Datasets:

---

amannagrawall002

/

github-issues

like 0

[]

[ "#self-assign", "Fixed by:\r\n- #4809" ]

[]

[ "Hi! Only the name of a particular split (\"train\", \"test\", ...) is supported as a split pattern if `streaming=True`. We plan to address this limitation soon.", "Hi, have you addressed this yet?", "yes, same error occurs.\r\n```python\r\nfrom datasets import load_dataset\r\n\r\n# error\r\nrepo = \"nateraw/ade20k-tiny\"\r\ndataset = load_dataset(repo, split=\"train+validation\", streaming=True)\r\n```\r\n\r\n```python\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n[<ipython-input-3-a6ae02d63899>](https://localhost:8080/#) in <cell line: 5>()\r\n 3 # error\r\n 4 repo = \"nateraw/ade20k-tiny\"\r\n----> 5 dataset = load_dataset(repo, split=\"train+validation\", streaming=True)\r\n\r\n1 frames\r\n[/usr/local/lib/python3.10/dist-packages/datasets/builder.py](https://localhost:8080/#) in as_streaming_dataset(self, split, base_path)\r\n 1265 splits_generator = splits_generators[split]\r\n 1266 else:\r\n-> 1267 raise ValueError(f\"Bad split: {split}. Available splits: {list(splits_generators)}\")\r\n 1268 \r\n 1269 # Create a dataset for each of the given splits\r\n\r\nValueError: Bad split: train+validation. Available splits: ['train', 'validation']\r\n```\r\n\r\ngoogle colab, `datasets==2.12.0`\r\n```\r\n- huggingface_hub version: 0.14.1\r\n- Platform: Linux-5.10.147+-x86_64-with-glibc2.31\r\n- Python version: 3.10.11\r\n- Running in iPython ?: No\r\n- Running in notebook ?: No\r\n- Running in Google Colab ?: No\r\n- Token path ?: /root/.cache/huggingface/token\r\n- Has saved token ?: False\r\n- Configured git credential helpers: \r\n- FastAI: 2.7.12\r\n- Tensorflow: 2.12.0\r\n- Torch: 2.0.0+cu118\r\n- Jinja2: 3.1.2\r\n- Graphviz: 0.20.1\r\n- Pydot: 1.4.2\r\n- Pillow: 8.4.0\r\n- hf_transfer: N/A\r\n- gradio: N/A\r\n- ENDPOINT: https://huggingface.co/\r\n- HUGGINGFACE_HUB_CACHE: /root/.cache/huggingface/hub\r\n- HUGGINGFACE_ASSETS_CACHE: /root/.cache/huggingface/assets\r\n- HF_TOKEN_PATH: /root/.cache/huggingface/token\r\n- HF_HUB_OFFLINE: False\r\n- HF_HUB_DISABLE_TELEMETRY: False\r\n- HF_HUB_DISABLE_PROGRESS_BARS: None\r\n- HF_HUB_DISABLE_SYMLINKS_WARNING: False\r\n- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False\r\n- HF_HUB_DISABLE_IMPLICIT_TOKEN: False\r\n- HF_HUB_ENABLE_HF_TRANSFER: False\r\n```\r\n", "Hi!, still not fixed this, the truth is that it is an important update for what we want to train the entire dataset because we want to train fast, also should be enabled the function \"[train:18%]\" for streaming" ]

[ "Now: the preview (first-rows) works, but not the conversion to parquet. See https://huggingface.co/datasets/wikipedia/viewer/20220301.de/train\r\n\r\n```\r\n_split_generators() missing 1 required positional argument: 'pipeline'\r\n\r\nError code: UnexpectedError\r\n```" ]

[ "Hi! You can get a `torch.Tensor` if you do the following:\r\n```python\r\nraw = load_dataset(\"beans\", split=\"train\")\r\nraw = raw.select(range(100))\r\n\r\npreprocessor = AutoFeatureExtractor.from_pretrained(\"nateraw/vit-base-beans\")\r\n\r\nfrom datasets import Array3D\r\nfeatures = raw.features.copy()\r\nfeatures[\"pixel_values\"] = datasets.Array3D(shape=(3, 224, 224), dtype=\"float32\")\r\n\r\ndef preprocess_func(examples):\r\n imgs = [img.convert(\"RGB\") for img in examples[\"image\"]]\r\n return preprocessor(imgs)\r\n\r\ndata = raw.map(preprocess_func, batched=True, features=features)\r\n\r\nprint(type(data[0][\"pixel_values\"]))\r\n\r\ndata = data.with_format(\"torch\", columns=[\"pixel_values\"])\r\n\r\nprint(type(data[0][\"pixel_values\"]))\r\n```\r\n\r\nThe reason for this \"inconsistency\" in the default case is the way PyArrow infers the type of multi-dim arrays (in this case, the `pixel_values` column). If the type is not specified manually, PyArrow assumes it is a dynamic-length sequence (it needs to know the type before writing the first batch to a cache file, and it can't be sure the array is fixed ahead of time; `ArrayXD` is our way of telling that the dims are fixed), so it already fails to convert the corresponding array to NumPy properly (you get an array of `np.object` arrays). And `with_format(\"torch\")` replaces NumPy arrays with Torch tensors, so this bad formatting propagates." ]

[ "Hi! We've just started discussing the video support in `datasets` (decoding backends, video feature type, etc.), so I believe we should have something tangible by the end of this year.\r\n\r\nAlso, if you have additional video features in mind that you would like to see, feel free to let us know", "Coool thanks @mariosasko ", "Hey @mariosasko, I was wondering if there's a way to load video data currently in the library? \r\nAlternatively is there a way I could hack it through the dataset.from_dict() method? I tried to hack it, but the issue I run into is that earlier I was doing a `cast_column()` call for the `Image` feature, but now I'm not sure about to do if I want the dataset to have the following keys when I call from_dict on it:\r\n`{\"caption\":[list of text captions], \"video_frames\": [list of image lists with one image list corresponding to one video]}`\r\n\r\nMaybe something like `cast_column(\"video_frames\", List(Image))` ..\r\n(This is assuming I have already extracted frames from video)" ]

[ "@mariosasko I'm getting a similar issue when creating a Dataset from a Pandas dataframe, like so:\r\n\r\n```\r\nfrom datasets import Dataset, Features, Image, Value\r\nimport pandas as pd\r\nimport requests\r\nimport PIL\r\n\r\n# we need to define the features ourselves\r\nfeatures = Features({\r\n 'a': Value(dtype='int32'),\r\n 'b': Image(),\r\n})\r\n\r\nurl = \"http://images.cocodataset.org/val2017/000000039769.jpg\"\r\nimage = PIL.Image.open(requests.get(url, stream=True).raw)\r\n\r\ndf = pd.DataFrame({\"a\": [1, 2], \r\n \"b\": [image, image]})\r\n\r\ndataset = Dataset.from_pandas(df, features=features) \r\n```\r\nresults in \r\n\r\n```\r\nArrowInvalid: ('Could not convert <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=640x480 at 0x7F7991A15C10> with type JpegImageFile: did not recognize Python value type when inferring an Arrow data type', 'Conversion failed for column b with type object')\r\n```\r\n\r\nWill the PR linked above also fix that?", "I would expect this to work, but it doesn't. Shouldn't be too hard to fix tho (in a subsequent PR).", "Hi @mariosasko just wanted to check in if there is a PR to follow for this. I was looking to create a demo app using this. If it's not working I can just use byte encoded images in the dataset which are not displayed. ", "Hi @darraghdog! No PR yet, but I plan to fix this before the next release.", "I was just pointed here by @mariosasko, meanwhile I found a workaround using `encode_example` like so:\r\n\r\n```\r\nfrom datasets import load_from_disk, Dataset\r\nDATASET_PATH = \"/hf/m4-master/data/cm4/cm4-10000-v0.1\"\r\nds1 = load_from_disk(DATASET_PATH)\r\nds2 = Dataset.from_dict(mapping={k: [] for k in ds1[99].keys()},\r\n features=ds1.features\r\n)\r\nfor i in range(2):\r\n # could add several representative items here\r\n row = ds1[99]\r\n row_encoded = ds2.features.encode_example(row)\r\n ds2 = ds2.add_item(row_encoded)\r\n```", "Hmm, interesting. If I create the dataset on the fly:\r\n\r\n```\r\nfrom datasets import load_from_disk, Dataset\r\nDATASET_PATH = \"/hf/m4-master/data/cm4/cm4-10000-v0.1\"\r\nds1 = load_from_disk(DATASET_PATH)\r\nds2 = Dataset.from_dict(mapping={k: [v]*2 for k, v in ds1[99].items()},\r\n features=ds1.features)\r\n```\r\n\r\nit doesn't fail with the error in the OP, as `from_dict` performs `encode_batch`.\r\n\r\nHowever if I try to use this dataset it fails now with:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/multiprocess/pool.py\", line 125, in worker\r\n result = (True, func(*args, **kwds))\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 557, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 524, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/fingerprint.py\", line 480, in wrapper\r\n out = func(self, *args, **kwargs)\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 2775, in _map_single\r\n batch = apply_function_on_filtered_inputs(\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 2655, in apply_function_on_filtered_inputs\r\n processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 2347, in decorated\r\n result = f(decorated_item, *args, **kwargs)\r\n File \"debug_leak2.py\", line 235, in split_pack_and_pad\r\n images.append(image_transform(image.convert(\"RGB\")))\r\nAttributeError: 'dict' object has no attribute 'convert'\r\n\"\"\"\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"debug_leak2.py\", line 418, in <module>\r\n train_loader, val_loader = get_dataloaders()\r\n File \"debug_leak2.py\", line 348, in get_dataloaders\r\n dataset = dataset.map(mapper, batch_size=32, batched=True, remove_columns=dataset.column_names, num_proc=4)\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 2500, in map\r\n transformed_shards[index] = async_result.get()\r\n File \"/home/stas/anaconda3/envs/py38-pt112/lib/python3.8/site-packages/multiprocess/pool.py\", line 771, in get\r\n raise self._value\r\nAttributeError: 'dict' object has no attribute 'convert'\r\n```\r\n\r\nbut if I create that same dataset one item at a time as in the previous comment's code snippet it doesn't fail.\r\n\r\nThe features of this dataset are set to:\r\n\r\n```\r\n{'texts': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None), \r\n'images': Sequence(feature=Image(decode=True, id=None), length=-1, id=None)}\r\n```", "> @mariosasko I'm getting a similar issue when creating a Dataset from a Pandas dataframe, like so:\r\n> \r\n> ```\r\n> from datasets import Dataset, Features, Image, Value\r\n> import pandas as pd\r\n> import requests\r\n> import PIL\r\n> \r\n> # we need to define the features ourselves\r\n> features = Features({\r\n> 'a': Value(dtype='int32'),\r\n> 'b': Image(),\r\n> })\r\n> \r\n> url = \"http://images.cocodataset.org/val2017/000000039769.jpg\"\r\n> image = PIL.Image.open(requests.get(url, stream=True).raw)\r\n> \r\n> df = pd.DataFrame({\"a\": [1, 2], \r\n> \"b\": [image, image]})\r\n> \r\n> dataset = Dataset.from_pandas(df, features=features) \r\n> ```\r\n> \r\n> results in\r\n> \r\n> ```\r\n> ArrowInvalid: ('Could not convert <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=640x480 at 0x7F7991A15C10> with type JpegImageFile: did not recognize Python value type when inferring an Arrow data type', 'Conversion failed for column b with type object')\r\n> ```\r\n> \r\n> Will the PR linked above also fix that?\r\n\r\nIt looks like the problem still exists.\r\nAny news ? Any good workaround ?\r\n\r\nThank you", "There is a workaround: \r\nCreate a loader python scrypt and upload the dataset to huggingface.\r\n\r\nHere is an example how to do that:\r\n\r\nhttps://huggingface.co/datasets/jamescalam/image-text-demo/tree/main\r\n\r\nand Here are videos with explanations:\r\n\r\nhttps://www.youtube.com/watch?v=lqK4ocAKveE and https://www.youtube.com/watch?v=ODdKC30dT8c", "cc @mariosasko gentle ping for a fix :)", "Any update on this? I'm still facing this issure. Any workaround?", "I was facing the same issue. Downgrading datasets from 2.11.0 to 2.4.0 solved the issue. ", "> Any update on this? I'm still facing this issure. Any workaround?\r\n\r\nI was able to resolve my issue with a quick workaround: \r\n\r\n```\r\nfrom collections import defaultdict\r\nfrom datasets import Dataset\r\n \r\ndata = defaultdict(list)\r\nfor idx in tqdm(range( len(dataloader)),desc=\"Captioning...\"):\r\n img = dataloader[idx]\r\n data['image'].append(img)\r\n data['text'].append(f\"{img_{idx}})\r\n \r\ndataset = Dataset.from_dict(data)\r\ndataset = dataset.filter(lambda example: example['image'] is not None)\r\ndataset = dataset.filter(lambda example: example['text'] is not None)\r\n \r\ndataset.push_to_hub(path-to-repo', private=False)\r\n```\r\n\r\nHope it helps!\r\nHappy coding", "> > Any update on this? I'm still facing this issure. Any workaround?\r\n> \r\n> I was able to resolve my issue with a quick workaround:\r\n> \r\n> ```\r\n> from collections import defaultdict\r\n> from datasets import Dataset\r\n> \r\n> data = defaultdict(list)\r\n> for idx in tqdm(range( len(dataloader)),desc=\"Captioning...\"):\r\n> img = dataloader[idx]\r\n> data['image'].append(img)\r\n> data['text'].append(f\"{img_{idx}})\r\n> \r\n> dataset = Dataset.from_dict(data)\r\n> dataset = dataset.filter(lambda example: example['image'] is not None)\r\n> dataset = dataset.filter(lambda example: example['text'] is not None)\r\n> \r\n> dataset.push_to_hub(path-to-repo', private=False)\r\n> ```\r\n> \r\n> Hope it helps! Happy coding\r\n\r\nIt works!! ", "> \r\n\r\nhow did this work, how to use this script or where to paste it?", "I had a similar issue to @NielsRogge where I was unable to create a dataset from a Pandas DataFrame containing PIL.Images.\r\n\r\nI found another workaround that works in this case which involves converting the DataFrame to a python dictionary, and then creating a dataset from said python dictionary.\r\n\r\nThis is a generic example of my workaround. The example assumes that you have your data in a Pandas DataFrame variable called \"dataframe\" plus a dictionary of your data's features in a variable called \"features\".\r\n```\r\nimport datasets\r\n\r\ndictionary = dataframe.to_dict(orient='list')\r\ndataset = datasets.Dataset.from_dict(dictionary, features=features)\r\n```", "cc @mariosasko this issue has been open for 2 years, would be great to resolve it :)", "I have the same issue, my current workaround is saving the dataframe to a csv and then loading the dataset from the csv. Would also appreciate it a fix :)" ]

