Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "Thanks for reporting.\r\n\r\nApparently this is a versioning issue: the lib downloaded the `bleurt` script from the master branch where we did this change recently. We'll fix that in a new release this week or early next week. Cc @thomwolf \r\n\r\nUntil that, I'd suggest you to download the right bleurt folder from github ([this one](https://github.com/huggingface/nlp/tree/0.4.0/metrics/bleurt)) and do\r\n\r\n```python\r\nfrom nlp import load_metric\r\n\r\nbleurt = load_metric(\"path/to/bleurt/folder\")\r\n```\r\n\r\nTo download it you can either clone the repo or download the `bleurt.py` file and place it in a folder named `bleurt` ", "Actually we can fix this on our side, this script didn't had to be updated. I'll do it in a few minutes" ]

[ "From my limited understanding, part of the issue seems related to the `prepare_module` and `download_and_prepare` functions each handling the case where no config is passed. For example, `prepare_module` does mutate the object passed and forces the flags `extract_compressed_file` and `force_extract` to `True`.\r\n\r\nSee:\r\n* https://github.com/huggingface/nlp/blob/5fb61e1012bda724a9b6b847307d90a1380abfa5/src/nlp/load.py#L227\r\n* https://github.com/huggingface/nlp/blob/5fb61e1012bda724a9b6b847307d90a1380abfa5/src/nlp/builder.py#L388\r\n\r\nMaybe a cleaner solution would be to always instantiate a default `DownloadConfig` object at the top-level, have it as non-optional for the lower-level functions and treat it as immutable. ", "Thanks for the report, I'll take a look.\r\n\r\nWhat is your specific use-case for providing a DownloadConfig object?\r\n", "Thanks. Our use case involves running a training job behind a corporate firewall with no access to any external resources (S3, GCP or other web resources).\r\n\r\nI was thinking about a 2-steps process:\r\n1) Download the resources / artifacts using some secure corporate channel, ie run `nlp.load_dataset()` without a specific `DownloadConfig`. After that, collect the files from the `$HF_HOME` folder\r\n2) Copy the `$HF_HOME` folder in the firewalled environment. Run `nlp.load_dataset()` with a custom config `DownloadConfig(local_files_only=True)`\r\n\r\nHowever this ends up a bit clunky in practice, even when solving the `DownloadConfig` issue above. For example, the `filename` hash computed in `get_from_cache()` differs in the `local_files_only=False` vs `local_files_only=True` case (local case defaults `etag` to `None`, which results in a different hash). So effectively step 2) above doesn't work because the hash computed differs from the hash in the cache folder. Some hacks / workaround are possible but this solution becomes very convoluted.\r\nhttps://github.com/huggingface/nlp/blob/c214aa5a4430c1df1bcd0619fd94d6abdf9d2da7/src/nlp/utils/file_utils.py#L417\r\n\r\nWould you recommend a different path?\r\n", "I see.\r\n\r\nProbably the easiest way for you would be that we add simple serialization/deserialization methods to the Dataset and DatasetDict objects once the data files have been downloaded and all the dataset is processed.\r\n\r\nWhat do you think @lhoestq ?", "This use-case will be solved with #571 ", "Thank you very much @thomwolf and @lhoestq we will give it a try" ]

[ "Indeed this is a known issue arising from the fact that we try to be compatible with cloupickle.\r\n\r\nDoes this also happen if you are installing in a virtual environment?", "> Indeed this is a know issue with the fact that we try to be compatible with cloupickle.\r\n> \r\n> Does this also happen if you are installing in a virtual environment?\r\n\r\nThen it would work, because the package is in a writable path.", "If it's fine for you then this is the recommended way to solve this issue.", "> If it's fine for you then this is the recommended way to solve this issue.\r\n\r\nI don't want to use a virtual environment, because Nix is fully reproducible, and virtual environments are not. And I am the maintainer of the `transformers` in nixpkgs, so sooner or later I will have to package `nlp`, since it is becoming a dependency of `transformers` ;).", "Ok interesting. We could have another check to see if it's possible to download and import the datasets script at another location than the module path. I think this would probably involve tweaking the python system path dynamically.\r\n\r\nI don't know anything about Nix so if you want to give this a try your self we can guide you or you can give us more information on your general project and how this works.\r\n\r\nRegarding `nlp` and `transformers`, we are not sure `nlp` will become a required dependency for `transformers`. It will probably be used a lot in the examples but I think it probably won't be a required dependency for the main package since we try to keep it as light as possible in terms of deps.\r\n\r\nHappy to help you make all these things work better for your use-case ", "@danieldk modules are now installed in a different location (by default in the cache directory of the lib, in `~/.cache/huggingface/modules`). You can also change that using the environment variable `HF_MODULES_PATH`\r\n\r\nFeel free to play with this change from the master branch for now, and let us know if it sounds good for you :)\r\nWe plan to do a release in the next coming days", "Awesome! I’ll hopefully have some time in the coming days to try this.", "> Feel free to play with this change from the master branch for now, and let us know if it sounds good for you :)\r\n> We plan to do a release in the next coming days\r\n\r\nThanks for making this change! I just packaged the latest commit on master and it works like a charm now! :partying_face: " ]

[ "When you load a text file for the first time with `nlp`, the file is converted into Apache Arrow format. Arrow allows to use memory-mapping, which means that you can load an arbitrary large dataset.\r\n\r\nNote that as soon as the conversion has been done once, the next time you'll load the dataset it will be much faster.\r\n\r\nHowever for a 1TB dataset, the conversion can indeed take time. You could try to load parts of it in parallel, and then use `nlp.concatenate_datasets` to get your full dataset.", "Humm, we can give a look at these large scale datasets indeed.\r\n\r\nDo you mind sharing a few stats on your dataset so I can try to test on a similar one?\r\n\r\nIn particular some orders of magnitudes for the number of files, number of lines per files, line lengths.", "@lhoestq Yes, I understand that the first time requires more time. The concatenate_datasets seems to be a workaround, but I believe a multi-processing method should be integrated into load_dataset to make it easier and more efficient for users.\r\n\r\n@thomwolf Sure, here are the statistics:\r\nNumber of lines: 4.2 Billion\r\nNumber of files: 6K\r\nNumber of tokens: 800 Billion\r\nThe number of lines is distributed equally across these 6k files.\r\nThe line length varies between 100 tokens to 40k tokens.\r\n", "@agemagician you can give a try at a multithreaded version if you want (currently on the #548).\r\n\r\nTo test it, you just need to copy the new `text` processing script which is [here](https://github.com/huggingface/nlp/blob/07d92a82b7594498ff702f3cca55c074e2052257/datasets/text/text.py) somewhere on your drive and give it's local path instead of `text` to `load_dataset`. E.g. in your example:\r\n```python\r\ntrain_files = glob.glob(\"xxx/*.txt\",recursive=True)\r\nrandom.shuffle(train_files)\r\n\r\nprint(train_files)\r\n\r\ndataset = nlp.load_dataset('./datasets/text.py', # path to where you've dowloaded the multi-threaded text loading script\r\n data_files=train_files,\r\n name=\"customDataset\",\r\n version=\"1.0.0\",\r\n cache_dir=\"xxx/nlp\")\r\n```", "I have already generated the dataset, but now I tried to reload it and it is still very slow.\r\n\r\nI also have installed your commit and it is slow, even after the dataset was already generated.\r\n`pip install git+https://github.com/huggingface/nlp.git@07d92a82b7594498ff702f3cca55c074e2052257`\r\n\r\nIt uses only a single thread.\r\n\r\nDid I miss something ?", "As mentioned in #548 , each time you call `load_dataset` with `data_files=`, they are hashed to get the cache directory name. Hashing can be too slow with 1TB of data. I feel like we should have a faster way of getting a hash that identifies the input data files", "I believe this is really a very important feature, otherwise, we will still have the issue of too slow loading problems even if the data cache generation is fast.", "Hmm ok then maybe it's the hashing step indeed.\r\n\r\nLet's see if we can improve this as well.\r\n\r\n(you will very likely have to regenerate your dataset if we change this part of the lib though since I expect modifications on this part of the lib to results in new hashes)", "Also, @agemagician you have to follow the step I indicate in my previous message [here](https://github.com/huggingface/nlp/issues/546#issuecomment-684648927) to use the new text loading script.\r\n\r\nJust doing `pip install git+https://github.com/huggingface/nlp.git@07d92a82b7594498ff702f3cca55c074e2052257` like you did won't use the new script (they are not inside the library but hosted on our hub).", "No problem, I will regenerate it. This will make us see if we solved both issues and now both the data generation step, as well as the hashing step, is fast.", "Any news for the hashing ?", "I'm working on it today :)", "Ok so now the text files won't be hashed.\r\n\r\nI also updated #548 to include this change.\r\nLet us know if it helps @agemagician :)", "Perfect thanks for your amazing work.", "Right now, for caching 18Gb data, it is taking 1 hour 10 minute. Is that proper expected time? @lhoestq @agemagician \r\nIn this rate (assuming large file will caching at the same rate) caching full mC4 (27TB) requires a month (~26 days). \r\n", "Hi ! Currently it is that slow because we haven't implemented parallelism for the dataset generation yet.\r\nThough we will definitely work on this :)\r\n\r\nFor now I'd recommend loading the dataset shard by shard in parallel, and then concatenate them:\r\n```python\r\n# in one process, load first 100 files for english\r\nshard1 = load_dataset(\"allenai/c4\", data_files=\"multilingual/c4-en.tfrecord-000**.json.gz\")\r\n# in another process load next 100 files for english\r\nshard2 = load_dataset(\"allenai/c4\", data_files=\"multilingual/c4-en.tfrecord-001**.json.gz\")\r\n\r\n# finally\r\nconcatenate_datasets([shard1, shard2, ...])", "Thanks for the help..!!!", "Sorry to write on a closed issue but, has there been any progress on parallelizing the `load_dataset` function?", "Hi ! No but this is in our plans (probably a few weeks)", "I'm literally crying waiting for the trainer to restart from checkpoint. It's getting stuck at `get_train_dataloader` and I think this is to do with the same issue... has there been any progress on this?", "> I'm literally crying waiting for the trainer to restart from checkpoint. It's getting stuck at get_train_dataloader and I think this is to do with the same issue...\r\n\r\nOnce the dataset is cached once, it's not regenerated again. Your issue seems different", "hmmm, yes. I'll come back with details on this, fairly easy to reproduce. Takes about 30 minutes to get from checkpoint loading to starting training...", "@lhoestq yo, any news in making it possible to download large datasets faster?", "@lhoestq For some reason [setting num_proc](https://discuss.huggingface.co/t/how-can-i-multithreadedly-download-a-huggingface-dataset/56178/2?u=kopyl) does not work at all... My dataset has 58 parquet files and i was hoping passing `num_proc` to `load_dataset` would spawn 58 Python processes each downloading its own parquet so I can load my dataset in 1 minutes instead of 50...", "It does spawn `num_proc` processes. Note that when you download in parallel you're often bounded by your bandwidth at one point, so 50 processes is unlikely to get you a x50 download speed up but a bit less", "> It does spawn `num_proc` processes. Note that when you download in parallel you're often bounded by your bandwidth at one point, so 50 processes is unlikely to get you a x50 download speed up but a bit less\r\n\r\nwhy then i see only 1 parquet file download progress bar at a time?", "Ah indeed parallel downloads are not enabled yet for some datasets.\r\nI opened https://github.com/huggingface/datasets/pull/6551 to fix this.\r\n", "@lhoestq thank you very much, i though for a second i'm just tripping." ]

[ "Update: release is planed mid-next week." ]

[ "I'll take a look!" ]

[ "Docs that explain how to train a tokenizer with `datasets` are available here: https://huggingface.co/docs/tokenizers/training_from_memory#using-the-datasets-library" ]

[ "Hi @gaguilar \r\n\r\nIf you want to take care of this, it very simple, you just need to regenerate the `dataset_infos.json` file as indicated [in the doc](https://huggingface.co/nlp/share_dataset.html#adding-metadata) by [installing from source](https://huggingface.co/nlp/installation.html#installing-from-source) and running the following command from the root of the repo:\r\n```bash\r\npython nlp-cli test ./datasets/lince --save_infos --all_configs\r\n```\r\nAnd then you can open a pull-request with the updated json file.\r\n\r\nOtherwise we'll do it sometime this week.", "Hi @thomwolf \r\n\r\nThanks for the details! I just created a PR with the updated `dataset_infos.json` file (#550).", "Thanks for updating the json file. Closing this one" ]

[ "`NonMatchingChecksumError` means that the checksum of the downloaded file is not the expected one.\r\nEither the file you downloaded was corrupted along the way, or the host updated the file.\r\nCould you try to clear your cache and run `load_dataset` again ? If the error is still there, it means that there was an update in the data, and we may have to update the expected checksum value.", "I just cleared the cache an run it again. The error persists ):\r\n\r\n```\r\n nlp (master) $ rm -rf /Users/abarbosa/.cache/huggingface/\r\n nlp (master) $ python\r\nPython 3.8.5 (default, Aug 5 2020, 03:39:04)\r\n[Clang 10.0.0 ] :: Anaconda, Inc. on darwin\r\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\r\n>>> import nlp\r\n>>> dataset = nlp.load_dataset(\"race\")\r\nDownloading: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4.39k/4.39k [00:00<00:00, 661kB/s]\r\nDownloading: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.81k/1.81k [00:00<00:00, 644kB/s]\r\nUsing custom data configuration default\r\nDownloading and preparing dataset race/default (download: 84.52 MiB, generated: 132.61 MiB, post-processed: Unknown size, total: 217.13 MiB) to /Users/abarbosa/.cache/huggingface/datasets/race/default/0.1.0/5461327f1a83549ca0d845a3159c806d2baf4f8d0d8f7d657157ce7cdf3899c2...\r\nDownloading: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 25.4M/25.4M [01:03<00:00, 401kB/s]\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/Users/abarbosa/Documents/nlp/src/nlp/load.py\", line 550, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/Users/abarbosa/Documents/nlp/src/nlp/builder.py\", line 471, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/Users/abarbosa/Documents/nlp/src/nlp/builder.py\", line 530, in _download_and_prepare\r\n verify_checksums(\r\n File \"/Users/abarbosa/Documents/nlp/src/nlp/utils/info_utils.py\", line 38, in verify_checksums\r\n raise NonMatchingChecksumError(error_msg + str(bad_urls))\r\nnlp.utils.info_utils.NonMatchingChecksumError: Checksums didn't match for dataset source files:\r\n['http://www.cs.cmu.edu/~glai1/data/race/RACE.tar.gz']\r\n>>>\r\n```", "Dealing with the same issue please update the checksum on nlp library end. The data seems to have changed on their end.", "We have a discussion on this datasets here: https://github.com/huggingface/nlp/pull/540\r\n\r\nFeel free to participate if you have some opinion on the scope of data which should be included in this dataset.", "At least for me, the file that was downloaded from CMU isn't the complete dataset, but a small subset of it (~25MB vs ~85MB). I've previously downloaded the dataset directly, so for my personal needs I could just swap out the corrupted file with the correct one. Perhaps you could host it like you do for the Wikipedia and BookCorpus datasets.\r\n\r\n", "> At least for me, the file that was downloaded from CMU isn't the complete dataset, but a small subset of it (~25MB vs ~85MB). I've previously downloaded the dataset directly, so for my personal needs I could just swap out the corrupted file with the correct one. Perhaps you could host it like you do for the Wikipedia and BookCorpus datasets.\r\n\r\nCould you upload this please?", "> > At least for me, the file that was downloaded from CMU isn't the complete dataset, but a small subset of it (~25MB vs ~85MB). I've previously downloaded the dataset directly, so for my personal needs I could just swap out the corrupted file with the correct one. Perhaps you could host it like you do for the Wikipedia and BookCorpus datasets.\r\n> \r\n> Could you upload this please?\r\n\r\nNot sure if I can upload it according to their license (\"You agree not to reproduce, duplicate, copy, sell, trade, resell or exploit for any commercial purpose, any portion of the contexts and any portion of derived data.\").", "I managed to fix it in #540 :)", "Closing since @540 is merged\r\n\r\nThanks again @abarbosa94 " ]

