Datasets:

---

Xiaowei

/

github-issues

like 0

[]

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4205). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4204). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4203). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4201). All of your documentation changes will be reflected on that endpoint.", "You can see rendering at: https://github.com/huggingface/datasets/blob/6b48fedbdafe12a42c7b6edcecc32820af1a4822/.github/ISSUE_TEMPLATE/dataset-viewer.yml" ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

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

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

[ "Hi ! Maybe one of the objects in the function is not deterministic across sessions ? You can read more about it and how to investigate here: https://huggingface.co/docs/datasets/about_cache", "Hi @apsdehal! Can you verify that replacing\r\n```python\r\ndef fetch_single_image(image_url, timeout=None, retries=0):\r\n for _ in range(retries + 1):\r\n try:\r\n request = urllib.request.Request(\r\n image_url,\r\n data=None,\r\n headers={\"user-agent\": get_datasets_user_agent()},\r\n )\r\n with urllib.request.urlopen(request, timeout=timeout) as req:\r\n image = PIL.Image.open(io.BytesIO(req.read()))\r\n break\r\n except Exception:\r\n image = None\r\n return image\r\n```\r\nwith \r\n```python\r\nUSER_AGENT = get_datasets_user_agent()\r\n\r\ndef fetch_single_image(image_url, timeout=None, retries=0):\r\n for _ in range(retries + 1):\r\n try:\r\n request = urllib.request.Request(\r\n image_url,\r\n data=None,\r\n headers={\"user-agent\": USER_AGENT},\r\n )\r\n with urllib.request.urlopen(request, timeout=timeout) as req:\r\n image = PIL.Image.open(io.BytesIO(req.read()))\r\n break\r\n except Exception:\r\n image = None\r\n return image\r\n```\r\nfixes the issue?", "Thanks @mariosasko. That does fix the issue. In general, I think these image downloading utilities since they are being used by a lot of image dataset should be provided as a part of `datasets` library right to keep the logic consistent and READMEs smaller? If they already exists, that is also great, please point me to those. I saw that `http_get` does exist." ]

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

[]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "Any reason why we can't just do `[inputs.copy()]` in this line for in-place operations to not have effects anymore:\r\nhttps://github.com/huggingface/datasets/blob/bf432011ff9155a5bc16c03956bc63e514baf80d/src/datasets/arrow_dataset.py#L2232.\r\n\r\n(in the `batched` case, we can also copy the inputs' values (list objects) to ignore in-place modifications to the inputs' columns)\r\n\r\nI think `remove_columns=True` has no meaning, so I'm not a fan of this change.", "@mariosasko copy does have a cost associated with it ... and plus you'll have to consider `deepcopy` Imagine columnds that are list of list of list of list .... Though I have to agree that `remove_columns=True` doesn't make sense (but, IMO, neither does it in its current use-case as it should refer to `input_columns`) ", "Okay closing this PR for the following reasons:\r\n - `remove_columns=True` was expected to keep the `.update`-like operator for `.map`. I initially thought it would be a good way to ignore function side effects and only keep output of that function (cf. PR description).\r\n - expected `remove_columns=True` is a bad API according to @mariosasko and introduces unecessary changes for little gain (strictly equivalent to `remove_columns=dset.column_names`)" ]

[]

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4194). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "Good catch, I updated the docstrings" ]

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

[]

[ "Hi! :)\r\n\r\nI believe that should work unless `dataset_infos.json` isn't actually a dataset. For Hugging Face datasets, there is usually a file named `dataset_infos.json` which contains metadata about the dataset (eg. the dataset citation, license, description, etc). Can you double-check that `dataset_infos.json` isn't just metadata please?", "Hi @ahf876828330, \r\n\r\nAs @stevhliu pointed out, the proper way to load a dataset is not trying to load its metadata file.\r\n\r\nIn your case, as the dataset script is local, you should better point to your local loading script:\r\n```python\r\ndataset = load_dataset(\"dataset/opus_books.py\")\r\n```\r\n\r\nPlease, feel free to re-open this issue if the previous code snippet does not work for you.", "> Hi! :)\r\n> \r\n> I believe that should work unless `dataset_infos.json` isn't actually a dataset. For Hugging Face datasets, there is usually a file named `dataset_infos.json` which contains metadata about the dataset (eg. the dataset citation, license, description, etc). Can you double-check that `dataset_infos.json` isn't just metadata please?\r\n\r\nYes，you are right!So if I have a metadata dataset local,How can I turn it to a dataset that can be used by the load_dataset() function？Are there some examples?" ]

