c3d, mpg and avi formats (the links are in the [FAQ](http://mocap.cs.cmu.edu/faqs.php) section), so it would be nice to have one config for each of these additional formats. \r\n\r\nAnd additionally, we can add a `Data Preprocessing` section to the card where we explain how to load/process the files. I can help with that.", "Hi @mariosasko ,\r\n\r\n1. Thanks for this, so we can add the file paths.\r\n2. Yes, I had already mailed the authors a couple of days back actually, asking for the metadata details[ i.e category, sub-category and motion description] . They are yet to respond though, I will wait for a couple of days and try to follow up with them again. :) Else we can use the workaround solution.\r\n\r\nYes. Supporting all the formats would be helpful. \r\n\r\n> And additionally, we can add a Data Preprocessing section to the card where we explain how to load/process the files. I can help with that.\r\n\r\nOkay. Got it." ]

[ "Hi ! I guess you are talking about the dataset cards that are in this repository on github ?\r\n\r\nI think github allows to submit a PR even for 1 line though the `Edit file` button on the page of the dataset card.\r\n\r\nMaybe let's mention this in `CONTRIBUTING.md` ?", "We now host all the datasets on the HF Hub, where you can easily edit them through UI (for single file changes) or Git workflow (for single/multiple file changes)" ]

[]

[ "Hi ! It's simply not added yet :)\r\n\r\nIf someone wants to contribute to add the `stratify` parameter I'd be happy to give some pointers.\r\n\r\nIn the meantime, I guess you can use `sklearn` or other tools to do a stratified train/test split over the **indices** of your dataset and then do\r\n```\r\ntrain_dataset = dataset.select(train_indices)\r\ntest_dataset = dataset.select(test_indices)\r\n```", "Hi @lhoestq I would like to add `stratify` parameter, can you give me some pointers for adding the same ?", "Hi ! Sure :)\r\n\r\nThe `train_test_split` method is defined here: \r\n\r\nhttps://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3253-L3253\r\n\r\nand inside `train_test_split ` we need to create the right `train_indices` and `test_indices` that are passed here to `.select()`:\r\n\r\nhttps://github.com/huggingface/datasets/blob/dc62232fa1b3bcfe2fbddcb721f2d141f8908943/src/datasets/arrow_dataset.py#L3450-L3464\r\n\r\nFor example if your dataset is like\r\n| | label |\r\n|---:|--------:|\r\n| 0 | 1 |\r\n| 1 | 1 |\r\n| 2 | 0 |\r\n| 3 | 0 |\r\n\r\nand the user passes `stratify=dataset[\"label\"]`, then you should get indices that look like this\r\n```\r\ntrain_indices = [0, 2]\r\ntest_indices = [1, 3]\r\n```\r\n\r\nthese indices will be passed to `.select` to return the stratified train and test splits :)\r\n\r\nFeel free to îng me if you have any question !", "@lhoestq \r\nI just added the implementation for `stratify` option here #4322 " ]

[ "Hi ! This is an issue with `datasets` 1.12. Sorry for the inconvenience. Can you update to `>=1.13` ?\r\nsee https://github.com/huggingface/datasets/issues/3190\r\n\r\nMaybe we should also backport the bug fix to `1.12` (in a new version `1.12.2`)" ]

[ "I was going through the codebase, and I believe the implementation of __add__() and __iadd__() will be similar to concatenate_datasets() after the elimination of code for arguments other than the list of datasets (info, split, axis). \r\n(Assuming elimination of axis means concatenating over axis 1.)", "Most data frame libraries (Polars, Pandas, etc.) override `__add__` to perform (mathematical) summation, so having different behavior could lead to confusion." ]

[ "Hi ! I think the issue comes from the dataset_infos.json file: it has the \"flat\" field twice.\r\n\r\nCan you try deleting this file and regenerating it please ?", "Thanks! That fixed that, but now I am getting:\r\nServer Error\r\nStatus code: 400\r\nException: KeyError\r\nMessage: 'feature'\r\n\r\nI checked the dataset_infos.json and pubmed_neg.py script, I don't use 'feature' anywhere as a key. Is the dataset viewer expecting that I do?", "It seems that the `feature` key is missing from some feature type definition in your dataset_infos.json:\r\n```json\r\n\t\t\t\"tokens\": {\r\n\t\t\t\t\"dtype\": \"list\",\r\n\t\t\t\t\"id\": null,\r\n\t\t\t\t\"_type\": \"Sequence\"\r\n\t\t\t},\r\n\t\t\t\"tags\": {\r\n\t\t\t\t\"dtype\": \"list\",\r\n\t\t\t\t\"id\": null,\r\n\t\t\t\t\"_type\": \"Sequence\"\r\n\t\t\t}\r\n```\r\nThey should be\r\n```json\r\n\t\t\t\"tokens\": {\r\n\t\t\t\t\"dtype\": \"list\",\r\n\t\t\t\t\"id\": null,\r\n\t\t\t\t\"_type\": \"Sequence\"\r\n \"feature\": {\"dtype\": \"string\", \"id\": null, \"_type\": \"Value\"}\r\n\t\t\t},\r\n\t\t\t\"tags\": {\r\n\t\t\t\t\"dtype\": \"list\",\r\n\t\t\t\t\"id\": null,\r\n\t\t\t\t\"_type\": \"Sequence\",\r\n \"feature\": {\"num_classes\": 5, \"names\": [\"-\", \"S\", \"H\", \"N\", \"C\"], \"names_file\": null, \"id\": null, \"_type\": \"ClassLabel\"}\r\n\t\t\t}\r\n```\r\n\r\nNote that you can generate the dataset_infos.json automatically to avoid mistakes:\r\n```bash\r\ndatasets-cli test ./path/to/dataset --save_infos\r\n```" ]

[ "Hi ! Indeed it says \"downloading and preparing\" but in your case it didn't need to download anything since you used local files (it would have thrown an error otherwise). I think we can improve the logging to make it clearer in this case", "@lhoestq Thank you for explaining. I am sorry but I was not clear about my intention. I didn't want to kill internet traffic; I wanted to kill all write activity. In other words, you can imagine that my storage has only read access but crashes on write.\r\n\r\nWhen run_clm.py is invoked with the same parameters, the hash in the cache directory \"datacache/trainpy.v2/json/default-471372bed4b51b53/0.0.0/...\" doesn't change, and my job can load cached data properly. This is great.\r\n\r\nUnfortunately, when params change (which happens sometimes), the hash changes and the old cache is invalid. datasets builder would create a new cache directory with the new hash and create JSON builder there, even though every JSON builder is the same. I didn't find a way to avoid such behavior.\r\n\r\nThis problem can be resolved when using datasets.map() for tokenizing and grouping text. This function allows me to specify output filenames with --cache_file_names, so that the cached files are always valid.\r\n\r\nThis is the code that I used to freeze cache filenames for tokenization. I wish I could do the same to datasets.load_dataset()\r\n```\r\n tokenized_datasets = raw_datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n num_proc=data_args.preprocessing_num_workers,\r\n remove_columns=column_names,\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n desc=\"Running tokenizer on dataset\",\r\n cache_file_names={k: os.path.join(model_args.cache_dir, f'{k}-tokenized') for k in raw_datasets},\r\n )\r\n```", "Hi ! `load_dataset` may re-generate your dataset if some parameters changed indeed. If you want to freeze a dataset loaded with `load_dataset`, I think the best solution is just to save it somewhere on your disk with `.save_to_disk(my_dataset_dir)` and reload it with `load_from_disk(my_dataset_dir)`. This way you will be able to reload the dataset without having to run `load_dataset`" ]

[ "Hi ! What's your question ?" ]

[ "Yes I agree, these should be as aligned as possible. Maybe we can also check the feedback in the survey at http://hf.co/oss-survey and see if people mentioned related things on the API (in particular if we go the breaking change way, it would be good to be sure we are taking the right direction for the community).", "I like this proposal.\r\n\r\n> There is also an important difference in terms of behavior:\r\nDataset.map adds new columns (with dict.update)\r\nBUT\r\nIterableDataset discards previous columns (it overwrites the dict)\r\nIMO the two methods should have the same behavior. This would be an important breaking change though.\r\n\r\n> The main breaking change would be the change of behavior of IterableDataset.map, because currently it discards all the previous columns instead of keeping them.\r\n\r\nYes, this behavior of `IterableDataset.map` was surprising to me the first time I used it because I was expecting the same behavior as `Dataset.map`, so I'm OK with the breaking change here.\r\n\r\n> IterableDataset only supports \"torch\" (it misses tf, jax, pandas, arrow) and is missing the parameters: columns, output_all_columns and format_kwargs\r\n\r\n\\+ it's also missing the actual formatting code (we return unformatted tensors)\r\n> We could have a completely aligned map method if both methods were lazy by default, but this is a very big breaking change so I'm not sure we can consider doing that.\r\n\r\n> For information, TFDS does lazy map by default, and has an additional .cache() method.\r\n\r\nIf I understand this part correctly, the idea would be for `Dataset.map` to behave similarly to `Dataset.with_transform` (lazy processing) and to have an option to cache processed data (with `.cache()`). This idea is really nice because it can also be applied to `IterableDataset` to fix https://github.com/huggingface/datasets/issues/3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?) \r\n> If the two APIs are more aligned it would be awesome for the examples in transformers, and it would create a satisfactory experience for users that want to switch from one mode to the other.\r\n\r\nYes, it would be amazing to have an option to easily switch between these two modes.\r\n\r\nI agree with the rest.\r\n", "> If I understand this part correctly, the idea would be for Dataset.map to behave similarly to Dataset.with_transform (lazy processing) and to have an option to cache processed data (with .cache()). This idea is really nice because it can also be applied to IterableDataset to fix #3142 (again we get the aligned APIs). However, this change would break a lot of things, so I'm still not sure if this is a step in the right direction (maybe it's OK for Datasets 2.0?)\r\n\r\nYea this is too big of a change in my opinion. Anyway it's fine as it is right now with streaming=lazy and regular=eager.", "Hi, IterableDataset is also missing set_format.", "Yes indeed, thanks. I added it to the list of methods to align in the first post", "I just encountered the problem of the missing `fn_kwargs` parameter in the `map` method. I am commenting to give a workaround in case someone has the same problem and does not find a solution.\r\nYou can wrap your function call inside a class that contains the other parameters needed by the function called by map, like this:\r\n\r\n```python\r\ndef my_func(x, y, z):\r\n # Do things\r\n\r\nclass MyFuncWrapper:\r\n def __init__(self, y, z):\r\n self.y = y\r\n self.z = z\r\n\r\n def __call__(self, x):\r\n return my_func(x, self.y, self.z)\r\n```\r\n\r\nThen, give an instance of the `MyFuncWrapper` to the map function.", "Any update on this? It's almost 2024😂 @lhoestq ", "The main differences have been addressed (map, formatting) but there are still a few things to implement like Dataset.take, Dataset.skip, IterableDataset.set_format, IterableDataset.formatted_as, IterableDataset.reset_format.\r\n\r\nThe rest cannot be implemented for the general case. E.g. train_test_split and select can only work on an iterable dataset if the underlying dataset format allows it (we need to know the number of rows and have some sort of random access)" ]

[ "I'll take this one if no one hasn't yet!" ]