[ "Thanks for reporting !\r\nThis has been fixed in #475 and the fix will be available in the next release", "What you can do instead to get the list of the datasets is call\r\n\r\n```python\r\nprint([dataset.id for dataset in nlp.list_datasets()])\r\n```", "Thanks @lhoestq . " ]

[ "I am facing probably facing similar issues with \r\n\r\n`wiki40b_en_100_0`", "Could you try to run `dataset = load_dataset(\"text\", data_files=file_path, split=\"train\")` once before calling the script ?\r\n\r\nIt looks like several processes try to create the dataset in arrow format at the same time. If the dataset is already created it should be fine", "Thanks! I tested on 328MB text data on `n1-standard-8 (8 vCPUs, 30 GB memory)`. The main script ran without any issue, but it seems to require a huge space in the drive.\r\n\r\nAs suggested, I ran the following script before running the pre-training command with `xla_spawn.py`.\r\n\r\n```python\r\nfrom nlp import load_dataset\r\n\r\nfile_path=\"your_file_name\"\r\nload_dataset(\"text\", data_files=file_path, split=\"train\")\r\n```\r\nThis will create `text-train.arrow` under the default cache directory. Then, I run the script with `xla_spawn.py`. It will load data from the cached file. My understanding is that there's no other way but to do this two-step process with the current version (0.4) of `nlp`.\r\n\r\nDuring another caching process that happens in the main script:\r\n\r\n```\r\n08/26/2020 09:19:51 - INFO - nlp.utils.info_utils - All the checksums matched successfully for post processing resources\r\n08/26/2020 09:19:53 - INFO - nlp.arrow_dataset - Caching processed dataset at /home/*****/.cache/huggingface/datasets/text/default-b0932b2bdbb63283/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d/cache-f90f341e5308a7469\r\n8d872bcc88f9c0e.arrow\r\n```\r\n\r\n`nlp` generates a temporary file per core, each of which is three times larger than the original text data. If each process is actually writing on the disk, you will need a huge amount of space in your drive. (Maybe I'm missing something.)\r\n\r\n```\r\n-rw-r--r-- 1 ***** ***** 674 Aug 26 09:19 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 09:19 LICENSE\r\n-rw-r--r-- 1 ***** ***** 332M Aug 26 09:10 text-train.arrow\r\n-rw------- 1 ***** ***** 940M Aug 26 09:31 tmp0k43sazw\r\n-rw------- 1 ***** ***** 940M Aug 26 09:31 tmp7sxs9mj5\r\n-rw------- 1 ***** ***** 939M Aug 26 09:31 tmpbbiqw2vp\r\n-rw------- 1 ***** ***** 937M Aug 26 09:31 tmpjxb5ptyu\r\n-rw------- 1 ***** ***** 933M Aug 26 09:31 tmpk3hkdh0e\r\n-rw------- 1 ***** ***** 944M Aug 26 09:31 tmpnoalwftz\r\n-rw------- 1 ***** ***** 931M Aug 26 09:31 tmpuxdr_dz3\r\n-rw------- 1 ***** ***** 945M Aug 26 09:31 tmpxjyuy6dk\r\n```\r\nAfter the caching process, they seem to be merged into one file.\r\n\r\n```\r\n-rw------- 1 ***** ***** 989M Aug 26 09:32 cache-f90f341e5308a74698d872bcc88f9c0e.arrow\r\n-rw-r--r-- 1 ***** ***** 674 Aug 26 09:19 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 09:19 LICENSE\r\n-rw-r--r-- 1 ***** ***** 332M Aug 26 09:10 text-train.arrow\r\n```", "Again it looks like every process tries to tokenize the full dataset at the same time.\r\nIf you do the tokenization before calling `xla_spawn.py` once, then each process will then use the tokenized cached file `cache-f90f341e5308a74698d872bcc88f9c0e.arrow` and not recompute it.\r\n\r\nNot sure if there's a better way to do that cc @julien-c @thomwolf ", "I wrote a separate script just for preparing a cached file, including tokenization. Each process did use the tokenized cached file.\r\n\r\nCurrently I'm testing the pipeline on 24GB text data. It took about 1.5 hour to create a cached file on `n1-highmem-16 (16 vCPUs, 104 GB memory)`. I assume loading this cached file in the main script with `xla_spawn.py` won't be an issue (even if there are 8 processes).\r\n\r\n```\r\ntotal 98G\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 26 13:38 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 26 12:24 ..\r\n-rw------- 1 ***** ***** 74G Aug 26 13:38 cache-a7aa04134ba7b1aff5d9710f14a4e334.arrow\r\n-rw-r--r-- 1 ***** ***** 681 Aug 26 12:24 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 12:24 LICENSE\r\n-rw-r--r-- 1 ***** ***** 25G Aug 26 12:24 text-train.arrow\r\n```", "Yes loading the cached file should be fine from different processes", "Sorry, I thought it was working, but actually the second call doesn't use the cached file that was generated separately, and it will generate another cache-****.arrorw file with a different name. If I run the training script again (with `xla_spawn.py`), it will use the second cached file, which was generated by the training script itself in the previous run.\r\n\r\n```\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 26 15:35 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 26 15:29 ..\r\n-rw------- 1 ***** ***** 99M Aug 26 15:35 cache-0d77dfce704493dbe63f071eed6a5431.arrow\r\n-rw------- 1 ***** ***** 99M Aug 26 15:29 cache-69633651476e943b93c89ace715f9487.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 26 15:33 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 15:33 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 26 15:29 text-train.arrow\r\n```", "So if I understand correctly it means that the cached file generated by your separated script is different by the one used by the training script ?", "Yes.\r\n\r\n1. `cache-69633651476e943b93c89ace715f9487.arrow` was generated with a separate script. \r\n2. I ran the entire script with `xla_spawn.py`.\r\n3. `cache-69633651476e943b93c89ace715f9487.arrow` is not used.\r\n4. `cache-0d77dfce704493dbe63f071eed6a5431.arrow` is created.\r\n5. training starts...\r\n\r\nNow, if I kill the process at step 5, and do the step 2 again, it will use `cache-0d77dfce704493dbe63f071eed6a5431.arrow` (cached file created at step 4) without any issue.\r\n\r\nI used the following to generate the first cached file.\r\n```python\r\ndataset = load_dataset(\"text\", data_files=file_path, split=\"train\")\r\ndataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=args.block_size), batched=True)\r\ndataset.set_format(type='torch', columns=['input_ids'])\r\n```", "1. Here's the log from the first step.\r\n```\r\nDownloading and preparing dataset text/default-e84dd29acc4ad9ef (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/\r\n447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d...\r\nDataset text downloaded and prepared to /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d. Subsequent calls will reuse this data.\r\n```\r\nThere's a file named `cache-7b1440ba7077af0f0d9035b5a55d01fc.arrow`, so it did create a cached file.\r\n```\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 26 15:59 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 26 15:58 ..\r\n-rw------- 1 ***** ***** 99M Aug 26 15:59 cache-7b1440ba7077af0f0d9035b5a55d01fc.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 26 15:58 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 15:58 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 26 15:58 text-train.arrow\r\n```\r\n2. Ideally, `cache-7b1440ba7077af0f0d9035b5a55d01fc.arrow` should be used in `run_language_modeling.py` (modified version using `nlp`) with `xla_spawn.py`. But it looks like it's creating a new cached file.\r\n\r\n```\r\n08/26/2020 16:13:03 - INFO - filelock - Lock 139635836351096 released on /home/*****/.cache/huggingface/datasets/3e34209a2741375a1db1ff03bf1abba1a9bd0e6016912d3ead0114b9d1ca2685.202fa4f84f552bff1f5400ae012663839c61efb3de068c6c8722d34ac0ea6192\r\n.py.lock\r\n08/26/2020 16:13:03 - WARNING - nlp.builder - Using custom data configuration default\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Overwrite dataset info from restored data version.\r\n08/26/2020 16:13:03 - INFO - nlp.info - Loading Dataset info from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Reusing dataset text (/home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d)\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Constructing Dataset for split train, from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:13:03 - INFO - nlp.utils.info_utils - All the checksums matched successfully for post processing resources\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Overwrite dataset info from restored data version.\r\n08/26/2020 16:13:03 - INFO - nlp.info - Loading Dataset info from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Reusing dataset text (/home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d)\r\n08/26/2020 16:13:03 - INFO - nlp.builder - Constructing Dataset for split train, from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:13:03 - INFO - nlp.utils.info_utils - All the checksums matched successfully for post processing resources\r\n08/26/2020 16:13:05 - INFO - nlp.arrow_dataset - Caching processed dataset at /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d/cache-0d77dfce704493dbe\r\n63f071eed6a5431.arrow\r\n^M 0%| | 0/100 [00:00<?, ?it/s]08/26/2020 16:13:05 - INFO - nlp.arrow_dataset - Caching processed dataset at /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6\r\nfe661fe4d070d380d/cache-0d77dfce704493dbe63f071eed6a5431.arrow\r\n```\r\n\r\nThere are two cached files in the directory:\r\n\r\n```\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 26 16:14 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 26 15:58 ..\r\n-rw------- 1 ***** ***** 99M Aug 26 16:14 cache-0d77dfce704493dbe63f071eed6a5431.arrow\r\n-rw------- 1 ***** ***** 99M Aug 26 15:59 cache-7b1440ba7077af0f0d9035b5a55d01fc.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 26 16:13 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 26 16:13 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 26 15:58 text-train.arrow\r\n```\r\n\r\nIf I kill the process, and run it again, it will use the second cached file.\r\n\r\n```\r\n08/26/2020 16:19:52 - WARNING - nlp.builder - Using custom data configuration default\r\n08/26/2020 16:19:52 - INFO - nlp.builder - Overwrite dataset info from restored data version.\r\n08/26/2020 16:19:52 - INFO - nlp.info - Loading Dataset info from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:19:52 - INFO - nlp.builder - Reusing dataset text (/home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d)\r\n08/26/2020 16:19:52 - INFO - nlp.builder - Constructing Dataset for split train, from /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\n08/26/2020 16:19:52 - INFO - nlp.utils.info_utils - All the checksums matched successfully for post processing resources\r\n08/26/2020 16:19:53 - INFO - nlp.arrow_dataset - Loading cached processed dataset at /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d/cache-0d77dfce70\r\n4493dbe63f071eed6a5431.arrow\r\n08/26/2020 16:19:53 - INFO - nlp.arrow_dataset - Set __getitem__(key) output type to torch for ['input_ids'] columns (when key is int or slice) and don't output other (un-formatted) columns.\r\n```", "Thanks for all the details.\r\nThe two cached files are supposed to be the same. I suspect that the caching has a problem with the tokenizer.\r\nWhich tokenizer did you use ?", "I trained a byte-level BPE tokenizer on my data with `tokenziers` library following this [example](https://github.com/huggingface/tokenizers/blob/master/bindings/python/examples/train_bytelevel_bpe.py).\r\n\r\nAnd I put these model files in a directory named `\"model_name\"`. I also put config.json, which is the original RoBERTa config file.\r\n\r\n```bash\r\n%ls model_name\r\nconfig.json merges.txt vocab.json\r\n```\r\n\r\n[This](https://github.com/huggingface/transformers/blob/4bd7be9a4268221d2a0000c7e8033aaeb365c03b/examples/language-modeling/run_language_modeling.py#L196) is the line where `run_language_modeling.py` loads the tokenier.\r\n\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)\r\n```\r\n\r\nI use `\"model_name\"` for `model_args.tokenizer_name`. I don't specify `model_args.cache_dir`. It is 'None' by default.", "In my separated script for caching, I'm using `use_fast=True` when initializing a tokenizer.\r\n\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(args.config_name, use_fast=True)\r\n```\r\nI wasn't using that option in the main script. That could be the reason...", "Yea it could definitely explain why you have two different cache files.\r\nLet me know if using the same tokenizers on both sides fixes the issue", "It still creates a new file even if I remove `use_fast=True`... \r\n\r\nHere's the script used to create a cached file.\r\n```python \r\n#!/usr/bin/env python3\r\n\r\nimport argparse\r\n\r\nfrom transformers import AutoTokenizer\r\n\r\nfrom nlp import load_dataset\r\n\r\n\r\ndef main():\r\n parser = argparse.ArgumentParser(description='description')\r\n parser.add_argument('--config_name', type=str, help='Pretrained config name or path if not the same as model_name')\r\n parser.add_argument('--data_file', type=str, help='The input data file (a text file).')\r\n parser.add_argument('--block_size', type=int, default=-1, help='The training dataset will be truncated in block of this size for training')\r\n args = parser.parse_args()\r\n\r\n tokenizer = AutoTokenizer.from_pretrained(args.config_name)\r\n\r\n dataset = load_dataset(\"text\", data_files=args.data_file, split=\"train\")\r\n dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=args.block_size), batched=True)\r\n dataset.set_format(type='torch', columns=['input_ids'])\r\n\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n```\r\n\r\nHere's how the data is loaded in the modified `run_language_modeling.py`. [[original function](https://github.com/huggingface/transformers/blob/971d1802d009d9996b36a34a34477cee849ef39f/examples/language-modeling/run_language_modeling.py#L128-L135)]\r\n\r\n```python\r\ndef get_dataset(args: DataTrainingArguments, tokenizer: PreTrainedTokenizer, evaluate=False):\r\n file_path = args.eval_data_file if evaluate else args.train_data_file\r\n split = \"validation\" if evaluate else \"train\"\r\n if args.line_by_line:\r\n # return LineByLineTextDataset(tokenizer=tokenizer, file_path=file_path, block_size=args.block_size)\r\n dataset = load_dataset(\"text\", data_files=file_path, split=\"train\")\r\n dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=args.block_size), batched=True)\r\n dataset.set_format(type='torch', columns=['input_ids'])\r\n return dataset\r\n\r\n else:\r\n return TextDataset(\r\n tokenizer=tokenizer, file_path=file_path, block_size=args.block_size, overwrite_cache=args.overwrite_cache\r\n )\r\n```\r\n\r\nProbably I don't need this part in the main script,\r\n\r\n```python\r\ndataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=args.block_size), batched=True)\r\n dataset.set_format(type='torch', columns=['input_ids'])\r\n```\r\nand simply do this?\r\n```python\r\ndataset = load_dataset(\"text\", data_files=file_path, split=\"train\")\r\nreturn dataset\r\n```", "You need this part in the main script or it will use the dataset that is not tokenized\r\n\r\n", "I can see that the tokenizer in `run_language_modeling.py` is not instantiated the same way as in your separated script.\r\nIndeed we can see L196:\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)\r\n```\r\nCould you try to make it so they are instantiated the exact same way please ?", "I updated my separated script, but it's creating a cached file again. If I don't use the `model_args.cache_dir`, both will get `None`, so they should be the same.\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nimport argparse\r\n\r\nfrom transformers import AutoTokenizer\r\nfrom nlp import load_dataset\r\n\r\ndef main():\r\n parser = argparse.ArgumentParser(description='description')\r\n parser.add_argument('--tokenizer_name', type=str, help='Pretrained tokenizer name or path if not the same as model_name')\r\n parser.add_argument('--data_file', type=str, help='The input data file (a text file).')\r\n parser.add_argument('--cache_dir', type=str, default=None, help='Where do you want to store the pretrained models downloaded from s3')\r\n parser.add_argument('--block_size', type=int, default=-1, help='The training dataset will be truncated in block of this size for training')\r\n\r\n model_args = parser.parse_args()\r\n\r\n tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)\r\n\r\n dataset = load_dataset(\"text\", data_files=model_args.data_file, split=\"train\")\r\n dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=model_args.block_size), batched=True)\r\n dataset.set_format(type='torch', columns=['input_ids'])\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n```\r\n\r\nIs there a way to specify the cache file to load, and skip the re-computation?", "Could you also check that the `args.block_size` used in the lambda function is the same as well ?", "Here's a minimal working example to reproduce this issue.\r\n\r\nAssumption:\r\n- You have access to TPU.\r\n- You have installed `transformers` and `nlp`.\r\n- You have tokenizer files (`config.json`, `merges.txt`, `vocab.json`) under the directory named `model_name`.\r\n- You have `xla_spawn.py` (Download from https://github.com/huggingface/transformers/blob/master/examples/xla_spawn.py).\r\n- You have saved the following script as `prepare_cached_dataset.py`.\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nimport argparse\r\nfrom transformers import AutoTokenizer\r\nfrom nlp import load_dataset\r\n\r\ndef main():\r\n parser = argparse.ArgumentParser(description='description')\r\n parser.add_argument('--tokenizer_name', type=str, help='Pretrained tokenizer name or path if not the same as model_name')\r\n parser.add_argument('--data_file', type=str, help='The input data file (a text file).')\r\n parser.add_argument('--cache_dir', type=str, default=None, help='Where do you want to store the pretrained models downloaded from s3')\r\n parser.add_argument('--block_size', type=int, default=-1, help='The training dataset will be truncated in block of this size for training')\r\n parser.add_argument('--tpu_num_cores', type=int, default=1, help='Number of TPU cores to use (1 or 8). For xla_apwan.py')\r\n model_args = parser.parse_args()\r\n \r\n tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=True)\r\n \r\n dataset = load_dataset(\"text\", data_files=model_args.data_file, split=\"train\")\r\n dataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=model_args.block_size), batched=True)\r\n dataset.set_format(type='torch', columns=['input_ids'])\r\n\r\ndef _mp_fn(index):\r\n # For xla_spawn (TPUs)\r\n main()\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n```\r\n\r\n- Run the following command. Replace `your_training_data` with some text file.\r\n\r\n```bash\r\nexport TRAIN_DATA=your_training_data\r\n\r\npython prepare_cached_dataset.py \\\r\n--tokenizer_name=model_name \\\r\n--block_size=512 \\\r\n--data_file=$TRAIN_DATA\r\n```\r\n- Check the cached directory.\r\n```bash\r\nls -lha /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\ntotal 132M\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 28 13:08 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 28 13:08 ..\r\n-rw------- 1 ***** ***** 99M Aug 28 13:08 cache-bfc7cb0702426d19242db5e8c079f04b.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 28 13:08 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 28 13:08 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 28 13:08 text-train.arrow\r\n```\r\n\r\n- Run the same script again. (The output should be just `Using custom data configuration default`.)\r\n```\r\npython prepare_cached_dataset.py \\\r\n--tokenizer_name=model_name \\\r\n--block_size=512 \\\r\n--data_file=$TRAIN_DATA\r\n```\r\n- Check the cached directory.\r\n```bash\r\nls -lha /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\ntotal 132M\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 28 13:08 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 28 13:08 ..\r\n-rw------- 1 ***** ***** 99M Aug 28 13:08 cache-bfc7cb0702426d19242db5e8c079f04b.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 28 13:20 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 28 13:20 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 28 13:08 text-train.arrow\r\n```\r\n- The cached file (`cache-bfc7cb0702426d19242db5e8c079f04b.arrow`) is reused.\r\n- Now, run this script with `xla_spawn.py`. Ideally, it should reuse the cached file, however, you will see each process is creating a cache file again.\r\n\r\n```bash\r\npython xla_spawn.py --num_cores 8 \\\r\nprepare_cached_dataset.py \\\r\n--tokenizer_name=model_name \\\r\n--block_size=512 \\\r\n--data_file=$TRAIN_DATA\r\n```\r\n\r\n- Check the cached directory. There are two arrrow files.\r\n```bash\r\nls -lha /home/*****/.cache/huggingface/datasets/text/default-e84dd29acc4ad9ef/0.0.0/447f2bcfa2a721a37bc8fdf23800eade1523cf07f7eada6fe661fe4d070d380d\r\ntotal 230M\r\ndrwxr-xr-x 2 ***** ***** 4.0K Aug 28 13:25 .\r\ndrwxr-xr-x 3 ***** ***** 4.0K Aug 28 13:08 ..\r\n-rw------- 1 ***** ***** 99M Aug 28 13:08 cache-bfc7cb0702426d19242db5e8c079f04b.arrow\r\n-rw------- 1 ***** ***** 99M Aug 28 13:25 cache-e0e2313e49c8a110aafcc8133154c19a.arrow\r\n-rw-r--r-- 1 ***** ***** 670 Aug 28 13:24 dataset_info.json\r\n-rw-r--r-- 1 ***** ***** 0 Aug 28 13:24 LICENSE\r\n-rw-r--r-- 1 ***** ***** 33M Aug 28 13:08 text-train.arrow\r\n```\r\n", "I ended up specifying the `cache_file_name` argument when I call `map` function.\r\n\r\n```python\r\ndataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True, truncation=True, max_length=args.block_size),\r\n batched=True,\r\n cache_file_name=cache_file_name)\r\n```\r\n\r\nNote:\r\n- `text` dataset in `nlp` does not strip `\"\\n\"`. If you want the same output as in [`LineByLineTextDataset`](https://github.com/huggingface/transformers/blob/afc4ece462ad83a090af620ff4da099a0272e171/src/transformers/data/datasets/language_modeling.py#L88-L111), you would need to create your own dataset class where you replace `line` to `line.strip()` [here](https://github.com/huggingface/nlp/blob/master/datasets/text/text.py#L35).\r\n" ]