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

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

[ "Hi @SaulLu, thanks for your proposal.\r\n\r\nJust I got a bit confused about the dimensions...\r\n- For the 2D case, you mention it is possible to create an `Array2D` from a list of arrays of dimension 1\r\n- However, you give an example of creating an `Array2D` from arrays of dimension 2:\r\n - the values of `data_map` are arrays of dimension 2\r\n - the outer list in `prepare_dataset_2D` should not be taken into account in the dimension counting, as it is used because in `map` you pass `batched=True`\r\n\r\nNote that for the 3D alternatives you mention:\r\n- In `prepare_dataset_3D_ter`, you create an `Array3D` from arrays of dimension 3:\r\n - the array `data_map[index][np.newaxis, :, :]` has dimension 3\r\n - the outer list in `prepare_dataset_3D_ter` is the one used by `batched=True`\r\n- In `prepare_dataset_3D_bis`, you create an `Array3D` from a list of list of lists:\r\n - the value of `data_map[index].tolist()` is a list of lists\r\n - it is enclosed by another list `[data_map[index].tolist()]`, thus giving a list of list of lists\r\n - the outer list is the one used by `batched=True`\r\n\r\nTherefore, if I understand correctly, your request would be to be able to create an `Array3D` from a list of an array of dimension 2:\r\n- In `prepare_dataset_3D`, `data_map[index]` is an array of dimension 2\r\n- it is enclosed by a list `[data_map[index]]`, thus giving a list of an array of dimension 2\r\n- the outer list is the one used by `batched=True`\r\n\r\nPlease, feel free to tell me if I did not understand you correctly.", "Hi @albertvillanova ,\r\n\r\nIndeed my message was confusing and you guessed right :smile: : I think would be interesting to be able to create an Array3D from a list of an array of dimension 2. \r\n\r\nFor the 2D case I should have given as a \"similar\" example:\r\n```python\r\n\r\ndata_map_1D = {\r\n 1: np.array([0.2, 0.4]),\r\n 2: np.array([0.1, 0.4]),\r\n}\r\n\r\ndef prepare_dataset_2D(batch):\r\n batch[\"pixel_values\"] = [[data_map_1D[index]] for index in batch[\"id\"]]\r\n return batch\r\n \r\nds_2D = ds.map(\r\n prepare_dataset_2D, \r\n batched=True, \r\n remove_columns=ds.column_names, \r\n features=features.Features({\"pixel_values\": features.Array2D(shape=(1, 2), dtype=\"float32\")})\r\n)\r\n```" ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4189). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "I'm not familiar with the concept of streaming vs non-streaming in HF datasets. I just wonder that you have the distinction here. Why doesn't it work to always make use of `bytes`? \"using a local file - which is often required for audio\" - why would that be?\r\n\r\nThe `path` would always point to some location in the `cache_dir`? I think this can be problematic. I would have expected that after I did `dataset.save_to_disk(...)` that I can remove the cache dir. But maybe just because I'm not familiar with HF. Or maybe the docs can be improved to clarify this.\r\n", "We could always load every data file into `bytes` and save it this way the audio as bytes in `arrow` format, but the problem then would be that it makes the `file` column useless, *i.e.* people cannot inspect the audio file locally anymore or else they would need to first save bytes as a file which is not evident. This either breaks backwards compatibility or forces the user to stored 2x the required size locally. There was a longer discussion here: https://github.com/huggingface/datasets/issues/3663\r\n\r\nIt's a good argument though that `dataset.save_to_disk(...)` should save everything that is needed to the disk and should be independent of other folders, but I do think the arguments of #3663 to not break backwards compatibility and to allow people to inspect the downloaded audio files locally are a bit more important here. \r\n\r\nBut maybe, we could add a flag, `save_files_as_bytes` or `make_independent`, `make_self_contained` or a better name to `save_to_disk(...)` and `push_to_hub(...)` that would allow to make the resulting folder completely independent. ", "What do you think @mariosasko @lhoestq @polinaeterna @anton-l ?\r\n", "For context: you can either store the path to local images or audio files, or the bytes of those files.\r\n\r\nIf your images and audio files are local files, then the arrow file from `save_to_disk` will store paths to these files.\r\nIf you want to include the bytes or your images or audio files instead, you must `read()` those files first.\r\nThis can be done by storing the \"bytes\" instead of the \"path\" of the images or audio files.\r\n\r\nOn the other hand, the resulting Parquet files from `push_to_hub` are self-contained, so that anyone can reload the dataset from the Hub. If your dataset contains image or audio data, the Parquet files will store the bytes of your images or audio files.\r\n\r\nFor now I just updated the documentation: https://github.com/huggingface/datasets/pull/4193. Maybe we can also embed the image and audio bytes in `save_to_disk` when we implement sharding, so that is can be done as efficiently as `push_to_hub`.\r\n\r\nAnyway, merging this one :)" ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "(@patrickvonplaten )", "Also cc @lhoestq here", "The documentation in the code is definitely outdated - thanks for letting me know, I'll remove it in https://github.com/huggingface/datasets/pull/4184 .\r\n\r\nYou're exactly right `audio` `array` already decodes the audio file to the correct waveform. This is done on the fly, which is also why one should **not** do `ds[\"audio\"][\"array\"][0]` as this will decode all dataset samples, but instead `ds[0][\"audio\"][\"array\"]` see: https://huggingface.co/docs/datasets/audio_process#audio-datasets\r\n\r\n", "So, again to clarify: On disk, only the raw flac file content is stored? Is this also the case after `save_to_disk`?\r\n\r\nAnd is it simple to also store it re-encoded as ogg or mp3 instead?\r\n", "Hey, \r\n\r\nSorry yeah I was just about to look into this! We actually had an outdated version of Librispeech ASR that didn't save any files, but instead converted the audio files to a byte string, then was then decoded on-the-fly. This however is not very user-friendly so we recently decided to instead show the full path of the audio files with the `path` parameter.\r\n\r\nI'm currently changing this for Librispeech here: https://github.com/huggingface/datasets/pull/4184 .\r\nYou should be able to see the audio file in the original `flac` format under `path` then. I don't think it's a good idea to convert to MP3 out-of-the-box, but we could maybe think about some kind of convert function for audio datasets cc @lhoestq ? ", "> I don't think it's a good idea to convert to MP3 out-of-the-box, but we could maybe think about some kind of convert function for audio datasets cc @lhoestq ?\r\n\r\nSure, I would expect that `load_dataset(\"librispeech_asr\")` would give you the original (not re-encoded) data (flac or already decoded). So such re-encoding logic would be some separate generic function. So I could do sth like `dataset.reencode_as_ogg(**ogg_encode_opts).save_to_disk(...)` or so.\r\n", "A follow-up question: I wonder whether a Parquet dataset is maybe more what we actually want to have? (Following also my comment here: https://github.com/huggingface/datasets/pull/4184#issuecomment-1105045491.) Because I think we actually would prefer to embed the data content in the dataset.\r\n\r\nSo, instead of `save_to_disk`/`load_from_disk`, we would use `to_parquet`,`from_parquet`? Is there any downside? Are arrow files more efficient?\r\n\r\nRelated is also the doc update in #4193.\r\n", "`save_to_disk` saves the dataset as an Arrow file, which is the format we use to load a dataset using memory mapping. This way the dataset does not fill your RAM, but is read from your disk instead.\r\n\r\nTherefore you can directly reload a dataset saved with `save_to_disk` using `load_from_disk`.\r\n\r\nParquet files are used for cold storage: to use memory mapping on a Parquet dataset, you first have to convert it to Arrow. We use Parquet to reduce the I/O when pushing/downloading data from the Hugging face Hub. When you load a Parquet file from the Hub, it is converted to Arrow on the fly during the download." ]