[ "Hi ! What version of `datasets` are you using ? Can you also provide the logs you get before it raises the error ?" ]

[ "This dataset is now available on the Hub here: https://huggingface.co/datasets/MLCommons/peoples_speech" ]

[]

[ "Hi ! `load_dataset` accepts as input either a local dataset directory or a dataset name from the Hugging Face Hub.\r\n\r\nSo here you are getting this error the second time because it tries to load the local `wiki_dpr` directory, instead of `wiki_dpr` from the Hub. It doesn't work since it's a **cache** directory, not a **dataset** directory in itself.\r\n\r\nTo fix that you can use another cache directory like `cache_dir=\"/data2/whr/lzy/open_domain_data/retrieval/cache\"`", "thx a lot" ]

[ "maybe related to https://github.com/huggingface/datasets/issues/2859\r\n", "It looks like it's currently calling `HfFileSystem.ls()` ~8 times at the root and for each subdirectory:\r\n- \"\"\r\n- \"en.noblocklist\"\r\n- \"en.noclean\"\r\n- \"en\"\r\n- \"multilingual\"\r\n- \"realnewslike\"\r\n\r\nCurrently `ls` is slow because it iterates on all the files inside the repository.\r\n\r\nAn easy optimization would be to cache the result of each call to `ls`.\r\nWe can also optimize `ls` by using a tree structure per directory instead of a list of all the files.\r\n", "ok\r\n" ]

[ "Hi ! Thanks for reporting :)\r\n\r\nWhen using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.\r\n\r\nWe can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens.", "> Hi ! Thanks for reporting :)\r\n> \r\n> When using a PyTorch's data loader with `num_workers>1` and an iterable dataset, each worker streams the exact same data by default, resulting in duplicate data when iterating using the data loader.\r\n> \r\n> We can probably fix this in `datasets` by checking `torch.utils.data.get_worker_info()` which gives the worker id if it happens.\r\nHi ! Thanks for reply\r\n\r\nDo u have some plans to fix the problem?\r\n", "Isn’t that somehow a bug on PyTorch side? (Just asking because this behavior seems quite general and maybe not what would be intended)", "From PyTorch's documentation [here](https://pytorch.org/docs/stable/data.html#dataset-types):\r\n\r\n> When using an IterableDataset with multi-process data loading. The same dataset object is replicated on each worker process, and thus the replicas must be configured differently to avoid duplicated data. See [IterableDataset](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset) documentations for how to achieve this.\r\n\r\nIt looks like an intended behavior from PyTorch\r\n\r\nAs suggested in the [docstring of the IterableDataset class](https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset), we could pass a `worker_init_fn` to the DataLoader to fix this. It could be called `streaming_worker_init_fn` for example.\r\n\r\nHowever, while this solution works, I'm worried that many users simply don't know about this parameter and just start their training with duplicate data without knowing it. That's why I'm more in favor of integrating the check on the worker id directly in `datasets` in our implementation of `IterableDataset.__iter__`.", "Fixed by https://github.com/huggingface/datasets/pull/4375", "> Fixed by #4375\r\n\r\nThanks!", "Hi there @lhoestq @cloudyuyuyu \r\nI met that problem recently, and #4375 is really useful because I finally found out I am training with duplicate data.\r\nHowever, in multi-GPU training, I'm using DDP mode and IterableDataset, which still yields duplicate data for each progress. And this is dangerous because users maybe not realize this behavior.", "If the worker_info.id is unique per process it should work fine, could you check that they're unique ?\r\n\r\nThe code to get the worker_info in each worker is `torch.utils.data.get_worker_info()`", "test.py\r\n```python\r\nimport json\r\nimport os\r\n\r\nimport torch\r\nfrom torch.utils.data import IterableDataset, DataLoader\r\nfrom transformers import PreTrainedTokenizer, TrainingArguments\r\n\r\nfrom common.arguments import DataTrainingArguments, ModelArguments\r\n\r\n\r\nclass MyIterableDataset(IterableDataset):\r\n def __iter__(self):\r\n worker_info = torch.utils.data.get_worker_info()\r\n print(worker_info)\r\n return iter(range(3))\r\n\r\n\r\nif __name__ == '__main__':\r\n dataset = MyIterableDataset()\r\n dataloader = DataLoader(dataset, num_workers=1)\r\n for i in dataloader:\r\n print(i)\r\n\r\n```\r\n\r\n\r\n```sh\r\n$ python3 -m torch.distributed.launch \\\r\n --nproc_per_node=2 test.py\r\nWorkerInfo(id=0, num_workers=1, seed=5545685212307804959, dataset=<__main__.MyIterableDataset object at 0x7f92648cf6a0>)\r\nWorkerInfo(id=0, num_workers=1, seed=3174108029709729025, dataset=<__main__.MyIterableDataset object at 0x7f19ab961670>)\r\ntensor([0])\r\ntensor([1])\r\ntensor([2])\r\ntensor([0])\r\ntensor([1])\r\ntensor([2])\r\n```\r\n\r\n@lhoestq they are not unique", "It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?\r\n\r\nI guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is", "> It looks like a bug from pytorch no ? How can we know which data should go in which process when using DDP ?\r\n> \r\n> I guess we need to check `torch.distributed.get_world_size()` and `torch.distributed.get_rank()` as well. Not fan of the design here tbh, but that's how it is\r\n\r\nMaybe we should document it?", "Never mind. After reading the code, `IterableDatasetShard` has solved this problem.", "I'm re-opening this one since I think it should be supported by `datasets` natively", "hmm actually let me open a new issue on DDP - original post was for single node" ]

[ "Hi ! I wasn't able to reproduce your error.\r\n\r\nCan you try to clear your cache at `~/.cache/huggingface/modules` and try again ?" ]

[ "Hi ! It's slower indeed because a datasets on which `select`/`shard`/`train_test_split`/`shuffle` has been called has to do additional steps to retrieve the data of the dataset table in the right order.\r\n\r\nIndeed, if you call `dataset.select([0, 5, 10])`, the underlying table of the dataset is not altered to keep the examples at index 0, 5, and 10. Instead, an indices mapping is added on top of the table, that says that the first example is at index 0, the second at index 5 and the last one at index 10.\r\n\r\nTherefore accessing the examples of the dataset is slower because of the additional step that uses the indices mapping.\r\n\r\nThe step that takes the most time is to query the dataset table from a list of indices here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/047dc756ed20fbf06e6bcaf910464aba0e20610a/src/datasets/formatting/formatting.py#L61-L63\r\n\r\nIn your case it can be made significantly faster by checking if the indices are contiguous. If they're contiguous, we could pass a python `slice` or `range` instead of a list of integers to `_query_table`. This way `_query_table` will do only one lookup to get the queried batch instead of `batch_size` lookups.\r\n\r\nGiven that calling `select` with contiguous indices is a common use case I'm in favor of implementing such an optimization :)\r\nLet me know what you think", "Hi, thanks for the response!\r\nI still don't understand why it is so much slower than iterating and saving:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\noriginal = load_dataset(\"squad\", split=\"train\")\r\noriginal.to_json(\"from_original.json\") # Takes 0 seconds\r\n\r\nselected_subset1 = original.select([i for i in range(len(original))])\r\nselected_subset1.to_json(\"from_select1.json\") # Takes 99 seconds\r\n\r\nselected_subset2 = original.select([i for i in range(int(len(original) / 2))])\r\nselected_subset2.to_json(\"from_select2.json\") # Takes 47 seconds\r\n\r\nselected_subset3 = original.select([i for i in range(len(original)) if i % 2 == 0])\r\nselected_subset3.to_json(\"from_select3.json\") # Takes 49 seconds\r\n\r\nimport json\r\nimport time\r\ndef fast_to_json(dataset, path):\r\n start = time.time()\r\n with open(path, mode=\"w\") as f:\r\n for example in dataset:\r\n f.write(json.dumps(example, separators=(',', ':')) + \"\\n\")\r\n end = time.time()\r\n print(f\"Saved {len(dataset)} examples to {path} in {end - start} seconds.\")\r\n\r\nfast_to_json(original, \"from_original_fast.json\")\r\nfast_to_json(selected_subset1, \"from_select1_fast.json\")\r\nfast_to_json(selected_subset2, \"from_select2_fast.json\")\r\nfast_to_json(selected_subset3, \"from_select3_fast.json\")\r\n```\r\n```\r\nSaved 87599 examples to from_original_fast.json in 8 seconds.\r\nSaved 87599 examples to from_select1_fast.json in 10 seconds.\r\nSaved 43799 examples to from_select2_fast.json in 6 seconds.\r\nSaved 43800 examples to from_select3_fast.json in 5 seconds.\r\n```", "There are slight differences between what you're doing and what `to_json` is actually doing.\r\nIn particular `to_json` currently converts batches of rows (as an arrow table) to a pandas dataframe, and then to JSON Lines. From your benchmark it looks like it's faster if we don't use pandas.\r\n\r\nThanks for investigating, I think we can optimize `to_json` significantly thanks to your test.", "Thanks for your observations, @eladsegal! I spent some time with this and tried different approaches. Turns out that https://github.com/huggingface/datasets/blob/bb13373637b1acc55f8a468a8927a56cf4732230/src/datasets/io/json.py#L100 is giving the problem when we use `to_json` after `select`. This is when `indices` parameter in `query_table` is not `None` (if it is `None` then `to_json` should work as expected)\r\n\r\nIn order to circumvent this problem, I found out instead of doing Arrow Table -> Pandas-> JSON we can directly go to JSON by using `to_pydict()` which is a little slower than the current approach but at least `select` works properly now. Lmk what you guys think of it @lhoestq, @eladsegal?", "Sounds good to me ! Feel free to also share your benchmarks for reference @bhavitvyamalik ", "Posting it in @eladsegal's format:\r\n\r\nFor `squad`:\r\nSaving examples using current `to_json` in 3.63 secs\r\nSaving examples to `from_select1_fast.json` in 5.00 secs\r\nSaving examples to `from_select2_fast.json` in 2.45 secs\r\nSaving examples to `from_select3_fast.json` in 2.50 secs\r\n\r\nFor `squad_v2`:\r\nSaving examples using current `to_json` in 5.26 secs\r\nSaving examples to `from_select1_fast.json` in 7.54 secs\r\nSaving examples to `from_select2_fast.json` in 3.80 secs\r\nSaving examples to `from_select3_fast.json` in 3.67 secs" ]

[ "Hi, thanks for reporting! This is a duplicate of https://github.com/huggingface/datasets/issues/3240. We are working on a fix.\r\n\r\n" ]

[ "I think it might come from two different issues:\r\n1. Google Drive is an unreliable host, mainly because of quota limitations\r\n2. the staging environment can sometimes raise some errors\r\n\r\nFor Google Drive tests we could set up some retries with backup URLs if necessary I guess.\r\nFor staging on the other hand, I guess we can investigate what causes this and discuss with the back-end team", "Closed by:\r\n- #3982" ]

[ "@LysandreJik not so sure who to @\r\nCould you help?", "Hi @hyusterr, I believe it is @wlhgtc from https://github.com/huggingface/transformers/pull/9887" ]

[ "Fixed, thanks\r\n<img width=\"661\" alt=\"Capture d’écran 2021-12-22 à 12 02 30\" src=\"https://user-images.githubusercontent.com/1676121/147082881-cf700e8d-0511-4431-b214-d6cf8137db10.png\">\r\n" ]