[ "Thanks for creating the issue :)\r\nThe download link for wmt-en-de raw looks like a mirror. We should use that instead of the current url.\r\nIs this mirror official ?\r\n\r\nAlso it looks like for `ro-en` it tried to download other languages. If we manage to only download the one that is asked it'd be cool\r\n\r\nAlso cc @patrickvonplaten ", "Mirror is not official.", "Shall we host the files ourselves or it is fine to use this mirror in your opinion ?", "Should we add an argument in `load_dataset` to override some URL with a custom URL (e.g. mirror) or a local path?\r\n\r\nThis could also be used to provide local files instead of the original files as requested by some users (e.g. when you made a dataset with the same format than SQuAD and what to use it instead of the official dataset files).", "@lhoestq I think we should host it ourselves. I'll put the subset of wmt (without preprocessed files) that we need on s3 and post a link over the weekend.", "Is there a solution yet? The download speed is still too slow. 60-70kbps download for wmt16 and around 100kbps for wmt19. @sshleifer ", "I'm working on mirror links which will provide high download speed :)\r\nSee https://github.com/huggingface/datasets/issues/1892", "Resolved via https://github.com/huggingface/datasets/pull/1912" ]

[ "Indeed, good catch!" ]

[ "Thanks!" ]

[ "Update - maybe this is only failing on bleu because I was not tokenizing inputs to the metric", "Closing - seems to be just forgetting to tokenize. And found the helpful discussion in huggingface/evaluate#105 " ]

[ "Any updates on this?", "This request is still an open issue waiting to be addressed by any community member, @GuillemGSubies." ]

[ "This seems to be fixed in #513 for the filter function, replacing `cache_file_name` with `indices_cache_file_name` in the assert. Although not for the `map()` function @thomwolf ", "Maybe I'm a bit tired but I fail to see the issue here.\r\n\r\nSince `cache_file_name` is `None` by default, if you set `keep_in_memory` to `True`, the assert should pass, no?", "I failed to realise that this only applies to `shuffle()`. Whenever `keep_in_memory` is set to True, this is passed on to the `select()` function. However, if `cache_file_name` is None, it will be defined in the `shuffle()` function before it is passed on to `select()`. \r\n\r\nThus, `select()` is called with `keep_in_memory=True` and a not None value for `cache_file_name`. \r\nThis is essentially fixed in #513 \r\n\r\nEasily reproducible:\r\n```python\r\n>>> import nlp\r\n>>> data = nlp.load_dataset(\"cosmos_qa\", split=\"train\")\r\nUsing custom data configuration default\r\n>>> data.shuffle(keep_in_memory=True)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/nlp/arrow_dataset.py\", line 1398, in shuffle\r\n verbose=verbose,\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/nlp/arrow_dataset.py\", line 1178, in select\r\n ), \"Please use either `keep_in_memory` or `cache_file_name` but not both.\"\r\nAssertionError: Please use either `keep_in_memory` or `cache_file_name` but not both.\r\n>>>data.select([0], keep_in_memory=True)\r\n# No error\r\n```", "Oh yes ok got it thanks. Should be fixed if we are happy with #513 indeed.", "My bad. This is actually not fixed in #513. Sorry about that...\r\nThe new `indices_cache_file_name` is set to a non-None value in the new `shuffle()` as well. \r\n\r\nThe buffer and caching mechanisms used in the `select()` function are too intricate for me to understand why the check is there at all. I've removed it in my local build and it seems to be working fine for my project, without really considering other implications of the change. \r\n\r\n", "Ok I'll investigate and add a series of tests on the `keep_in_memory=True` settings which is under-tested atm", "Hey, still seeing this issue with the latest version.", "The same :(", "These are the steps needed to fix this issue:\r\n1. add the following check to `Dataset.shuffle`:\r\n```python\r\nif keep_in_memory and indices_cache_file_name is not None:\r\n raise ValueError(\"Please use either `keep_in_memory` or `indices_cache_file_name` but not both.\")\r\n```\r\n2. set `indices_cache_file_name` to `None` if `keep_in_memory` is True in the call to `select`\r\n3. add a test with `shuffle(keep_in_memory=True)`", "Hi @mariosasko , I have opened this PR #5082 " ]

[ "Hi @vegarab yes feel free to open a discussion here.\r\n\r\nThis design choice was not very much thought about.\r\n\r\nSince `dataset.select()` (like all the method without a trailing underscore) is non-destructive and returns a new dataset it has most of its properties initialized from scratch (except the table and infos).\r\n\r\nThinking about it I don't see a strong reason against transmitting the format from the parent dataset to its newly created child. It's probably what's expected by the user in most cases. What do you think @lhoestq?\r\n\r\nBy the way, I've been working today on a refactoring of all the samples re-ordering/selection methods (`select`, `sort`, `shuffle`, `shard`, `train_test_split`). The idea is to speed them up by a lot (like, really a lot) by working as much as possible with an indices mapping table instead of doing a deep copy of the full dataset as we've been doing currently. You can give it a look and try it here: https://github.com/huggingface/nlp/pull/513\r\nFeedbacks are very much welcome", "I think it's ok to keep the format.\r\nIf we want to have this behavior for `.map` too we just have to make sure it doesn't keep a column that's been removed.", "Shall we have this in the coming release by the way @lhoestq ?", "Yes sure !", "Since datasets 1.0.0 the format is not reset anymore.\r\nClosing this one, but feel free to re-open if you have other questions" ]

[ "Seems like this method was added in 1.17. I'll add a requirement on this.", "Thank you so much. After upgrading the numpy library, it worked." ]

[ "Do you want to convert a dataset script to the tfds format ?\r\nIf so, we currently have a comversion script nlp/commands/convert.py but it is a conversion script that goes from tfds to nlp.\r\nI think it shouldn't be too hard to do the changes in reverse (at some manual adjustments).\r\nIf you manage to make it work in reverse, feel free to open a PR to share it with the community :)", "In our docs: [Using a Dataset with PyTorch/Tensorflow](https://huggingface.co/docs/datasets/torch_tensorflow.html)." ]

[ "Which version of Apache Beam do you have (can you copy your full environment info here)?", "apache-beam==2.23.0\r\nnlp==0.4.0\r\n\r\nFor me this was resolved by running the same python script on Linux (or really WSL). ", "Do you manage to run a dummy beam pipeline with python on windows ? \r\nYou can test a dummy pipeline with [this code](https://github.com/apache/beam/blob/master/sdks/python/apache_beam/examples/wordcount_minimal.py)\r\n\r\nIf you get the same error, it means that the issue comes from apache beam.\r\nOtherwise we'll investigate what went wrong here", "Still, same error, so I guess it is on apache beam then. \r\nThanks for the investigation.", "Thanks for trying\r\nLet us know if you find clues of what caused this issue, or if you find a fix" ]

[ "Looks like an issue with 3.0.2 transformers version. Works fine when I use \"master\" version of transformers." ]

[ "Thanks for reporting !\r\n\r\nTo store the cache file, we compute a hash of the function given in `.map`, using our own hashing function.\r\nThe hash doesn't seem to stay the same over sessions for the tokenizer.\r\nApparently this is because of the regex at `tokenizer.pat` is not well supported by our hashing function.\r\n\r\nI'm working on a fix", "Thanks everyone. Works great now.", "Hi. I believe the fix was for the nlp library. Is there a solution to handle compiled regex expressions in .map() with the caching. I want to run a simple regex pattern on a big dataset, but I am running into the issue of compiled expression not being cached. \r\n\r\nInstead of opening a new issue, I thought I would put my query here. Let me know if a new issue would be more suitable. Thanks", "Hi @MaveriQ! This fix is also included in the `datasets` library. Can you provide a reproducer?" ]

[ "In 0.4.0, the assertion in `concatenate_datasets ` is on the features, and not the schema.\r\nCould you try to update `nlp` ?\r\n\r\nAlso, since 0.4.0, you can use `dset_wikipedia.cast_(dset_books.features)` to avoid the schema cast hack.", "Or maybe the assertion comes from elsewhere ?", "I'm using the master branch. The assertion failure comes from the underlying `pa.concat_tables()`, which is in the pyarrow package. That method does check schemas.\r\n\r\nSince `features.type` does not contain information about nullable vs non-nullable features, the `cast_()` method won't resolve the schema mismatch. There is information in a schema which is not stored in features.", "I'm doing a refactor of type inference in #363 . Both text fields should match after that", "By default nullable will be set to True", "It should be good now. I was able to run\r\n\r\n```python\r\n>>> from nlp import concatenate_datasets, load_dataset\r\n>>>\r\n>>> bookcorpus = load_dataset(\"bookcorpus\", split=\"train\")\r\n>>> wiki = load_dataset(\"wikipedia\", \"20200501.en\", split=\"train\")\r\n>>> wiki.remove_columns_(\"title\") # only keep the text\r\n>>>\r\n>>> assert bookcorpus.features.type == wiki.features.type\r\n>>> bert_dataset = concatenate_datasets([bookcorpus, wiki])\r\n```", "Thanks!" ]

[ "I did the release on github, and updated the doc :)\r\nSorry for the delay", "Thanks!" ]

[]

[ "whoops", "please delete this" ]

[ "I found `UNv1.0.en-ru.tar.gz` here: https://conferences.unite.un.org/uncorpus/en/downloadoverview, so it can be reconstructed with:\r\n```\r\nwget -c https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.00\r\nwget -c https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.01\r\nwget -c https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-ru.tar.gz.02\r\ncat UNv1.0.en-ru.tar.gz.0* > UNv1.0.en-ru.tar.gz\r\n```\r\nit has other languages as well, in case https://storage.googleapis.com/tfdataset-data/downloadataset/uncorpus/ is gone", "Further, `nlp.load_dataset('wmt19', 'ru-en')` has only the `train` and `val` datasets. `test` is missing.\r\n\r\nFixed locally for summarization needs, by running:\r\n```\r\npip install sacrebleu\r\nsacrebleu -t wmt19 -l ru-en --echo src > test.source\r\nsacrebleu -t wmt19 -l ru-en --echo ref > test.target\r\n```\r\nh/t @sshleifer ", "Fixed in https://github.com/huggingface/datasets/pull/1912" ]