[]

[]

[ "Fix https://github.com/huggingface/datasets/issues/4179", "_The documentation is not available anymore as the PR was closed or merged._", "Just that I understand: With this change, simply doing `load_dataset(\"librispeech_asr\")` is possible and returns the whole dataset?\r\n\r\nAnd to get the subsets, I do sth like:\r\n```python\r\nds = load_dataset(\"librispeech_asr\")\r\ntrain_ds = ds[\"train\"]\r\ndev_clean_ds = ds[\"dev-clean\"]\r\ndev_other_ds = ds[\"dev-other\"]\r\ntest_clean_ds = ds[\"test-clean\"]\r\ntest_other_ds = ds[\"test-other\"]\r\n```\r\n?\r\n", "> Just that I understand: With this change, simply doing `load_dataset(\"librispeech_asr\")` is possible and returns the whole dataset?\r\n> \r\n> And to get the subsets, I do sth like:\r\n> \r\n> ```python\r\n> ds = load_dataset(\"librispeech_asr\")\r\n> train_ds = ds[\"train\"]\r\n> dev_clean_ds = ds[\"dev-clean\"]\r\n> dev_other_ds = ds[\"dev-other\"]\r\n> test_clean_ds = ds[\"test-clean\"]\r\n> test_other_ds = ds[\"test-other\"]\r\n> ```\r\n> \r\n> ?\r\n\r\nYou could do:\r\n\r\n\r\n```python\r\nds = load_dataset(\"librispeech_asr\", \"all\") # <- note that we have to pass a config\r\ntrain_ds = ds[\"train\"]\r\ndev_clean_ds = ds[\"dev-clean\"]\r\ndev_other_ds = ds[\"dev-other\"]\r\ntest_clean_ds = ds[\"test-clean\"]\r\ntest_other_ds = ds[\"test-other\"]\r\n```", "So, `load_dataset(\"librispeech_asr\")` is not possible, it must be `load_dataset(\"librispeech_asr\", \"all\")`?\r\n\r\nWhy is that?\r\n\r\nThe docs say:\r\n```\r\nname: `str` name, optional configuration for the dataset that affects the data generated on disk. Different\r\n `builder_config`s will have their own subdirectories and versions.\r\n If not provided, uses the first configuration in self.BUILDER_CONFIGS\r\n```\r\nhttps://github.com/huggingface/datasets/blob/cd3ce34ab1604118351e1978d26402de57188901/src/datasets/builder.py#L228\r\n\r\nOr maybe you could just define `DEFAULT_CONFIG_NAME`?\r\n", "> If not provided, uses the first configuration in self.BUILDER_CONFIGS\r\n\r\nOh crap this is outdated documentation. No it doesn't take the first config by default.\r\n\r\nEDIT: opened a PR to fix this: https://github.com/huggingface/datasets/pull/4186", "> No it doesn't take the first config by default.\r\n\r\nBut defining `DEFAULT_CONFIG_NAME` would work?\r\n\r\nSo should we define `DEFAULT_CONFIG_NAME = \"all\"` here as well? I think this is a reasonable default config.\r\n\r\nDon't most datasets have some default config?\r\n", "> But defining DEFAULT_CONFIG_NAME would work?\r\n>\r\n> So should we define DEFAULT_CONFIG_NAME = \"all\" here as well? I think this is a reasonable default config.\r\n\r\nYes that would work, and I also find it reasonable to do it :)\r\n\r\n> Don't most datasets have some default config?\r\n\r\nMost datasets only have one configuration, so the single configuration is the default one. Then other datasets gave several configurations, and whether they have a default one is decided case-by-case.\r\n\r\ne.g. `glue` is a benchmark and doesn't have a default task, one must choose which task of `glue` they want to use explicitely.", "Thanks a lot for the feedback! \r\n\r\nUsing `\"all\"` now as the default config. I changed the layout a bit so that there is not a single \"train\", but instead we have multiple \"train.clean.100\", \"train.clean.360\", \"train.other.500\". This way we don't even need to do filtering and it's also cleaner IMO.\r\n\r\n@albertz - you should now be able to do the following:\r\n\r\n```python\r\nload_dataset(\"librispeech_asr\") # <- run this once to download, prepare dataset and cache everything\r\n\r\n# The following operations will be very fast since all the downloading and processing is already cached\r\ntrain_1 = load_dataset(\"librispeech_asr\", split=\"train.clean.100\")\r\nprint(train_1)\r\ntrain_2 = load_dataset(\"librispeech_asr\", split=\"train.clean.100+train.clean.360\")\r\nprint(train_2)\r\ntrain_full = load_dataset(\"librispeech_asr\", split=\"train.clean.100+train.clean.360+train.other.500\")\r\nprint(train_full)\r\ndev_clean_ds = load_dataset(\"librispeech_asr\", split=\"validation.clean\")\r\nprint(dev_clean_ds)\r\ndev_other_ds = load_dataset(\"librispeech_asr\", split=\"validation.other\")\r\nprint(dev_other_ds)\r\ntest_clean_ds = load_dataset(\"librispeech_asr\", split=\"test.clean\")\r\nprint(test_clean_ds)\r\ntest_other_ds = load_dataset(\"librispeech_asr\", split=\"test.other\")\r\nprint(test_other_ds)\r\n```\r\n\r\n\r\n", "Think this way we have the best of both worlds. Also @lhoestq, I think we could highlight better in the docs that it's possible to combine different splits. We do this actually quite a lot for speech. For Common Voice many people include \"validation\" in the training if the data is too small, e.g.: https://github.com/huggingface/transformers/blob/ff06b177917384137af2d9585697d2d76c40cdfc/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L147\r\n\r\nShould we maybe add a short section to the loading tutorial here: https://huggingface.co/docs/datasets/v2.1.0/en/loading#hugging-face-hub ? (Happy to do it)", "Is there any advantage or difference in calling `load_dataset` multiple times for each split? Or why not just call `load_dataset` once and then access each split?\r\n\r\nNote in our case, we cannot really use the caching mechanism because we have a recipe pipeline used by multiple users (and I think a common cache dir for all users might end up in problems) and we basically would use `load_dataset(\"librispeech_asr\").save_to_disk(...)` and then later `load_from_disk(...)`. (See here: https://github.com/rwth-i6/i6_core/pull/253)\r\n\r\nSo with `load_from_disk`, we cannot really provide the split this way, so we anyway would do sth like:\r\n```python\r\nds = datasets.load_from_disk(...)\r\ntrain = ds[\"train\"]\r\n```\r\nOr with your latest proposal, it would look like:\r\n```python\r\nds = datasets.load_from_disk(...)\r\ntrain_ds = datasets.concatenate_datasets(\r\n [ds[\"train.clean.100\"], ds[\"train.clean.360\"], ds[\"train.other.500\"]])\r\n```\r\nright?\r\n", "> Is there any advantage or difference in calling `load_dataset` multiple times for each split? Or why not just call `load_dataset` once and then access each split?\r\n> \r\n> Note in our case, we cannot really use the caching mechanism because we have a recipe pipeline used by multiple users (and I think a common cache dir for all users might end up in problems) and we basically would use `load_dataset(\"librispeech_asr\").save_to_disk(...)` and then later `load_from_disk(...)`. (See here: [rwth-i6/i6_core#253](https://github.com/rwth-i6/i6_core/pull/253))\r\n> \r\n> So with `load_from_disk`, we cannot really provide the split this way, so we anyway would do sth like:\r\n> \r\n> ```python\r\n> ds = datasets.load_from_disk(...)\r\n> train = ds[\"train\"]\r\n> ```\r\n> \r\n> Or with your latest proposal, it would look like:\r\n> \r\n> ```python\r\n> ds = datasets.load_from_disk(...)\r\n> train_ds = datasets.concatenate_datasets(\r\n> [ds[\"train.clean.100\"], ds[\"train.clean.360\"], ds[\"train.other.500\"]])\r\n> ```\r\n> \r\n> right?