[]

[]

[ "Hi ! Can you try updating `fsspec` ? The minimum version is `2021.05.0`", "hey @lhoestq. I'm using `fsspec-2021.11.1` but still getting that error.", "Maybe this discussion can help:\r\n\r\nhttps://github.com/fsspec/filesystem_spec/issues/597#issuecomment-958646964", "Thanks @lhoestq. Downgrading `fsspec and s3fs` to `2021.10` fixed this issue!" ]

[ "As we are planning to stop using Apache Beam (our `datasets.BeamBasedBuilder`) for the generation of some datasets (including [Wikipedia](https://huggingface.co/datasets/wikipedia/blob/main/wikipedia.py)), I have been working on [wikimedia/wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) to:\r\n- Port the Wikipedia generation script to use `datasets.GeneratorBasedBuilder` instead and place it under the \"script\" branch: https://huggingface.co/datasets/wikimedia/wikipedia/tree/script\r\n- Improve the efficiency of the code and make it highly parellizable. See:\r\n - [Parallelize dataset generation over multistreams](https://huggingface.co/datasets/wikimedia/wikipedia/commit/610c55864586dbdad7ac5a13c21a367bb000a1d3)\r\n - [Parallelize data downloading](https://huggingface.co/datasets/wikimedia/wikipedia/commit/b35d406bd9e81f08c68e7bf95d130d2f506dfe77)\r\n\r\n With these improvements, I can generate the English Wikipedia in 5h using 16 processors in a machine without needing a huge amount of RAM (the machine had 32 GB, but I think less can be used as well):\r\n ```python\r\n ds = load_dataset(\"wikimedia/wikipedia\", revision=\"script\", date=\"20231101\", language=\"en\", host=\"https://mirror.accum.se/mirror/wikimedia.org/dumps\", split=\"train\", num_proc=16)\r\n ```\r\n- Pre-process all Wikipedia languages for the latest 2023-11-01 dump and make them available to the entire community for easy use:\r\n ```python\r\n ds = load_dataset(\"wikimedia/wikipedia\", \"20231101.en\", split=\"train\", num_proc=16)\r\n ```\r\nCC: @geohci " ]

[ "Thanks! See https://public.paws.wmcloud.org/User:Isaac_(WMF)/HuggingFace%20Wikipedia%20Processing.ipynb for more implementation details / some data around the overhead induced by adding the extra preprocessing steps (stripping link prefixes and magic words)", "Closed by:\r\n- #3435" ]

[ "See notebook by @geohci: https://public.paws.wmcloud.org/User:Isaac_(WMF)/HuggingFace%20Wikisource%20Processing.ipynb" ]

[ "@geohci, I think the field \"url\" does not appear in the Wikimedia dumps. Therefore I guess we should generate it, using the \"title\" field and making some transformation of it (replacing spaces with underscores) and prepending the domain (created using the language)?", "Indeed:\r\n\r\n> To re-distribute text on Wikipedia in any form, provide credit to the authors either by including a) a [hyperlink](https://en.wikipedia.org/wiki/Hyperlink) (where possible) or [URL](https://en.wikipedia.org/wiki/URL) to the page or pages you are re-using, b) a hyperlink (where possible) or URL to an alternative, stable online copy which is freely accessible, which conforms with the license, and which provides credit to the authors in a manner equivalent to the credit given on this website, or c) a list of all authors. (Any list of authors may be filtered to exclude very small or irrelevant contributions.) This applies to text developed by the Wikipedia community. Text from external sources may attach additional attribution requirements to the work, which should be indicated on an article's face or on its talk page. For example, a page may have a banner or other notation indicating that some or all of its content was originally published somewhere else. Where such notations are visible in the page itself, they should generally be preserved by re-users.\r\n\r\nsource: https://en.wikipedia.org/wiki/Wikipedia:Copyrights\r\n\r\nI guess it's fine to add the URL field - it can be constructed easily from the title page IIRC.", "yep, sorry forgot that that wasn't already in the dumps. specifically `f\"https://{language}.wikipedia.org/wiki/{title.replace(' ', '_')}` should do it", "Thanks @geohci.\r\n\r\nI had already been looking for information about the conversion from title to URL and I found that apart from replacing blanks with underscores, some other special character must also be percent-encoded (e.g. `\"` to `%22`): https://meta.wikimedia.org/wiki/Help:URL\r\n\r\nTherefore, I have finally used `urllib.parse.quote` function. This additionally percent-encodes non-ASCII characters, but Wikimedia docs say these are equivalent:\r\n> For the other characters either the code or the character can be used in internal and external links, they are equivalent. The system does a conversion when needed.\r\n> [[%C3%80_propos_de_M%C3%A9ta]]\r\n> is rendered as [À_propos_de_Méta](https://meta.wikimedia.org/wiki/%C3%80_propos_de_M%C3%A9ta), almost like [À propos de Méta](https://meta.wikimedia.org/wiki/%C3%80_propos_de_M%C3%A9ta), which leads to this page on Meta with in the address bar the URL\r\n> [http://meta.wikipedia.org/wiki/%C3%80_propos_de_M%C3%A9ta](https://meta.wikipedia.org/wiki/%C3%80_propos_de_M%C3%A9ta)\r\n> while [http://meta.wikipedia.org/wiki/À_propos_de_Méta](https://meta.wikipedia.org/wiki/%C3%80_propos_de_M%C3%A9ta) leads to the same. ", "Closed by:\r\n- #3789 " ]

[ "https://huggingface.co/datasets/projecte-aina/parlament_parla -> works (but we still have to show an audio player)\r\n\r\nhttps://huggingface.co/datasets/openslr -> another issue: `Message: [Errno 2] No such file or directory: '/home/hf/datasets-preview-backend/zip:/asr_javanese/data/00/00004fe6aa.flac'`", "Done", "https://huggingface.co/datasets/projecte-aina/parlament_parla/viewer/clean/train works\r\n\r\n<img width=\"1535\" alt=\"Capture d’écran 2022-04-12 à 13 58 47\" src=\"https://user-images.githubusercontent.com/1676121/162957855-cb3d9e2e-4b61-488c-99ca-8065cd8fe377.png\">\r\n", "But https://huggingface.co/datasets/openslr/viewer does not work\r\n\r\n<img width=\"678\" alt=\"Capture d’écran 2022-04-12 à 13 59 46\" src=\"https://user-images.githubusercontent.com/1676121/162958013-e31ef2ae-f886-47b7-9f27-664ed3d4b5a1.png\">\r\n\r\nSame issue as #4126:\r\n\r\n```\r\nStatus code: 400\r\nException: TypeError\r\nMessage: __init__() got an unexpected keyword argument 'audio_column'\r\n```", "Fixed:\r\n<img width=\"1561\" alt=\"Capture d’écran 2022-04-25 à 18 11 51\" src=\"https://user-images.githubusercontent.com/1676121/165129813-018ece9e-8b20-4544-844d-4e88148e738f.png\">\r\n" ]

[ "According to this [comment in the forum](https://discuss.huggingface.co/t/save-datasetdict-to-huggingface-hub/12075/8?u=lhoestq), using `push_to_hub` on a dataset with `ClassLabel` can also make the feature simply disappear when it's reloaded !", "Maybe we can also fix https://github.com/huggingface/datasets/issues/3035 while working on this because, as pointed out in my initial post, `save_to_disk` also saves the `dataset_info.json` file." ]

[]

[ "This issue was fixed by me calling `all_datasets.push_to_hub(\"hackernews_hiring_posts\")`.\r\n\r\nThe previous problems were from calling `all_datasets.save_to_disk` and then pushing with `my_repo.git_add` and `my_repo.push_to_hub`.\r\n" ]

[ "duplicate of #2655" ]

[ "Got solved it with push_to_hub, closing" ]

[ "cc @juliensimon ", "Datasets are not tracked in this repository anymore. But you can make your own dataset in the huggingface hub" ]

[ "Hi ! Thanks for the suggestion :)\r\nIt makes sense to me, and it can surely be implemented by wrapping the user's function to make it a batched function. However I'm not a big fan of the inconsistency it would create with `map`: `with_transform` is batched by default while `map` isn't.\r\n\r\nIs there something you would like to contribute ? I can give you some pointers if you want", "Hi @lhoestq ,\r\nSorry I missed your reply.\r\n\r\nI would love to contribute. But I don't know which solution would be the best for this repo.\r\n\r\n> However I'm not a big fan of the inconsistency it would create with map: with_transform is batched by default while map isn't.\r\n\r\nI agree. What do you think about the alternative solutions?\r\n\r\n> * Convert a non-batched transform function to batched one myself.\r\n\r\nThis won't be able to use torch loader multi-worker.\r\n\r\n> * Wrap a 🤗 Dataset with torch Dataset, and add a __getitem__. 🙄\r\n\r\nThis is actually pretty simple.\r\n\r\n```python\r\nimport torch\r\n\r\nclass LazyMapTorchDataset(torch.utils.data.Dataset):\r\n def __init__(self, ds, fn):\r\n self.ds = ds\r\n self.fn = fn\r\n def __getitem__(self, i):\r\n return self.fn(self.ds[i])\r\n\r\nd = [{1:2, 2:3}, {1:3, 2:4}]\r\nds = LazyMapTorchDataset(d, lambda x:{k:v*2 for k,v in x.items()})\r\nfor i in range(2):\r\n print(f'before {d[i]}')\r\n print(f'after {ds[i]}')\r\n```\r\n```\r\nbefore {1: 2, 2: 3}\r\nafter {1: 4, 2: 6}\r\nbefore {1: 3, 2: 4}\r\nafter {1: 6, 2: 8}\r\n```\r\n\r\nBut this requires converting data to torch tensor myself. And this is really similar to `.map()`, why not just use it? So I have the next solution.\r\n\r\n> * Have lazy=False in Dataset.map, and returns a LazyDataset if lazy=True. This way the same map interface can be used, and existing code can be updated with one argument change.\r\n\r\nI think I like this solution best. Because `.with_transform` is entangled with `.with_format`, so seems more flexible to modify the `.map` than to modify `.with_transform`.\r\n\r\nThe usage looks nice, too.\r\n```python\r\n# lazy, one to one, can be parallelized via torch loader, no need to set `num_worker` beforehand.\r\ndataset = dataset.map(fn, lazy=True, batched=False)\r\n# collate_fn\r\ndataloader = Dataloader(dataset.with_format('torch'), collate_fn=collate_fn, num_workers=...) \r\n```\r\n\r\nThere are some minor decisions like whether a lazy map should be allowed before another map, but I think we can work it out later. The implementation can probably borrow from `IterableDataset`.", "I like the idea of lazy map. On the other hand we should only have either lazy map or `with_transform` (not both). That's why I'd rather stick with `with_transform` for now (but maybe we can consider it for later major releases like `datasets` v2).\r\n\r\nI understand the issue with `with_transform` and `with_format` being exclusive, maybe we can separate them: first transform, them format.\r\n\r\nFinally I think what's also going to be important in the end will be the addition of multiprocessing to transforms" ]