[ "Yes indeed it looks like some `'` and spaces are missing (for example in `dont` or `didnt`).\r\nDo you know if there exist some copies without this issue ?\r\nHow would you fix this issue on the current data exactly ? I can see that the data is raw text (not tokenized) so I'm not sure I understand how you would do it. Could you provide more details ?", "I'm afraid that I don't know how to obtain the original BookCorpus data. I believe this version came from an anonymous Google Drive link posted in another issue.\r\n\r\nGoing through the raw text in this version, it's apparent that NLTK's TreebankWordTokenizer was applied on it (I gave some examples in my original post), followed by:\r\n`' '.join(tokens)`\r\nYou can retrieve the tokenization by splitting on whitespace. You can then \"detokenize\" it with TreebankWordDetokenizer class of NLTK (though, as I suggested, use the fixed version in my repo). This will bring the text closer to its original form, but some steps of TreebankWordTokenizer are destructive, so it wouldn't be one-to-one. Something along the lines of the following should work:\r\n```\r\ntreebank_detokenizer = nltk.tokenize.treebank.TreebankWordDetokenizer()\r\ndb = nlp.load_dataset('bookcorpus', split=nlp.Split.TRAIN)\r\ndb = db.map(lambda x: treebank_detokenizer.detokenize(x['text'].split()))\r\n```\r\n\r\nRegarding other issues beyond the above, I'm afraid that I can't help with that.", "Ok I get it, that would be very cool indeed\r\n\r\nWhat kinds of patterns the detokenizer can't retrieve ?", "The TreebankTokenizer makes some assumptions about whitespace, parentheses, quotation marks, etc. For instance, while tokenizing the following text:\r\n```\r\nDwayne \"The Rock\" Johnson\r\n```\r\nwill result in:\r\n```\r\nDwayne `` The Rock '' Johnson\r\n```\r\nwhere the left and right quotation marks are turned into distinct symbols. Upon reconstruction, we can attach the left part to its token on the right, and respectively for the right part. However, the following texts would be tokenized exactly the same:\r\n```\r\nDwayne \" The Rock \" Johnson\r\nDwayne \" The Rock\" Johnson\r\nDwayne \" The Rock\" Johnson\r\n...\r\n```\r\nIn the above examples, the detokenizer would correct these inputs into the canonical text\r\n```\r\nDwayne \"The Rock\" Johnson\r\n```\r\nHowever, there are cases where there the solution cannot easily be inferred (at least without a true LM - this tokenizer is just a bunch of regexes). For instance, in cases where you have a fragment that contains the end of quote, but not its beginning, plus an accidental space:\r\n```\r\n... and it sounds fantastic, \" he said.\r\n```\r\nIn the above case, the tokenizer would assume that the quotes refer to the next token, and so upon detokenization it will result in the following mistake:\r\n```\r\n... and it sounds fantastic, \"he said.\r\n```\r\n\r\nWhile these are all odd edge cases (the basic assumptions do make sense), in noisy data they can occur, which is why I mentioned that the detokenizer cannot restore the original perfectly.\r\n", "To confirm, since this is preprocessed, this was not the exact version of the Book Corpus used to actually train the models described here (particularly Distilbert)? https://huggingface.co/datasets/bookcorpus\r\n\r\nOr does this preprocessing exactly match that of the papers?", "I believe these are just artifacts of this particular source. It might be better to crawl it again, or use another preprocessed source, as found here: https://github.com/soskek/bookcorpus ", "Yes actually the BookCorpus on hugginface is based on [this](https://github.com/soskek/bookcorpus/issues/24#issuecomment-643933352). And I kind of regret naming it as \"BookCorpus\" instead of something like \"BookCorpusLike\".\r\n\r\nBut there is a good news ! @shawwn has replicated BookCorpus in his way, and also provided a link to download the plain text files. see [here](https://github.com/soskek/bookcorpus/issues/27). There is chance we can have a \"OpenBookCorpus\" !", "Resolved via #856" ]

[]

[ "found a mirror: https://storage.googleapis.com/seldon-datasets/sentence_polarity_v1/rt-polaritydata.tar.gz", "fixed in #484 ", "Closing this one. Thanks again @jxmorris12 for taking care of this :)" ]

[ "This comes from an overflow in pyarrow's array.\r\nIt is stuck inside the loop that reduces the batch size to avoid the overflow.\r\nI'll take a look", "I created a PR to fix the issue.\r\nIt was due to an overflow check that handled badly an empty list.\r\n\r\nYou can try the changes by using \r\n```\r\n!pip install git+https://github.com/huggingface/nlp.git@fix-bad-type-in-overflow-check\r\n```\r\n\r\nAlso I noticed that the first 1000 examples have an empty list in the `title_urls` field. The feature type inference in `.map` will consider it `null` because of that, and it will crash when it encounter the next example with a `title_urls` that is not empty.\r\n\r\nTherefore to fix that, what you can do for now is increase the writer batch size so that the feature inference will take into account at least one example with a non-empty `title_urls`:\r\n\r\n```python\r\n# default batch size is 1_000 and it's not enough for feature type inference because of empty lists\r\nvalid_dataset = valid_dataset.map(make_input_target, writer_batch_size=3_000) \r\n```\r\n\r\nI was able to run the frozen cell with these changes.", "@lhoestq Perfect and thank you very much!!\r\nClose the issue.", "@lhoestq mapping the function `make_input_target` was passed by your fixing.\r\n\r\nHowever, there is another error in the final step of `valid_dataset.map(convert_to_features, batched=True)`\r\n\r\n`ArrowInvalid: Could not convert Thepiratebay.vg with type str: converting to null type`\r\n(The [same colab notebook above with new error message](https://colab.research.google.com/drive/14wttOTv3ky74B_c0kv5WrbgQjCF2fYQk?usp=sharing#scrollTo=5sRrJ3_C8rLt))\r\n\r\nDo you have some ideas? (I am really sorry I could not debug it by myself since I never used `pyarrow` before) \r\nNote that `train_dataset.map(convert_to_features, batched=True)` can be run successfully even though train_dataset is 27x bigger than `valid_dataset` so I believe the problem lies in some field of `valid_dataset` again .", "I got this issue too and fixed it by specifying `writer_batch_size=3_000` in `.map`.\r\nThis is because Arrow didn't expect `Thepiratebay.vg` in `title_urls `, as all previous examples have empty lists in `title_urls `", "I am clear now . Thank so much again Quentin!", "I'm getting a hanging `dataset.map()` when running a gradio app with `gradio` for auto-reloading instead of `python`", "Maybe this is an issue with gradio, could you open an issue on their repo ? `Dataset.map` simply uses `multiprocess.Pool` for multiprocessing\r\n\r\nIf you interrupt the program mayeb the stack trace would give some information of where it was hanging in the code (maybe a lock somewhere ?)" ]

[ "Nevermind, I restarted my python session and it worked fine...\r\n\r\n---\r\n\r\nI had an authentification error, and I authenticated from another terminal. After that, no more error but it was not working. Restarting the sessions makes it work :)" ]

[ "Thanks for reporting this issue\r\n\r\nThere was a bug where numpy arrays would get returned instead of tensorflow tensors.\r\nThis is fixed on master.\r\n\r\nI tried to re-run the colab and encountered this error instead:\r\n\r\n```\r\nAttributeError: 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'to_tensor'\r\n```\r\n\r\nThis is because the dataset returns a Tensor and not a RaggedTensor.\r\nBut I think we should always return a RaggedTensor unless the length of the sequence is fixed (it that case they can be stack into a Tensor).", "Hi, I got another error (on Colab):\r\n\r\n```python\r\n# You can read a few attributes of the datasets before loading them (they are python dataclasses)\r\nfrom dataclasses import asdict\r\n\r\nfor key, value in asdict(datasets[6]).items():\r\n print('👉 ' + key + ': ' + str(value))\r\n\r\n---------------------------------------------------------------------------\r\n\r\nTypeError Traceback (most recent call last)\r\n\r\n<ipython-input-6-b8ace6c227a2> in <module>()\r\n 2 from dataclasses import asdict\r\n 3 \r\n----> 4 for key, value in asdict(datasets[6]).items():\r\n 5 print('👉 ' + key + ': ' + str(value))\r\n\r\n/usr/local/lib/python3.6/dist-packages/dataclasses.py in asdict(obj, dict_factory)\r\n 1008 \"\"\"\r\n 1009 if not _is_dataclass_instance(obj):\r\n-> 1010 raise TypeError(\"asdict() should be called on dataclass instances\")\r\n 1011 return _asdict_inner(obj, dict_factory)\r\n 1012 \r\n\r\nTypeError: asdict() should be called on dataclass instances\r\n```", "Indeed we'll update the cola with the new release coming up this week." ]

[ "The `data_dir` parameter has been removed. Now the error is `ValueError: Config name is missing`\r\n\r\nAs mentioned in #470 I think we can have one test with the first config of BUILDER_CONFIGS, and another test that runs all of the configs in BUILDER_CONFIGS", "This was fixed in #527 \r\n\r\nClosing this one, but feel free to re-open if you have other questions" ]

[ "Hi ! Did you try to set the output format to pytorch ? (or tensorflow if you're using tensorflow)\r\nIt can be done with `dataset.set_format(\"torch\", columns=columns)` (or \"tensorflow\").\r\n\r\nNote that for pytorch, string columns can't be converted to `torch.Tensor`, so you have to specify in `columns=` the list of columns you want to keep (`input_ids` for example)", "Hello . Yes, I did set the output format as below for the two columns \r\n\r\n `train_dataset.set_format('torch',columns=['Text','Label'])`\r\n ", "I think you're having this issue because you try to format strings as pytorch tensors, which is not possible.\r\nIndeed by having \"Text\" in `columns=['Text','Label']`, you try to convert the text values to pytorch tensors.\r\n\r\nInstead I recommend you to first tokenize your dataset using a tokenizer from transformers. For example\r\n\r\n```python\r\nfrom transformers import BertTokenizer\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\n\r\ntrain_dataset.map(lambda x: tokenizer(x[\"Text\"]), batched=True)\r\ntrain_dataset.set_format(\"torch\", column=[\"input_ids\"])\r\n```\r\n\r\nAnother way to fix your issue would be to not set the format to pytorch, and leave the dataset as it is by default. In that case, the strings are returned normally when you get examples from your dataloader. It means that you would have to tokenize the examples in the training loop (or using a data collator) though.\r\n\r\nLet me know if you have other questions", "Hi, actually the thing is I am getting the same error and even after tokenizing them I am passing them through batch_encode_plus.\r\nI dont know what seems to be the problem is. I even converted it into 'pt' while passing them through batch_encode_plus but when I am evaluating my model , i am getting this error\r\n\r\n\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\n<ipython-input-145-ca218223c9fc> in <module>()\r\n----> 1 val_loss, predictions, true_val = evaluate(dataloader_validation)\r\n 2 val_f1 = f1_score_func(predictions, true_val)\r\n 3 tqdm.write(f'Validation loss: {val_loss}')\r\n 4 tqdm.write(f'F1 Score (Weighted): {val_f1}')\r\n\r\n6 frames\r\n/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataset.py in <genexpr>(.0)\r\n 160 \r\n 161 def __getitem__(self, index):\r\n--> 162 return tuple(tensor[index] for tensor in self.tensors)\r\n 163 \r\n 164 def __len__(self):\r\n\r\nTypeError: new(): invalid data type 'str' ", "> Hi, actually the thing is I am getting the same error and even after tokenizing them I am passing them through batch_encode_plus.\r\n> I dont know what seems to be the problem is. I even converted it into 'pt' while passing them through batch_encode_plus but when I am evaluating my model , i am getting this error\r\n> \r\n> TypeError Traceback (most recent call last)\r\n> in ()\r\n> ----> 1 val_loss, predictions, true_val = evaluate(dataloader_validation)\r\n> 2 val_f1 = f1_score_func(predictions, true_val)\r\n> 3 tqdm.write(f'Validation loss: {val_loss}')\r\n> 4 tqdm.write(f'F1 Score (Weighted): {val_f1}')\r\n> \r\n> 6 frames\r\n> /usr/local/lib/python3.6/dist-packages/torch/utils/data/dataset.py in (.0)\r\n> 160\r\n> 161 def **getitem**(self, index):\r\n> --> 162 return tuple(tensor[index] for tensor in self.tensors)\r\n> 163\r\n> 164 def **len**(self):\r\n> \r\n> TypeError: new(): invalid data type 'str'\r\n\r\nI got the same error and fix it .\r\nyou can check your input where there may be string contained.\r\nsuch as\r\n```\r\na = [1,2,3,4,'<unk>']\r\ntorch.tensor(a)\r\n```", "I didn't know tokenizers could return strings in the token ids. Which tokenizer are you using to get this @Doragd ?", "> I didn't know tokenizers could return strings in the token ids. Which tokenizer are you using to get this @Doragd ?\r\n\r\ni'm sorry that i met this issue in another place (not in huggingface repo). ", "@akhilkapil do you have strings in your dataset ? When you set the dataset format to \"pytorch\" you should exclude columns with strings as pytorch can't make tensors out of strings", "Closing due to inactivity." ]

[ "Indeed. Solution 1 is the simplest.\r\n\r\nThis is actually a recurring problem.\r\nI think we should scan all the datasets with regexpr to fix the use of `open()` without encodings.\r\nAnd probably add a test in the CI to forbid using this in the future.", "I'm happy to tackle the broader problem - will open a PR when it's ready!", "That would be awesome!", "I've created a simple function that seems to do the trick:\r\n\r\n```python\r\ndef apply_encoding_on_file_open(filepath: str):\r\n \"\"\"Apply UTF-8 encoding for all instances where a non-binary file is opened.\"\"\"\r\n \r\n with open(filepath, 'r', encoding='utf-8') as input_file:\r\n regexp = re.compile(r\"\"\"\r\n (?!.*\\b(?:encoding|rb|wb|wb+|ab|ab+)\\b)\r\n (open)\r\n \\((.*)\\)\r\n \"\"\")\r\n input_text = input_file.read()\r\n match = regexp.search(input_text)\r\n \r\n if match:\r\n print('Found match!', match.group())\r\n # append utf-8 encoding to matching groups in-place\r\n output = regexp.sub(lambda m: m.group()[:-1]+', encoding=\"utf-8\")', input_text)\r\n with open(filepath, 'w', encoding='utf-8') as output_file:\r\n output_file.write(output)\r\n else:\r\n print(\"No match found!\")\r\n```\r\n\r\nThe regexp does a negative lookahead to avoid matching on cases where the encoding is already specified or when binary files are involved.\r\n\r\nFrom an implementation perspective:\r\n\r\n* Would it make sense to include this function in `nlp-cli` so that we can run something like\r\n```\r\nnlp-cli fix_encoding path/to/folder\r\n```\r\nand the command recursively fixes all files in the target?\r\n* What is the desired behaviour in the CI test? Here we could either have a simple script that we run as a `job` in the CI and raises an error if a missing encoding is detected. Alternatively we could incorporate this behaviour into the CLI and run that in the CI.\r\n\r\nPlease let me know what you prefer among the alternatives.\r\n", "I realised I was overthinking the problem, so decided to just run the regexp over the codebase and make the PR. In other words, we can ignore my comments about using the CLI 😸 " ]

[ "This issue was resolved by #447 " ]

[ "Same here !", "This is the only fix I could come up with without touching the repo's code.\r\n```python\r\nfrom nlp.builder import FORCE_REDOWNLOAD\r\ndataset = load_dataset('csv', data_file='./a.csv', download_mode=FORCE_REDOWNLOAD, version='0.0.1')\r\n```\r\nYou'll have to change the version each time you want to load a different csv file.\r\nIf you're willing to add a ```print```, you can go to ```nlp.load``` and add ```print(builder_instance.cache_dir)``` right before the ```return ds``` in the ```load_dataset``` method. It'll print the cache folder, and you'll just have to erase it (and then you won't need the change here above)." ]

[ "This seems to be fixed in a non-released version. \r\n\r\nInstalling nlp from source\r\n```\r\ngit clone https://github.com/huggingface/nlp\r\ncd nlp\r\npip install .\r\n```\r\nsolves the issue. " ]

[]

[ "`DPRContextEncoder` and `DPRContextEncoderTokenizer` will be available in the next release of `transformers`.\r\n\r\nRight now you can experiment with it by installing `transformers` from the master branch.\r\nYou can also check the docs of DPR [here](https://huggingface.co/transformers/master/model_doc/dpr.html).\r\n\r\nMoreover all the indexing features will also be available in the next release of `nlp`.", "@lhoestq Thanks for the info ", "@lhoestq I tried installing transformer from the master branch. Python imports for DPR again didnt' work. Anyways, Looking forward to trying it in the next release of nlp ", "@nsankar have you tried with the latest version of the library?", "@yjernite it worked. Thanks" ]

[ "Thanks Sam, we now have a very detailed tutorial and template on how to add a new dataset to the library. It typically take 1-2 hours to add one. Do you want to give it a try ?\r\nThe tutorial on writing a new dataset loading script is here: https://huggingface.co/nlp/add_dataset.html\r\nAnd the part on how to share a new dataset is here: https://huggingface.co/nlp/share_dataset.html", "Hi @sshleifer, I'm trying to add IWSLT using the link you provided but the download urls are not working. Only `[en, de]` pair is working. For others language pairs it throws a `404` error.\r\n\r\n" ]