\r\n\r\nI see the use case! The only advantage by calling `datasets` multiple times is that one can easily \"merge\" splits with `\"+\"`, but yeah you can do the exact same with `concatenate`.\r\n\r\n@lhoestq what do you think is the best approach with `load_from_disk`? \r\n\r\n@albertz, you could also define the `cache_dir` when doing `load_dataset(...)` which will then put all the relevant `arrow` files int the cache dir that you defined, e.g.:\r\n\r\n```python\r\nload_dataset(\"librispeech_asr\", cache_dir=\"/easy/to/access/directory\")\r\n```", "@albertz, I took a read through https://github.com/rwth-i6/i6_core/pull/253 . \r\n\r\nI think the best would be the following:\r\n\r\n1. Do `ds = load_dataset(..., cache_dir=\"/dir/that/is/easy/to/access\")` <- having merged this PR, this will save all the original `.flac` files in the `cache_dir`\r\n2. Do `ds.save_to_disk(\"local/path\")` this should then only save the `arrow.format` with a `path` string to the audio files which are located in `cache_dir` <- this won't require a lot of memory after https://github.com/huggingface/datasets/pull/4184#discussion_r854132740 is fixed and can be done for each person individually.\r\n3. `ds = datasets.load_from_disk(\"local/path\")` can the be used. An object of `ds` will then have a `path` variable that links to the original audio files in the `cache_dir`. You can change these audio files then easily to `.mp3. You could do this with the `.map(...)` function, e.g. define a function that maps through all audio files, load them and then save them on disk afterward.", "@lhoestq - I think this one is good to go", "> @albertz, I took a read through [rwth-i6/i6_core#253](https://github.com/rwth-i6/i6_core/pull/253) .\r\n> \r\n> I think the best would be the following:\r\n> \r\n> 1. Do `ds = load_dataset(..., cache_dir=\"/dir/that/is/easy/to/access\")` <- having merged this PR, this will save all the original `.flac` files in the `cache_dir`\r\n> 2. Do `ds.save_to_disk(\"local/path\")` this should then only save the `arrow.format` with a `path` string to the audio files which are located in `cache_dir` <- this won't require a lot of memory after [[Librispeech] Add 'all' config #4184 (comment)](https://github.com/huggingface/datasets/pull/4184#discussion_r854132740) is fixed and can be done for each person individually.\r\n> 3. `ds = datasets.load_from_disk(\"local/path\")` can the be used. An object of `ds` will then have a `path` variable that links to the original audio files in the `cache_dir`. You can change these audio files then easily to `.mp3. You could do this with the `.map(...)` function, e.g. define a function that maps through all audio files, load them and then save them on disk afterward.\r\n\r\nOh, so you say that our current implementation in https://github.com/rwth-i6/i6_core/pull/253 is broken? Because our cache dir is just some temp directory which will be removed afterwards, and we just store what we get out of `save_to_disk`. I think it would be good to clarify that in the doc of `save_to_disk`, that this is not enough and can depend on files from the cache dir. (@dthulke)\r\n\r\nSo, you say we anyway need to share the cache dir among users? But we would want to make sure that after the initial download and preparation of the data, this is set to readonly, because we want to make sure that other people will not modify the data in any way. Right?\r\n\r\nBut then, we don't really need the `save_to_disk` and `load_from_disk` at all, right?\r\n", "@albertz \r\n\r\n> Oh, so you say that our current implementation in https://github.com/rwth-i6/i6_core/pull/253 is broken? Because our cache dir is just some temp directory which will be removed afterwards, and we just store what we get out of save_to_disk. I think it would be good to clarify that in the doc of save_to_disk, that this is not enough and can depend on files from the cache dir. (@dthulke)\r\n\r\nOh, I wasn't aware that audio files are handled this way. Then we should have the cache directory as an additional job output, so that we keep the audio files. \r\n\r\n> So, you say we anyway need to share the cache dir among users?\r\n\r\nNo, the cache dir can still be a directory in the job output folder. Then the audio paths in the corresponding dataset column correspond to the flac files in that directory. This way the \"output\" of the job is contained into the job directory and we don't write files to a global cache directory that is independent of the sisyphus graph.\r\n\r\nIf we want to share the audio data between different users, we can just link to a central instance of the job (similar to how we do it with the `DownloadLibriSpeechCorpusJob`).", "@dthulke - that's a good point actually! So you can do both things:\r\n\r\n1. Convert all audio files to bytes. Bytes can be saved by `arrow` so in this case you can do `save_to_disk(...)`, but then you cannot really inspect the audio files locally as they'll just be saved within a large arrow file (this actually used to be the default case but we're changing this now). The problem of this is summarized here a bit: https://github.com/huggingface/datasets/issues/3663 . You can still do this if you'd like, e.g. you could do:\r\n\r\n```python\r\nds = load_dataset(\"librispeech_asr\")\r\n\r\ndef read_file(batch):\r\n with open(batch[\"file\"], \"r\") as f:\r\n batch[\"bytes\"] = f.read() \r\n return batch\r\n\r\nds = ds.map(read_file)\r\nds.save_to_disk(\"/path\") <- the saved arrow object will now contain everything you need\r\n```\r\n\r\nhowever this is not recommend - it's should be much easier to just save the path to the downloaded audio files.\r\n\r\n2. Not convert audio files to bytes, but just leave them in their original file format. Then only the path to the original files will be save in arrow. This will be the default case. This means that when you do `load_dataset(...)` both the orginal audio data and the arrow file will be saved in the `cache_dir` (which can be saved locally for every user or in a shared cache - we actually use a shared cache quite a bit at Hugging Face). When do you do `save_to_disk(...)` now only the `path` will be saved in `arrow` format (after this PR is merged, you'll see that the `arrow files should be very light weight` meaning that `save_to_disk(...)` can be done for every user, but has a dependency on the `cache_dir` (because the audio files live there).\r\n\r\n=> Now what you could do as well would be to simply move all the audio files to the folder you want (the `save_to_disk(...)` folder) and then change the path of every sample to this folder (maybe with `map(...)`) and then this folder would be self contained. I do however think it's better to just specific a `cache_dir` and re-use `load_dataset(...)` every time instead of `load_from_disk` or `save_to_disk(...)`. Note that you can even pass the relevant cache files to `load_dataset(...)