[ "Thanks for reporting this as well, @stadlerb.\r\n\r\nI suggest waiting for the answer of the data owners... ", "@albertvillanova The first author of the paper responded to the upstream issue:\r\n> Task IDs 11-510 are the 500 test problems. We use 90 problems (511-600) for validation and then remaining 374 for fine-tuning (601-974). The other problems can be used as desired, either for training or few-shot prompting (although this should be specified).", "Thanks for the follow-up, @stadlerb.\r\n\r\nWould you be willing to open a Pull Request to address this issue? :wink: ", "Opened a [PR](https://github.com/huggingface/datasets/pull/4943) to implement this--lmk if you have any feedback" ]

[ "Thanks for proposing, @NielsRogge.\r\n\r\nPlease, note this dataset requires registering in their website and their Terms and Conditions state we cannot distribute their URL:\r\n```\r\n1. You will NOT distribute the download URLs\r\n...\r\n```" ]

[ "Thanks for reporting. It's a known issue that should be fixed soon. Meanwhile, I had to manually trigger the dataset viewer. It's OK now.\r\nNote that the extreme aspect ratio of the images generates another issue, that we're inspecting." ]

[]

[]

[]

[ "Hi @osanseviero I would like to add this dataset.", "Hey @nandwalritik! Thanks for offering to help!\r\n\r\nThis dataset might be somewhat complex and I'm concerned about it being 65 TB, which would be quite expensive to host. @lhoestq @mariosasko I would love your input if you think it's worth adding this dataset.", "Thanks for proposing this interesting dataset, @osanseviero.\r\n\r\nPlease note that the data files are already hosted in a third-party server: e.g. the index of data files for entity \"6795937\" is at https://fb-baas-f32eacb9-8abb-11eb-b2b8-4857dd089e15.s3.amazonaws.com/MugsyDataRelease/v0.0/identities/6795937/index.html \r\n- audio.tar: https://fb-baas-f32eacb9-8abb-11eb-b2b8-4857dd089e15.s3.amazonaws.com/MugsyDataRelease/v0.0/identities/6795937/audio.tar\r\n- ...\r\n\r\nTherefore, in principle, we don't need to host them on our Hub: it would be enough to just implement a loading script in the corresponding Hub dataset repo, e.g. \"facebook/multiface\"..." ]

[ "Thanks for reporting @conceptofmind.\r\n\r\nCould you please give details about your environment? \r\n```\r\n## Environment info\r\n<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->\r\n- `datasets` version:\r\n- Platform:\r\n- Python version:\r\n- PyArrow version:\r\n```", "Hi @albertvillanova ,\r\n\r\nHere is the environment information:\r\n```\r\n- `datasets` version: 2.3.2\r\n- Platform: Linux-5.4.0-122-generic-x86_64-with-glibc2.27\r\n- Python version: 3.9.12\r\n- PyArrow version: 7.0.0\r\n- Pandas version: 1.4.2\r\n```\r\nThanks,\r\n\r\nEnrico", "I think this issue is solved here https://discuss.huggingface.co/t/minhash-deduplication/19992/12?u=loubnabnl, this only happens for very large datasets we will update it in CodeParrot code", "Hi @loubnabnl,\r\n\r\nYes, the issue is solved in the discussion thread.\r\n\r\nI will close this issue.\r\n\r\nThank you again for all of your help.\r\n\r\nEnrico", "Thanks @loubnabnl for pointing out the solution to this issue." ]

[]

[ "Requiring torchaudio<0.12.0 isn't really a viable solution because that implies torch<0.12.0 which means no sm_86 CUDA support which means no RTX 3090 support in PyTorch.\r\n\r\nBut in my case, the error only occurs if `_fallback_load` resolves to `_fail_load` inside torchaudio 0.12.0 which is only the case if FFMPEG initialization failed: https://github.com/pytorch/audio/blob/b1f510fa5681e92ee82bdc6b2d1ed896799fc32c/torchaudio/backend/sox_io_backend.py#L36-L47\r\n\r\nThat means the proper solution for torchaudio>=0.12.0 is to check `torchaudio._extension._FFMPEG_INITIALIZED` and if it is False, then we need to remind the user to install a dynamically linked ffmpeg 4.1.8 and then maybe call `torchaudio._extension._init_ffmpeg()` to force a user-visible exception showing the missing ffmpeg dynamic library name.\r\n\r\nOn my system, installing \r\n\r\n- libavcodec.so.58 \r\n- libavdevice.so.58 \r\n- libavfilter.so.7 \r\n- libavformat.so.58 \r\n- libavutil.so.56 \r\n- libswresample.so.3 \r\n- libswscale.so.5\r\n\r\nfrom ffmpeg 4.1.8 made HF datasets 2.3.2 work just fine with torchaudio 0.12.1+cu116:\r\n\r\n```python3\r\nimport sox, torchaudio, datasets\r\nprint('torchaudio', torchaudio.__version__)\r\nprint('datasets', datasets.__version__)\r\ntorchaudio._extension._init_ffmpeg()\r\nprint(torchaudio._extension._FFMPEG_INITIALIZED)\r\nwaveform, sample_rate = torchaudio.load('/workspace/.cache/huggingface/datasets/downloads/extracted/8e5aa88585efa2a4c74c6664b576550d32b7ff9c3d1d17cc04f44f11338c3dc6/cv-corpus-8.0-2022-01-19/en/clips/common_voice_en_100038.mp3', format='mp3')\r\nprint(waveform.shape)\r\n```\r\n\r\n```\r\ntorchaudio 0.12.1+cu116\r\ndatasets 2.3.2\r\nTrue\r\ntorch.Size([1, 369792])\r\n```", "Related: https://github.com/huggingface/datasets/issues/4889", "Closing as we no longer use `torchaudio` for decoding MP3 files." ]

[ "Thanks for reporting @NitishkKarra.\r\n\r\nThe root source of the issue is that streaming mode is not supported out-of-the-box for that dataset, because it contains a TAR file.\r\n\r\nWe have opened a discussion in the corresponding Hub dataset page, pointing out this issue: https://huggingface.co/datasets/Theivaprakasham/wildreceipt/discussions/1\r\n\r\nI'm closing this issue here, so this discussion is transferred there instead." ]

[]

[ "cc @Rocketknight1 ", "Hi @lehrig, this is caused by the data collator renaming \"label\" to \"labels\". If you set `label_cols=[\"labels\"]` in the call it will work correctly. However, I agree that the cause of the bug is not obvious, so I'll see if I can make a PR to clarify things when the collator renames columns.", "Thanks - and wow, that appears like a strange side-effect of the data collator. Is that really needed?\r\n\r\nWhy not make it more explicit? For example, extend `DefaultDataCollator` with an optional property `label_col_name` to be used as label column; only when it is not provided default to `labels` (and document that this happens) for backwards-compatibility? ", "Haha, I honestly have no idea why our data collators rename `\"label\"` (the standard label column name in our datasets) to `\"labels\"` (the standard label column name input to our models). It's been a pain point when I design TF data pipelines, though, because I don't want to hardcode things like that - especially in `datasets`, because the renaming is something that happens purely at the `transformers` end. I don't think I could make the change in the data collators themselves at this point, because it would break backward compatibility for everything in PyTorch as well as TF.\r\n\r\nIn the most recent version of `transformers` we added a [prepare_tf_dataset](https://huggingface.co/docs/transformers/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset) method to our models which takes care of these details for you, and even chooses appropriate columns and labels for the model you're using. In future we might make that the officially recommended way to convert HF datasets to `tf.data.Dataset`.", "Interesting, that'd be great especially for clarity. https://huggingface.co/docs/datasets/use_with_tensorflow#data-loading already improved clarity, yet, all those options will still confuse people. Looking forward to those advances in the hope there'll be only 1 way in the future ;)\r\n\r\nAnyways, I am happy for the time being with the work-around you provided. Thank you!" ]

[]