[ "Indeed, we’ll make a new PyPi release next week to solve this. Cc @lhoestq ", "+1! this is the reason our tests are failing at [TextAttack](https://github.com/QData/TextAttack) \r\n\r\n(Though it's worth noting if we fixed the version number of pyarrow to 0.16.0 that would fix our problem too. But in this case we'll just wait for you all to update)", "Came to raise this issue, great to see other already have and it's being fixed so soon!\r\n\r\nAs an aside, since no one wrote this already, it seems like the version check only looks at the second part of the version number making sure it is >16, but pyarrow newest version is 1.0.0 so the second past is 0!", "> Indeed, we’ll make a new PyPi release next week to solve this. Cc @lhoestq\r\n\r\nYes definitely", "please fix this on pypi! @lhoestq ", "Is this issue fixed ?", "We’ll release the new version later today. Apologies for the delay.", "I just pushed the new version on pypi :)", "Thanks for the update." ]

[ "This was fixed in #434 \r\nWe'll do a release later this week to include this fix.\r\nThanks for reporting", "I dont know if the fix was made but the problem is still present : \r\nInstaled with pip : NLP 0.3.0 // pyarrow 1.0.0 \r\nOS : archlinux with kernel zen 5.8.5", "Yes it was fixed in `nlp>=0.4.0`\r\nYou can update with pip", "Sorry, I didn't got the updated version, all is now working perfectly thanks" ]

[ "Hi @ArneBinder, we have a few \"generic\" datasets which are intended to load data files with a predefined format:\r\n- csv: https://github.com/huggingface/nlp/tree/master/datasets/csv\r\n- json: https://github.com/huggingface/nlp/tree/master/datasets/json\r\n- text: https://github.com/huggingface/nlp/tree/master/datasets/text\r\n\r\nYou can find more details about this way to load datasets here in the documentation: https://huggingface.co/nlp/loading_datasets.html#from-local-files\r\n\r\nMaybe your brat loading script could be shared in a similar fashion?", "> Maybe your brat loading script could be shared in a similar fashion?\r\n\r\n@thomwolf that was also my first idea and I think I will tackle that in the next days. I separated the code and created a real abstract class `AbstractBrat` to allow to inherit from that (I've just seen that the dataset_loader loads the first non abstract class), now `Brat` is very similar in its functionality to https://github.com/huggingface/nlp/tree/master/datasets/text but inherits from `AbstractBrat`.\r\n\r\nHowever, it is still not clear to me how to add a specific dataset (as explained in https://huggingface.co/nlp/add_dataset.html) to your repo that uses this format/abstract class, i.e. re-using the `features` entry of the `DatasetInfo` object and `_generate_examples()`. Again, by doing so, the only remaining entries/functions to define would be `_DESCRIPTION`, `_CITATION`, `homepage` and `_URL` (which is all copy-paste stuff) and `_split_generators()`.\r\n \r\nIn a lack of better ideas, I tried sth like below, but of course it does not work outside `nlp` (`AbstractBrat` is currently defined in [datasets/brat.py](https://github.com/ArneBinder/nlp/blob/5e81fb8710546ee7be3353a7f02a3045e9a8351e/datasets/brat/brat.py)):\r\n```python\r\nfrom __future__ import absolute_import, division, print_function\r\n\r\nimport os\r\n\r\nimport nlp\r\n\r\nfrom datasets.brat.brat import AbstractBrat\r\n\r\n_CITATION = \"\"\"\r\n@inproceedings{lauscher2018b,\r\n title = {An argument-annotated corpus of scientific publications},\r\n booktitle = {Proceedings of the 5th Workshop on Mining Argumentation},\r\n publisher = {Association for Computational Linguistics},\r\n author = {Lauscher, Anne and Glava\\v{s}, Goran and Ponzetto, Simone Paolo},\r\n address = {Brussels, Belgium},\r\n year = {2018},\r\n pages = {40–46}\r\n}\r\n\"\"\"\r\n\r\n_DESCRIPTION = \"\"\"\\\r\nThis dataset is an extension of the Dr. Inventor corpus (Fisas et al., 2015, 2016) with an annotation layer containing \r\nfine-grained argumentative components and relations. It is the first argument-annotated corpus of scientific \r\npublications (in English), which allows for joint analyses of argumentation and other rhetorical dimensions of \r\nscientific writing.\r\n\"\"\"\r\n\r\n_URL = \"http://data.dws.informatik.uni-mannheim.de/sci-arg/compiled_corpus.zip\"\r\n\r\n\r\nclass Sciarg(AbstractBrat):\r\n\r\n VERSION = nlp.Version(\"1.0.0\")\r\n\r\n def _info(self):\r\n\r\n brat_features = super()._info().features\r\n return nlp.DatasetInfo(\r\n # This is the description that will appear on the datasets page.\r\n description=_DESCRIPTION,\r\n # nlp.features.FeatureConnectors\r\n features=brat_features,\r\n # If there's a common (input, target) tuple from the features,\r\n # specify them here. They'll be used if as_supervised=True in\r\n # builder.as_dataset.\r\n #supervised_keys=None,\r\n # Homepage of the dataset for documentation\r\n homepage=\"https://github.com/anlausch/ArguminSci\",\r\n citation=_CITATION,\r\n )\r\n\r\n def _split_generators(self, dl_manager):\r\n \"\"\"Returns SplitGenerators.\"\"\"\r\n # TODO: Downloads the data and defines the splits\r\n # dl_manager is a nlp.download.DownloadManager that can be used to\r\n # download and extract URLs\r\n dl_dir = dl_manager.download_and_extract(_URL)\r\n data_dir = os.path.join(dl_dir, \"compiled_corpus\")\r\n print(f'data_dir: {data_dir}')\r\n return [\r\n nlp.SplitGenerator(\r\n name=nlp.Split.TRAIN,\r\n # These kwargs will be passed to _generate_examples\r\n gen_kwargs={\r\n \"directory\": data_dir,\r\n },\r\n ),\r\n ]\r\n``` \r\n\r\nNevertheless, many thanks for tackling the dataset accessibility problem with this great library!", "As temporary fix I've created [ArneBinder/nlp-formats](https://github.com/ArneBinder/nlp-formats) (contributions welcome).", "Hi! You can either copy&paste the builder script and import the builder from there or use `datasets.load_dataset_builder` inside the script and call the methods of the returned builder object." ]

[ "Yes that's definitely something we plan to add ^^", "Yes, that would be nice. We could take a look at what tensorflow `tf.data` does under the hood for instance.", "So `tf.data.Dataset.map()` returns a `ParallelMapDataset` if `num_parallel_calls is not None` [link](https://github.com/tensorflow/tensorflow/blob/2b96f3662bd776e277f86997659e61046b56c315/tensorflow/python/data/ops/dataset_ops.py#L1623).\r\n\r\nThere, `num_parallel_calls` is turned into a tensor and and fed to `gen_dataset_ops.parallel_map_dataset` where it looks like tensorflow takes over.\r\n\r\nWe could start with something simple like a thread or process pool that `imap`s over some shards.\r\n ", "Multiprocessing was added in #552 . You can set the number of processes with `.map(..., num_proc=...)`. It also works for `filter`\r\n\r\nClosing this one, but feel free to reo-open if you have other questions", "@lhoestq Great feature implemented! Do you have plans to add it to official tutorials [Processing data in a Dataset](https://huggingface.co/docs/datasets/processing.html?highlight=save#augmenting-the-dataset)? It took me sometime to find this parallel processing api.", "Thanks for the heads up !\r\n\r\nI just added a paragraph about multiprocessing:\r\nhttps://huggingface.co/docs/datasets/master/processing.html#multiprocessing" ]

[ "Thanks for noticing ! This looks more reasonable indeed.\r\nFeel free to open a PR", "Hi @lhoestq \r\nI made the proposed changes to the `xtreme.py` script. I noticed that I also need to change the schema in the `dataset_infos.json` file. More specifically the `\"features\"` part of the PAN-X.LANG dataset:\r\n\r\n```json\r\n\"features\":{\r\n \"word\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n },\r\n \"ner_tag\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n },\r\n \"lang\":{\r\n \"dtype\":\"string\",\r\n \"id\":null,\r\n \"_type\":\"Value\"\r\n }\r\n}\r\n```\r\nTo fit the code above the fields `\"word\"`, `\"ner_tag\"`, and `\"lang\"` would become `\"words\"`, `ner_tags\"` and `\"langs\"`. In addition the `dtype` should be changed from `\"string\"` to `\"list\"`.\r\n\r\n I made this changes but when trying to test this locally with `dataset = load_dataset(\"xtreme\", \"PAN-X.en\", data_dir='./data')` I face the issue that the `dataset_info.json` file is always overwritten by a downloaded version with the old settings, which then throws an error because the schema does not match. This makes it hard to test the changes locally. Do you have any suggestions on how to deal with that?\r\n", "Hi !\r\n\r\nYou have to point to your local script.\r\nFirst clone the repo and then:\r\n\r\n```python\r\ndataset = load_dataset(\"./datasets/xtreme\", \"PAN-X.en\")\r\n```\r\nThe \"xtreme\" directory contains \"xtreme.py\".\r\n\r\nYou also have to change the features definition in the `_info` method. You could use:\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": [nlp.Value(\"string\")],\r\n \"ner_tags\": [nlp.Value(\"string\")],\r\n \"langs\": [nlp.Value(\"string\")],\r\n})\r\n```\r\n\r\nHope this helps !\r\nLet me know if you have other questions.", "Thanks, I am making progress. I got a new error `NonMatchingSplitsSizesError ` (see traceback below), which I suspect is due to the fact that number of rows in the dataset changed (one row per word --> one row per sentence) as well as the number of bytes due to the slightly updated data structure. \r\n\r\n```python\r\nNonMatchingSplitsSizesError: [{'expected': SplitInfo(name='validation', num_bytes=1756492, num_examples=80536, dataset_name='xtreme'), 'recorded': SplitInfo(name='validation', num_bytes=1837109, num_examples=10000, dataset_name='xtreme')}, {'expected': SplitInfo(name='test', num_bytes=1752572, num_examples=80326, dataset_name='xtreme'), 'recorded': SplitInfo(name='test', num_bytes=1833214, num_examples=10000, dataset_name='xtreme')}, {'expected': SplitInfo(name='train', num_bytes=3496832, num_examples=160394, dataset_name='xtreme'), 'recorded': SplitInfo(name='train', num_bytes=3658428, num_examples=20000, dataset_name='xtreme')}]\r\n```\r\nI can fix the error by replacing the values in the `datasets_infos.json` file, which I tested for English. However, to update this for all 40 datasets manually is slightly painful. Is there a better way to update the expected values for all datasets?", "You can update the json file by calling\r\n```\r\nnlp-cli test ./datasets/xtreme --save_infos --all_configs\r\n```", "One more thing about features. I mentioned\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": [nlp.Value(\"string\")],\r\n \"ner_tags\": [nlp.Value(\"string\")],\r\n \"langs\": [nlp.Value(\"string\")],\r\n})\r\n```\r\n\r\nbut it's actually not consistent with the way we write datasets. Something like this is simpler to read and more consistent with the way we define datasets:\r\n\r\n```python\r\nfeatures = nlp.Features({\r\n \"words\": nlp.Sequence(nlp.Value(\"string\")),\r\n \"ner_tags\": nlp.Sequence(nlp.Value(\"string\")),\r\n \"langs\": nlp.Sequence(nlp.Value(\"string\")),\r\n})\r\n```\r\n\r\nSorry about that", "Closing this since PR #437 fixed the problem and has been merged to `master`. " ]

[ "I think the problem is the way you wrote your urls. Try the following structure to see `https://drive.google.com/uc?export=download&id=your_file_id` . \r\n\r\n@lhoestq ", "Oh sorry, I think `_get_drive_url` is doing that. \r\n\r\nHave you tried to use `dl_manager.download_and_extract(_get_drive_url(_TRAIN_URL)`? it should work with google drive links.\r\n", "Yes it worked, thank you!" ]

[]

[ "`from_dict` was added in #350 that was unfortunately not included in the 0.3.0 release. It's going to be included in the next release that will be out pretty soon though.\r\nRight now if you want to use `from_dict` you have to install the package from the master branch\r\n```\r\npip install git+https://github.com/huggingface/nlp.git\r\n```", "> `from_dict` was added in #350 that was unfortunately not included in the 0.3.0 release. It's going to be included in the next release that will be out pretty soon though.\r\n> Right now if you want to use `from_dict` you have to install the package from the master branch\r\n> \r\n> ```\r\n> pip install git+https://github.com/huggingface/nlp.git\r\n> ```\r\nOK, thank you.\r\n" ]

[ "Unfortunately it's not possible to download only the dev set of NQ.\r\n\r\nI think we could add a way to download only the test set by adding a custom configuration to the processing script though.", "Ok, got it. I think this could be a valuable feature - especially for large datasets like NQ, but potentially also others. \r\nFor us, it will in this case make the difference of using the library or keeping the old downloads of the raw dev datasets. \r\nHowever, I don't know if that fits into your plans with the library and can also understand if you don't want to support this.", "I don't think we could force this behavior generally since the dataset script authors are free to organize the file download as they want (sometimes the mapping between split and files can be very much nontrivial) but we can add an additional configuration for Natural Question indeed as @lhoestq indicate." ]

[ "Hi @lewtun, you have to provide the full path to the downloaded file for example `/home/lewtum/..`", "I was able to repro. Opening a PR to fix that.\r\nThanks for reporting this issue !", "Thanks for the rapid fix @lhoestq!" ]

[ "It was fixed in 2ddd18d139d3047c9c3abe96e1e7d05bb360132c.\r\nCould you pull the latest changes from master @morganmcg1 ?", "Thanks @lhoestq, works fine now!" ]

[ "Fixed. Could you try again @mitchellgordon95 ?\r\nIt was due a file not being updated on S3.\r\n\r\nWe need to make sure all the datasets scripts get updated properly @julien-c ", "Works for me! Thanks.", "I found the same issue with almost any language other than English. (For English, it works). Will someone need to update the file on S3 again?", "This is because only some languages are already preprocessed (en, de, fr, it) and stored on our google storage.\r\nWe plan to have a systematic way to preprocess more wikipedia languages in the future.\r\n\r\nFor the other languages you have to process them on your side using apache beam. That's why the lib asks for a Beam runner." ]