` here: https://huggingface.co/docs/datasets/v2.1.0/en/package_reference/loading_methods#datasets.load_dataset.data_files in which case you can be 100% sure that nothing is redownloaded. \r\n\r\nWe discussed storing audio files quite a bit, e.g. see: https://github.com/huggingface/datasets/issues/3663 and had (too many) changes around this topic recently, but we've come to the conclusion that the best is to leave the audio format in the format it was originally (`.flac` for Librispeech) so that the user can easily inspect it / understand the data. Arrow cannot save data is `.flac` so we'll just save a path to the original data. Curious to hear your guys opinion on this as well.", "So what I would suggest here is to do the following:\r\n\r\n1. Do `load_dataset(..., cache_dir=/a/read-only/folder)`\r\n2. \r\n- Either just re-use `load_dataset(..., cache_dir=...)` which should always re-use the data in the `cache_dir` since the hash of the url matches - so there should never be any duplicated downloading \r\n\r\nor \r\n\r\n- If you want to store the files in MP3 locally, first convert the files to MP3 in the read-only folder, then take do `ds.save_to_disk(/some/path)` which will save the correct path to the read-only folder to MP3 and then you can easily re-use the small arrow dataset that is saved in `/some/path`", "> So what I would suggest here is to do the following:\r\n> \r\n> 1. Do `load_dataset(..., cache_dir=/a/read-only/folder)`\r\n> \r\n> * Either just re-use `load_dataset(..., cache_dir=...)` which should always re-use the data in the `cache_dir` since the hash of the url matches - so there should never be any duplicated downloading\r\n> \r\n> or\r\n> \r\n> * If you want to store the files in MP3 locally, first convert the files to MP3 in the read-only folder, then take do `ds.save_to_disk(/some/path)` which will save the correct path to the read-only folder to MP3 and then you can easily re-use the small arrow dataset that is saved in `/some/path`\r\n\r\nAlso relevant here: https://github.com/huggingface/datasets/issues/3663", "I also added some documentation about how `save_to_disk` handles audio files here: https://github.com/huggingface/datasets/pull/4193", "> > So, you say we anyway need to share the cache dir among users?\r\n> \r\n> No, the cache dir can still be a directory in the job output folder.\r\n\r\n@dthulke But this is what I mean. When we share the job output folder, it means we share the cache dir among users.\r\n\r\nI wonder if `load_dataset(..., cache_dir=job_output_cache_dir)` is always save to do then, that it really would not modify the `job_output_cache_dir`.\r\n\r\nWe could enforce that by making the `job_output_cache_dir` read-only afterwards. We currently don't do this.\r\n\r\n@patrickvonplaten @dthulke But in any case, we actually prefer the data content to be inside the dataset (the arrow files). Lots of small files would be very problematic for our cache manager. We have one main copy of the data on NFS, but accessing the NFS directly by all computing nodes is not feasible, so the cache manager will have copies of the files on the nodes. So it means, whenever we access some file, we query the cache manager DB whether the file is already cached somewhere (some other computing node) and if so, it copies it from the other computing node and not from NFS. This works very well when there are not too many files (but the files can be big). So, we want to have only a few but big files. Even for NFS access this is much better.\r\n\r\nI also commented in #3663.\r\n", "Hey @albertz @dthulke,\r\n\r\nThanks a lot for your input! \r\n\r\nWe've discussed quite a bit with @lhoestq and we think the best approach is the following:\r\n\r\n\r\na)\r\n`load_dataset(...)` will not store both bytes and the files because this would mean that 3x the size of the dataset would often be needed (1. the compressed `tar.gz` file, 2. the extracted file b, 3. the raw bytes in arrow format). \r\n\r\nFor canonical datasets like librispeech and common voice I think we want to keep the dataset filenames because of i) no breaking changes and ii) reasons explained in #3663\r\n\r\nHowever it's also trivial to write your own datasetset downloading script of librispeech and just not extract the folder e.g. this line: https://huggingface.co/datasets/common_voice/blob/main/common_voice.py#L671\r\n\r\nAnd then it'll be allowed to save the bytes and the dataset will be self-contained out-of-the-box when using `load_dataset(...)`\r\n\r\nb) Now, one major problem that you guys uncovered is that `save_to_disk(...)` is currently not necessarily saving a dataset to be self-contained. We will change that asap. This means that after we've corrected this when you do download the canonical librispeech dataset the following will work:\r\n\r\n```python\r\nds = load_dataset(\"....\") # <- here we have a dependency on the filepathes\r\nds[0][\"audio\"][\"bytes\"] # <- will not work\r\n\r\nds.save_to_disk(\"/local/path\") # <- now we want to have a self-contained dataset in arrow format, so we load the files into bytes and save it in arrow format\r\n\r\n# now you can delete everything besides \"/local/path\"\r\n\r\nds = load_from_disk(\"/local/path\") # <- this will work\r\n```\r\n\r\nSo either option a) where you define your own librispeech data downloading script (you guys could just sign up here: https://huggingface.co/join) and upload a dataset loading script in private mode so that no one can see it and you would always store the audio as bytes or b) where you first load then save to disk then delete cache would work. \r\n\r\nHope that fits in your vision :-)\r\n\r\ncc @lhoestq @mariosasko ", "@patrickvonplaten sounds like a good approach to me. For b) this could even be configurable with a parameter like `embed_external_files` as you have for `push_to_hub` (if people prefer to keep separate audio files).\r\n", "> However it's also trivial to write your own datasetset downloading script of librispeech and just not extract the folder\r\n\r\nI don't exactly understand. In all cases, we need to extract it to prepare the dataset, or not? No matter if we want to store the raw bytes inside the dataset or leaving them as local files. Just in the first case, we can safely delete the extracted files after the dataset preparation.\r\n\r\n> `save_to_disk(...)` is currently not necessarily saving a dataset to be self-contained. We will change that asap.\r\n\r\nFor us, this sounds exactly like what we want.\r\n\r\nBut regarding not introducing breaking changes, wouldn't this maybe also break some setups for users who don't expect this new behavior?\r\n", "@albertz I would suggest to move the discussion on implementation details on our side to the following issue: rwth-i6/i6_core/issues/257", "I like the idea of adding `embed_external_files` and set it to True by default to `save_to_disk`.\r\nIt's indeed a kind of breaking change since some users will get bigger Arrow files when updating the lib, but the advantages are nice:\r\n1. I like the idea of having it self contained, in case you want to delete your cache\r\n2. users also upload these Arrow files to cloud storage via the `fs` parameter, and in this case they would expect to upload a self-contained dataset\r\n3. consistency with `push_to_hub`\r\n\r\nIf it sounds good to you I'll open an issue to discuss this and track the advancements" ]