[ "And I don't see any speed up when increasing the number of GPUs while calling `get_nearest_examples_batch`.", "Hi ! Yes search_batch uses FAISS search which happens in parallel across the GPUs\r\n\r\n> And I don't see any speed up when increasing the number of GPUs while calling get_nearest_examples_batch.\r\n\r\nThat's unexpected, can you share the code you're running ?", "here is the code snippet\r\n\r\n```python\r\n\r\n# add faiss index\r\nsource_dataset = load_dataset(source_path)\r\nqueries = load_dataset(query_path)\r\ngpu = [0,1,2,3]\r\nsource_dataset.add_faiss_index(\r\n \"embedding\",\r\n device=gpu,\r\n )\r\n\r\n\r\n# batch query\r\nbatch_size = 32\r\nfor i in tqdm(range(0, len(queries), batch_size)):\r\n if i + batch_size >= len(queries):\r\n batched_queries = queries[i:]\r\n else:\r\n batched_queries = queries[i:i+batch_size]\r\n\r\n batched_query_embeddings = np.stack([i for i in batched_queries['embedding']], axis=0)\r\n scores, candidates = source_dataset.get_nearest_examples_batch(\r\n \"embedding\",\r\n batched_query_embeddings,\r\n k=5\r\n )\r\n```", "My version of datasets is `2.4.1.dev0`.", "The code looks all good to me, do you see all the GPUs being utilized ? What version of faiss are you using ?", "I can see the memory usage of all the GPUs.\r\nMy version of `faiss-gpu` is `1.7.2`", "It looks all good to me then ^^ though you said you didn't experienced speed improvements by adding more GPUs ? What size is your source dataset and what time differences did you experience ?", "query set: 1e6\r\nsource dataset: 1e6\r\nembedding size: 768\r\nindex: Flat\r\ntopk: 20\r\nGPU: V100\r\n\r\nThe time taken to traverse the query set once is about 1.5h, which is almost not influenced by the value of query batch size or the number of GPUs according to my experiments.", "Hmmm the number of GPUs should divide the time, something is going wrong. Can you check that adding more GPU does divide the memory used per GPU ? Maybe it can be worth looking at similar issues in the FAISS repository or create a noew issue over there to understand what's going on", "> Can you check that adding more GPU does divide the memory used per GPU \r\n\r\nThe memory used per GPU is unchanged while adding more GPU. Is this unexpected?\r\n\r\nI used to think that every GPU loads all the source vectors and the data parallelism is at the query level. 😆 ", "> I used to think that every GPU loads all the source vectors and the data parallelism is at the query level. 😆\r\n\r\nOh indeed that's possible, I wasn't sure. Anyway you can check that calling get_nearest_examples_batch simply calls search under the hood: \r\n\r\nhttps://github.com/huggingface/datasets/blob/f90f71fbbb33889fe75a3ffc101cdf16a88a3453/src/datasets/search.py#L375", "Here is a runnable script. \r\nMulti-GPU searching still does not work in my experiments.\r\n\r\n\r\n```python\r\nimport os\r\nfrom tqdm import tqdm\r\nimport numpy as np\r\nimport datasets\r\nfrom datasets import Dataset\r\n\r\nclass DPRSelector:\r\n\r\n def __init__(self, source, target, index_name, gpu=None):\r\n self.source = source\r\n self.target = target\r\n self.index_name = index_name\r\n\r\n cache_path = 'embedding.faiss'\r\n\r\n if not os.path.exists(cache_path):\r\n self.source.add_faiss_index(\r\n column=\"embedding\",\r\n index_name=index_name,\r\n device=gpu,\r\n )\r\n self.source.save_faiss_index(index_name, cache_path)\r\n else:\r\n self.source.load_faiss_index(\r\n index_name,\r\n cache_path,\r\n device=gpu\r\n )\r\n print('index builded!')\r\n\r\n def build_dataset(self, top_k, batch_size):\r\n print('start search')\r\n\r\n for i in tqdm(range(0, len(self.target), batch_size)):\r\n if i + batch_size >= len(self.target):\r\n batched_queries = self.target[i:]\r\n else:\r\n batched_queries = self.target[i:i+batch_size]\r\n\r\n\r\n batched_query_embeddings = np.stack([i for i in batched_queries['embedding']], axis=0)\r\n search_res = self.source.get_nearest_examples_batch(\r\n self.index_name,\r\n batched_query_embeddings,\r\n k=top_k\r\n )\r\n \r\n print('finish search')\r\n\r\n\r\ndef get_pseudo_dataset():\r\n pseudo_dict = {\"embedding\": np.zeros((1000000, 768), dtype=np.float32)}\r\n print('generate pseudo data')\r\n\r\n dataset = Dataset.from_dict(pseudo_dict)\r\n def list_to_array(data):\r\n return {\"embedding\": [np.array(vector, dtype=np.float32) for vector in data[\"embedding\"]]} \r\n dataset.set_transform(list_to_array, columns='embedding', output_all_columns=True)\r\n\r\n print('build dataset')\r\n return dataset\r\n\r\n\r\n\r\nif __name__==\"__main__\":\r\n\r\n np.random.seed(42)\r\n\r\n\r\n source_dataset = get_pseudo_dataset()\r\n target_dataset = get_pseudo_dataset()\r\n\r\n gpu = [0,1,2,3,4,5,6,7]\r\n selector = DPRSelector(source_dataset, target_dataset, \"embedding\", gpu=gpu)\r\n\r\n selector.build_dataset(top_k=20, batch_size=32)\r\n```", "@lhoestq Hi, could you please test the code above if you have time? 😄 ", "Maybe @albertvillanova you can take a look ? I won't be available in the following days", "@albertvillanova Hi, can you help with this issue?", "Hi @xwwwwww I'm investigating it, but I'm not an expert in Faiss. In principle, it is weird that your code does not work properly because it seems right...", "Have you tried passing `gpu=-1` and check if there is a speedup?", "> Have you tried passing `gpu=-1` and check if there is a speedup?\r\n\r\nyes, there is a speed up using GPU compared with CPU. ", "When passing `device=-1`, ALL existing GPUs are used (multi GPU): this is the maximum speedup you can get. To know the number of total GPUs:\r\n```\r\nimport faiss\r\n\r\nngpus = faiss.get_num_gpus()\r\nprint(ngpus)\r\n```\r\n\r\nWhen passing a list of integers to `device`, then only that number of GPUs are used (multi GPU as well)\r\n- the speedup should be proportional (more or less) to the ratio of the number of elements passed to `device` over `ngpus`\r\n- if this is not the case, then there is an issue in the implementation of this use case (however, I have reviewed the code and in principle I can't find any evident bug)\r\n\r\nWhen passing a positive integer to `device`, then only a single GPU is used.\r\n- this time should be more or less proportional to the time when passing `device=-1` over `ngpus`", "Thanks for your help!\r\nHave you run the code and replicated the same experimental results (i.e., no speedup while increasing the number of GPUs)?", "@albertvillanova @lhoestq Sorry for the bother, is there any progress on this issue? 😃 ", "I can confirm `add_faiss_index` calls `index = faiss.index_cpu_to_gpus_list(index, gpus=list(device))`.\r\n\r\nCould this be an issue with your environment ? Could you try running with 1 and 8 GPUs with a code similar to[ this one from the FAISS examples](https://github.com/facebookresearch/faiss/blob/main/tutorial/python/5-Multiple-GPUs.py) but using `gpu_index = faiss.index_cpu_to_gpus_list(cpu_index, gpus=list(device))`, and see if the speed changes ?", "Hi, I test the FAISS example and the speed indeed changes. I set `nb=1000000`, `nq=1000000` and `d=64`\r\n\r\n| num GPUS | time cost |\r\n| -------- | --------- |\r\n| 1 | 28.53 |\r\n| 5 | 7.16 |\r\n\r\n\r\n\r\n", "Ok the benchmark is great, not sure why it doesn't speed up the index in your case though. You can try running the benchmark with the same settings as your actual dataset\r\n```\r\nquery set: 1e6\r\nsource dataset: 1e6\r\nembedding size: 768\r\nindex: Flat\r\ntopk: 20\r\nGPU: V100\r\n```\r\n\r\nNote that you can still pass a FAISS index you built yourself to a dataset using https://huggingface.co/docs/datasets/v2.4.0/en/package_reference/main_classes#datasets.Dataset.add_faiss_index_from_external_arrays", "> Here is a runnable script. Multi-GPU searching still does not work in my experiments.\r\n> \r\n> ```python\r\n> import os\r\n> from tqdm import tqdm\r\n> import numpy as np\r\n> import datasets\r\n> from datasets import Dataset\r\n> \r\n> class DPRSelector:\r\n> \r\n> def __init__(self, source, target, index_name, gpu=None):\r\n> self.source = source\r\n> self.target = target\r\n> self.index_name = index_name\r\n> \r\n> cache_path = 'embedding.faiss'\r\n> \r\n> if not os.path.exists(cache_path):\r\n> self.source.add_faiss_index(\r\n> column=\"embedding\",\r\n> index_name=index_name,\r\n> device=gpu,\r\n> )\r\n> self.source.save_faiss_index(index_name, cache_path)\r\n> else:\r\n> self.source.load_faiss_index(\r\n> index_name,\r\n> cache_path,\r\n> device=gpu\r\n> )\r\n> print('index builded!')\r\n> \r\n> def build_dataset(self, top_k, batch_size):\r\n> print('start search')\r\n> \r\n> for i in tqdm(range(0, len(self.target), batch_size)):\r\n> if i + batch_size >= len(self.target):\r\n> batched_queries = self.target[i:]\r\n> else:\r\n> batched_queries = self.target[i:i+batch_size]\r\n> \r\n> \r\n> batched_query_embeddings = np.stack([i for i in batched_queries['embedding']], axis=0)\r\n> search_res = self.source.get_nearest_examples_batch(\r\n> self.index_name,\r\n> batched_query_embeddings,\r\n> k=top_k\r\n> )\r\n> \r\n> print('finish search')\r\n> \r\n> \r\n> def get_pseudo_dataset():\r\n> pseudo_dict = {\"embedding\": np.zeros((1000000, 768), dtype=np.float32)}\r\n> print('generate pseudo data')\r\n> \r\n> dataset = Dataset.from_dict(pseudo_dict)\r\n> def list_to_array(data):\r\n> return {\"embedding\": [np.array(vector, dtype=np.float32) for vector in data[\"embedding\"]]} \r\n> dataset.set_transform(list_to_array, columns='embedding', output_all_columns=True)\r\n> \r\n> print('build dataset')\r\n> return dataset\r\n> \r\n> \r\n> \r\n> if __name__==\"__main__\":\r\n> \r\n> np.random.seed(42)\r\n> \r\n> \r\n> source_dataset = get_pseudo_dataset()\r\n> target_dataset = get_pseudo_dataset()\r\n> \r\n> gpu = [0,1,2,3,4,5,6,7]\r\n> selector = DPRSelector(source_dataset, target_dataset, \"embedding\", gpu=gpu)\r\n> \r\n> selector.build_dataset(top_k=20, batch_size=32)\r\n> ```\r\n\r\nBy the way, have you run this toy example and replicated my experiment results? I think it is a more direct way to figure this out :)", "Hi,\r\n\r\nI have a similar question and would like to know if there's any progress in this issue. \r\n\r\n`dataset.add_faiss_index(column=\"embedding\")`, this takes around 5minutes to add the index.\r\n\r\n`dataset.add_faiss_index(column=\"embedding\", device=-1)`, this ran for more than 10minutes and still didn't complete execution. \r\n\r\nNow, I don't understand why that's the case as I expected for GPU the indexing should be faster" ]

[ "Hi @SaulLu, thanks for reporting.\r\n\r\nI think offline mode is not supported for datasets containing only data files (without any loading script). I'm having a look into this...", "Thanks for your feedback! \r\n\r\nTo give you a little more info, if you don't set the offline mode flag, the script will load the cache. I first noticed this behavior with the `evaluate` library, and while trying to understand the downloading flow I realized that I had a similar error with datasets.", "This is an issue we have to fix.", "This is related to https://github.com/huggingface/datasets/issues/3547", "Still not fixed? ......", "#5331 will be helpful to fix this, as it updates the cache directory template to be aligned with the other datasets", "Any updates ?", "I'm facing the same problem", "This issue has been fixed in `datasets` 2.16 by https://github.com/huggingface/datasets/pull/6493. The cache is now working properly :)\r\n\r\nYou just have to update `datasets`:\r\n\r\n```\r\npip install -U datasets\r\n```", "I'm on version 2.17.0, and this exact problem is still persisting.", "Can you share some code to reproduce your issue ?\r\n\r\nAlso make sure your cache was populated with recent versions of `datasets`. Datasets cached with old versions may not be reloadable in offline mode, though we did our best to keep as much backward compatibility as possible.", "I'm not sure if this is related @lhoestq but I am experiencing a similar issue when using offline mode:\r\n\r\n```bash\r\n$ python -c \"from datasets import load_dataset; load_dataset('openai_humaneval', split='test')\"\r\n$ HF_DATASETS_OFFLINE=1 python -c \"from datasets import load_dataset; load_dataset('openai_humaneval', split='test')\"\r\nUsing the latest cached version of the dataset since openai_humaneval couldn't be found on the Hugging Face Hub (offline mode is enabled).\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/dodrio/scratch/projects/2023_071/alignment-handbook/.venv/lib/python3.10/site-packages/datasets/load.py\", line 2556, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/dodrio/scratch/projects/2023_071/alignment-handbook/.venv/lib/python3.10/site-packages/datasets/load.py\", line 2265, in load_dataset_builder\r\n builder_instance: DatasetBuilder = builder_cls(\r\n File \"/dodrio/scratch/projects/2023_071/alignment-handbook/.venv/lib/python3.10/site-packages/datasets/packaged_modules/cache/cache.py\", line 122, in __init__\r\n config_name, version, hash = _find_hash_in_cache(\r\n File \"/dodrio/scratch/projects/2023_071/alignment-handbook/.venv/lib/python3.10/site-packages/datasets/packaged_modules/cache/cache.py\", line 48, in _find_hash_in_cache\r\n raise ValueError(\r\nValueError: Couldn't find cache for openai_humaneval for config 'default'\r\nAvailable configs in the cache: ['openai_humaneval']\r\n```", "Thanks for reporting @BramVanroy, I managed to reproduce and I opened a fix here: https://github.com/huggingface/datasets/pull/6741", "Awesome, thanks for the quick fix @lhoestq! Looking forward to update my dependency version list.", "> Thanks for reporting @BramVanroy, I managed to reproduce and I opened a fix here: #6741\r\n\r\nThanks a lot！ I have faced the same problem. Can I use your fix code to directly replace the existing version code? I noticed that this fix has not been merged yet. Will it affect other functionalities?\r\n", "I just merged the fix, you can install `datasets` from source or wait for the patch release which will be out in the coming days" ]