[ "Hi ! Feel free to access `batch[\"audio\"][\"array\"]` and `batch[\"audio\"][\"sampling_rate\"]` instead\r\n\r\n`datasets` 1.16 introduced some changes in `common_voice` and now the `path` field is no longer a path to a local file (but rather the path to the file in the archive it's extracted from)", "Thanks for the information. It works.", "Cool ! Closing this issue then" ]

[]

[ "Seems like the issue still exists,:\r\n`Downloading and preparing dataset clue/chid (download: 127.15 MiB, generated: 259.71 MiB, post-processed: Unknown size, total: 386.86 MiB) to /mnt/cache/tanhaochen/.cache/huggingface/datasets/clue/chid/1.0.0/e55b490cb7809dcd8db31b9a87119f2e2ec87cdc060da8a9ac070b070ca3e379...\r\nTraceback (most recent call last):\r\n File \"/mnt/cache/tanhaochen/PromptCLUE/test_datasets.py\", line 3, in <module>\r\n cluewsc2020 = datasets.load_dataset(\"clue\",\"chid\")\r\n File \"/mnt/cache/tanhaochen/dependencies/datasets/src/datasets/load.py\", line 1667, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/mnt/cache/tanhaochen/dependencies/datasets/src/datasets/builder.py\", line 593, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/mnt/cache/tanhaochen/dependencies/datasets/src/datasets/builder.py\", line 663, in _download_and_prepare\r\n verify_checksums(\r\n File \"/mnt/cache/tanhaochen/dependencies/datasets/src/datasets/utils/info_utils.py\", line 40, in verify_checksums\r\n raise NonMatchingChecksumError(error_msg + str(bad_urls))\r\ndatasets.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['https://storage.googleapis.com/cluebenchmark/tasks/chid_public.zip']\r\n`", "Hi,\r\n\r\nthe fix hasn't been merged yet (it should be merged early next week)." ]

[]

[]

[ "This requires implementing `cast_column` for iterable datasets, it could be a very nice addition !\r\n\r\n<s>It can also be useful to be able to disable the audio/image decoding for the dataset viewer (see PR https://github.com/huggingface/datasets/pull/3430) cc @severo </s>\r\nEDIT: actually following https://github.com/huggingface/datasets/issues/3145 the dataset viewer might not need it anymore", "Just to clarify a bit. This feature is **always** needed when using the common voice dataset in streaming mode. So I think it's quite important" ]

[]

[ "The Hub pulls these tags from [here](https://github.com/huggingface/hub-docs/blob/main/js/src/lib/interfaces/Types.ts) (allows multimodal tasks) now, so I'm closing this issue." ]

[ "Just a side note: duplicate #3264" ]

[ "Hi, \r\n\r\ncould you please post this question on our [Forum](https://discuss.huggingface.co/) as we keep issues for bugs and feature requests? ", "> Hi,\r\n> \r\n> could you please post this question on our [Forum](https://discuss.huggingface.co/) as we keep issues for bugs and feature requests?\r\n\r\nok." ]

[ "Hi ! Your function looks fine, I used to map `squad` locally and it indeed added the `example_id` field correctly.\r\n\r\nHowever I think that in the `compute_loss` method only a subset of the fields are available: the model inputs. Since `example_id` is not a model input (it's not passed as a parameter to the model), the data loader doesn't need to return it by default.\r\n\r\nHowever you can disable this behavior by setting `remove_unused_columns` to `False` to your training arguments. In this case in `compute_loss` you will get the full item with all the fields.\r\n\r\nNote that since the model doesn't take `example_id` as input, you will have to remove it from the inputs when `model(**inputs)` is called", "Hi, I have set **args.remove_unused_columns=False** and **training_args.remove_unused_columns=False**, but the field doesn't been contained yet.\r\n```\r\ndef main():\r\n argp = HfArgumentParser(TrainingArguments)\r\n # The HfArgumentParser object collects command-line arguments into an object (and provides default values for unspecified arguments).\r\n # In particular, TrainingArguments has several keys that you'll need/want to specify (when you call run.py from the command line):\r\n # --do_train\r\n # When included, this argument tells the script to train a model.\r\n # See docstrings for \"--task\" and \"--dataset\" for how the training dataset is selected.\r\n # --do_eval\r\n # When included, this argument tells the script to evaluate the trained/loaded model on the validation split of the selected dataset.\r\n # --per_device_train_batch_size <int, default=8>\r\n # This is the training batch size.\r\n # If you're running on GPU, you should try to make this as large as you can without getting CUDA out-of-memory errors.\r\n # For reference, with --max_length=128 and the default ELECTRA-small model, a batch size of 32 should fit in 4gb of GPU memory.\r\n # --num_train_epochs <float, default=3.0>\r\n # How many passes to do through the training data.\r\n # --output_dir <path>\r\n # Where to put the trained model checkpoint(s) and any eval predictions.\r\n # *This argument is required*.\r\n\r\n argp.add_argument('--model', type=str,\r\n default='google/electra-small-discriminator',\r\n help=\"\"\"This argument specifies the base model to fine-tune.\r\n This should either be a HuggingFace model ID (see https://huggingface.co/models)\r\n or a path to a saved model checkpoint (a folder containing config.json and pytorch_model.bin).\"\"\")\r\n argp.add_argument('--task', type=str, choices=['nli', 'qa'], required=True,\r\n help=\"\"\"This argument specifies which task to train/evaluate on.\r\n Pass \"nli\" for natural language inference or \"qa\" for question answering.\r\n By default, \"nli\" will use the SNLI dataset, and \"qa\" will use the SQuAD dataset.\"\"\")\r\n argp.add_argument('--dataset', type=str, default=None,\r\n help=\"\"\"This argument overrides the default dataset used for the specified task.\"\"\")\r\n argp.add_argument('--max_length', type=int, default=128,\r\n help=\"\"\"This argument limits the maximum sequence length used during training/evaluation.\r\n Shorter sequence lengths need less memory and computation time, but some examples may end up getting truncated.\"\"\")\r\n argp.add_argument('--max_train_samples', type=int, default=None,\r\n help='Limit the number of examples to train on.')\r\n argp.add_argument('--max_eval_samples', type=int, default=None,\r\n help='Limit the number of examples to evaluate on.')\r\n\r\n argp.remove_unused_columns = False\r\n training_args, args = argp.parse_args_into_dataclasses()\r\n args.remove_unused_columns=False\r\n training_args.remove_unused_columns=False\r\n```\r\n\r\n\r\n```\r\n**************** train_dataset: Dataset({\r\n features: ['id', 'title', 'context', 'question', 'answers'],\r\n num_rows: 87599\r\n})\r\n\r\n\r\n**************** train_dataset_featurized: Dataset({\r\n features: ['attention_mask', 'end_positions', 'input_ids', 'start_positions', 'token_type_ids'],\r\n num_rows: 87714\r\n})\r\n```", "Hi, I print the value, all are set to False, but don't work.\r\n```\r\n********************* training_args: TrainingArguments(\r\n_n_gpu=1,\r\nadafactor=False,\r\nadam_beta1=0.9,\r\nadam_beta2=0.999,\r\nadam_epsilon=1e-08,\r\ndataloader_drop_last=False,\r\ndataloader_num_workers=0,\r\ndataloader_pin_memory=True,\r\nddp_find_unused_parameters=None,\r\ndebug=[],\r\ndeepspeed=None,\r\ndisable_tqdm=False,\r\ndo_eval=False,\r\ndo_predict=False,\r\ndo_train=True,\r\neval_accumulation_steps=None,\r\neval_steps=None,\r\nevaluation_strategy=IntervalStrategy.NO,\r\nfp16=False,\r\nfp16_backend=auto,\r\nfp16_full_eval=False,\r\nfp16_opt_level=O1,\r\ngradient_accumulation_steps=1,\r\ngreater_is_better=None,\r\ngroup_by_length=False,\r\nignore_data_skip=False,\r\nlabel_names=None,\r\nlabel_smoothing_factor=0.0,\r\nlearning_rate=5e-05,\r\nlength_column_name=length,\r\nload_best_model_at_end=False,\r\nlocal_rank=-1,\r\nlog_level=-1,\r\nlog_level_replica=-1,\r\nlog_on_each_node=True,\r\nlogging_dir=./re_trained_model/runs/Dec01_14-15-08_399b9290604c,\r\nlogging_first_step=False,\r\nlogging_steps=500,\r\nlogging_strategy=IntervalStrategy.STEPS,\r\nlr_scheduler_type=SchedulerType.LINEAR,\r\nmax_grad_norm=1.0,\r\nmax_steps=-1,\r\nmetric_for_best_model=None,\r\nmp_parameters=,\r\nno_cuda=False,\r\nnum_train_epochs=3.0,\r\noutput_dir=./re_trained_model,\r\noverwrite_output_dir=False,\r\npast_index=-1,\r\nper_device_eval_batch_size=8,\r\nper_device_train_batch_size=8,\r\nprediction_loss_only=False,\r\npush_to_hub=False,\r\npush_to_hub_model_id=re_trained_model,\r\npush_to_hub_organization=None,\r\npush_to_hub_token=None,\r\nremove_unused_columns=False,\r\nreport_to=['tensorboard'],\r\nresume_from_checkpoint=None,\r\nrun_name=./re_trained_model,\r\nsave_on_each_node=False,\r\nsave_steps=500,\r\nsave_strategy=IntervalStrategy.STEPS,\r\nsave_total_limit=None,\r\nseed=42,\r\nsharded_ddp=[],\r\nskip_memory_metrics=True,\r\ntpu_metrics_debug=False,\r\ntpu_num_cores=None,\r\nuse_legacy_prediction_loop=False,\r\nwarmup_ratio=0.0,\r\nwarmup_steps=0,\r\nweight_decay=0.0,\r\n)\r\n```\r\n```\r\n********************* args: Namespace(dataset='squad', max_eval_samples=None, max_length=128, max_train_samples=None, model='google/electra-small-discriminator', remove_unused_columns=False, task='qa')\r\n2021-12-01 14:15:10,048 - WARNING - datasets.builder - Reusing dataset squad (/root/.cache/huggingface/datasets/squad/plain_text/1.0.0/d6ec3ceb99ca480ce37cdd35555d6cb2511d223b9150cce08a837ef62ffea453)\r\nSome weights of the model checkpoint at google/electra-small-discriminator were not used when initializing ElectraForQuestionAnswering: ['discriminator_predictions.dense_prediction.weight', 'discriminator_predictions.dense_prediction.bias', 'discriminator_predictions.dense.weight', 'discriminator_predictions.dense.bias']\r\n- This IS expected if you are initializing ElectraForQuestionAnswering from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing ElectraForQuestionAnswering from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of ElectraForQuestionAnswering were not initialized from the model checkpoint at google/electra-small-discriminator and are newly initialized: ['qa_outputs.bias', 'qa_outputs.weight']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\nPreprocessing data... (this takes a little bit, should only happen once per dataset)\r\n```", "Hmmm, it might be because the default data collator removes all the fields with `string` type:\r\n\r\nhttps://github.com/huggingface/transformers/blob/4c0dd199c8305903564c2edeae23d294edd4b321/src/transformers/data/data_collator.py#L107-L112\r\n\r\nI guess you also need a custom data collator that doesn't remove them.", "can you give a tutorial about how to do this?", "I overwrite **get_train_dataloader**, and remove **_remove_unused_columns**, but it doesn't work.\r\n\r\n```\r\n def get_train_dataloader(self) -> DataLoader:\r\n \"\"\"\r\n Returns the training :class:`~torch.utils.data.DataLoader`.\r\n\r\n Will use no sampler if :obj:`self.train_dataset` does not implement :obj:`__len__`, a random sampler (adapted\r\n to distributed training if necessary) otherwise.\r\n\r\n Subclass and override this method if you want to inject some custom behavior.\r\n \"\"\"\r\n if self.train_dataset is None:\r\n raise ValueError(\"Trainer: training requires a train_dataset.\")\r\n\r\n train_dataset = self.train_dataset\r\n # if is_datasets_available() and isinstance(train_dataset, datasets.Dataset):\r\n # train_dataset = self._remove_unused_columns(train_dataset, description=\"training\")\r\n\r\n if isinstance(train_dataset, torch.utils.data.IterableDataset):\r\n if self.args.world_size > 1:\r\n train_dataset = IterableDatasetShard(\r\n train_dataset,\r\n batch_size=self.args.train_batch_size,\r\n drop_last=self.args.dataloader_drop_last,\r\n num_processes=self.args.world_size,\r\n process_index=self.args.process_index,\r\n )\r\n\r\n return DataLoader(\r\n train_dataset,\r\n batch_size=self.args.train_batch_size,\r\n collate_fn=self.data_collator,\r\n num_workers=self.args.dataloader_num_workers,\r\n pin_memory=self.args.dataloader_pin_memory,\r\n )\r\n\r\n train_sampler = self._get_train_sampler()\r\n\r\n return DataLoader(\r\n train_dataset,\r\n batch_size=self.args.train_batch_size,\r\n sampler=train_sampler,\r\n collate_fn=self.data_collator,\r\n drop_last=self.args.dataloader_drop_last,\r\n num_workers=self.args.dataloader_num_workers,\r\n pin_memory=self.args.dataloader_pin_memory,\r\n )\r\n```", "Hi, it works now, thank you.\r\n1. **args.remove_unused_columns=False** and **training_args.remove_unused_columns=False**\r\n2. overwrite **get_train_dataloader**, and remove **_remove_unused_columns**\r\n3. add new fields, and can be got in **inputs**. " ]