[]

[]

[ "I think the main purpose of #4179 was how to be able to load both configs into one, so should we maybe add this part of the code: https://github.com/huggingface/datasets/issues/4179#issuecomment-1102383717 \r\n\r\nto the doc? \r\n\r\nActually @lhoestq would this work given that they have different split names: https://huggingface.co/datasets/librispeech_asr#data-splits ? ", "This doc extension does not explain why I can't simply load the whole dataset. Or what workaround I need to get the whole dataset, which is what people usually want for Librispeech.", "_The documentation is not available anymore as the PR was closed or merged._", "@lhoestq, I can add a `\"all\"` config to Librispeech have the datasets already cached somewhere ", "I'm closing this PR then, feel free to continue the discussion in https://github.com/huggingface/datasets/issues/4179\r\n" ]

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

[]

[ "[Off topic but related: Is the uptime of S3 provably better than Zenodo's?]", "Hi @dkajtoch, please note that at HuggingFace we are not hosting this dataset: we are just using a script to download their data file and create a dataset from it.\r\n\r\nIt was the dataset owners decision to host their data at Zenodo. You can see this on their website: https://marcobaroni.org/composes/sick.html\r\n\r\nAnd yes, you are right: Zenodo is currently having some incidents and people are reporting problems from it.\r\n\r\nOn the other hand, we could contact the data owners and propose them to host their data at our Hugging Face Hub.\r\n\r\n@julien-c I guess so.\r\n", "Thanks @albertvillanova. I know that the problem lies in the source data. I just wanted to point out that these kind of problems are unavoidable without having one place where data sources are cached. Websites may go down or data sources may move. Having a copy in Hugging Face Hub would be a great solution. ", "Definitely, @dkajtoch! But we have to ask permission to the data owners. And many dataset licenses directly forbid data redistribution: in those cases we are not allowed to host their data on our Hub.", "Ahhh good point! License is the problem :(" ]

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

[]

[ "Yes, you just have to use `dl_manager.iter_archive` instead of `dl_manager.download_and_extract`.\r\n\r\nThat's because `download_and_extract` doesn't support TAR archives in streaming mode." ]

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

[]