[ "I refreshed the dataset viewer manually, it's fixed now. Sorry for the inconvenience.\r\n<img width=\"1557\" alt=\"Capture d’écran 2022-07-28 à 09 17 39\" src=\"https://user-images.githubusercontent.com/1676121/181514666-92d7f8e1-ddc1-4769-84f3-f1edfdb902e8.png\">\r\n\r\n" ]

[]

[ "I've built a minimal example that shows this bug without `n_proc`. It seems like it's a problem any way of using **tokenizers, `overflow_to_sample_mapping`, and Dataset.map, with a small batch size**:\r\n\r\n```\r\nimport datasets\r\nimport transformers\r\npretrained = 'deepset/tinyroberta-squad2'\r\ntokenizer = transformers.AutoTokenizer.from_pretrained(pretrained)\r\n\r\nquestions = ['Can you tell me why?', 'What time is it?']\r\ncontexts = ['This is context zero', 'Another paragraph goes here'] \r\n\r\ndef tok(questions, contexts):\r\n return tokenizer(text=questions,\r\n text_pair=contexts,\r\n truncation='only_second',\r\n return_overflowing_tokens=True,\r\n )\r\nprint(tok(questions, contexts)['overflow_to_sample_mapping'])\r\nassert tok(questions, contexts)['overflow_to_sample_mapping'] == [0, 1] # PASSES\r\n\r\ndef tok2(d):\r\n return tok(d['question'], d['context'])\r\n\r\ndef tok2(d):\r\n return tok(d['question'], d['context'])\r\n\r\nds = datasets.Dataset.from_dict({'question': questions, 'context': contexts})\r\ntokens = ds.map(tok2, batched=True, batch_size=1)\r\nprint(tokens['overflow_to_sample_mapping'])\r\nassert tokens['overflow_to_sample_mapping'] == [0, 1] # FAILS produces [0,0]\r\n```\r\n\r\nNote that even if the batch size would be larger, there will be instances where we will not have a lot of data, and end up using small batches. This can occur e.g. if `n_proc` causes batches to be underfill. I imagine it can also occur in other ways, e.g. the final leftover batch at the end.", "A larger batch size does _not_ have this behavior:\r\n\r\n```\r\ndef tok2(d):\r\n return tok(d['question'], d['context'])\r\n\r\nds = datasets.Dataset.from_dict({'question': questions, 'context': contexts})\r\ntokens = ds.map(tok2, batched=True, batch_size=2)\r\nprint(tokens['overflow_to_sample_mapping'])\r\nassert tokens['overflow_to_sample_mapping'] == [0, 1] # PASSES\r\n```", "I was trying the [Question answering](https://huggingface.co/learn/nlp-course/chapter7/7#question-answering) tutorial on Hugging face when i faced the same problem. The preprocessing step is [here](https://huggingface.co/learn/nlp-course/chapter7/7#processing-the-validation-data). i have changed ```max_length=200, stride=50```,\r\n\r\n```\r\nvalidation_dataset = raw_datasets['validation'].select(range(8)).map(\r\n preprocess_validation_examples,\r\n batched=True,\r\n remove_columns=raw_datasets[\"validation\"].column_names,\r\n num_proc=1\r\n)\r\nprint(validation_dataset['overflow_to_sample_mapping'])\r\nprint(validation_dataset['example_id'])\r\n```\r\nresult\r\n\r\n```\r\n[0, 1, 2, 3, 4, 5, 6, 7]\r\n['56be4db0acb8001400a502ec', '56be4db0acb8001400a502ed', '56be4db0acb8001400a502ee', \r\n'56be4db0acb8001400a502ef', '56be4db0acb8001400a502f0', '56be8e613aeaaa14008c90d1', \r\n'56be8e613aeaaa14008c90d2', '56be8e613aeaaa14008c90d3']\r\n```\r\nwhen ```num_proc=2```, result - \r\n\r\n```\r\n[0, 1, 2, 3, 0, 1, 2, 3]\r\n['56be4db0acb8001400a502ec', '56be4db0acb8001400a502ed', '56be4db0acb8001400a502ee', \r\n'56be4db0acb8001400a502ef', '56be4db0acb8001400a502f0', '56be8e613aeaaa14008c90d1', \r\n'56be8e613aeaaa14008c90d2', '56be8e613aeaaa14008c90d3']\r\n```\r\n\r\nwhen ```num_proc=3```, result - \r\n\r\n```\r\n[0, 1, 2, 0, 1, 2, 0, 1]\r\n['56be4db0acb8001400a502ec', '56be4db0acb8001400a502ed', '56be4db0acb8001400a502ee', \r\n'56be4db0acb8001400a502ef', '56be4db0acb8001400a502f0', '56be8e613aeaaa14008c90d1', \r\n'56be8e613aeaaa14008c90d2', '56be8e613aeaaa14008c90d3']\r\n```\r\n\r\nThe```overflow_to_sample_mapping``` changes with ```num_proc```, but ```example_id``` field remains the same . It seems that each process in ```map``` has its own counter for overflow_to_sample_mapping. If you are using ```overflow_to_sample_mapping``` inside the ```preprocess_validation_examples``` function, then there is no issue." ]

[ "I've narrowed down the issue to the `dataset_module_factory` which already creates a `dataset_infos.json` file down in the `.cache/modules/dataset_modules/..` folder. That JSON file already contains the wrong task_templates for `unfiltered`.", "Ugh. Found the issue: apparently `datasets` was reusing the already existing `dataset_infos.json` that is inside `datasets/datasets/hebban-reviews`! Is this desired behavior?\r\n\r\nPerhaps when `--save_infos` and `--all_configs` are given, an existing `dataset_infos.json` file should first be deleted before continuing with the test? Because that would assume that the user wants to create a new infos file for all configs anyway.", "Hi! I think this is a reasonable solution. Would you be interested in submitting a PR?" ]

[ "Hi ! I think the simplest is to copy paste the `_split_generators` code from the other datasets and do a bunch of if-else, as in the glue dataset: https://huggingface.co/datasets/glue/blob/main/glue.py#L467", "Ok, I see. Thank you" ]

[ "The dataset is empty, as far as I can tell: there are no files in the repository at https://huggingface.co/datasets/yanekyuk/wikikey/tree/main\r\n\r\nMaybe the viewer can display a better message for empty datasets", "OK. Closing as it's not an error. We will work on making the error message a lot clearer." ]

[ "Thanks for opening this issue :)\r\n\r\nIf it can help, I think you can already use `huggingface_hub` to achieve this:\r\n```python\r\n>>> from huggingface_hub import HfApi\r\n>>> [ds_info.id for ds_info in HfApi().list_datasets(use_auth_token=token) if ds_info.private]\r\n['bigscience/xxxx', 'bigscience-catalogue-data/xxxxxxx', ... ]\r\n```\r\n\r\n---------\r\n\r\nThough the latest versions of `huggingface_hub` that contain this feature are not available on python 3.6, so maybe we should first drop support for python 3.6 (see #4460) to update `list_datasets` in `datasets` as well (or we would have to copy/paste some `huggingface_hub` code)", "Great, thanks @lhoestq the workaround works! I think it would be intuitive to have the support directly in `datasets` but it makes sense to wait given that the workaround exists :)", "i also think that going forward we should replace more and more implementations inside datasets with the corresponding ones from `huggingface_hub` (same as we're doing in `transformers`)", "`datasets.list_datasets` is now deprecated in favor of `huggingface_hub.list_datasets` (returns private datasets when `token` is present), so I'm closing this issue." ]

[ "Note that for me personally, conceptually all the dummy data (even for \"canonical\" datasets) should be superseded by `datasets-server`, which performs some kind of CI/CD of datasets (including the canonical ones)", "I totally agree: next step should be rethinking if dummy data makes sense for canonical datasets (once we have datasets-server) and eventually remove it.\r\n\r\nBut for now, we could at least start by removing the indication to generate dummy data from our docs." ]

[ "Hi @BramVanroy, thanks for reporting.\r\n\r\nFirst of all, please note that you do not need the dummy data: this was the case when we were adding datasets to the `datasets` library (on this GitHub repo), so that we could test the correct loading of all datasets with our CI. However, this is no longer the case for datasets on the Hub.\r\n- We should definitely update our docs.\r\n\r\nSecond, the dummy data is generated locally:\r\n- in your case, the dummy data will be generated inside the directory: `./datasets/hebban-reviews/dummy`\r\n- please note the preceding `./datasets` directory: the reason for this is that the command to generate the dummy data was specifically created for our `datasets` library, and therefore assumes our directory structure: commands are run from the root directory of our GitHub repo, and datasets scripts are under `./datasets` \r\n\r\n\r\n ", "I have opened an Issue to update the instructions on dummy data generation:\r\n- #4744", "Dummy data generation is deprecated now, so I think we can close this issue." ]

[ "Hi! The server with the data seems to be down. I've reported this issue (https://github.com/CSAILVision/sceneparsing/issues/34) in the dataset repo. ", "The URL seems to work now, and therefore the script as well." ]

[ "Thanks for reporting. You're right, workers were under-provisioned due to a manual error, and the job queue was full. It's fixed now." ]

[ "We have added a comment on an existing issue opened in their repo: https://github.com/google-research/google-research/issues/1212#issuecomment-1192267130\r\n- https://github.com/google-research/google-research/issues/1212" ]

[ "Fixed by:\r\n- #4735" ]

[ "Thanks for the feedback!\r\n\r\nI think since this issue is closely related to loading, I can add a clearer explanation under [Load > local loading script](https://huggingface.co/docs/datasets/main/en/loading#local-loading-script).", "That makes sense but I think having a line about it under https://huggingface.co/docs/datasets/installation#source the \"source\" header here would be useful. My mental model of `pip install -e .` does not include the fact that the source files aren't actually being used. ", "Thanks for sharing your perspective. I think the `load_dataset` function is the only one that pulls from GitHub, and since this use-case is very specific, I don't think we need to include such a broad clarification in the Installation section.\r\n\r\nFeel free to check out the linked PR and let me know if it needs any additional explanation 😊" ]

[ "My bad, `482 * 418 * 50000 * 3 / 1000000 = 30221 MB` ( https://stackoverflow.com/a/42979315 ).\r\n\r\nMeanwhile `256 * 256 * 50000 * 3 / 1000000 = 9830 MB`. We are loading the non-cropped images and that is why we take so much RAM." ]

[ "Hi @rohitvincent, thanks for reporting.\r\n\r\nUnfortunately I'm not able to reproduce your issue:\r\n```python\r\nIn [2]: from datasets import load_dataset, DownloadMode\r\n ...: load_dataset(path='financial_phrasebank',name='sentences_allagree', download_mode=\"force_redownload\")\r\nDownloading builder script: 6.04kB [00:00, 2.87MB/s] \r\nDownloading metadata: 13.7kB [00:00, 7.24MB/s] \r\nDownloading and preparing dataset financial_phrasebank/sentences_allagree (download: 665.91 KiB, generated: 296.26 KiB, post-processed: Unknown size, total: 962.17 KiB) to .../.cache/huggingface/datasets/financial_phrasebank/sentences_allagree/1.0.0/550bde12e6c30e2674da973a55f57edde5181d53f5a5a34c1531c53f93b7e141...\r\nDownloading data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 682k/682k [00:00<00:00, 7.66MB/s]\r\nDataset financial_phrasebank downloaded and prepared to .../.cache/huggingface/datasets/financial_phrasebank/sentences_allagree/1.0.0/550bde12e6c30e2674da973a55f57edde5181d53f5a5a34c1531c53f93b7e141. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 918.80it/s]\r\nOut[2]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['sentence', 'label'],\r\n num_rows: 2264\r\n })\r\n})\r\n```\r\n\r\nAre you able to access the link? https://www.researchgate.net/profile/Pekka-Malo/publication/251231364_FinancialPhraseBank-v10/data/0c96051eee4fb1d56e000000/FinancialPhraseBank-v10.zip", "Yes was able to download from the link manually. But still, get the same error when I use load_dataset.", "Fixed once data files are hosted on the Hub:\r\n- #4598" ]

[ "The preview is working OK:\r\n\r\n\r\n\r\n" ]