[ "I think the slowness here probably come from the fact that we are copying from and to python.\r\n\r\n@lhoestq for all the `select`-based methods I think we should stay in Arrow format and update the writer so that it can accept Arrow tables or batches as well. What do you think?", "> @lhoestq for all the `select`-based methods I think we should stay in Arrow format and update the writer so that it can accept Arrow tables or batches as well. What do you think?\r\n\r\nI just tried with `writer.write_table` with tables of 1000 elements and it's slower that the solution in #405 \r\n\r\nOn my side (select 10 000 examples):\r\n- Original implementation: 12s\r\n- Batched solution: 100ms\r\n- solution using arrow tables: 350ms\r\n\r\nI'll try with arrays and record batches to see if we can make it work.", "I tried using `.take` from pyarrow recordbatches but it doesn't improve the speed that much:\r\n```python\r\nimport nlp\r\nimport numpy as np\r\n\r\ndset = nlp.Dataset.from_file(\"dummy_test_select.arrow\") # dummy dataset with 100000 examples like {\"a\": \"h\"*512}\r\nindices = np.random.randint(0, 100_000, 1000_000)\r\n```\r\n\r\n```python\r\n%%time\r\nbatch_size = 10_000\r\nwriter = ArrowWriter(schema=dset.schema, path=\"dummy_path\",\r\n writer_batch_size=1000, disable_nullable=False)\r\nfor i in tqdm(range(0, len(indices), batch_size)):\r\n table = pa.concat_tables(dset._data.slice(int(i), 1) for i in indices[i : min(len(indices), i + batch_size)])\r\n batch = table.to_pydict()\r\n writer.write_batch(batch)\r\nwriter.finalize()\r\n# 9.12s\r\n```\r\n\r\n\r\n```python\r\n%%time\r\nbatch_size = 10_000\r\nwriter = ArrowWriter(schema=dset.schema, path=\"dummy_path\", \r\n writer_batch_size=1000, disable_nullable=False)\r\nfor i in tqdm(range(0, len(indices), batch_size)):\r\n batch_indices = indices[i : min(len(indices), i + batch_size)]\r\n # First, extract only the indices that we need with a mask\r\n mask = [False] * len(dset)\r\n for k in batch_indices:\r\n mask[k] = True\r\n t_batch = dset._data.filter(pa.array(mask))\r\n # Second, build the list of indices for the filtered table, and taking care of duplicates\r\n rev_positions = {}\r\n duplicates = 0\r\n for i, j in enumerate(sorted(batch_indices)):\r\n if j in rev_positions:\r\n duplicates += 1\r\n else:\r\n rev_positions[j] = i - duplicates\r\n rev_map = [rev_positions[j] for j in batch_indices]\r\n # Third, use `.take` from the combined recordbatch\r\n t_combined = t_batch.combine_chunks() # load in memory\r\n recordbatch = t_combined.to_batches()[0]\r\n table = pa.Table.from_arrays(\r\n [recordbatch[c].take(pa.array(rev_map)) for c in range(len(dset._data.column_names))],\r\n schema=writer.schema\r\n )\r\n writer.write_table(table)\r\nwriter.finalize()\r\n# 3.2s\r\n```\r\n", "Shuffling is now significantly faster thanks to #513 \r\nFeel free to play with it now :)\r\n\r\nClosing this one, but feel free to re-open if you have other questions", "> Shuffling is now significantly faster thanks to #513 Feel free to play with it now :)\r\n> \r\n> Closing this one, but feel free to re-open if you have other questions\r\n\r\nI have a similar issue. My code is \r\n\r\n```\r\n for batch_num in range(num_batches):\r\n print(f'--> {batch_num=}\\n') if verbose else None\r\n # - Get batch\r\n shuffled_dataset = dataset.shuffle(buffer_size=buffer_size, seed=seed)\r\n raw_text_batch = shuffled_dataset.take(batch_size)\r\n tokenized_batch = map(raw_text_batch)\r\n if verbose:\r\n time_start = time.time()\r\n print(f'{raw_text_batch=}')\r\n print(f'{tokenized_batch=}')\r\n print(f'{next(iter(raw_text_batch))=}')\r\n print(f'{next(iter(tokenized_batch))=}')\r\n print(f'Time it took: {time.time() - time_start} seconds \\a\\n')\r\n ```\r\n\r\nwithout the suffle it takes 4.5 secs with it takes 87.1 secs. Is this difference expected? my dataset version is:\r\n```\r\n(beyond_scale) brando9@ampere1:~/beyond-scale-language-data-diversity$ pip list | grep dataset\r\ndatasets 2.14.3\r\n```\r\n\r\n@lhoestq thoughts?\r\n\r\n", "still slow even with update to `2.14.4` most recent as of this writing\r\n\r\n```\r\nTime it took: 4.301205635070801 seconds \r\n\r\n--> batch_num=0\r\n\r\nraw_text_batch=<datasets.iterable_dataset.IterableDataset object at 0x7f1fea7c2a40>\r\ntokenized_batch=<datasets.iterable_dataset.IterableDataset object at 0x7f1fea7c2f20>\r\nnext(iter(raw_text_batch))={'text': \"No matter which style you choose, you can be sure of one thing: our quality and craftsmanship are the best in the business. It's who we are and what we believe in. And it's evident every day on the factory floor where our dedicated teams take great pride in every stitch.\", 'timestamp': '2019-04-19T06:43:50Z', 'url': 'https://institchescustoms.com/katzkin.html'}\r\nnext(iter(tokenized_batch))={'input_ids': tensor([ 2949, 2300, 543, 3918, 345, 3853, 11, 345, 460, 307,\r\n 1654, 286, 530, 1517, 25, 674, 3081, 290, 5977, 49820,\r\n 389, 262, 1266, 287, 262, 1597, 13, 632, 338, 508,\r\n 356, 389, 290, 644, 356, 1975, 287, 13, 843, 340,\r\n 338, 10678, 790, 1110, 319, 262, 8860, 4314, 810, 674,\r\n 7256, 3466, 1011, 1049, 11293, 287, 790, 24695, 13, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,\r\n 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256]), 'attention_mask': tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0])}\r\nTime it took: 102.99258613586426 seconds\r\n```", "Shuffling leads to doing random access in many different locations on disk which is slower than reading contiguous data.\r\n\r\nThere is a super fast approximate shuffling algorithm implemented for iterable datasets though:\r\n\r\n```python\r\niterable_dataset = dataset.to_iterable_dataset(num_shards=1024)\r\nshuffled_dataset = iterable_dataset.shuffle(buffer_size=1000)\r\n```\r\n\r\n(the first batch might be a bit slow to get because the algorithm first fills a buffer before returning the first batch, see the [docs](https://huggingface.co/docs/datasets/v2.14.4/en/stream#shuffle) for more info)" ]

[]

[ "similar slow download speed here for nlp.load_dataset('wmt14', 'fr-en')\r\n`\r\nDownloading: 100%|██████████████████████████████████████████████████████████| 658M/658M [1:00:42<00:00, 181kB/s]\r\nDownloading: 100%|██████████████████████████████████████████████████████████| 918M/918M [1:39:38<00:00, 154kB/s]\r\nDownloading: 2%|▉ | 40.9M/2.37G [04:48<5:03:06, 128kB/s]\r\n`\r\nCould we just download a specific subdataset in 'wmt14', such as 'newstest14'? ", "> The code runs but the download speed is extremely slow, the same behaviour is not observed on wmt16 and wmt18\r\n\r\nThe original source for the files may provide slow download speeds.\r\nWe can probably host these files ourselves.\r\n\r\n> When trying to download wmt17 zh-en, I got the following error:\r\n> ConnectionError: Couldn't reach https://storage.googleapis.com/tfdataset-data/downloadataset/uncorpus/UNv1.0.en-zh.tar.gz\r\n\r\nLooks like the file`UNv1.0.en-zh.tar.gz` is missing, or the url changed. We need to fix that\r\n\r\n> Could we just download a specific subdataset in 'wmt14', such as 'newstest14'?\r\n\r\nRight now I don't think it's possible. Maybe @patrickvonplaten knows more about it\r\n", "Yeah, the download speed is sadly always extremely slow :-/. \r\nI will try to check out the `wmt17 zh-en` bug :-) ", "Maybe this can be used - https://stuncorpusprod.blob.core.windows.net/corpusfiles/UNv1.0.en-zh.tar.gz.00 ", "These issues seem to be fixed now." ]

[ "To convert from arrow type we have three options: to_numpy, to_pandas and to_pydict/to_pylist.\r\n\r\n- to_numpy and to_pandas return numpy arrays instead of lists but are very fast.\r\n- to_pydict/to_pylist can be 100x slower and become the bottleneck for reading data, but at least they return lists.\r\n\r\nMaybe we can have to_pydict/to_pylist as the default and use to_numpy or to_pandas when the format (set by `set_format`) is 'numpy' or 'pandas'" ]

[]

[]

[ "Same for the RACE dataset: https://github.com/huggingface/nlp/blob/master/datasets/race/race.py\r\n\r\nShould we scan all the datasets to remove this pattern of un-necessary nesting?", "You're right, I think we don't need to use the nested dictionary. \r\n" ]

[]

[ "**Edit**: other topic previously in this message moved to a new issue: https://github.com/huggingface/nlp/issues/387", "Could you try to update pyarrow to >=0.17.0 ? It should fix the `to_pandas` bug\r\n\r\nAlso I'm not sure that structures like list<struct> are fully supported in the lib (none of the datasets use that).\r\nIt can cause issues when using dataset transforms like `filter` for example" ]

[ "I am not able to reproduce this issue on my side.\r\nCould you give us more details about the inputs you used ?\r\n\r\nI do get another error though:\r\n```\r\n~/.virtualenvs/hf-datasets/lib/python3.7/site-packages/bert_score/utils.py in bert_cos_score_idf(model, refs, hyps, tokenizer, idf_dict, verbose, batch_size, device, all_layers)\r\n 371 return sorted(list(set(l)), key=lambda x: len(x.split(\" \")))\r\n 372 \r\n--> 373 sentences = dedup_and_sort(refs + hyps)\r\n 374 embs = []\r\n 375 iter_range = range(0, len(sentences), batch_size)\r\n\r\nValueError: operands could not be broadcast together with shapes (0,) (2,)\r\n```\r\nThat's because it gets numpy arrays as input and not lists. See #387 ", "The other issue was fixed by #403 \r\n\r\nDo you still get this issue @willywsm1013 ?\r\n" ]

[ "I've seen this sort of thing before -- it might help to delete the directory -- I've also noticed that there is an error with the json Dataloader for any data I've tried to load. I've replaced it with this, which skips over the data feature population step:\r\n\r\n\r\n```python\r\nimport os\r\n\r\nimport pyarrow.json as paj\r\n\r\nimport nlp as hf_nlp\r\n\r\nfrom nlp import DatasetInfo, BuilderConfig, SplitGenerator, Split, utils\r\nfrom nlp.arrow_writer import ArrowWriter\r\n\r\n\r\nclass JSONDatasetBuilder(hf_nlp.ArrowBasedBuilder):\r\n BUILDER_CONFIG_CLASS = BuilderConfig\r\n\r\n def _info(self):\r\n return DatasetInfo()\r\n\r\n def _split_generators(self, dl_manager):\r\n \"\"\" We handle string, list and dicts in datafiles\r\n \"\"\"\r\n if isinstance(self.config.data_files, (str, list, tuple)):\r\n files = self.config.data_files\r\n if isinstance(files, str):\r\n files = [files]\r\n return [SplitGenerator(name=Split.TRAIN, gen_kwargs={\"files\": files})]\r\n splits = []\r\n for split_name in [Split.TRAIN, Split.VALIDATION, Split.TEST]:\r\n if split_name in self.config.data_files:\r\n files = self.config.data_files[split_name]\r\n if isinstance(files, str):\r\n files = [files]\r\n splits.append(SplitGenerator(name=split_name, gen_kwargs={\"files\": files}))\r\n return splits\r\n\r\n def _prepare_split(self, split_generator):\r\n fname = \"{}-{}.arrow\".format(self.name, split_generator.name)\r\n fpath = os.path.join(self._cache_dir, fname)\r\n\r\n writer = ArrowWriter(path=fpath)\r\n\r\n generator = self._generate_tables(**split_generator.gen_kwargs)\r\n for key, table in utils.tqdm(generator, unit=\" tables\", leave=False):\r\n writer.write_table(table)\r\n num_examples, num_bytes = writer.finalize()\r\n\r\n split_generator.split_info.num_examples = num_examples\r\n split_generator.split_info.num_bytes = num_bytes\r\n\r\n def _generate_tables(self, files):\r\n for i, file in enumerate(files):\r\n pa_table = paj.read_json(\r\n file\r\n )\r\n yield i, pa_table\r\n\r\n```", "Yes, deleting the directory solves the error whenever I try to rerun.\r\n\r\nBy replacing the json-loader, you mean the cached file in my `site-packages` directory? e.g. `/home/XXX/.cache/lib/python3.7/site-packages/nlp/datasets/json/(...)/json.py` \r\n\r\nWhen I was testing this out before the #372 PR was merged I had issues installing it properly locally. Since the `json.py` script was downloaded instead of actually using the one provided in the local install. Manually updating that file seemed to solve it, but it didn't seem like a proper solution. Especially when having to run this on a remote compute cluster with no access to that directory.", "I see, diving in the JSON file for SQuAD it's a pretty complex structure.\r\n\r\nThe best solution for you, if you have a dataset really similar to SQuAD would be to copy and modify the SQuAD data processing script. We will probably add soon an option to be able to specify file path to use instead of the automatic URL encoded in the script but in the meantime you can:\r\n- copy the [squad script](https://github.com/huggingface/nlp/blob/master/datasets/squad/squad.py) in a new script for your dataset\r\n- in the new script replace [these `urls_to_download `](https://github.com/huggingface/nlp/blob/master/datasets/squad/squad.py#L99-L102) by `urls_to_download=self.config.data_files`\r\n- load the dataset with `dataset = load_dataset('path/to/your/new/script', data_files={nlp.Split.TRAIN: \"./datasets/train-v2.0.json\"})`\r\n\r\nThis way you can reuse all the processing logic of the SQuAD loading script.", "This seems like a more sensible solution! Thanks, @thomwolf. It's been a little daunting to understand what these scripts actually do, due to the level of abstraction and central documentation.\r\n\r\nAm I correct in assuming that the `_generate_examples()` function is the actual procedure for how the data is loaded from file? Meaning that essentially with a file containing another format, that is the only function that requires re-implementation? I'm working with a lot of datasets that, due to licensing and privacy, cannot be published. As this library is so neatly integrated with the transformers library and gives easy access to public sets such as SQUAD and increased performance, it is very neat to be able to load my private sets as well. As of now, I have just been working on scripts for translating all my data into the SQUAD-format before using the json script, but I see that it might not be necessary after all. ", "Yes `_generate_examples()` is the main entry point. If you change the shape of the returned dictionary you also need to update the `features` in the `_info`.\r\n\r\nI'm currently writing the doc so it should be easier soon to use the library and know how to add your datasets.\r\n", "Could you try to update pyarrow to >=0.17.0 @vegarab ?\r\nI don't have any segmentation fault with my version of pyarrow (0.17.1)\r\n\r\nI tested with\r\n```python\r\nimport nlp\r\ns = nlp.load_dataset(\"json\", data_files=\"train-v2.0.json\", field=\"data\", split=\"train\")\r\ns[0]\r\n# {'title': 'Normans', 'paragraphs': [{'qas': [{'question': 'In what country is Normandy located?', 'id':...\r\n```", "Also if you want to have your own dataset script, we now have a new documentation !\r\nSee here:\r\nhttps://huggingface.co/nlp/add_dataset.html", "@lhoestq \r\nFor some reason, I am not able to reproduce the segmentation fault, on pyarrow==0.16.0. Using the exact same environment and file.\r\n\r\nAnyhow, I discovered that pyarrow>=0.17.0 is required to read in a JSON file where the pandas structs contain lists. Otherwise, pyarrow complains when attempting to cast the struct:\r\n```py\r\nimport nlp\r\n>>> s = nlp.load_dataset(\"json\", data_files=\"datasets/train-v2.0.json\", field=\"data\", split=\"train\")\r\nUsing custom data configuration default\r\n>>> s[0]\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/nlp/arrow_dataset.py\", line 558, in __getitem__\r\n format_kwargs=self._format_kwargs,\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/nlp/arrow_dataset.py\", line 498, in _getitem\r\n outputs = self._unnest(self._data.slice(key, 1).to_pandas().to_dict(\"list\"))\r\n File \"pyarrow/array.pxi\", line 559, in pyarrow.lib._PandasConvertible.to_pandas\r\n File \"pyarrow/table.pxi\", line 1367, in pyarrow.lib.Table._to_pandas\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/pyarrow/pandas_compat.py\", line 766, in table_to_blockmanager\r\n blocks = _table_to_blocks(options, table, categories, ext_columns_dtypes)\r\n File \"/home/vegarab/.conda/envs/torch/lib/python3.7/site-packages/pyarrow/pandas_compat.py\", line 1101, in _table_to_blocks\r\n list(extension_columns.keys()))\r\n File \"pyarrow/table.pxi\", line 881, in pyarrow.lib.table_to_blocks\r\n File \"pyarrow/error.pxi\", line 105, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowNotImplementedError: Not implemented type for Arrow list to pandas: struct<qas: list<item: struct<question: string, id: string, answers: list<item: struct<text: string, answer_start: int64>>, is_impossible: bool, plausible_answers: list<item: struct<text: string, answer_start: int64>>>>, context: string>\r\n>>> s\r\nDataset(schema: {'title': 'string', 'paragraphs': 'list<item: struct<qas: list<item: struct<question: string, id: string, answers: list<item: struct<text: string, answer_start: int64>>, is_impossible: bool, plausible_answers: list<item: struct<text: string, answer_start: int64>>>>, context: string>>'}, num_rows: 35)\r\n```\r\n\r\nUpgrading to >=0.17.0 provides the same dataset structure, but accessing the records is possible without the same exception. \r\n\r\n", "Very happy to see some extended documentation! ", "#376 seems to be reporting the same issue as mentioned above. ", "This issue helped me a lot, thanks.\r\nHope this issue will be fixed soon." ]