[ "Thanks for the suggestion ! I agree it would be nice to have something directly in `datasets` to do something as simple as that\r\n\r\ncc @albertvillanova @mariosasko @polinaeterna What do you think if we have something similar to pandas `Series` that wouldn't bring everything in memory when doing `dataset[\"audio\"]` ? Currently it returns a list with all the decoded audio data in memory.\r\n\r\nIt would be a breaking change though, since `isinstance(dataset[\"audio\"], list)` wouldn't work anymore, but we could implement a `Sequence` so that `dataset[\"audio\"][0]` still works and only loads one item in memory.\r\n\r\nYour alternative suggestion with `iterate` is also sensible, though maybe less satisfactory in terms of experience IMO", "I agree that current behavior (decoding all audio file sin the dataset when accessing `dataset[\"audio\"]`) is not useful, IMHO. Indeed in our docs, we are constantly warning our collaborators not to do that.\r\n\r\nTherefore I upvote for a \"useful\" behavior of `dataset[\"audio\"]`. I don't think the breaking change is important in this case, as I guess no many people use it with its current behavior. Therefore, for me it seems reasonable to return a generator (instead of an in-memeory list) for \"special\" features, like Audio/Image.\r\n\r\n@lhoestq on the other hand I don't understand your proposal about Pandas-like... ", "I recall I had the same idea while working on the `Image` feature, so I agree implementing something similar to `pd.Series` that lazily brings elements in memory would be beneficial.", "@lhoestq @mariosasko Could you please give a link to that new feature of `pandas.Series`? As far as I remember since I worked with pandas for more than 6 years, there was no lazy in-memory feature; it was everything in-memory; that was the reason why other frameworks were created, like Vaex or Dask, e.g. ", "Yea pandas doesn't do lazy loading. I was referring to pandas.Series to say that they have a dedicated class to represent a column ;)" ]

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

[]

[ "@patrickvonplaten Hi! I saw that you prepared this? :)", "Another thing, but maybe this should be a separate issue: As I see from the code, it would try to use up to 16 simultaneous downloads? This is problematic for Librispeech or anything on OpenSLR. On [the homepage](https://www.openslr.org/), it says:\r\n\r\n> If you want to download things from this site, please download them one at a time, and please don't use any fancy software-- just download things from your browser or use 'wget'. We have a firewall rule to drop connections from hosts with more than 5 simultaneous connections, and certain types of download software may activate this rule.\r\n\r\nRelated: https://github.com/tensorflow/datasets/issues/3885", "Hey @albertz,\r\n\r\nNice to see you here! It's been a while ;-) ", "Sorry maybe the docs haven't been super clear here. By `split` we mean one of `train.500`, `train.360`, `train.100`, `validation`, `test`. For Librispeech, you'll have to specific a config (either `other` or `clean`) though:\r\n\r\n```py\r\ndatasets.load_dataset(\"librispeech_asr\", \"clean\")\r\n```\r\n\r\nshould work and give you all splits (being \"train\", \"test\", ...) for the clean config of the dataset.\r\n", "If you need both `\"clean\"` and `\"other\"` I think you'll have to do concatenate them as follows: \r\n\r\n```py\r\nfrom datasets import concatenate_datasets, load_dataset\r\n\r\nother = load_dataset(\"librispeech_asr\", \"other\")\r\nclean = load_dataset(\"librispeech_asr\", \"clean\")\r\n\r\nlibrispeech = concatenate_datasets([other, clean])\r\n```\r\n\r\nSee https://huggingface.co/docs/datasets/v2.1.0/en/process#concatenate", "Downloading one split would be:\r\n\r\n```py\r\nfrom datasets import load_dataset\r\n\r\nother = load_dataset(\"librispeech_asr\", \"other\", split=\"train.500\")\r\n```\r\n\r\n\r\n", "cc @lhoestq FYI maybe the docs can be improved here", "Ah thanks. But wouldn't it be easier/nicer (and more canonical) to just make it in a way that simply `load_dataset(\"librispeech_asr\")` works?", "Pinging @lhoestq here, think this could make sense! Not sure however how the dictionary would then look like", "Would it make sense to have `clean` as the default config ?\r\n\r\nAlso I think `load_dataset(\"librispeech_asr\")` should have raised you an error that says that you need to specify a config\r\n\r\nI also opened a PR to improve the doc: https://github.com/huggingface/datasets/pull/4183", "> Would it make sense to have `clean` as the default config ?\r\n\r\nI think a user would expect that the default would give you the full dataset.\r\n\r\n> Also I think `load_dataset(\"librispeech_asr\")` should have raised you an error that says that you need to specify a config\r\n\r\nIt does raise an error, but this error confused me because I did not understand why I needed a config, or why I could not simply download the whole dataset, which is what people usually do with Librispeech.\r\n", "+1 for @albertz. Also think lots of people download the whole dataset (`\"clean\"` + `\"other\"`) for Librispeech.\r\n\r\nThink there are also some people though who:\r\n- a) Don't have the memory to store the whole dataset\r\n- b) Just want to evaluate on one of the two configs", "Ok ! Adding the \"all\" configuration would do the job then, thanks ! In the \"all\" configuration we can merge all the train.xxx splits into one \"train\" split, or keep them separate depending on what's the most practical to use (probably put everything in \"train\" no ?)", "I'm not too familiar with how to work with HuggingFace datasets, but people often do some curriculum learning scheme, where they start with train.100, later go over to train.100 + train.360, and then later use the whole train (960h). It would be good if this is easily possible.\r\n", "Hey @albertz, \r\n\r\nopened a PR here. Think by adding the \"subdataset\" class to each split \"train\", \"dev\", \"other\" as shown here: https://github.com/huggingface/datasets/pull/4184/files#r853272727 it should be easily possible (e.g. with the filter function https://huggingface.co/docs/datasets/v2.1.0/en/package_reference/main_classes#datasets.Dataset.filter )", "But also since everything is cached one could also just do:\r\n\r\n```python\r\nload_dataset(\"librispeech\", \"clean\", \"train.100\")\r\nload_dataset(\"librispeech\", \"clean\", \"train.100+train.360\")\r\nload_dataset(\"librispeech\" \"all\", \"train\") \r\n```" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4178). All of your documentation changes will be reflected on that endpoint.", "Thanks for the comments. I believe I have addressed all of them and also decreased the size of the dummy data file, so it should be ready for a re-review. I also made a change to allow adding synset mapping and valprep script in config in case we add ImageNet 21k some time later. ", "@lhoestq I have updated the PR to address all of the review comments." ]

[]

[]

[]