[ "Thanks for reporting, @TrentBrick. We are addressing the change with their data host server.\r\n\r\nOn the meantime, if you would like to work with your fixed local copy of the_pile script, you should use:\r\n```python\r\nload_dataset(\"path/to/your/local/the_pile/the_pile.py\",...\r\n```\r\ninstead of just `load_dataset(\"the_pile\",...`.\r\n\r\nThe latter downloads a copy of `the_pile.py` from our GitHub, caches it locally (inside `~/.cache/huggingface/modules`) and uses that.", "@TrentBrick, I have checked the URLs and both hosts work, the original (https://the-eye.eu/) and the mirror (https://mystic.the-eye.eu/). See e.g.:\r\n- https://mystic.the-eye.eu/public/AI/pile/\r\n- https://mystic.the-eye.eu/public/AI/pile_preliminary_components/\r\n\r\nPlease, let me know if you still find any issue loading this dataset by using current server URLs.", "Great this is working now. Re the download from GitHub... I'm sure thought went into doing this but could it be made more clear maybe here? https://huggingface.co/docs/datasets/installation for example under installing from source? I spent over an hour questioning my sanity as I kept trying to edit this file, uninstall and reinstall the repo, git reset to previous versions of the file etc.", "Thanks for the quick reply and help too\r\n", "Thanks @TrentBrick for the suggestion about improving our docs: we should definitely do this if you find they are not clear enough.\r\n\r\nCurrently, our docs explain how to load a dataset from a local loading script here: [Load > Local loading script](https://huggingface.co/docs/datasets/loading#local-loading-script)\r\n\r\nI've opened an issue here:\r\n- #4732\r\n\r\nFeel free to comment on it any additional explanation/suggestion/requirement related to this problem." ]

[ "Hi ! It looks like a bug in `pyarrow`. If you manage to end up with only one chunk per parquet file it should workaround this issue.\r\n\r\nTo achieve that you can try to lower the value of `max_shard_size` and also don't use `map` before `push_to_hub`.\r\n\r\nDo you have a minimum reproducible example that we can share with the Arrow team for further debugging ?", "> If you manage to end up with only one chunk per parquet file it should workaround this issue.\r\n\r\nYup, I did not encounter this bug when I was testing my script with a slice of <1000 samples for my dataset.\r\n\r\n> Do you have a minimum reproducible example...\r\n\r\nNot sure if I can get more minimal than the script I shared above. Are you asking for a sample json file?\r\nJust generate a random manifest list, I can add that to the above script if that's what you mean?\r\n", "Actually this is probably linked to this open issue: https://issues.apache.org/jira/browse/ARROW-5030.\r\n\r\nsetting `max_shard_size=\"2GB\"` should do the job (or `max_shard_size=\"1GB\"` if you want to be on the safe side, especially given that there can be some variance in the shard sizes if the dataset is not evenly distributed)" ]

[ "It seems like the list of splits could not be obtained:\r\n\r\n```python\r\n>>> from datasets import get_dataset_split_names\r\n>>> get_dataset_split_names(\"shamikbose89/lancaster_newsbooks\", \"default\")\r\nUsing custom data configuration default\r\nTraceback (most recent call last):\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 354, in get_dataset_config_info\r\n for split_generator in builder._split_generators(\r\n File \"/home/slesage/.cache/huggingface/modules/datasets_modules/datasets/shamikbose89--lancaster_newsbooks/2d1c63d269bf7b9342accce0a95960b1710ab4bc774248878bd80eb96c1afaf7/lancaster_newsbooks.py\", line 73, in _split_generators\r\n data_dir = dl_manager.download_and_extract(_URL)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 916, in download_and_extract\r\n return self.extract(self.download(url_or_urls))\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 879, in extract\r\n urlpaths = map_nested(self._extract, path_or_paths, map_tuple=True)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/utils/py_utils.py\", line 348, in map_nested\r\n return function(data_struct)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 884, in _extract\r\n protocol = _get_extraction_protocol(urlpath, use_auth_token=self.download_config.use_auth_token)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 388, in _get_extraction_protocol\r\n return _get_extraction_protocol_with_magic_number(f)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/download/streaming_download_manager.py\", line 354, in _get_extraction_protocol_with_magic_number\r\n f.seek(0)\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 684, in seek\r\n raise ValueError(\"Cannot seek streaming HTTP file\")\r\nValueError: Cannot seek streaming HTTP file\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 404, in get_dataset_split_names\r\n info = get_dataset_config_info(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 359, in get_dataset_config_info\r\n raise SplitsNotFoundError(\"The split names could not be parsed from the dataset config.\") from err\r\ndatasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.\r\n```\r\n\r\nping @huggingface/datasets ", "Oh, I removed the 'split' key from `kwargs`. I put it back in, but there's still the same error", "It looks like the data host doesn't support http range requests, which is necessary to glob inside a ZIP archive in streaming mode. Can you try hosting the dataset elsewhere ? Or download each file separately from https://ota.bodleian.ox.ac.uk/repository/xmlui/handle/20.500.12024/2531 ?", "@lhoestq Thanks! That seems to have solved it. I can get the splits with the `get_dataset_split_names()` function. The dataset viewer is still not loading properly, though. The new error is\r\n```\r\nStatus code: 400\r\nException: BadZipFile\r\nMessage: File is not a zip file\r\n```\r\n\r\nPS. The dataset loads properly and can be accessed" ]

[ "I am also getting a ValueError: 'Couldn't cast' at the bottom. Is this because of some delimiter issue? My dataset is on the Huggingface Hub. If you could look at it, that would be greatly appreciated.", "Hi @thenerd31, thanks for reporting.\r\n\r\nPlease note that your issue is not caused by the Hugging Face Datasets library, but it has to do with the specific implementation of your dataset on the Hub.\r\n\r\nTherefore, I'm transferring this discussion to your own dataset Community tab: https://huggingface.co/datasets/TheNoob3131/mosquito-data/discussions/1" ]

[ "It's currently working, as far as I understand\r\n\r\nhttps://huggingface.co/datasets/LawalAfeez/englishreview-ds-mini/viewer/LawalAfeez--englishreview-ds-mini/train\r\n\r\n<img width=\"1556\" alt=\"Capture d’écran 2022-07-19 à 09 24 01\" src=\"https://user-images.githubusercontent.com/1676121/179761130-2d7980b9-c0f6-4093-8b1d-f0a3872fef3f.png\">\r\n\r\n---\r\n\r\nWhat was your issue?" ]

[ "I'm closing this issue as both the Audio and Image sections now have a \"Create dataset\" page that contains the info about writing the loading script version of a dataset." ]

[ "Hi ! That would be awesome to have them indeed, thanks for opening this issue\r\n\r\nI just added you to the WMT org on the HF Hub if you're interested in adding those datasets.\r\n\r\nFeel free to create a dataset repository for each dataset and upload the data files there :) preferably in ZIP archives instead of TAR archives (the current WMT scripts don't support streaming TAR archives, so it would break the dataset preview). We've also had issues with the `statmt.org` host (data unavailable, slow download speed), that's why I think it's better if we re-host the files on the Hub.\r\n\r\n`wmt21` (and wmt22) can be added <s>in this GitHub repository I think</s> on the HF Hub under the `WMT` org (we'll move the previous ones to this org soon as well).\r\nTo add it, you can copy paste the code of the previous one (e.g. wmt19), and add the new data:\r\n- in wmt_utils.py, add the new data subsets. You need to provide the download URLs, as well as the target and source languages\r\n- in wmt21.py (renamed from wmt19.py), you can specify the subsets that WMT21 uses (i.e. the one you just added)\r\n- in wmt_utils.py, define the python function that must be used to parse the subsets you added. To do so, you must go in `_generate_examples` and chose the proper `sub_generator` based on the subset name. For example, the `paracrawl_v3` subset uses the `_parse_tmx` function:\r\n\r\nhttps://github.com/huggingface/datasets/blob/ede72d3f9796339701ec59899c7c31d2427046fb/datasets/wmt19/wmt_utils.py#L834-L835\r\n\r\nHopefully the data is in a format that is already supported and there's no need to write a new `_parse_*` function for the new subsets. Let me know if you have questions or if I can help :)", "@Muennighoff , @lhoestq let me know if you want me to look into this. Happy to help bring WMT21 & WMT22 datasets into 🤗 ! ", "Hi @srhrshr :) Sure, feel free to create a dataset repository on the Hub and start from the implementation of WMT19 if you want. Then we can move the dataset under the WMT org (we'll move the other ones there as well).\r\n\r\nLet me know if you have questions or if I can help", "#self-assign", "Hello @lhoestq ,\r\n\r\nWould it be possible for me to be granted in the WMT organization (on hf ofc) in order to facilitate dataset uploads? I've already initiated the joining process at this link: https://huggingface.co/wmt\r\n\r\nI appreciate your help with this. Thank you!", "Hi ! Cool I just added you" ]

[ "Thanks for reporting. I refreshed the dataset viewer and it now works as expected.\r\n\r\nhttps://huggingface.co/datasets/TheNoob3131/mosquito-data\r\n\r\n<img width=\"1135\" alt=\"Capture d’écran 2022-07-18 à 13 15 22\" src=\"https://user-images.githubusercontent.com/1676121/179566497-e47f1a27-fd84-4a8d-9d7f-2e0f2da803df.png\">\r\n\r\nWe will investigate why it occurred in the first place\r\n", "By chance, could you provide some details about the operations done on the dataset: was it private? gated?", "Yes, it was a private dataset, and when I made it public, the Dataset Preview did not work. \r\n\r\nHowever, now when I make the dataset private, it says that the Dataset Preview has been disabled. Why is this?", "Thanks for the details. For now, the dataset viewer is always disabled on private datasets (see https://huggingface.co/docs/hub/datasets-viewer for more details)", "Hi, it was working fine for a few hours, but then I can't see the dataset viewer again (public dataset). Why is this still happening?\r\nIt's the same error too:\r\n\r\n", "OK? This is a bug, thanks for help spotting and reproducing it (it occurs when a dataset is switched to private, then to public). We will be working on it, meanwhile, I've restored the dataset viewer manually again." ]

[ "Hi! I added a link to this issue in our internal request for adding keywords/topics to the Hub, which is identical to the `topic tags` solution. The `collections` solution seems too complex (as you point out). Regarding the `domain tags` solution, we primarily focus on machine learning, so I'm not sure if it's a good idea to make our current taxonomy more complex.", "> Hi! I added a link to this issue in our internal request for adding keywords/topics to the Hub, which is identical to the `topic tags` solution. The `collections` solution seems too complex (as you point out). Regarding the `domain tags` solution, we primarily focus on machine learning, so I'm not sure if it's a good idea to make our current taxonomy more complex.\r\n\r\nThanks, for letting me know. Will you allow the topic tags to be user-generated or only chosen from a list?", "Thanks for opening this issue @davanstrien.\r\n\r\nAs we discussed last week, the tag approach would be in principle the simpler to be implemented, either the domain tag (with closed vocabulary: more reliable but also more rigid), or the topic tag (with open vocabulary: more flexible for user needs)", "Hi @davanstrien If i remember correctly this was also discussed inside a hf.co Discussion, would you be able to link it here too?\r\n\r\n(where i suggested using `tags: - foo - bar` IIRC.\r\n\r\nThanks a ton!", "> Hi @davanstrien If i remember correctly this was also discussed inside a hf.co Discussion, would you be able to link it here too?\r\n> \r\n> (where i suggested using `tags: - foo - bar` IIRC.\r\n> \r\n> Thanks a ton!\r\n\r\nThis doesn't ring a bell - I did a quick search of https://discuss.huggingface.co but didn't find anything. \r\n\r\nThe `tags: ` approach sounds like a good option for this. It would be especially nice if these could suggest existing tags, but this probably won't be easily possible through the current interface. \r\n", "I opened a PR to add \"tags\" to the YAML validator:\r\nhttps://github.com/huggingface/datasets/pull/4716\r\n\r\nI also added \"tags\" to the [tagging app](https://huggingface.co/spaces/huggingface/datasets-tagging), with suggestions like \"bio\" or \"newspapers\"", "Thanks @lhoestq for the initiative.\r\n \r\nJust one question: are \"tags\" already supported on the Hub? \r\n\r\nI think they aren't. Thus, the Hub should support them so that they are properly displayed.", "I think they're not displayed, but at least it should enable users to filter by tag in using `huggingface_hub` or using the appropriate query params on the website (not sure if it's possible yet though)", "> I think they're not displayed, but at least it should enable users to filter by tag in using `huggingface_hub` or using the appropriate query params on the website (not sure if it's possible yet though)\r\n\r\nI think this would already be a helpful start. I'm happy to try this out with the datasets added to https://huggingface.co/organizations/biglam and use the `huggingface_hub` to filter those datasets using the tags. ", "Is this abandoned? \r\nI'm looking for a transport logistics dataset; how can I find one?", "@younes-io Full text search is probably your best bet: https://huggingface.co/search/full-text?type=dataset" ]