[ "I am able to reproduce this with the official SQuAD `train-v2.0.json` file downloaded directly from https://rajpurkar.github.io/SQuAD-explorer/", "I am facing this issue in transformers library 3.0.2 while reading a csv using datasets.\r\nIs this fixed in latest version? \r\nI updated the latest version 4.0.1 but still getting this error. What could cause this error?" ]

[ "I found that, in the same process (or the same interactive session), if I do\r\n\r\nimport nlp\r\n\r\nm1 = nlp.load_metric('glue', 'mrpc')\r\nm2 = nlp.load_metric('glue', 'sst2')\r\n\r\nI will get the same error `ValueError: Cannot acquire lock, caching file might be used by another process, you should setup a unique 'experiment_id'`." ]

[ "Using batched map is probably the easiest way at the moment.\r\nWhat kind of augmentation would you like to do ?", "Some samples in the dataset are too long, I want to divide them in several samples.", "Using batched map is the way to go then.\r\nWe'll make it clearer in the docs that map could be used for augmentation.\r\n\r\nLet me know if you think there should be another way to do it. Or feel free to close the issue otherwise.", "It just feels awkward to use map to augment data. Also it means it's not possible to augment data in a non-batched way.\r\n\r\nBut to be honest I have no idea of a good API...", "Or for non-batched samples, how about returning a tuple ?\r\n\r\n```python\r\ndef aug(sample):\r\n # Simply copy the existing data to have x2 amount of data\r\n return sample, sample\r\n\r\ndataset = dataset.map(aug)\r\n```\r\n\r\nIt feels really natural and easy, but :\r\n\r\n* it means the behavior with batched data is different\r\n* I don't know how doable it is backend-wise\r\n\r\n@lhoestq ", "As we're working with arrow's columnar format we prefer to play with batches that are dictionaries instead of tuples.\r\nIf we have tuple it implies to re-format the data each time we want to write to arrow, which can lower the speed of map for example.\r\n\r\nIt's also a matter of coherence, as we don't want users to be confused whether they have to return dictionaries for some functions and tuples for others when they're doing batches." ]

[ "You're right, thanks for pointing it out. We will update it " ]

[ "Hi, can you give a full self-contained example to reproduce this behavior?", "> Hi, can you give a full self-contained example to reproduce this behavior?\r\n\r\nThere is a notebook in the post ;)", "> If I run the ROUGE metric 2 times, with same predictions / references, the scores are slightly different.\r\n> \r\n> Refer to [this Colab notebook](https://colab.research.google.com/drive/1wRssNXgb9ldcp4ulwj-hMJn0ywhDOiDy?usp=sharing) for reproducing the problem.\r\n> \r\n> Example of F-score for ROUGE-1, ROUGE-2, ROUGE-L in 2 differents run :\r\n> \r\n> > ['0.3350', '0.1470', '0.2329']\r\n> > ['0.3358', '0.1451', '0.2332']\r\n> \r\n> Why ROUGE is not deterministic ?\r\n\r\nThis is because of rouge's `BootstrapAggregator` that uses sampling to get confidence intervals (low, mid, high).\r\nYou can get deterministic scores per sentence pair by using\r\n```python\r\nscore = rouge.compute(rouge_types=[\"rouge1\", \"rouge2\", \"rougeL\"], use_aggregator=False)\r\n```\r\nOr you can set numpy's random seed if you still want to use the aggregator.", "Maybe we can set all the random seeds of numpy/torch etc. while running `metric.compute` ?", "We should probably indeed!", "Now if you re-run the notebook, the two printed results are the same @colanim\r\n```\r\n['0.3356', '0.1466', '0.2318']\r\n['0.3356', '0.1466', '0.2318']\r\n```\r\nHowever across sessions, the results may change (as numpy's random seed can be different). You can prevent that by setting your seed:\r\n```python\r\nrouge = nlp.load_metric('rouge', seed=42)\r\n```", "> \r\n\r\nMinor nit: Note that \"aggregator\" is misspelled in this command. Should be `use_aggregator=False`. ", "Thanks, I fixed the code snippet" ]

[ "Actually `map(batched=True)` can already change the size of the dataset.\r\nIt can accept examples of length `N` and returns a batch of length `M` (can be null or greater than `N`).\r\n\r\nI'll make that explicit in the doc that I'm currently writing.", "You're two steps ahead of me :) In my testing, it also works if `M` < `N`.\r\n\r\nA batched map of different length seems to work if you directly overwrite all of the original keys, but fails if any of the original keys are preserved.\r\n\r\nFor example,\r\n```python\r\n# Create a dummy dataset\r\ndset = load_dataset(\"wikitext\", \"wikitext-2-raw-v1\")[\"test\"]\r\ndset = dset.map(lambda ex: {\"length\": len(ex[\"text\"]), \"foo\": 1})\r\n\r\n# Do an allreduce on each batch, overwriting both keys\r\ndset.map(lambda batch: {\"length\": [sum(batch[\"length\"])], \"foo\": [1]})\r\n# Dataset(schema: {'length': 'int64', 'foo': 'int64'}, num_rows: 5)\r\n\r\n# Now attempt an allreduce without touching the `foo` key\r\ndset.map(lambda batch: {\"length\": [sum(batch[\"length\"])]})\r\n# This fails with the error message below\r\n```\r\n\r\n```bash\r\n File \"/path/to/nlp/src/nlp/arrow_dataset.py\", line 728, in map\r\n arrow_schema = pa.Table.from_pydict(test_output).schema\r\n File \"pyarrow/io.pxi\", line 1532, in pyarrow.lib.Codec.detect\r\n File \"pyarrow/table.pxi\", line 1503, in pyarrow.lib.Table.from_arrays\r\n File \"pyarrow/public-api.pxi\", line 390, in pyarrow.lib.pyarrow_wrap_table\r\n File \"pyarrow/error.pxi\", line 85, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Column 1 named foo expected length 1 but got length 2\r\n```\r\n\r\nAdding the `remove_columns=[\"length\", \"foo\"]` argument to `map()` solves the issue. Leaving the above error for future visitors. Perfect, thank you!" ]

[ "Hi, it depends on what it is in your `dataset_builder.py` file. Can you share it?\r\n\r\nIf you are just loading `json` files, you can also directly use the `json` script (which will find the schema/features from your JSON structure):\r\n\r\n```python\r\nfrom nlp import load_dataset\r\nds = load_dataset(\"json\", data_files=rel_datafiles)\r\n```", "The behavior I'm seeing is from the `json` script. \r\nI hacked this together to overcome the error with the `JSON` dataloader\r\n\r\n```\r\nclass DatasetBuilder(hf_nlp.ArrowBasedBuilder):\r\n BUILDER_CONFIG_CLASS = BuilderConfig\r\n\r\n def _info(self):\r\n return DatasetInfo()\r\n\r\n def _split_generators(self, dl_manager):\r\n \"\"\" We handle string, list and dicts in datafiles\r\n \"\"\"\r\n if isinstance(self.config.data_files, (str, list, tuple)):\r\n files = self.config.data_files\r\n if isinstance(files, str):\r\n files = [files]\r\n return [SplitGenerator(name=Split.TRAIN, gen_kwargs={\"files\": files})]\r\n splits = []\r\n for split_name in [Split.TRAIN, Split.VALIDATION, Split.TEST]:\r\n if split_name in self.config.data_files:\r\n files = self.config.data_files[split_name]\r\n if isinstance(files, str):\r\n files = [files]\r\n splits.append(SplitGenerator(name=split_name, gen_kwargs={\"files\": files}))\r\n return splits\r\n\r\n def _prepare_split(self, split_generator):\r\n fname = \"{}-{}.arrow\".format(self.name, split_generator.name)\r\n fpath = os.path.join(self._cache_dir, fname)\r\n\r\n writer = ArrowWriter(path=fpath)\r\n\r\n generator = self._generate_tables(**split_generator.gen_kwargs)\r\n for key, table in utils.tqdm(generator, unit=\" tables\", leave=False):\r\n writer.write_table(table)\r\n num_examples, num_bytes = writer.finalize()\r\n\r\n split_generator.split_info.num_examples = num_examples\r\n split_generator.split_info.num_bytes = num_bytes\r\n # this is where the error is coming from\r\n # def parse_schema(schema, schema_dict):\r\n # for field in schema:\r\n # if pa.types.is_struct(field.type):\r\n # schema_dict[field.name] = {}\r\n # parse_schema(field.type, schema_dict[field.name])\r\n # elif pa.types.is_list(field.type) and pa.types.is_struct(field.type.value_type):\r\n # schema_dict[field.name] = {}\r\n # parse_schema(field.type.value_type, schema_dict[field.name])\r\n # else:\r\n # schema_dict[field.name] = Value(str(field.type))\r\n # \r\n # parse_schema(writer.schema, features)\r\n # self.info.features = Features(features)\r\n\r\n def _generate_tables(self, files):\r\n for i, file in enumerate(files):\r\n pa_table = paj.read_json(\r\n file\r\n )\r\n yield i, pa_table\r\n```\r\n\r\nSo I basically just don't populate the `self.info.features` though this doesn't seem to cause any problems in my downstream applications. \r\n\r\nThe other workaround I was doing was to just use pyarrow.json to build a table and then to create the Dataset with its constructor or from_table methods. `load_dataset` has nice split logic, so I'd prefer to use that.\r\n\r\n", "Also noticed that if you for example in a loader script\r\n\r\n```\r\nfrom nlp import ArrowBasedBuilder\r\n\r\nclass MyBuilder(ArrowBasedBuilder):\r\n...\r\n\r\n```\r\nand use that in the subclass, it will be on the module's __dict__ and will be selected before the `MyBuilder` subclass, and it will raise `NotImplementedError` on its `_generate_examples` method... In the code it check for abstract classes but Builder and ArrowBasedBuilder aren't abstract classes, they're regular classes with `@abstract_methods`.", "Indeed this is part of a more general limitation which is the fact that we should generate and update the `features` from the auto-inferred Arrow schema when they are not provided (also happen when a user change the schema using `map()`, the features should be auto-generated and guessed as much as possible to keep the `features` synced with the underlying Arrow table schema).\r\n\r\nWe will try to solve this soon." ]

[ "I just added the processed files of `snli` on our google storage, so that when you do `load_dataset` it can download the processed files from there :)\r\n\r\nWe are thinking about having available those processed files for more datasets in the future, because sometimes files aren't available (like for `snli`), or the download speed is too slow, or sometimes the files take time to be processed.", "Closing this one. Feel free to re-open if you have other questions :)", "Thank you!" ]

[ "Pinging @mariamabarham as well", "- `nlp` has MR! It's called `rotten_tomatoes`\r\n- SST is part of GLUE, or is that just SST-2?\r\n- `nlp` also has `ag_news`, a popular news classification dataset\r\n\r\nI'd also like to see:\r\n- the Yahoo Answers topic classification dataset\r\n- the Kaggle Fake News classification dataset", "Thanks @jxmorris12 for pointing this out. \r\n\r\nIn glue we only have SST-2 maybe we can add separately SST-1.\r\n", "This is the homepage for the Amazon dataset: https://www.kaggle.com/datafiniti/consumer-reviews-of-amazon-products\r\n\r\nIs there an easy way to download kaggle datasets programmatically? If so, I can add this one!", "Hi @jxmorris12 for now I think our `dl_manager` does not download from Kaggle.\r\n@thomwolf , @lhoestq", "Pretty sure the quora dataset is the same one I implemented here: https://github.com/huggingface/nlp/pull/366", "Great list. Any idea if Amazon Reviews has been added?\r\n\r\n- ~40 GB of text (sadly no emoji)\r\n- popular MLM pre-training dataset before bigger datasets like WebText https://arxiv.org/abs/1808.01371\r\n- turns out that binarizing the 1-5 star rating leads to great Pos/Neg/Neutral dataset, T5 paper claims to get very high accuracy (98%!) on this with small amount of finetuning https://arxiv.org/abs/2004.14546\r\n\r\nApologies if it's been included (great to see where) and if not, it's one of the better medium/large NLP dataset for semi-supervised learning, albeit a bit out of date. \r\n\r\nThanks!! \r\n\r\ncc @sshleifer ", "On the Amazon Reviews dataset, the original UCSD website has noted these are now updated to include product reviews through 2018 -- actually quite recent compared to many other datasets. Almost certainly the largest NLP dataset out there with labels!\r\nhttps://jmcauley.ucsd.edu/data/amazon/ \r\n\r\nAny chance someone has time to onboard this dataset in a HF way?\r\n\r\ncc @sshleifer ", "@albertvillanova How up to date is this issue? I see that some of these datasets are now on huggingface but have not been checked off the list" ]

[ "This is probably a Windows issue, we need to specify the encoding when `load_dataset()` reads the original CSV file.\r\nTry to find the `open()` statement called by `load_dataset()` and add an `encoding='utf-8'` parameter.\r\nSee issues #242 and #307 ", "It should be in `xtreme.py:L755`:\r\n```python\r\n if self.config.name == \"tydiqa\" or self.config.name.startswith(\"MLQA\") or self.config.name == \"SQuAD\":\r\n with open(filepath) as f:\r\n data = json.load(f)\r\n```\r\n\r\nCould you try to add the encoding parameter:\r\n```python\r\nopen(filepath, encoding='utf-8')\r\n```", "Hello @jerryIsHere :) Did it work ?\r\nIf so we may change the dataset script to force the utf-8 encoding", "@lhoestq sorry for being that late, I found 4 copy of xtreme.py. I did the changes as what has been told to all of them.\r\nThe problem is not solved", "Could you provide a better error message so that we can make sure it comes from the opening of the `tydiqa`'s json files ?\r\n", "@lhoestq \r\nThe error message is same as before:\r\nException has occurred: UnicodeDecodeError\r\n'cp950' codec can't decode byte 0xe2 in position 111: illegal multibyte sequence\r\n File \"D:\\python\\test\\test.py\", line 3, in <module>\r\n dataset = load_dataset('xtreme', 'tydiqa')\r\n\r\n\r\n\r\nI said that I found 4 copy of xtreme.py and add the 「, encoding='utf-8'」 parameter to the open() function\r\nthese python script was found under this directory\r\nC:\\Users\\USER\\AppData\\Local\\Programs\\Python\\Python37\\Lib\\site-packages\\nlp\\datasets\\xtreme\r\n", "Hi there !\r\nI encountered the same issue with the IMDB dataset on windows. It threw an error about charmap not being able to decode a symbol during the first time I tried to download it. I checked on a remote linux machine I have, and it can't be reproduced.\r\nI added ```encoding='UTF-8'``` to both lines that have ```open``` in ```imdb.py``` (108 and 114) and it worked for me.\r\nThank you !", "> Hi there !\r\n> I encountered the same issue with the IMDB dataset on windows. It threw an error about charmap not being able to decode a symbol during the first time I tried to download it. I checked on a remote linux machine I have, and it can't be reproduced.\r\n> I added `encoding='UTF-8'` to both lines that have `open` in `imdb.py` (108 and 114) and it worked for me.\r\n> Thank you !\r\n\r\nHello !\r\nGlad you managed to fix this issue on your side.\r\nDo you mind opening a PR for IMDB ?", "> This is probably a Windows issue, we need to specify the encoding when `load_dataset()` reads the original CSV file.\r\n> Try to find the `open()` statement called by `load_dataset()` and add an `encoding='utf-8'` parameter.\r\n> See issues #242 and #307\r\n\r\nSorry for not responding for about a month.\r\nI have just found that it is necessary to change / add the environment variable as what was told in #242.\r\nEverything works after I add the new environment variable and restart my PC.\r\n\r\nI think the encoding issue for windows isn't limited to the open() function call specific to few dataset, but actually in the entire library, depends on the machine / os you use.", "Since #481 we shouldn't have other issues with encodings as they need to be set to \"utf-8\" be default.\r\n\r\nClosing this one, but feel free to re-open if you gave other questions" ]