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

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4175). All of your documentation changes will be reflected on that endpoint.", "Hi! Coming in late with some context.\r\n\r\nThere are two versions of the WIT dataset:\r\n1. The original source dataset managed by Wikimedia. It has more information, raw image representations, and each row corresponds to an image linked to all of its captions wherever it happens in Wikipedia (in multiple languages)\r\n2. The Google version, corresponding to the data script in this PR, which duplicates image instances and requires the user to download the images themselves from the provided URL (note that a basic implementation will have them download the same picture several time. @thomasw21 using our download manager instead of `urllib` could help with that, but it wouldn't be required if people had access to the first version)\r\n\r\nThe Wikimedia folks were really interested in us hosting a ready-to-go streaming version of this dataset where users don't have to download the version themselves, which is why we have the pre-processed versions on an HF bucket, with the raw images and a pre-computed embedding (don't remember the model, we can keep it ). That's the data script currently in https://github.com/huggingface/datasets/pull/2981 . It's nearly ready to go, the one thing we should do is move the raw data from our HF google Cloud bucket to the Hub.\r\n\r\nHow do you want to move forward? IMO the best way would be to have a WIT dataset under the Wikimedia org with both configurations, but it depends on everyone's timelines", "Okay after offline discussion. We'll improve this versions and push it to the hub under `google` namespace. \r\n\r\n> which duplicates image instances and requires the user to download the images themselves from the provided URL (note that a basic implementation will have them download the same picture several time. @thomasw21 using our download manager instead of urllib could help with that, but it wouldn't be required if people had access to the first version)\r\n\r\nAh interesting wasn't aware of this duplication issue, concretely it'll just mean that our dataset in bigger than expected ... I think this should be handled after this loading script (though I have to figure our how to spawn a dl_manager).\r\n\r\n> The Wikimedia folks were really interested in us hosting a ready-to-go streaming version of this dataset where users don't have to download the version themselves, which is why we have the pre-processed versions on an HF bucket, with the raw images and a pre-computed embedding (don't remember the model, we can keep it ). That's the data script currently in https://github.com/huggingface/datasets/pull/2981 . It's nearly ready to go, the one thing we should do is move the raw data from our HF google Cloud bucket to the Hub.\r\n\r\nSimilarly a script will be written and pushed to `wikimedia` organisation." ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4173). All of your documentation changes will be reflected on that endpoint.", "oops looks like some tests are failing sorry, will fix them tomorrow\r\n\r\nEDIT: not today but asap hopefully" ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4170). All of your documentation changes will be reflected on that endpoint.", "[Magic is now banned](https://www.youtube.com/watch?v=WIn58XoY728#t=36s) by decree of @sgugger. This is honestly much cleaner, and the functionality will make much more sense in `transformers` anyway!", "@gante I renamed the default collator to `minimal_tf_collate_fn`!" ]

[]

[]

[ "Thanks for reporting. The bug has already been detected, and we hope to fix it soon.", "TIMIT is now a dataset that requires manual download, see #4145 \r\n\r\nTherefore it might take a bit more time to fix it", "> TIMIT is now a dataset that requires manual download, see #4145\r\n> \r\n> Therefore it might take a bit more time to fix it\r\n\r\nThank you for your quickly response. Exactly, I also found the manual download issue in the morning. But when I used *list_datasets()* to check the available datasets, *'timit_asr'* is still in the list. So I am a little bit confused. If *'timit_asr'* need to be manually downloaded, does that mean we can **not** automatically download it **any more** in the future?", "Yes exactly. If you try to load the dataset it will ask you to download it manually first, and to pass the downloaded and extracted data like `load_dataset(\"timir_asr\", data_dir=\"path/to/extracted/data\")`\r\n\r\nThe URL we were using was coming from a host that doesn't have the permission to redistribute the data, and the dataset owners (LDC) notified us about it." ]

[ { "id": 1935892861, "node_id": "MDU6TGFiZWwxOTM1ODkyODYx", "url": "https://api.github.com/repos/huggingface/datasets/labels/documentation", "name": "documentation", "color": "0075ca", "default": true, "description": "Improvements or additions to documentation" } ]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4168). All of your documentation changes will be reflected on that endpoint.", "> Do you think it is clearer to make every code example fully reproducible so when users copy the code they can actually run it and get an output? This seems quite repetitive - maybe even unnecessary - but it is definitely clearer.\r\n\r\nI think it's ok to be repetitive to get more clarity. Many users come from `transformers` and may have little experience with some processing methods (especially torch users).\r\n\r\n> Should we showcase a function with more than one parameter to highlight different use-cases (it's pretty basic right now, but I'd be happy to add more)?\r\n\r\nMaybe let's do it case by case, depending on whether there are parameters that are likely to be used often ?\r\n\r\n> For the class_encode_column function, let me know if there is a simpler dataset with fewer columns (currently using winograd_wsc) so it is easier for users to see what changed.\r\n\r\nYou can try with `boolq`, it has a boolean column that can be converted to labels\r\n\r\n> Where possible, I try to show the input before and the output after using a function like flatten for example. Do you think this is too much and just showing the usage (ie, >>> ds.flatten()) will be sufficient?\r\n\r\nNo I don't think it's too much, it's nice this way thanks :)", "Updated each code example so they are fully reproducible (where applicable)! The next step will be to identify some functions where we can show off some parameters that are useful or commonly used. Some useful parameters can be:\r\n\r\n- use `map(batched=True)` to process batches of examples.\r\n- set a seed in `shuffle`.\r\n- set `shuffle` and `seed` in `train_test_split`.\r\n\r\nLet me know if you think of anything else related to the functions in `arrow_dataset.py`!", "Cool thanks ! I think you can also do `num_proc` for `map`" ]

[]

[]

[ "I also set GIT_LFS_SKIP_SMUDGE=1 to speed up git clones", "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4166). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_4165). All of your documentation changes will be reflected on that endpoint." ]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

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

[]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "Looks like your dummy_data.zip file is not in the right location ;)\r\ndatasets/datasets/conceptual_12m/dummy/default/0.0.0/dummy_data.zip\r\n->\r\ndatasets/conceptual_12m/dummy/default/0.0.0/dummy_data.zip" ]