[ "Hi @frgfm, thanks for reporting.\r\n\r\nAs the error message says, streaming mode is not supported out of the box when the dataset contains TAR archive files.\r\n\r\nTo make the dataset streamable, you have to use `dl_manager.iter_archive`.\r\n\r\nThere are several examples in other datasets, e.g. food101: https://huggingface.co/datasets/food101/blob/main/food101.py\r\n\r\nAnd yes, as the link you pointed out, for the streaming to be possible, the metadata file must be loaded before all of the images:\r\n- either this is the case when iterating the archive (and you get the metadata file before the images)\r\n- or you have to extract the metadata file by hand and upload it separately to the Hub", "Hi @albertvillanova :wave:\r\n\r\nThanks! Yeah I saw that but since I didn't have any metadata, I wasn't sure whether I should create them myself.\r\n\r\nSo one last question:\r\nWhat is the metadata supposed to be for archives? The relative path of all files in it?\r\n_(Sorry I'm a bit confused since it's quite hard to debug using the single error message from the data preview :sweat_smile: )_", "Hi @frgfm, streaming a dataset that contains a TAR file requires some tweaks because (contrary to ZIP files), tha TAR archive does not allow random access to any of the contained member files. Instead they have to be accessed sequentially (in the order in which they were put into the TAR file when created) and yielded.\r\n\r\nSo when iterating over the TAR file content, when an image file is found, we need to yield it (and not keeping it in memory, which will require huge RAM memory for large datasets). But when yielding an image file, we also need to yield with it what we call \"metadata\": the class label, and other textual information (for example, for audio files, sometimes we also add info such as the speaker ID, their sex, their age,...).\r\n\r\nAll this information usually is stored in what we call the metadata file: either a JSON or a CSV/TSV file.\r\n\r\nBut if this is also inside the TAR archive, we need to find this file in the first place when iterating the TAR archive, so that we already have this information when we find an image file and we can yield the image file and its metadata info.\r\n\r\nTherefore:\r\n- either the TAR archive contains the metadata file as the first member when iterating it (something we cannot change as it is done at the creation of the TAR file)\r\n- or if not, then we need to have the metadata file elsewhere\r\n - in these cases, what we do (if the dataset license allows it) is:\r\n - we download the TAR file locally, we extract the metadata file and we host the metadata on the Hub\r\n - we modify the dataset loading script so that it first downloads the metadata file (and reads it) and only then starts iterating the content of the TAR archive file\r\n\r\nSee an example of this process we recently did for \"google/fleurs\" (their metadata files for \"train\" were at the end of the TAR archives, after all audio files): https://huggingface.co/datasets/google/fleurs/discussions/4\r\n- we uploaded the metadata file to the Hub\r\n- we adapted the loading script to use it", "Hi @albertvillanova :wave: \r\n\r\nThanks, since my last message, I went through the repo of https://huggingface.co/datasets/food101/blob/main/food101.py and managed to get it to work in the end :pray: \r\n\r\nHere it is: https://huggingface.co/datasets/frgfm/imagenette\r\n\r\nI appreciate you opening an issue to document the process, it might help a few!", "Great to see that you manage to make your dataset streamable. :rocket: \r\n\r\nI'm closing this issue, as for the docs update there is another issue opened:\r\n- #4711" ]

[ "Hi @finiteautomata, thanks for reporting.\r\n\r\nUnfortunately, I'm not able to reproduce your issue:\r\n```python\r\nIn [1]: from datasets import load_dataset\r\n ...: dataset = load_dataset(\"lince\", \"ner_spaeng\")\r\nDownloading builder script: 20.8kB [00:00, 9.09MB/s] \r\nDownloading metadata: 31.2kB [00:00, 13.5MB/s] \r\nDownloading and preparing dataset lince/ner_spaeng (download: 2.93 MiB, generated: 18.45 MiB, post-processed: Unknown size, total: 21.38 MiB) to .../.cache/huggingface/datasets/lince/ner_spaeng/1.0.0/10d41747f55f0849fa84ac579ea1acfa7df49aa2015b60426bc459c111b3d589...\r\nDownloading data: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3.08M/3.08M [00:01<00:00, 2.73MB/s]\r\nDataset lince downloaded and prepared to .../.cache/huggingface/datasets/lince/ner_spaeng/1.0.0/10d41747f55f0849fa84ac579ea1acfa7df49aa2015b60426bc459c111b3d589. Subsequent calls will reuse this data.\r\n100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 630.66it/s]\r\n\r\nIn [2]: dataset\r\nOut[2]: \r\nDatasetDict({\r\n train: Dataset({\r\n features: ['idx', 'words', 'lid', 'ner'],\r\n num_rows: 33611\r\n })\r\n validation: Dataset({\r\n features: ['idx', 'words', 'lid', 'ner'],\r\n num_rows: 10085\r\n })\r\n test: Dataset({\r\n features: ['idx', 'words', 'lid', 'ner'],\r\n num_rows: 23527\r\n })\r\n})\r\n``` \r\n\r\nPlease note that for this dataset, the original data files are not hosted on the Hugging Face Hub, but on https://ritual.uh.edu\r\nAnd sometimes, the server might be temporarily unavailable, as your error message said (trying to connect to the server timed out):\r\n```\r\nConnectionError: Couldn't reach https://ritual.uh.edu/lince/libaccess/eyJ1c2VybmFtZSI6ICJodWdnaW5nZmFjZSBubHAiLCAidXNlcl9pZCI6IDExMSwgImVtYWlsIjogImR1bW15QGVtYWlsLmNvbSJ9/ner_spaeng.zip (ConnectTimeout(MaxRetryError(\"HTTPSConnectionPool(host='ritual.uh.edu', port=443): Max retries exceeded with url: /lince/libaccess/eyJ1c2VybmFtZSI6ICJodWdnaW5nZmFjZSBubHAiLCAidXNlcl9pZCI6IDExMSwgImVtYWlsIjogImR1bW15QGVtYWlsLmNvbSJ9/ner_spaeng.zip (Caused by ConnectTimeoutError(<urllib3.connection.HTTPSConnection object at 0x7feb1c45a690>, 'Connection to ritual.uh.edu timed out. (connect timeout=100)'))\")))\r\n```\r\nIn these cases you could:\r\n- either contact the owners of the data server where the data is hosted to inform them about the issue in their server\r\n- or re-try after waiting some time: usually these issues are just temporary", "Great, thanks for checking out!" ]

[ "Hi ! According to https://huggingface.co/docs/datasets/use_with_pytorch#stream-data you can use the pytorch DataLoader with `num_workers>0` to distribute the shards across your workers (it uses `torch.utils.data.get_worker_info()` to get the worker ID and select the right subsets of shards to use)\r\n\r\n<s> EDIT: here is a code example </s>\r\n```python\r\n# ds = ds.with_format(\"torch\")\r\n# dataloader = DataLoader(ds, num_workers=num_workers)\r\n```\r\n\r\nEDIT: `with_format(\"torch\")` is not required, now you can just do\r\n```python\r\ndataloader = DataLoader(ds, num_workers=num_workers)\r\n```", "@cyk1337 does streaming datasets with multi-gpu works for you? I am testing on one node with multiple gpus, but this is freezing, https://github.com/huggingface/datasets/issues/5123 \r\nIn case you could make this work, could you share with me your data-loading codes?\r\nthank you", "+1", "This has been implemented in `datasets` 2.8:\r\n```python\r\nfrom datasets.distributed import split_dataset_by_node\r\n\r\nds = split_dataset_by_node(ds, rank=rank, world_size=world_size)\r\n```\r\n\r\ndocs: https://huggingface.co/docs/datasets/use_with_pytorch#distributed", "i'm having hanging issues with this when using DDP and allocating the datasets with `split_dataset_by_node` 🤔\r\n\r\n--- \r\n### edit\r\nI don't want to pollute this thread, but for the sake of following up, I observed hanging close to the final iteration of the dataloader. I think this was happening on the final shard. First, I removed the final shard and things worked. Then (including all shards), I reordered the list of shards: `load_dataset('json', data_files=reordered, streaming=True)` and no hang. \r\n\r\nI won't open an issue yet bc I am not quite sure about this observation.", "@wconnell would you mind opening a different bug issue and giving more details?\r\nhttps://github.com/huggingface/datasets/issues/new?assignees=&labels=&template=bug-report.yml\r\n\r\nThanks." ]

[ "Hi, thanks for reporting! A PR (https://github.com/huggingface/datasets/pull/4614) has already been opened to address this issue." ]

[ "Hi, thanks for reporting. I triggered a refresh of the preview for this dataset, and it works now. I'm not sure what occurred.\r\n<img width=\"1019\" alt=\"Capture d’écran 2022-07-18 à 11 01 52\" src=\"https://user-images.githubusercontent.com/1676121/179541327-f62ecd5e-a18a-4d91-b316-9e2ebde77a28.png\">\r\n\r\n" ]

[ "Hi! One option is use `map` with a function that overwrites the labels (`dset = dset.map(lamba _: {\"label\": 0}, features=dset.features`)). Or you can use the `remove_column` + `add_column` combination (`dset = dset.remove_columns(\"label\").add_column(\"label\", [0]*len(data)).cast(dset.features)`, but note that this approach creates an in-memory table for the added column instead of writing to disk, which could be problematic for large datasets.", "Hi! I tried your proposed solution, but it does not solve my problem unfortunately. I am working with a set of protein sequences that have been tokenized with ESM, but some sequences are longer than `max_length`, they have been truncated in the tokenization. So now I want to truncate my labels as well, but that does not work with a mapping (e.g. `dset.map` as you suggested). Specifically, what I did was the following:\r\n\r\n```\r\ndef postprocess_tokenize(tokenized_data):\r\n \"\"\"\r\n adjust label lengths if they dont match.\r\n \"\"\"\r\n if len(tokenized_data['input_ids']) < len(tokenized_data['labels']):\r\n new_labels = tokenized_data['labels'][:len(tokenized_data['input_ids'])]\r\n tokenized_data[\"labels\"] = new_labels\r\n return tokenized_data\r\n\r\ntokenized_data = tokenized_data.map(postprocess_tokenize, batched=True) # this does not adjust the labels...\r\n```\r\n\r\nAny tips on how to do this properly?\r\n\r\nMore generally, I am wondering why the DataCollator supports padding but does not support truncation? Seems odd to me.\r\n\r\nThanks in advance!" ]

[]

[ "Hi, thanks for reporting! If there are no examples in ImageFolder, the `label` column is of type `ClassLabel(names=[])`, which leads to an error in [this line](https://github.com/huggingface/datasets/blob/c15b391942764152f6060b59921b09cacc5f22a6/src/datasets/arrow_writer.py#L387) as `asdict(info)` calls `Features({..., \"label\": {'num_classes': 0, 'names': [], 'id': None, '_type': 'ClassLabel'}})`, which then calls `require_decoding` [here](https://github.com/huggingface/datasets/blob/c15b391942764152f6060b59921b09cacc5f22a6/src/datasets/features/features.py#L1516) on the dict value it does not expect.\r\n\r\nI see two ways to fix this:\r\n* custom `asdict` where `dict_factory` is also applied on the `dict` object itself besides dataclasses (the built-in implementation calls `type(dict_obj)` - this means we also need to fix `Features.to_dict` btw) \r\n* implement `DatasetInfo.to_dict` (though adding `to_dict` to a data class is a bit weird IMO)\r\n\r\n@lhoestq Which one of these approaches do you like more?\r\n", "Small pref for the first option, it feels weird to know that `Features()` can be called with a dictionary of types defined as dictionaries instead of type instances." ]