[ "Scratch that, probably the old and incompatible usage of dataset builder from promptsource.", "Actually, re-opening this issue cause the error persists\r\n\r\n```python\r\n>>> load_dataset(\"qed\")\r\nDownloading and preparing dataset qed/qed (download: 13.43 MiB, generated: 9.70 MiB, post-processed: Unknown size, total: 23.14 MiB) to /home/victor_huggingface_co/.cache/huggingface/datasets/qed/qed/1.0.0/47d8b6f033393aa520a8402d4baf2d6bdc1b2fbde3dc156e595d2ef34caf7d75...\r\n100%|███████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2228.64it/s]\r\nTraceback (most recent call last): \r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/load.py\", line 1669, in load_dataset\r\n use_auth_token=use_auth_token,\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/builder.py\", line 594, in download_and_prepare\r\n dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/builder.py\", line 681, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/builder.py\", line 1083, in _prepare_split\r\n num_examples, num_bytes = writer.finalize()\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/arrow_writer.py\", line 468, in finalize\r\n self.write_examples_on_file()\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/arrow_writer.py\", line 339, in write_examples_on_file\r\n pa_array = pa.array(typed_sequence)\r\n File \"pyarrow/array.pxi\", line 229, in pyarrow.lib.array\r\n File \"pyarrow/array.pxi\", line 110, in pyarrow.lib._handle_arrow_array_protocol\r\n File \"/home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages/datasets/arrow_writer.py\", line 125, in __arrow_array__\r\n out = pa.array(cast_to_python_objects(self.data, only_1d_for_numpy=True), type=type)\r\n File \"pyarrow/array.pxi\", line 315, in pyarrow.lib.array\r\n File \"pyarrow/array.pxi\", line 39, in pyarrow.lib._sequence_to_array\r\n File \"pyarrow/error.pxi\", line 143, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 99, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Could not convert 'in' with type str: tried to convert to boolean\r\n```\r\n\r\nEnvironment (datasets and pyarrow):\r\n\r\n```bash\r\n(promptsource) victor_huggingface_co@victor-dev:~/promptsource$ datasets-cli env\r\n\r\nCopy-and-paste the text below in your GitHub issue.\r\n\r\n- `datasets` version: 1.16.1\r\n- Platform: Linux-5.0.0-1020-gcp-x86_64-with-debian-buster-sid\r\n- Python version: 3.7.11\r\n- PyArrow version: 6.0.1\r\n```\r\n```bash\r\n(promptsource) victor_huggingface_co@victor-dev:~/promptsource$ pip show pyarrow\r\nName: pyarrow\r\nVersion: 6.0.1\r\nSummary: Python library for Apache Arrow\r\nHome-page: https://arrow.apache.org/\r\nAuthor: \r\nAuthor-email: \r\nLicense: Apache License, Version 2.0\r\nLocation: /home/victor_huggingface_co/miniconda3/envs/promptsource/lib/python3.7/site-packages\r\nRequires: numpy\r\nRequired-by: streamlit, datasets\r\n```" ]

[ "Hi,\r\n\r\nthe dataset is downloaded normally on my machine. Maybe the URL was down at the time of your download. Could you try again?", "> Hi,\r\n> \r\n> the dataset is downloaded normally on my machine. Maybe the URL was down at the time of your download. Could you try again?\r\n\r\nI have tried that many times with both load_dataset() and a browser almost simultaneously. The browser always works for me while load_dataset() fails.", "@mariosasko \r\n> the dataset is downloaded normally on my machine. Maybe the URL was down at the time of your download. Could you try again?\r\n\r\nI've tried yet again just a moment ago. This time I realize that, the step `(... post-processed: Unknown size, total: 20.89 MiB) to /Users/mike/.cache/huggingface/datasets/species800/species_800/1.0.0/532167f0bb8fbc0d77d6d03c4fd642c8c55527b9c5f2b1da77f3d00b0e559976...` and the one after seem unstable. If I want to retry, I will have to delete it (and probably other cache lock files). It **_sometimes_** works.\r\n\r\nBut I didn't try `download_mode=\"force_redownload\"` yet.\r\n\r\nAnyway, I suppose this isn't really a pressing issue for the time being, so I'm going to close this. Thank you.\r\n\r\n" ]

[ "I created a GitHub project to keep track of what needs to be done:\r\nhttps://github.com/huggingface/datasets/projects/3\r\n\r\nI also store my code in a (private for now) repository at https://github.com/huggingface/mirror_canonical_datasets_on_hub", "I understand that the datasets are mirrored on the Hub now, right? Might I close @lhoestq @SBrandeis?" ]

[ "This might be related to https://github.com/tensorflow/tensorflow/issues/38762 , what do you think @Rocketknight1 ?\r\n> Dataset.from_generator is expected to not work with TPUs as it uses py_function underneath which is incompatible with Cloud TPU 2VM setup. If you would like to read from large datasets, maybe try to materialize it on disk and use TFRecordDataest instead.", "Hi @lhoestq @nbroad1881, I think it's very similar, yes. Unfortunately `to_tf_dataset` uses `tf.numpy_function` which can't be compiled - this is a necessary evil to load from the underlying Arrow dataset. We need to update the notebooks/examples to clarify that this won't work, or to identify a workaround. You may be able to get it to work on an actual cloud TPU VM, but those are quite new and we haven't tested it yet. ", "Thank you for the explanation. I didn't realize the nuances of `tf.numpy_function`. In this scenario, would it be better to use `export(format='tfrecord')` ? It's not quite the same, but for very large datasets that don't fit in memory it looks like it is the only option. I haven't used `export` before, but I do recall reading that there are suggestions for how big and how many tfrecords there should be to not bottleneck the TPU. It might be nice if there were a way for the `export` method to split the files up into appropriate chunk sizes depending on the size of the dataset and the number of devices. And if that is too much, it would be nice to be able to specify the number of files that would be created when using `export`. Well... maybe the user should just do the chunking themselves and call `export` a bunch of times. Whatever the case, you have been helpful. Thanks Tensorflow boy ;-) ", "Yeah, this is something we really should have a proper guide on. I'll make a note to test some things and make a 'TF TPU best practices' notebook at some point, but in the meantime I think your solution of exporting TFRecords will probably work. ", "Also: I knew that tweet would haunt me" ]

[ "Hi ! Thanks for the suggestion, I didn't know about this decorator.\r\n\r\nIf you are interesting in contributing, feel free to open a pull request to add the overload methods for each typing combination :) To assign you to this issue, you can comment `#self-assign` in this thread.\r\n\r\n`Dataset.__getitem__` is defined right here: https://github.com/huggingface/datasets/blob/e6f1352fe19679de897f3d962e616936a17094f5/src/datasets/arrow_dataset.py#L1840", "#self-assign" ]

[ "If possible, a neat API could be something like `Dataset.to_polars()`, as well as `Dataset.set_format(\"polars\")`", "Note they use a \"custom\" implementation of Arrow: [Arrow2](https://github.com/jorgecarleitao/arrow2).", "Polars has grown rapidly in popularity over the last year - could you consider integrating the Polars functionality again?\r\n\r\nI don't think the \"custom\" implementation should be a barrier, it still conforms to the Arrow specification ", "Is there some direction regarding this from the HF team @lewtun ? Can conversion from polars to HF dataset be implemented with limited/zero copy? So, something like ``Dataset.from_polars()`` and ``Dataset.to_polars()`` like you mentioned. Happy to contribute if I can get some pointers on how this may be implemented.", "Hi, is there any updates? Thanks!" ]

[ "Hi ! The message you're getting is not an error. It simply says that your JSON dataset is being prepared to a location in `/root/.cache/huggingface/datasets`", "> \r\n\r\nbut I want to load local JSON file by command\r\n`python3 run.py --do_train --task qa --dataset squad-retrain-data/train-v2.0.json --output_dir ./re_trained_model/`\r\n\r\n**squad-retrain-data/train-v2.0.json** is the local JSON file, how to load it and map it to a special structure?", "You can load it with `dataset = datasets.load_dataset('json', data_files=args.dataset)` as you said.\r\nThen if you need to apply additional processing to map it to a special structure, you can use rename columns or use `dataset.map`. For more information, you can check the documentation here: https://huggingface.co/docs/datasets/process.html\r\n\r\nAlso feel free to share your `run.py` code so we can take a look", "```\r\n# Dataset selection\r\n if args.dataset.endswith('.json') or args.dataset.endswith('.jsonl'):\r\n dataset_id = None\r\n # Load from local json/jsonl file\r\n dataset = datasets.load_dataset('json', data_files=args.dataset)\r\n # By default, the \"json\" dataset loader places all examples in the train split,\r\n # so if we want to use a jsonl file for evaluation we need to get the \"train\" split\r\n # from the loaded dataset\r\n eval_split = 'train'\r\n else:\r\n default_datasets = {'qa': ('squad',), 'nli': ('snli',)}\r\n dataset_id = tuple(args.dataset.split(':')) if args.dataset is not None else \\\r\n default_datasets[args.task]\r\n # MNLI has two validation splits (one with matched domains and one with mismatched domains). Most datasets just have one \"validation\" split\r\n eval_split = 'validation_matched' if dataset_id == ('glue', 'mnli') else 'validation'\r\n # Load the raw data\r\n dataset = datasets.load_dataset(*dataset_id)\r\n```\r\n\r\nI want to load JSON squad dataset instead `dataset = datasets.load_dataset('squad')` to retrain the model. \r\n", "If your JSON has the same format as the SQuAD dataset, then you need to pass `field=\"data\"` to `load_dataset`, since the SQuAD format is one big JSON object in which the \"data\" field contains the list of questions and answers.\r\n```python\r\ndataset = datasets.load_dataset('json', data_files=args.dataset, field=\"data\")\r\n```\r\n\r\nLet me know if that helps :)\r\n\r\n", "Yes, code works. but the format is not as expected.\r\n```\r\ndataset = datasets.load_dataset('json', data_files=args.dataset, field=\"data\")\r\n```\r\n```\r\npython3 run.py --do_train --task qa --dataset squad --output_dir ./re_trained_model/\r\n```\r\n************ train_dataset: Dataset({\r\n features: ['id', 'title', 'context', 'question', 'answers'],\r\n num_rows: 87599\r\n})\r\n\r\n\r\n```\r\npython3 run.py --do_train --task qa --dataset squad-retrain-data/train-v2.0.json --output_dir ./re_trained_model/\r\n```\r\n************ train_dataset: Dataset({\r\n features: ['title', 'paragraphs'],\r\n num_rows: 442\r\n})\r\n\r\nI want the JSON to have the same format as before features. https://github.com/huggingface/datasets/blob/master/datasets/squad_v2/squad_v2.py is the script dealing with **squad** but how can I apply it by using JSON? ", "Ok I see, you have the paragraphs so you just need to process them to extract the questions and answers. I think you can process the SQuAD-like data this way:\r\n```python\r\ndef process_squad(articles):\r\n out = {\r\n \"title\": [],\r\n \"context\": [],\r\n \"question\": [],\r\n \"id\": [],\r\n \"answers\": [],\r\n }\r\n for title, paragraphs in zip(articles[\"title\"], articles[\"paragraphs\"]):\r\n for paragraph in paragraphs:\r\n for qa in paragraph[\"qas\"]:\r\n out[\"title\"].append(title)\r\n out[\"context\"].append(paragraph[\"context\"])\r\n out[\"question\"].append(qa[\"question\"])\r\n out[\"id\"].append(qa[\"id\"])\r\n out[\"answers\"].append({\r\n \"answer_start\": [answer[\"answer_start\"] for answer in qa[\"answers\"]],\r\n \"text\": [answer[\"text\"] for answer in qa[\"answers\"]],\r\n })\r\n return out\r\n\r\ndataset = dataset.map(process_squad, batched=True, remove_columns=[\"paragraphs\"])\r\n```\r\n\r\nI adapted the code from [squad.py](https://github.com/huggingface/datasets/blob/master/datasets/squad/squad.py). The code takes as input a batch of articles (title + paragraphs) and gets all the questions and answers from the JSON structure.\r\n\r\nThe output is a dataset with `features: ['answers', 'context', 'id', 'question', 'title']`\r\n\r\nLet me know if that helps !\r\n", "Yes, this works. But how to get the training output during training the squad by **Trainer** \r\nfor example https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/trainer_qa.py \r\nI want the training inputs, labels, outputs for every epoch and step to produce the training dynamic graph", "I think you may need to implement your own Trainer, from the `QuestionAnsweringTrainer` for example.\r\nThis way you can have the flexibility of saving all the inputs/output used at each step", "does there have any function to be overwritten to do this?", "> does there have any function to be overwritten to do this?\r\n\r\nok, I overwrote the compute_loss, thank you.", "Hi, I add one field **example_id**, but I can't see it in the **comput_loss** function, how can I do this? below is the information of inputs\r\n\r\n```\r\n*********************** inputs: {'attention_mask': tensor([[1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n ...,\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0],\r\n [1, 1, 1, ..., 0, 0, 0]], device='cuda:0'), 'end_positions': tensor([ 25, 97, 93, 44, 25, 112, 109, 134], device='cuda:0'), 'input_ids': tensor([[ 101, 2054, 2390, ..., 0, 0, 0],\r\n [ 101, 2054, 2515, ..., 0, 0, 0],\r\n [ 101, 2054, 2106, ..., 0, 0, 0],\r\n ...,\r\n [ 101, 2339, 2001, ..., 0, 0, 0],\r\n [ 101, 2054, 2515, ..., 0, 0, 0],\r\n [ 101, 2054, 2003, ..., 0, 0, 0]], device='cuda:0'), 'start_positions': tensor([ 20, 90, 89, 41, 25, 96, 106, 132], device='cuda:0'), 'token_type_ids': tensor([[0, 0, 0, ..., 0, 0, 0],\r\n [0, 0, 0, ..., 0, 0, 0],\r\n [0, 0, 0, ..., 0, 0, 0],\r\n ...,\r\n [0, 0, 0, ..., 0, 0, 0],\r\n [0, 0, 0, ..., 0, 0, 0],\r\n [0, 0, 0, ..., 0, 0, 0]], device='cuda:0')} \r\n```\r\n\r\n```\r\n# This function preprocesses a question answering dataset, tokenizing the question and context text\r\n# and finding the right offsets for the answer spans in the tokenized context (to use as labels).\r\n# Adapted from https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_qa.py\r\ndef prepare_train_dataset_qa(examples, tokenizer, max_seq_length=None):\r\n questions = [q.lstrip() for q in examples[\"question\"]]\r\n max_seq_length = tokenizer.model_max_length\r\n # tokenize both questions and the corresponding context\r\n # if the context length is longer than max_length, we split it to several\r\n # chunks of max_length\r\n tokenized_examples = tokenizer(\r\n questions,\r\n examples[\"context\"],\r\n truncation=\"only_second\",\r\n max_length=max_seq_length,\r\n stride=min(max_seq_length // 2, 128),\r\n return_overflowing_tokens=True,\r\n return_offsets_mapping=True,\r\n padding=\"max_length\"\r\n )\r\n\r\n # Since one example might give us several features if it has a long context,\r\n # we need a map from a feature to its corresponding example.\r\n sample_mapping = tokenized_examples.pop(\"overflow_to_sample_mapping\")\r\n # The offset mappings will give us a map from token to character position\r\n # in the original context. This will help us compute the start_positions\r\n # and end_positions to get the final answer string.\r\n offset_mapping = tokenized_examples.pop(\"offset_mapping\")\r\n\r\n tokenized_examples[\"start_positions\"] = []\r\n tokenized_examples[\"end_positions\"] = []\r\n\r\n tokenized_examples[\"example_id\"] = []\r\n\r\n for i, offsets in enumerate(offset_mapping):\r\n input_ids = tokenized_examples[\"input_ids\"][i]\r\n # We will label features not containing the answer the index of the CLS token.\r\n cls_index = input_ids.index(tokenizer.cls_token_id)\r\n sequence_ids = tokenized_examples.sequence_ids(i)\r\n # from the feature idx to sample idx\r\n sample_index = sample_mapping[i]\r\n # get the answer for a feature\r\n answers = examples[\"answers\"][sample_index]\r\n\r\n tokenized_examples[\"example_id\"].append(examples[\"id\"][sample_index])\r\n\r\n if len(answers[\"answer_start\"]) == 0:\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Start/end character index of the answer in the text.\r\n start_char = answers[\"answer_start\"][0]\r\n end_char = start_char + len(answers[\"text\"][0])\r\n\r\n # Start token index of the current span in the text.\r\n token_start_index = 0\r\n while sequence_ids[token_start_index] != 1:\r\n token_start_index += 1\r\n\r\n # End token index of the current span in the text.\r\n token_end_index = len(input_ids) - 1\r\n while sequence_ids[token_end_index] != 1:\r\n token_end_index -= 1\r\n\r\n # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).\r\n if not (offsets[token_start_index][0] <= start_char and\r\n offsets[token_end_index][1] >= end_char):\r\n tokenized_examples[\"start_positions\"].append(cls_index)\r\n tokenized_examples[\"end_positions\"].append(cls_index)\r\n else:\r\n # Otherwise move the token_start_index and token_end_index to the two ends of the answer.\r\n # Note: we could go after the last offset if the answer is the last word (edge case).\r\n while token_start_index < len(offsets) and \\\r\n offsets[token_start_index][0] <= start_char:\r\n token_start_index += 1\r\n tokenized_examples[\"start_positions\"].append(\r\n token_start_index - 1)\r\n while offsets[token_end_index][1] >= end_char:\r\n token_end_index -= 1\r\n tokenized_examples[\"end_positions\"].append(token_end_index + 1)\r\n\r\n return tokenized_examples\r\n```" ]

[ "Hi,\r\n\r\nthe fix was merged and will be available in the next release of `datasets`.\r\nIn the meantime, you can use it by installing `datasets` directly from master as follows:\r\n```\r\npip install git+https://github.com/huggingface/datasets.git\r\n```" ]

[ "Hi, thanks for reporting.\r\n\r\nIt would be really helpful if you could provide the actual code of the `text_numbers_to_int` function so we can reproduce the error.", "```\r\ndef text_numbers_to_int(text, column=\"\"):\r\n \"\"\"\r\n Convert text numbers to int.\r\n\r\n :param text: text numbers\r\n :return: int\r\n \"\"\"\r\n try:\r\n numbers = find_numbers(text)\r\n if not numbers:\r\n return text\r\n result = \"\"\r\n i, j = 0, 0\r\n while i < len(text):\r\n if j < len(numbers) and i == numbers[j][1]:\r\n n = int(numbers[j][0]) if numbers[j][0] % 1 == 0 else float(numbers[j][0])\r\n result += str(n)\r\n i = numbers[j][2] #end\r\n j += 1\r\n else:\r\n result += text[i]\r\n i += 1\r\n if column:\r\n return{column: result}\r\n else:\r\n return {column: result}\r\n except Exception as e:\r\n print(e)\r\n return {column: result}\r\n```", "Maybe this is because of the `return text` line ? I think it should return a dictionary rather than a string", "Yes that was it, good catch! Thanks" ]

[ "#3323 " ]

[]

[]

[ "Hi,\r\n\r\n`Dataset.map` allows more generic transforms compared to `Dataset.filter`, which purpose is very specific (to filter examples based on a condition). That's why I don't think we need the `desc` parameter there for consistency. #3196 has added descriptions to the `Dataset` methods that call `.map` internally, but not for the `filter` method, so we should do that.\r\n\r\nDo you have a description in mind? Maybe `\"Filtering the dataset\"` or `\"Filtering the indices\"`? If yes, feel free to open a PR.", "I'm personally ok with adding the `desc` parameter actually. Let's say you have different filters, it can be nice to differentiate between the different filters when they're running no ?", "@mariosasko the use case is filtering of a dataset prior to tokenization and subsequent training. As the dataset is huge it's just a matter of giving a user (model trainer) some feedback on what's going on. Otherwise, feedback is given for all steps in training preparation and not for filtering and the filtering in my use case lasts about 4-5 minutes. And yes, if there are more filtering stages, as @lhoestq pointed out, it would be nice to give some feedback. I thought desc is there already and got confused when I got the script error. ", "I don't have a strong opinion on that, so having `desc` as a parameter is also OK." ]

[]