[ "There seems to be an issue with the `C++-snippet-level` config:\r\n\r\n```python\r\n>>> from datasets import get_dataset_split_names\r\n>>> get_dataset_split_names(\"codeparrot/xlcost-text-to-code\", \"C++-snippet-level\")\r\nTraceback (most recent call last):\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 352, in get_dataset_config_info\r\n info.splits = {\r\nTypeError: 'NoneType' object is not iterable\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 404, in get_dataset_split_names\r\n info = get_dataset_config_info(\r\n File \"/home/slesage/hf/datasets-server/services/worker/.venv/lib/python3.9/site-packages/datasets/inspect.py\", line 359, in get_dataset_config_info\r\n raise SplitsNotFoundError(\"The split names could not be parsed from the dataset config.\") from err\r\ndatasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.\r\n```\r\n\r\nI remove the dataset-viewer tag since it's not directly related.\r\n\r\nPinging @huggingface/datasets ", "Thanks I found that this subset wasn't properly defined the the config, I fixed it. Now I can see the subsets but I get this error for the viewer\r\n````\r\nStatus code: 400\r\nException: Status400Error\r\nMessage: The split cache is empty.\r\n```", "Yes, the cache is being refreshed, hopefully, it will work in some minutes for all the splits. Some are already here:\r\n\r\nhttps://huggingface.co/datasets/codeparrot/xlcost-text-to-code/viewer/Python-snippet-level/train\r\n\r\n<img width=\"1533\" alt=\"Capture d’écran 2022-07-18 à 12 04 06\" src=\"https://user-images.githubusercontent.com/1676121/179553933-64d874fa-ada9-4b82-900e-082619523c20.png\">\r\n", "I think all the splits are working as expected now", "Perfect, thank you!" ]

[ "Hi! Calling `take` on an iterable/streamable dataset makes it not possible to shard the dataset, which in turn disables multi-process loading (attempts to split the workload over the shards), so to go past this limitation, you can either use single-process loading in `DataLoader` (`num_workers=None`) or fetch the first `50_000/batch_size` batches in the loop." ]

[ "did you ever fix this? I'm experiencing the same", "I am having the same issue. Even the simple example from the documentation gives me the 400 Error\r\n\r\n\r\n> from datasets import load_dataset\r\n> \r\n> dataset = load_dataset(\"stevhliu/demo\")\r\n> dataset.push_to_hub(\"processed_demo\")\r\n\r\n\r\n`requests.exceptions.HTTPError: 400 Client Error: Bad Request for url: https://huggingface.co/api/datasets/REDACTED/commit/main (Request ID: e-tPnYTiCdB5KPmSL86dQ)`\r\n\r\nI \"fixed\" it by initializing a new virtual environment with only datasets==2.5.2 installed.\r\n\r\nThe workaround consists of saving to disk then loading from disk and pushing to hub but from the new clean virtual environment." ]

[ "Are you able to reproduce this? My example is small enough that it should be easy to try.", "Hi! Thanks for reporting and providing a reproducible example. Indeed, by default, `datasets` performs an expensive cast on the values returned by `map` to convert them to one of the types supported by PyArrow (the underlying storage format used by `datasets`). This cast is not needed on NumPy arrays as PyArrow supports them natively, so one way to make this transform faster is to add `return_tensors=\"np\"` to the tokenizer call. \r\n\r\nI think we should mention this in the docs (cc @stevhliu)", "I tested this tokenize function and indeed noticed a casting. However it seems to only concerns the `offset_mapping` field, which contains a list of tuples, that is converted to a list of lists. Since `pyarrow` also supports tuples, we actually don't need to convert the tuples to lists. \r\n\r\nI think this can be changed here: \r\n\r\nhttps://github.com/huggingface/datasets/blob/ede72d3f9796339701ec59899c7c31d2427046fb/src/datasets/features/features.py#L382-L383\r\n\r\n```diff\r\n- if isinstance(obj, list): \r\n+ if isinstance(obj, (list, tuple)): \r\n```\r\n\r\nand here: \r\n\r\nhttps://github.com/huggingface/datasets/blob/ede72d3f9796339701ec59899c7c31d2427046fb/src/datasets/features/features.py#L386-L387\r\n\r\n```diff\r\n- return obj if isinstance(obj, list) else [], isinstance(obj, tuple)\r\n+ return obj, False\r\n```\r\n\r\n@srobertjames can you try applying these changes and let us know if it helps ? If so, feel free to open a Pull Request to contribute this improvement if you want :)", "Wow, adding `return_tensors=\"np\"` sped up my example by a **factor 17x** of and completely eliminated the casting! I'd recommend not only to document it, but to make that the default.\r\n\r\nThe code at https://github.com/huggingface/notebooks/blob/main/examples/question_answering.ipynb does not specify `return_tensors=\"np\"` but yet avoids the casting penalty. How does it do that? (The ntbk seems to do `return_overflowing_tokens=True, return_offsets_mapping=True,`).\r\n\r\nAlso, surprisingly enough, using `return_tensors=\"pt\"` (which is my eventual application) yields this error:\r\n```\r\nTypeError: Provided `function` which is applied to all elements of table returns a `dict` of types \r\n[<class 'torch.Tensor'>, <class 'torch.Tensor'>, <class 'torch.Tensor'>, <class 'torch.Tensor'>]. \r\nWhen using `batched=True`, make sure provided `function` returns a `dict` of types like \r\n`(<class 'list'>, <class 'numpy.ndarray'>)`.\r\n```", "Setting the output to `\"np\"` makes the whole pipeline fast because it moves the data buffers from rust to python to arrow using zero-copy, and also because it does eliminate the casting completely ;)\r\n\r\nHave you had a chance to try eliminating the tuple casting using the trick above ?", "@lhoestq I just benchmarked the two edits to `features.py` above, and they appear to solve the problem, bringing my original example to within 20% the speed of the output `\"np\"` example. Nice!\r\n\r\nFor a pull request, do you suggest simply following https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md ?", "Cool ! Sure feel free to follow these instructions to open a PR :) thanks !", "#take", "Resolved via https://github.com/huggingface/datasets/pull/4993." ]

[ "Hi! `para_crawl` has a single column of type `Translation`, which stores translation dictionaries. These dictionaries can be stored in a NumPy array but not in a PyTorch tensor since PyTorch only supports numeric types. In `datasets`, the conversion to `torch` works as follows: \r\n1. convert PyArrow table to NumPy arrays \r\n2. convert NumPy arrays to Torch tensors. \r\n\r\nThe 2nd step is problematic for your case as `datasets` attempts to convert the array of dictionaries to a PyTorch tensor. One way to fix this is to use the [preprocessing logic](https://github.com/huggingface/transformers/blob/8581a798c0a48fca07b29ce2ca2ef55adcae8c7e/examples/pytorch/translation/run_translation.py#L440-L458) from the Transformers translation script. And on our side, I think we can replace a NumPy array of dicts with a dict of NumPy array if the feature type is `Translation`/`TranslationVariableLanguages` (one array for each language) to get the official PyTorch error message for strings in such case." ]

[ "Hi @margotwagner, thanks for reporting.\r\n\r\nUnfortunately, I'm not able to reproduce your bug: in an environment with datasets-2.3.2 and pyarrow-8.0.0, I can load the datasets without any problem:\r\n```python\r\n>>> ds = load_dataset(\"glue\", \"cola\")\r\n>>> ds\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['sentence', 'label', 'idx'],\r\n num_rows: 8551\r\n })\r\n validation: Dataset({\r\n features: ['sentence', 'label', 'idx'],\r\n num_rows: 1043\r\n })\r\n test: Dataset({\r\n features: ['sentence', 'label', 'idx'],\r\n num_rows: 1063\r\n })\r\n})\r\n\r\n>>> import pyarrow\r\n>>> pyarrow.__version__\r\n8.0.0\r\n>>> from pyarrow.lib import IpcReadOptions\r\n>>> IpcReadOptions\r\npyarrow.lib.IpcReadOptions\r\n```\r\n\r\nI think you may have a problem in your Python environment: maybe you have also an old version of pyarrow that has precedence when importing it.\r\n\r\nCould you please check this (just after you tried to load the dataset and got the error)?\r\n```python\r\n>>> import pyarrow\r\n>>> pyarrow.__version__\r\n``` " ]

[ "Hi @courtneysprouse, thanks for reporting.\r\n\r\nYes, you are right: by default the \"csv\" loader loads all columns as strings. \r\n\r\nYou could tweak this behavior by passing the `feature` argument to `load_dataset`, but it is also true that currently it is not possible to perform some kind of casts, due to lacking of implementation in PyArrow. For example:\r\n```python\r\nimport datasets\r\n\r\nfeatures = datasets.Features(\r\n {\r\n \"tokens\": datasets.Sequence(datasets.Value(\"string\")),\r\n \"ner_tags\": datasets.Sequence(datasets.Value(\"int32\")),\r\n }\r\n)\r\n\r\nnew_conll = datasets.load_dataset(\"csv\", data_files=\"ner_conll.csv\", features=features)\r\n```\r\ngives `ArrowNotImplementedError` error:\r\n```\r\n/usr/local/lib/python3.7/dist-packages/pyarrow/error.pxi in pyarrow.lib.check_status()\r\n\r\nArrowNotImplementedError: Unsupported cast from string to list using function cast_list\r\n```\r\n\r\nOn the other hand, if you just would like to save and afterwards load your dataset, you could use `save_to_disk` and `load_from_disk` instead. These functions preserve all data types.\r\n```python\r\n>>> orig_conll.save_to_disk(\"ner_conll\")\r\n\r\n>>> from datasets import load_from_disk\r\n\r\n>>> new_conll = load_from_disk(\"ner_conll\")\r\n>>> new_conll\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['id', 'tokens', 'pos_tags', 'chunk_tags', 'ner_tags'],\r\n num_rows: 14042\r\n })\r\n validation: Dataset({\r\n features: ['id', 'tokens', 'pos_tags', 'chunk_tags', 'ner_tags'],\r\n num_rows: 3251\r\n })\r\n test: Dataset({\r\n features: ['id', 'tokens', 'pos_tags', 'chunk_tags', 'ner_tags'],\r\n num_rows: 3454\r\n })\r\n})\r\n>>> new_conll[\"train\"][0]\r\n{'chunk_tags': [11, 21, 11, 12, 21, 22, 11, 12, 0],\r\n 'id': '0',\r\n 'ner_tags': [3, 0, 7, 0, 0, 0, 7, 0, 0],\r\n 'pos_tags': [22, 42, 16, 21, 35, 37, 16, 21, 7],\r\n 'tokens': ['EU',\r\n 'rejects',\r\n 'German',\r\n 'call',\r\n 'to',\r\n 'boycott',\r\n 'British',\r\n 'lamb',\r\n '.']}\r\n>>> new_conll[\"train\"].features\r\n{'chunk_tags': Sequence(feature=ClassLabel(num_classes=23, names=['O', 'B-ADJP', 'I-ADJP', 'B-ADVP', 'I-ADVP', 'B-CONJP', 'I-CONJP', 'B-INTJ', 'I-INTJ', 'B-LST', 'I-LST', 'B-NP', 'I-NP', 'B-PP', 'I-PP', 'B-PRT', 'I-PRT', 'B-SBAR', 'I-SBAR', 'B-UCP', 'I-UCP', 'B-VP', 'I-VP'], id=None), length=-1, id=None),\r\n 'id': Value(dtype='string', id=None),\r\n 'ner_tags': Sequence(feature=ClassLabel(num_classes=9, names=['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC', 'B-MISC', 'I-MISC'], id=None), length=-1, id=None),\r\n 'pos_tags': Sequence(feature=ClassLabel(num_classes=47, names=['\"', \"''\", '#', '$', '(', ')', ',', '.', ':', '``', 'CC', 'CD', 'DT', 'EX', 'FW', 'IN', 'JJ', 'JJR', 'JJS', 'LS', 'MD', 'NN', 'NNP', 'NNPS', 'NNS', 'NN|SYM', 'PDT', 'POS', 'PRP', 'PRP$', 'RB', 'RBR', 'RBS', 'RP', 'SYM', 'TO', 'UH', 'VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ', 'WDT', 'WP', 'WP$', 'WRB'], id=None), length=-1, id=None),\r\n 'tokens': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)}\r\n```", "Hi @albertvillanova!\r\n\r\nThanks so much for your suggestions! That worked! " ]

[ "Thanks for reporting, @lewtun.\r\n\r\n~We can't load the dataset locally, so I think this is an issue with the loading script (not the viewer).~\r\n\r\n We are investigating...", "Recently, there was a merged PR related to this dataset:\r\n- #4554\r\n\r\nWe are looking at this...", "Indeed, the above mentioned PR fixed the loading script (it was not working before).\r\n\r\nI'm forcing the refresh of the Viewer.", "Please note that the above mentioned PR also made an enhancement in the `datasets` library, required by this loading script. This enhancement will only be available to the Viewer once we make our next release.", "OK, it's working now.\r\n\r\nhttps://huggingface.co/datasets/wmt16/viewer/ro-en/test\r\n\r\n<img width=\"1434\" alt=\"Capture d’écran 2022-09-08 à 10 15 55\" src=\"https://user-images.githubusercontent.com/1676121/189071665-17d2d149-9b22-42bf-93ac-1a966c3f637a.png\">\r\n", "Thank you @severo !!" ]