[ "Hi ! You're completely right, this must be mentioned in the dataset card.\r\nIf you're interesting in contributing, feel free to open a pull request to mention this in the `trivia_qa` dataset card in the \"Licensing Information\" section at https://github.com/huggingface/datasets/blob/master/datasets/trivia_qa/README.md" ]

[]

[ "Hi ! Are you having this issue only with this specific dataset, or it also happens with other ones like `squad` ?", "@lhoestq It happens also on `squad`. It successfully downloads the whole dataset and then crashes on: \r\n\r\n```\r\nFatal error condition occurred in D:\\bld\\aws-c-io_1633633258269\\work\\source\\event_loop.c:74: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS\r\nExiting Application\r\n```\r\n\r\nI tested it on Ubuntu and its working OK. Didn't test on non-preview version of Windows 11, `Windows-10-10.0.22504-SP0` is a preview version, not sure if this is causing it.", "I see the same error in Windows-10.0.19042 as of a few days ago:\r\n\r\n`Fatal error condition occurred in D:\\bld\\aws-c-io_1633633258269\\work\\source\\event_loop.c:74: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS`\r\n\r\npython 3.8.12 h7840368_2_cpython conda-forge\r\nboto3 1.20.11 pyhd8ed1ab_0 conda-forge\r\nbotocore 1.23.11 pyhd8ed1ab_0 conda-forge\r\n\r\n...but I am not using `datasets` (although I might take a look now that I know about it!)\r\n\r\nThe error has occurred a few times over the last two days, but not consistently enough for me to get it with DEBUG. If there is any interest I can report back here, but it seems not unique to `datasets`.", "I'm not sure what `datasets` has to do with a crash that seems related to `aws-c-io`, could it be an issue with your environment ?", "> I'm not sure what `datasets` has to do with a crash that seems related to `aws-c-io`, could it be an issue with your environment ?\r\n\r\nAgreed, this issue is not likely a bug in datasets, since I get the identical error without datasets installed.", "Will close this issue. Bug in `aws-c-io` shouldn't be in `datasets` repo. Nevertheless, it can be useful to know that it happens. Thanks @leehaust @lhoestq ", "I have also had this issue since a few days, when running scripts using PyCharm in particular, but it does not seem to affect the script from running, only reporting this error at the end of the run.", "I also get this issue, It appears after my script has finished running. I get the following error message\r\n```\r\nFatal error condition occurred in /home/conda/feedstock_root/build_artifacts/aws-c-io_1637179816120/work/source/event_loop.c:72: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS\r\nExiting Application\r\n################################################################################\r\nStack trace:\r\n################################################################################\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_backtrace_print+0x59) [0x2aabe0479579]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_fatal_assert+0x48) [0x2aabe04696c8]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././././libaws-c-io.so.1.0.0(+0x13ad3) [0x2aabe0624ad3]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_ref_count_release+0x1d) [0x2aabe047b60d]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././././libaws-c-io.so.1.0.0(+0x113ca) [0x2aabe06223ca]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_ref_count_release+0x1d) [0x2aabe047b60d]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-crt-cpp.so(_ZN3Aws3Crt2Io15ClientBootstrapD1Ev+0x3a) [0x2aabe041cf5a]\r\n/home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././libaws-cpp-sdk-core.so(+0x5f570) [0x2aabe00eb570]\r\n/lib64/libc.so.6(+0x39ce9) [0x2aaaab835ce9]\r\n/lib64/libc.so.6(+0x39d37) [0x2aaaab835d37]\r\n/lib64/libc.so.6(__libc_start_main+0xfc) [0x2aaaab81e55c]\r\npython(+0x1c721d) [0x55555571b21d]\r\nAborted\r\n```\r\nI don't get this issue when running my code in a container, and it seems more relevant to PyArrow but thought a more complete stack trace might be helpful to someone\r\n", "I created an issue on JIRA:\r\nhttps://issues.apache.org/jira/browse/ARROW-15141", "@CallumMcMahon Do you have a small reproducer for this problem on Linux? I can reproduce this on Windows but sadly not with linux.", "Any updates on this issue? I started receiving the same error a few days ago on the amazon reviews", "Hi,\r\n\r\nI also ran into this issue, Windows only. It caused our massive binary to minidump left and right, very annoying.\r\nWhen the program is doing an exit, the destructors in the exit-handlers want to do cleanup, leading to code in event_loop.c, on line 73-ish:\r\n\r\nAWS_FATAL_ASSERT(\r\n aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) ==\r\n AWS_OP_SUCCESS);\r\n\r\nThe fatal_assert end in an abort/minidump.\r\n\r\nDigging through the code, I found that aws_thread_launch in the Windows version (aws-c-common/source/windows/thread.c) has only ONE reason to return anything other than AWS_OP_SUCCESS:\r\n\r\nreturn aws_raise_error(AWS_ERROR_THREAD_INSUFFICIENT_RESOURCE);\r\n\r\non line 263, when CreateThread fails. Our conclusion was that, apparently, Windows dislikes launching a new thread while already handling the exit-handlers. And while I appreciate the the fatal_assert is there in case of problems, the cure here is worse than the problem.\r\n\r\nI \"fixed\" this in our (Windows) environment by (bluntly) removing the AWS_FATAL_ASSERT. If Windows cannot start a thread, the program is in deep trouble anyway and the chances of that actually happening are acceptable (to us).\r\nThe exit is going to clean up all resources anyway.\r\n\r\nA neater fix would probably be to detect somehow that the program is actually in the process of exiting and then not bother (on windows, anyway) to start a cleanup thread. Alternatively, try to start the thread but not fatal-assert when it fails during exit. Or perhaps Windows can be convinced somehow to start the thread under these circumstances?\r\n\r\n@xhochy : The problem is Windows-only, the aws_thread_launch has two implementations (posix and windows). The problem is in the windows CreateThread which fails.\r\n", "I also encountered the same problem, but I made an error in the multi gpu training environment on Linux, and the single gpu training environment will not make an error.\r\ni use accelerate package to do multi gpu training.", "> I also get this issue, It appears after my script has finished running. I get the following error message\r\n> \r\n> ```\r\n> Fatal error condition occurred in /home/conda/feedstock_root/build_artifacts/aws-c-io_1637179816120/work/source/event_loop.c:72: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS\r\n> Exiting Application\r\n> ################################################################################\r\n> Stack trace:\r\n> ################################################################################\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_backtrace_print+0x59) [0x2aabe0479579]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_fatal_assert+0x48) [0x2aabe04696c8]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././././libaws-c-io.so.1.0.0(+0x13ad3) [0x2aabe0624ad3]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_ref_count_release+0x1d) [0x2aabe047b60d]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././././libaws-c-io.so.1.0.0(+0x113ca) [0x2aabe06223ca]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-c-common.so.1(aws_ref_count_release+0x1d) [0x2aabe047b60d]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../../././libaws-crt-cpp.so(_ZN3Aws3Crt2Io15ClientBootstrapD1Ev+0x3a) [0x2aabe041cf5a]\r\n> /home/user_name/conda_envs/env_name/lib/python3.7/site-packages/pyarrow/../../.././libaws-cpp-sdk-core.so(+0x5f570) [0x2aabe00eb570]\r\n> /lib64/libc.so.6(+0x39ce9) [0x2aaaab835ce9]\r\n> /lib64/libc.so.6(+0x39d37) [0x2aaaab835d37]\r\n> /lib64/libc.so.6(__libc_start_main+0xfc) [0x2aaaab81e55c]\r\n> python(+0x1c721d) [0x55555571b21d]\r\n> Aborted\r\n> ```\r\n> \r\n> I don't get this issue when running my code in a container, and it seems more relevant to PyArrow but thought a more complete stack trace might be helpful to someone\r\n\r\nAny updates for your issue because I'm getting the same one ", "Potentially related AWS issue: https://github.com/aws/aws-sdk-cpp/issues/1809\r\n\r\nRan into this issue today while training a BPE tokenizer on a dataset.\r\n\r\nTrain code:\r\n\r\n```python\r\n\"\"\"Train a ByteLevelBPETokenizer based on a given dataset. The dataset must be on the HF Hub.\r\nThis script is adaptated from the Transformers example in https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling\r\n\"\"\"\r\nfrom os import PathLike\r\nfrom pathlib import Path\r\nfrom typing import Sequence, Union\r\n\r\nfrom datasets import load_dataset\r\nfrom tokenizers import ByteLevelBPETokenizer\r\n\r\n\r\ndef train_tokenizer(dataset_name: str = \"oscar\", dataset_config_name: str = \"unshuffled_deduplicated_nl\",\r\n dataset_split: str = \"train\", dataset_textcol: str = \"text\",\r\n vocab_size: int = 50265, min_frequency: int = 2,\r\n special_tokens: Sequence[str] = (\"<s>\", \"<pad>\", \"</s>\", \"<unk>\", \"<mask>\"),\r\n dout: Union[str, PathLike] = \".\"):\r\n # load dataset\r\n dataset = load_dataset(dataset_name, dataset_config_name, split=dataset_split)\r\n # Instantiate tokenizer\r\n tokenizer = ByteLevelBPETokenizer()\r\n\r\n def batch_iterator(batch_size=1024):\r\n for i in range(0, len(dataset), batch_size):\r\n yield dataset[i: i + batch_size][dataset_textcol]\r\n\r\n # Customized training\r\n tokenizer.train_from_iterator(batch_iterator(), vocab_size=vocab_size, min_frequency=min_frequency,\r\n special_tokens=special_tokens)\r\n\r\n # Save to disk\r\n pdout = Path(dout).resolve()\r\n pdout.mkdir(exist_ok=True, parents=True)\r\n tokenizer.save_model(str(pdout))\r\n\r\n\r\ndef main():\r\n import argparse\r\n cparser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.ArgumentDefaultsHelpFormatter)\r\n\r\n cparser.add_argument(\"dataset_name\", help=\"Name of dataset to use for tokenizer training\")\r\n cparser.add_argument(\"--dataset_config_name\", default=None,\r\n help=\"Name of the config to use for tokenizer training\")\r\n cparser.add_argument(\"--dataset_split\", default=None,\r\n help=\"Name of the split to use for tokenizer training (typically 'train')\")\r\n cparser.add_argument(\"--dataset_textcol\", default=\"text\",\r\n help=\"Name of the text column to use for tokenizer training\")\r\n cparser.add_argument(\"--vocab_size\", type=int, default=50265, help=\"Vocabulary size\")\r\n cparser.add_argument(\"--min_frequency\", type=int, default=2, help=\"Minimal frequency of tokens\")\r\n cparser.add_argument(\"--special_tokens\", nargs=\"+\", default=[\"<s>\", \"<pad>\", \"</s>\", \"<unk>\", \"<mask>\"],\r\n help=\"Special tokens to add. Useful for specific training objectives. Note that if you wish\"\r\n \" to use this tokenizer with a default transformers.BartConfig, then make sure that the\"\r\n \" order of at least these special tokens are correct: BOS (0), padding (1), EOS (2)\")\r\n cparser.add_argument(\"--dout\", default=\".