[ "Hi @bhavitvyamalik, thanks for reporting.\r\n\r\nYes, currently we do not support 7zip archive compression: I think we should.\r\n\r\nAs a workaround, you could uncompress it explicitly, like done in e.g. `samsum` dataset: \r\n\r\nhttps://github.com/huggingface/datasets/blob/fedf891a08bfc77041d575fad6c26091bc0fce52/datasets/samsum/samsum.py#L106-L110\r\n", "Related to this issue: https://github.com/huggingface/datasets/issues/3541", "Sure, let me look into and check what can be done. Will keep you guys updated here!", "Initially, I thought of solving this without any external dependency. Almost everywhere I saw `lzma` can be used for this but there is a caveat that lzma doesn’t work with 7z archives but only single files. In my case the 7z archive has multiple files so it didn't work. Is it fine to use external library here?", "Hi @bhavitvyamalik, thanks for your investigation.\r\n\r\nOn Monday, I started a PR that will eventually close this issue as well: I'm linking it to this.\r\n- #4672\r\n\r\nLet me know what you think. " ]

[ "I had to use the appropriate token for use_auth_token. Thank you." ]

[ "It seems like a private dataset. The viewer is currently not supported on the private datasets." ]

[]

[ "Hi! I agree we should improve the features equality checks to account for this particular case. However, your code fails due to `answer_start` having the dtype `int64` instead of `int32` after loading from JSON (it's not possible to embed type precision info into a JSON file; `save_to_disk` does that for arrow files), which would lead to the concatenation error as PyArrow does not support this sort of type promotion. This can be fixed as follows:\r\n```python\r\ntemp = load_dataset(\"json\", data_files={\"train\": \"output.jsonl\"}, features=squad[\"train\"].features)\r\n``` ", "That makes sense. I totally missed the `int64` and `int32` part. Thanks for pointing it out! Will close this issue for now." ]

[ "Hi @aleSuglia, thanks for reporting.\r\n\r\nWe are having a look at it. \r\n\r\nWe transfer this issue to the Community tab of the corresponding Hub dataset: https://huggingface.co/datasets/Heriot-WattUniversity/dialog-babi/discussions" ]

[ "I can confirm that if I run one job first that processes the dataset, then I can run any jobs in parallel with no problem (no write-concurrency anymore...). ", "Hi! That's weird. It seems like the error points to the `mkstemp` function, but the official docs state the following:\r\n```\r\nThere are no race conditions in the file’s creation, assuming that the platform properly implements the [os.O_EXCL](https://docs.python.org/3/library/os.html#os.O_EXCL) flag for [os.open()](https://docs.python.org/3/library/os.html#os.open)\r\n```\r\nSo this could mean your platform doesn't support that flag.\r\n\r\n~~Can you please check if wrapping the temp file creation (the line `tmp_file = tempfile.NamedTemporaryFile(\"wb\", dir=os.path.dirname(cache_file_name), delete=False)` in `_map_single`) with the `multiprocess.Lock` fixes the issue?~~\r\nPerhaps wrapping the temp file creation in `_map_single` with `filelock` could work:\r\n```python\r\nwith FileLock(lock_path):\r\n tmp_file = tempfile.NamedTemporaryFile(\"wb\", dir=os.path.dirname(cache_file_name), delete=False)\r\n```\r\nCan you please check if that helps?" ]

[]

[ "Hey, It's already present [here](https://huggingface.co/datasets/squad_v2) ", "Hi! This dataset is indeed already available on the Hub. Closing." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/Amazon-QA)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/simple-wiki)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/QQP_triplets)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/altlex)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/sentence-compression)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/flickr30k-captions)." ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/SPECTER)" ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/PAQ_pairs)" ]

[ "uploaded dataset [here](https://huggingface.co/datasets/embedding-data/WikiAnswers)" ]

[]

[ "Thanks for reporting @lewtun.\r\n\r\nI confirm there is an issue with streaming: it does not stream locally. ", "Oh, after investigation, the source of the issue is in the Hub dataset loading script.\r\n\r\nI'm opening a PR on the Hub dataset.", "I've opened a PR on their Hub dataset to support streaming: https://huggingface.co/datasets/ccdv/pubmed-summarization/discussions/2" ]

[ "Thanks for reporting, @lewtun.\r\n\r\nIt works locally in streaming mode:\r\n```\r\n{'bid': 27681,\r\n 'is_aggregate': True,\r\n 'source': 'cliffnotes',\r\n 'chapter_path': 'all_chapterized_books/27681-chapters/chapters_1_to_2.txt',\r\n 'summary_path': 'finished_summaries/cliffnotes/The Last of the Mohicans/section_1_part_0.txt',\r\n 'book_id': 'The Last of the Mohicans.chapters 1-2',\r\n 'summary_id': 'chapters 1-2',\r\n 'content': None,\r\n 'summary': '{\"name\": \"Chapters 1-2\", \"url\": \"https://web.archive.org/web/20201101053205/https://www.cliffsnotes.com/literature/l/the-last-of-the-mohicans/summary-and-analysis/chapters-12\", \"summary\": \"Before any characters appear, the time and geography are made clear. Though it is the last war that England and France waged for a country that neither would retain, the wilderness between the forces still has to be...\r\n```\r\n\r\nI'm forcing the refresh of the preview. ", "The preview appears as expected once the refresh forced.", "Thank you @albertvillanova 🤗 !" ]

[]

[ "Thanks for reporting @cakiki.\r\n\r\nYes, this is a bug. We are investigating it.", "@albertvillanova Nice! Let me know if it's something I can fix my self; would love to contribtue!", "@cakiki I was working on this but if you would like to contribute, go ahead. I will close my PR. ;)\r\n\r\nFor the moment I just pushed the test (to see if it impacts other tests).", "It impacted the test `test_generator_based_download_and_prepare` and I have fixed this.\r\n\r\nSo that you can copy the test I implemented in my PR and then implement a fix for this issue that passes the test `tests/test_builder.py::test_builder_as_streaming_dataset`.", "Hi @cakiki are you still interested in working on this? Are you planning to open a PR?", "Hi @albertvillanova ! Sorry it took so long; I wanted to spend this weekend working on it." ]

[]

[ "Thanks for reporting, @e-tornike \r\n\r\nSome context:\r\n- #4527 \r\n\r\nThe dataset loads locally in streaming mode:\r\n```python\r\nIn [2]: from datasets import load_dataset; ds = load_dataset(\"vadis/sv-ident\", split=\"validation\", streaming=True); item = next(iter(ds)); item\r\nUsing custom data configuration default\r\nOut[2]: \r\n{'sentence': 'Im Falle von Umweltbelastungen kann selten eindeutig entschieden werden, ob Unbedenklichkeitswerte bereits erreicht oder überschritten sind, die die menschliche Gesundheit oder andere Wohlfahrts»güter« beeinträchtigen.',\r\n 'is_variable': 0,\r\n 'variable': [],\r\n 'research_data': [],\r\n 'doc_id': '51971',\r\n 'uuid': 'ee3d7f88-1a3e-4a59-997f-e986b544a604',\r\n 'lang': 'de'}\r\n```", "~~I have forced the refresh of the split in the preview without success.~~\r\n\r\nI have forced the refresh of the split in the preview, and now it works.", "Preview seems to work now. \r\n\r\nhttps://huggingface.co/datasets/vadis/sv-ident/viewer/default/validation", "OK, thank you @e-tornike.\r\n\r\nApparently, after forcing the refresh, we just had to wait a little until it is effectively refreshed. ", "I'm closing this issue as it was solved after forcing the refresh of the split in the preview.", "Thanks a lot! :)" ]

[ "Could you provide the error details. It is difficult to debug otherwise. Also try other config. `ha` is not a valid." ]

[ "Hi ! `ds.sort()` does sort the full dataset, not just one column:\r\n```python\r\nfrom datasets import *\r\n\r\nds = Dataset.from_dict({\"foo\": [3, 2, 1], \"bar\": [\"c\", \"b\", \"a\"]})\r\nprint(d.sort(\"foo\").to_pandas()\r\n# foo bar\r\n# 0 1 a\r\n# 1 2 b\r\n# 2 3 c\r\n```\r\n\r\nWhat made you think it was not the case ? Did you experience a situation where it was only sorting one column ?", "Hi! thank you for your quick reply!\r\nI wanted to sort the `cnn_dailymail` dataset by the length of the labels (num of characters). I added a new column to the dataset (`ds.add_column`) with the lengths and then sorted by this new column. Only the new length column was sorted, the reset left in their original order. ", "That's unexpected, can you share the code you used to get this ?" ]

[]

[ "yep plus `license_details` also makes sense for this IMO" ]

[ "Hi ! We haven't implemented the conversion from images data to PyTorch tensors yet I think\r\n\r\ncc @mariosasko ", "So I understand:\r\n\r\nset_format() does not properly do the conversion to pytorch tensors from PIL images.\r\n\r\nSo that someone who stumbles on this can use the package:\r\n\r\n```python\r\ndataset = load_dataset(\"mnist\", split=\"train\")\r\ndef transform_func(examples):\r\n examples[\"image\"] = [np.array(img) for img in examples[\"image\"]]\r\n return examples\r\ndataset = dataset.with_transform(transform_func)\r\ndataset[0]\r\n``` ", "This then appears to work with pytorch dataloaders as:\r\n```\r\ndataloader=torch.utils.data.DataLoader(dataset,batch_size=1)\r\n```\r\n\r\nand tensorflow as:\r\n```\r\ndataset=dataset.to_tf_dataset(batch_size=1)\r\n```", "Hi! `set_transform`/`with_transform` is indeed the correct solution for the conversion. Improving this part of the API is one of the things I'm working on currently, so stay tuned!" ]

[]

[ "cc @mariosasko ", "Hi, thanks for reporting! I'm investigating the issue." ]

[ "If you add the line `dataset2.set_format('np')` before calling `dataset2[0]['tmp']` it should return `np.ndarray`.\r\nI believe internally it will not store it as a list, it is only returning a list when you index it.\r\n\r\n```\r\nIn [1]: import datasets, numpy as np\r\nIn [2]: dataset = datasets.load_dataset(\"glue\", \"mrpc\")[\"validation\"]\r\nIn [3]: dataset2 = dataset.map(lambda x: {\"tmp\": np.array([0.5])}, batched=False)\r\nIn [4]: dataset2[0][\"tmp\"]\r\nOut[4]: [0.5]\r\n\r\nIn [5]: dataset2.set_format('np')\r\n\r\nIn [6]: dataset2[0][\"tmp\"]\r\nOut[6]: array([0.5])\r\n```", "I see, thanks! Any idea if the default numpy → list conversion might cause precision loss?", "I'm not super familiar with our datasets works internally, but I think your `np` array will be stored in a `pyarrow` format, and then you take a view of this as a python array. In which case, I think the precision should be preserved." ]

[ "Hi! The `imagefolder` script is already quite complex, so a standalone script sounds better. Also, I suggest we create an org on the Hub (e.g. `hf-loaders`) and store such scripts there for easier maintenance rather than having them as packaged modules (IMO only very generic loaders should be packaged). WDYT @lhoestq @albertvillanova @polinaeterna?", "@mariosasko sounds good to me!\r\n", "Thank you for the suggestion @mariosasko . I agree with the point, but I have a few doubts\r\n\r\n1. How would the user access the script if it's not a part of the core codebase?\r\n2. Could you direct me as to what will be the tasks I have to do to contribute to the code? As per my understanding, it would be like\r\n 1. Create a new org \"hf-loaders\" and add you (and more HF people) to the org\r\n 2. Add data loader script as a (model?)\r\n 3. Test it with a dataset on HF hub\r\n3. We should maybe brainstorm as to which public datasets have this format (YOLO type) and are the most important ones to test the script with. We can even add the datasets on HF Hub alongside the script", "1. Like this: `load_dataset(\"hf-loaders/yolo\", data_files=...)`\r\n2. The steps would be:\r\n 1. Create a new org `hf-community-loaders` (IMO a better name than \"hf-loaders\") and add me (as an admin)\r\n 2. Create a new dataset repo `yolo` and add the loading script to it (`yolo.py`)\r\n 3. Open a discussion to request our review\r\n4. I like this idea. Another option is to add snippets that describe how to load such datasets using the `yolo` loader." ]