\", help=\"Path to directory to save tokenizer.json file\")\r\n\r\n train_tokenizer(**vars(cparser.parse_args()))\r\n\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n```\r\n\r\nCommand:\r\n\r\n```sh\r\n$WDIR=\"your_tokenizer\"\r\npython prepare_tokenizer.py dbrd --dataset_config_name plain_text --dataset_split unsupervised --dout $WDIR\r\n```\r\n\r\nOutput:\r\n\r\n```\r\nReusing dataset dbrd (cache/datasets/dbrd/plain_text/3.0.0/2b12e31348489dfe586c2d0f40694e5d9f9454c9468457ac9f1b51abf686eeb3)\r\n[00:00:30] Pre-processing sequences ████████ 0 / 0\r\n[00:00:00] Tokenize words ████████ 333319 / 333319\r\n[00:01:06] Count pairs ████████ 333319 / 333319\r\n[00:00:03] Compute merges ████████ 50004 / 50004\r\n\r\nFatal error condition occurred in /opt/vcpkg/buildtrees/aws-c-io/src/9e6648842a-364b708815.clean/source/event_loop.c:72: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS\r\nExiting Application\r\n################################################################################\r\nStack trace:\r\n################################################################################\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x200af06) [0x155106589f06]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x20028e5) [0x1551065818e5]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x1f27e09) [0x1551064a6e09]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x200ba3d) [0x15510658aa3d]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x1f25948) [0x1551064a4948]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x200ba3d) [0x15510658aa3d]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x1ee0b46) [0x15510645fb46]\r\nvenv/lib/python3.9/site-packages/pyarrow/libarrow.so.900(+0x194546a) [0x155105ec446a]\r\n/lib64/libc.so.6(+0x39b0c) [0x1551075b8b0c]\r\n/lib64/libc.so.6(on_exit+0) [0x1551075b8c40]\r\n/lib64/libc.so.6(__libc_start_main+0xfa) [0x1551075a249a]\r\npython(_start+0x2e) [0x4006ce]\r\nAborted (core dumped)\r\n```\r\n\r\nRunning on datasets==2.4.0 and pyarrow==9.0.0 on RHEL 8.\r\n", "There is also a discussion here https://issues.apache.org/jira/browse/ARROW-15141 where it is suggested for conda users to use an older version of aws-sdk-cpp: `aws-sdk-cpp=1.8.186`", "Downgrading pyarrow to 6.0.1 solves the issue for me.\r\n\r\n`pip install pyarrow==6.0.1`", "First of all, I’d never call a downgrade a solution, at most a (very) temporary workaround.\r\nFurthermore: This bug also happens outside pyarrow, I incorporate AWS in a standalone Windows C-program and that crashes during exit.\r\n\r\nFrom: Bo-Ru (Roy) Lu ***@***.***>\r\nSent: Thursday, 15 September 2022 01:12\r\nTo: huggingface/datasets ***@***.***>\r\nCc: Ruurd Beerstra ***@***.***>; Comment ***@***.***>\r\nSubject: Re: [huggingface/datasets] Fatal error condition occurred in aws-c-io (Issue #3310)\r\n\r\nSent by an external sender. Please be cautious about clicking on links and opening attachments.\r\n--------------------------------------------------------------------------------------------------------------------------------\r\n\r\n\r\nDowngrading pyarrow to 6.0.1 solves the issue.\r\n\r\n—\r\nReply to this email directly, view it on GitHub<https://github.com/huggingface/datasets/issues/3310#issuecomment-1247390774>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKYUE3WBCSMHKJOOA2RQELLV6JLSVANCNFSM5IQ3WG7Q>.\r\nYou are receiving this because you commented.Message ID: ***@***.******@***.***>>\r\n", "> First of all, I’d never call a downgrade a solution, at most a (very) temporary workaround.\r\n\r\nVery much so! It looks like an apparent fix for the underlying problem [might](https://github.com/awslabs/aws-c-io/pull/515) have landed, but it sounds like it might still be a bit of a [lift](https://github.com/aws/aws-sdk-cpp/issues/1809#issuecomment-1289859795) to get it into aws-sdk-cpp.\r\n\r\n> Downgrading pyarrow to 6.0.1 solves the issue for me.\r\n\r\nSidenote: On conda-forge side, all recent pyarrow releases (all the way up to v9 and soon v10) have carried the respective pin and will not run into this issue.\r\n\r\n```\r\nconda install -c conda-forge pyarrow\r\n```\r\n\r\n", "For pip people, I confirmed that installing the nightly version of pyarrow also solves this by: `pip install --extra-index-url https://pypi.fury.io/arrow-nightlies/ --prefer-binary --pre pyarrow --upgrade`. (See https://arrow.apache.org/docs/python/install.html#installing-nightly-packages)\r\nAny version after https://github.com/apache/arrow/pull/14157 would work fine.", "> Furthermore: This bug also happens outside pyarrow, I incorporate AWS in a standalone Windows C-program and that crashes during exit.\r\n\r\nDo you have a reproducer you could share? I'd like to test if the new versions that supposedly solve this actually do, but we don't have a way to test it...", "Hi,\r\n\r\nNo – sorry. It is part of a massive eco-system which cannot easily be shared.\r\nBut I think the problem was summarized quite clearly: Windows does not allow a CreateThread while doing ExitProcess.\r\nThe cleanup that gets called as part of the exit handler code tries to start a thread, the fatal-assert on that causes the crash, and in windows we get a very big dump file.\r\nThe fix I applied simply removes that fatal assert, that solves the problem for me.\r\nI did not delve into the what the thread was trying to achieve and if that might cause issues when not executed during exit of the process. We did not notice anything of the kind.\r\nHowever, we *did* notice the many, many gigabytes of accumulated dumps of hundreds of processes 😊\r\n\r\nI’ll try and upgrade to the latest AWS version and report my findings, but that will be after I return from a month of vacationing…\r\n\r\n\r\n * Regards – Ruurd Beerstra\r\n\r\n\r\nFrom: h-vetinari ***@***.***>\r\nSent: Friday, 28 October 2022 02:09\r\nTo: huggingface/datasets ***@***.***>\r\nCc: Ruurd Beerstra ***@***.***>; Comment ***@***.***>\r\nSubject: Re: [huggingface/datasets] Fatal error condition occurred in aws-c-io (Issue #3310)\r\n\r\nSent by an external sender. Please be cautious about clicking on links and opening attachments.\r\n--------------------------------------------------------------------------------------------------------------------------------\r\n\r\n\r\nFurthermore: This bug also happens outside pyarrow, I incorporate AWS in a standalone Windows C-program and that crashes during exit.\r\n\r\nDo you have a reproducer you could share? I'd like to test if the new versions that supposedly solve this actually do, but we don't have a way to test it...\r\n\r\n—\r\nReply to this email directly, view it on GitHub<https://github.com/huggingface/datasets/issues/3310#issuecomment-1294251331>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AKYUE3SHHPC5AT7KQ4GDAJDWFMKRTANCNFSM5IQ3WG7Q>.\r\nYou are receiving this because you commented.Message ID: ***@***.******@***.***>>\r\n", "> No – sorry. It is part of a massive eco-system which cannot easily be shared.\r\n\r\nOK, was worth a try...\r\n\r\n> The fix I applied simply removes that fatal assert, that solves the problem for me.\r\n\r\nThis seems to be what https://github.com/awslabs/aws-c-io/pull/515 did upstream.\r\n\r\n> I’ll try and upgrade to the latest AWS version and report my findings, but that will be after I return from a month of vacationing…\r\n\r\ncaution: aws-sdk-cpp hasn't yet upgraded its bundled(?) aws-c-io and hence doesn't contain the fix AFAICT", "Hi, I also encountered the same problem, but I made an error on Ubuntu without using `datasets` as @Crabzmatic he wrote.\r\n\r\nAt that time, I find my version of pyarrow is 9.0.0, which is different from as follow:\r\n> https://github.com/huggingface/datasets/issues/3310#issuecomment-1247390774\r\n> Downgrading pyarrow to 6.0.1 solves the issue for me.\r\n> \r\n> `pip install pyarrow==6.0.1`\r\n\r\nAs it happens, I found this error message when I introduced the [`Trainer`](https://huggingface.co/docs/transformers/main_classes/trainer) of HuggingFace\r\n\r\nFor example, I write following code:\r\n```python\r\nfrom transformers import Trainer\r\nprint('Hugging Face')\r\n```\r\n I get the following error message:\r\n```python\r\nHugging Face\r\nFatal error condition occurred in /opt/vcpkg/buildtrees/aws-c-io/src/9e6648842a-364b708815.clean/source/event_loop.c:72: aws_thread_launch(&cleanup_thread, s_event_loop_destroy_async_thread_fn, el_group, &thread_options) == AWS_OP_SUCCESS\r\nExiting Application\r\n################################################################################\r\nStack trace:\r\n################################################################################\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x200af06) [0x7fa9add1df06]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x20028e5) [0x7fa9add158e5]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x1f27e09) [0x7fa9adc3ae09]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x200ba3d) [0x7fa9add1ea3d]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x1f25948) [0x7fa9adc38948]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x200ba3d) [0x7fa9add1ea3d]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x1ee0b46) [0x7fa9adbf3b46]\r\n/home/ubuntu/anaconda3/envs/pytorch38/lib/python3.8/site-packages/pyarrow/libarrow.so.900(+0x194546a) [0x7fa9ad65846a]\r\n/lib/x86_64-linux-gnu/libc.so.6(+0x468d7) [0x7faa2fcfe8d7]\r\n/lib/x86_64-linux-gnu/libc.so.6(on_exit+0) [0x7faa2fcfea90]\r\n/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xfa) [0x7faa2fcdc0ba]\r\n/home/ubuntu/anaconda3/envs/pytorch38/bin/python(+0x1f9ad7) [0x5654571d1ad7]\r\n```\r\nBut, when I remove the `Trainer` module from transformers, **everthing is OK**.\r\n\r\nSo Why ?\r\n\r\n**Environment info**\r\n- Platform: Ubuntu 18\r\n- Python version: 3.8\r\n- PyArrow version: 9.0.0\r\n- transformers: 4.22.1\r\n- simpletransformers: 0.63.9", "> I get the following error message:\r\n\r\nNot sure what's going on, but that shouldn't happen, especially as we're pinning to a version that should avoid this.\r\n\r\nCan you please open an issue https://github.com/conda-forge/arrow-cpp-feedstock, including the requested output of `conda list` & `conda info`?", "pyarrow 10.0.1 was just released in conda-forge, which is the first release where we're building against aws-sdk-cpp 1.9.* again after more than a year. Since we cannot test the failure reported here on our infra, I'd be very grateful if someone could verify that the problem does or doesn't reappear. 🙃 \r\n\r\n```\r\nconda install -c conda-forge pyarrow=10\r\n```", "> pyarrow 10.0.1 was just released in conda-forge, which is the first release where we're building against aws-sdk-cpp 1.9.* again after more than a year. Since we cannot test the failure reported here on our infra, I'd be very grateful if someone could verify that the problem does or doesn't reappear. 🙃\r\n> \r\n> ```\r\n> conda install -c conda-forge pyarrow=10\r\n> ```\r\n\r\nThe problem is gone after I install the new version. Thanks!\r\npip install pyarrow==10", "@liuchaoqun, with `pip install pyarrow` you don't get aws-bindings, they're too complicated to package into wheels as far as I know. And even if they're packaged, at the time of the release of pyarrow 10 it would have still been pinned to aws 1.8 for the same reasons as in this issue." ]