[ "Hi @saverymax ! For licensing reasons, the original team was unable to release pre-processed CommonCrawl documents. Instead, they provided a script to re-create them from a CommonCrawl dump, but it unfortunately requires access to a medium-large size cluster:\r\nhttps://github.com/facebookresearch/ELI5#downloading-support-documents-from-the-commoncrawl\r\n\r\nIn order to make the task accessible to people who may not have access to this kind of infrastructure, we suggest to use Wikipedia as a knowledge source rather than the full CommonCrawl. The following blog post shows how you can create Wikipedia support documents and get a performance that is on par with a system that uses CommonCrawl pages.\r\nhttps://yjernite.github.io/lfqa.html#task_description\r\n\r\nHope that helps, using ElasticSearch to index Wiki40b and create the documents should take about 4 hours. Let us know if you have any trouble with the blog post though!", "Hi, thanks for the quick response. The blog post is quite an interesting working example, thanks for sharing it.\r\nTwo follow-up points/questions about my original question:\r\n\r\n1. Yes, I read that the facebook team could not share the CommonCrawl b/c of licensing reasons. They state \"No, we are not allowed to host processed Reddit or CommonCrawl data,\" which indicates they could also not share the Reddit data for licensing reasons. But it seems that HuggingFace is able to share the Reddit data, so why not a subset of CommonCrawl?\r\n\r\n2. Thanks for the suggestion about ElasticSearch and Wiki40b. This is good to know about performance. I definitely could do the indexing and querying myself. What I like about the ELI5 dataset though, at least what is suggested by the paper, is that to create the dataset they had already selected the top 100 web sources and made a single support document from those. Though it doesn't appear to be too sophisticated an approach, having a single support document pre-computed (without having to run the facebook code or a replacement with another dataset) is super useful for my work, especially since I'm not working on developing the latest and greatest retrieval model. Of course, I don't expect HF NLP datasets to be perfectly tailored to my use-case. I know there is overhead to any project, I'm just illustrating a use-case of ELI5 which is not possible with the data provided as-is. If it's for licensing reasons, that is perfectly acceptable a reason, and I appreciate your response." ]

[ "`dataset.column_names` are being updated but `dataset.features` aren't indeed..." ]

[]

[]

[ "I couldn't reproduce on my side.\r\nIt looks like you were not able to generate all the examples, and you have the problem for each split train-test-validation.\r\nCould you try to enable logging, try again and send the logs ?\r\n```python\r\nimport logging\r\nlogging.basicConfig(level=logging.INFO)\r\n```", "here's the log\r\n```\r\n>>> import nlp\r\nimport logging\r\nlogging.basicConfig(level=logging.INFO)\r\nnlp.load_dataset('cnn_dailymail', '3.0.0')\r\n>>> import logging\r\n>>> logging.basicConfig(level=logging.INFO)\r\n>>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\nINFO:nlp.load:Checking /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py for additional imports.\r\nINFO:filelock:Lock 140443095301136 acquired on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\nINFO:nlp.load:Found main folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail\r\nINFO:nlp.load:Found specific version folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\nINFO:nlp.load:Found script file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.py\r\nINFO:nlp.load:Updating dataset infos file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/dataset_infos.json to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/dataset_infos.json\r\nINFO:nlp.load:Found metadata file for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.json\r\nINFO:filelock:Lock 140443095301136 released on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\nINFO:nlp.info:Loading Dataset Infos from /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\nINFO:nlp.builder:Generating dataset cnn_dailymail (/u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0)\r\nINFO:nlp.builder:Dataset not on Hf google storage. Downloading and preparing it from source\r\nDownloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\nINFO:nlp.utils.info_utils:All the checksums matched successfully.\r\nINFO:nlp.builder:Generating split train\r\nINFO:nlp.arrow_writer:Done writing 285161 examples in 1240618482 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-train.arrow.\r\nINFO:nlp.builder:Generating split validation\r\nINFO:nlp.arrow_writer:Done writing 13255 examples in 56637485 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-validation.arrow.\r\nINFO:nlp.builder:Generating split test\r\nINFO:nlp.arrow_writer:Done writing 11379 examples in 48931393 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-test.arrow.\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/load.py\", line 520, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 431, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 488, in _download_and_prepare\r\n verify_splits(self.info.splits, split_dict)\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/utils/info_utils.py\", line 70, in verify_splits\r\n raise NonMatchingSplitsSizesError(str(bad_splits))\r\nnlp.utils.info_utils.NonMatchingSplitsSizesError: [{'expected': SplitInfo(name='test', num_bytes=49424491, num_examples=11490, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='test', num_bytes=48931393, num_examples=11379, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='train', num_bytes=1249178681, num_examples=287113, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='train', num_bytes=1240618482, num_examples=285161, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='validation', num_bytes=57149241, num_examples=13368, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='validation', num_bytes=56637485, num_examples=13255, dataset_name='cnn_dailymail')}]\r\n```", "> here's the log\r\n> \r\n> ```\r\n> >>> import nlp\r\n> import logging\r\n> logging.basicConfig(level=logging.INFO)\r\n> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n> >>> import logging\r\n> >>> logging.basicConfig(level=logging.INFO)\r\n> >>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n> INFO:nlp.load:Checking /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py for additional imports.\r\n> INFO:filelock:Lock 140443095301136 acquired on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\n> INFO:nlp.load:Found main folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail\r\n> INFO:nlp.load:Found specific version folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\n> INFO:nlp.load:Found script file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.py\r\n> INFO:nlp.load:Updating dataset infos file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/dataset_infos.json to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/dataset_infos.json\r\n> INFO:nlp.load:Found metadata file for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.json\r\n> INFO:filelock:Lock 140443095301136 released on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\n> INFO:nlp.info:Loading Dataset Infos from /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\n> INFO:nlp.builder:Generating dataset cnn_dailymail (/u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0)\r\n> INFO:nlp.builder:Dataset not on Hf google storage. Downloading and preparing it from source\r\n> Downloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\n> INFO:nlp.utils.info_utils:All the checksums matched successfully.\r\n> INFO:nlp.builder:Generating split train\r\n> INFO:nlp.arrow_writer:Done writing 285161 examples in 1240618482 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-train.arrow.\r\n> INFO:nlp.builder:Generating split validation\r\n> INFO:nlp.arrow_writer:Done writing 13255 examples in 56637485 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-validation.arrow.\r\n> INFO:nlp.builder:Generating split test\r\n> INFO:nlp.arrow_writer:Done writing 11379 examples in 48931393 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-test.arrow.\r\n> Traceback (most recent call last):\r\n> File \"<stdin>\", line 1, in <module>\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/load.py\", line 520, in load_dataset\r\n> builder_instance.download_and_prepare(\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 431, in download_and_prepare\r\n> self._download_and_prepare(\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 488, in _download_and_prepare\r\n> verify_splits(self.info.splits, split_dict)\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/utils/info_utils.py\", line 70, in verify_splits\r\n> raise NonMatchingSplitsSizesError(str(bad_splits))\r\n> nlp.utils.info_utils.NonMatchingSplitsSizesError: [{'expected': SplitInfo(name='test', num_bytes=49424491, num_examples=11490, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='test', num_bytes=48931393, num_examples=11379, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='train', num_bytes=1249178681, num_examples=287113, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='train', num_bytes=1240618482, num_examples=285161, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='validation', num_bytes=57149241, num_examples=13368, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='validation', num_bytes=56637485, num_examples=13255, dataset_name='cnn_dailymail')}]\r\n> ```\r\n\r\nWith `nlp == 0.3.0` version, I'm not able to reproduce this error on my side.\r\nWhich version are you using for reproducing your bug?\r\n\r\n```\r\n>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n\r\n8.90k/8.90k [00:18<00:00, 486B/s]\r\n\r\nDownloading: 100%\r\n9.37k/9.37k [00:00<00:00, 234kB/s]\r\n\r\nDownloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /root/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\nDownloading:\r\n159M/? [00:09<00:00, 16.7MB/s]\r\n\r\nDownloading:\r\n376M/? [00:06<00:00, 62.6MB/s]\r\n\r\nDownloading:\r\n2.11M/? [00:06<00:00, 333kB/s]\r\n\r\nDownloading:\r\n46.4M/? [00:02<00:00, 18.4MB/s]\r\n\r\nDownloading:\r\n2.43M/? [00:00<00:00, 2.62MB/s]\r\n\r\nDataset cnn_dailymail downloaded and prepared to /root/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0. Subsequent calls will reuse this data.\r\n{'test': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 11490),\r\n 'train': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 287113),\r\n 'validation': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 13368)}\r\n\r\n>> ...\r\n\r\n```", "In general if some examples are missing after processing (hence causing the `NonMatchingSplitsSizesError `), it is often due to either\r\n1) corrupted cached files\r\n2) decoding errors\r\n\r\nI just checked the dataset script for code that could lead to decoding errors but I couldn't find any. Before we try to dive more into the processing of the dataset, could you try to clear your cache ? Just to make sure that it isn't 1)", "Yes thanks for the support! I cleared out my cache folder and everything works fine now" ]

[ "Hi, can you give details on your environment/os/packages versions/etc?", "Environment is Ubuntu 18.04, Python 3.7.5, nlp==0.3.0, filelock=3.0.12.\r\n\r\nThe external volume is Amazon FSx for Lustre, and it by default creates files with limited permissions. My working theory is that FileLock creates a lockfile that isn't writable, and thus there's no way to acquire it by removing the .lock file. But Python is able to create new files and write to them outside of the FileLock package.\r\n\r\nWhen I attempt to use FileLock within a Docker container by writing to `/root/.cache/hello.txt`, it succeeds. So there's some permissions issue. But it's not a Docker configuration issue; I've replicated it without Docker.\r\n```bash\r\necho \"hello world\" >> hello.txt\r\nls -l\r\n\r\n-rw-rw-r-- 1 ubuntu ubuntu 10 Jun 30 19:52 hello.txt\r\n```", "Looks like the `flock` syscall does not work on Lustre filesystems by default: https://github.com/benediktschmitt/py-filelock/issues/67.\r\n\r\nI added the `-o flock` option when mounting the filesystem, as [described here](https://docs.aws.amazon.com/fsx/latest/LustreGuide/getting-started-step2.html), which fixed the issue.", "Awesome, thanks a lot for sharing your fix!", "I'm wondering if this can be revisited. In some managed environments the same person using HF cannot change the file-system mount flags, (and the organization may be unwilling to change these flags due to other concerns) but can ensure that there won't be concurrent writes, for example because HF is offline and the models/datasets were downloaded earlier. \r\n\r\nThe real fix would be to FileLock itself, which does not seem very active and seems to not deal with failed system flock calls , which would be one way to fix this, as they mention in the issue below also raised by @jarednielsen \r\n\r\nhttps://github.com/tox-dev/py-filelock/issues/67", "> I'm wondering if this can be revisited. In some managed environments the same person using HF cannot change the file-system mount flags, (and the organization may be unwilling to change these flags due to other concerns) but can ensure that there won't be concurrent writes, for example because HF is offline and the models/datasets were downloaded earlier.\r\n\r\nI am one of those users. Is there a work around for this?\r\n", "The machines I use have a shared FS which has the filelock problem as well as a local one that does not. Using some env vars (HF_HOME, which controls both models and datasets, and HF_DATASETS_OFFLINE) for both transformers and datasets library one can influence where these downloads happen, and whether the locks get taken. I think some of the relevant documentation is here https://huggingface.co/docs/transformers/installation#cache-setup. I do end up using different settings when I download the models and when I use them, and have to rsync the models to the local file system using a separate script. ", "Thanks @orm011 . These filesystems are such a pain. I'll dig around, looks like setting `cache_dir` to a non-lustre filesystem works for `transformers` but not `datasets`.", "Note I `export HF_HOME=` in the shell prior to running python (I do not use the `cache_dir` argument, I think I ran into similar issues with it, nor `HF_DATASETS_CACHE` , though maybe that works, or maybe you can set it in python prior to importing the library ), and I change no other variables. Then `datasets.load_dataset()` works without any additional flags, and they go into `HF_HOME/datasets/` and the models go into `HF_HOME/transformers/` (and the lock files are all there as well). ", "I am using a shared cluster with a lustre system that I can't change. I am unable to download or load datsets onto the filesystem because of file lock. @thomwolf can this issue be reopened? " ]

[ "To be able to generate the Arrow dataset you need to either use our csv or json utilities `load_dataset(\"json\", data_files=my_json_files)` OR write your own custom dataset script (you can find some inspiration from the [squad](https://github.com/huggingface/nlp/blob/master/datasets/squad/squad.py) script for example). Custom dataset scripts can be called locally with `nlp.load_dataset(path_to_my_script_directory)`.\r\n\r\nThis should help you get what you call \"Dataset1\".\r\n\r\nThen using some dataset transforms like `.map` for example you can get to \"DatasetNER\" and \"DatasetREL\".\r\n", "Thanks for the helpful advice, @lhoestq -- I wasn't quite able to get the json recipe working - \r\n\r\n```\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/pyarrow/ipc.py in __init__(self, source)\r\n 60 \r\n 61 def __init__(self, source):\r\n---> 62 self._open(source)\r\n 63 \r\n 64 \r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/pyarrow/ipc.pxi in pyarrow.lib._RecordBatchStreamReader._open()\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/pyarrow/error.pxi in pyarrow.lib.check_status()\r\nArrowInvalid: Tried reading schema message, was null or length 0\r\n```\r\n\r\nBut I'm going to give the generator_dataset_builder a try.\r\n\r\n1 more quick question -- can .map be used to output different length mappings -- could I skip one, or yield 2, can you map_batch ", "You can use `.map(my_func, batched=True)` and return less examples, or more examples if you want", "Thanks this answers my question. I think the issue I was having using the json loader were due to using gzipped jsonl files.\r\n\r\nThe error I get now is :\r\n\r\n```\r\n\r\nUsing custom data configuration test\r\n---------------------------------------------------------------------------\r\n\r\nValueError Traceback (most recent call last)\r\n\r\n<ipython-input-38-29082a31e5b2> in <module>\r\n 5 print(ner_datafiles)\r\n 6 \r\n----> 7 ds = nlp.load_dataset(\"json\", \"test\", data_files=ner_datafiles[0])\r\n 8 \r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)\r\n 522 download_mode=download_mode,\r\n 523 ignore_verifications=ignore_verifications,\r\n--> 524 save_infos=save_infos,\r\n 525 )\r\n 526 \r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/builder.py in download_and_prepare(self, download_config, download_mode, ignore_verifications, save_infos, try_from_hf_gcs, dl_manager, **download_and_prepare_kwargs)\r\n 430 verify_infos = not save_infos and not ignore_verifications\r\n 431 self._download_and_prepare(\r\n--> 432 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n 433 )\r\n 434 # Sync info\r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/builder.py in _download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)\r\n 481 try:\r\n 482 # Prepare split will record examples associated to the split\r\n--> 483 self._prepare_split(split_generator, **prepare_split_kwargs)\r\n 484 except OSError:\r\n 485 raise OSError(\"Cannot find data file. \" + (self.manual_download_instructions or \"\"))\r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/builder.py in _prepare_split(self, split_generator)\r\n 736 schema_dict[field.name] = Value(str(field.type))\r\n 737 \r\n--> 738 parse_schema(writer.schema, features)\r\n 739 self.info.features = Features(features)\r\n 740 \r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/builder.py in parse_schema(schema, schema_dict)\r\n 734 parse_schema(field.type.value_type, schema_dict[field.name])\r\n 735 else:\r\n--> 736 schema_dict[field.name] = Value(str(field.type))\r\n 737 \r\n 738 parse_schema(writer.schema, features)\r\n\r\n<string> in __init__(self, dtype, id, _type)\r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/features.py in __post_init__(self)\r\n 55 \r\n 56 def __post_init__(self):\r\n---> 57 self.pa_type = string_to_arrow(self.dtype)\r\n 58 \r\n 59 def __call__(self):\r\n\r\n~/.virtualenvs/inv-text2struct/lib/python3.6/site-packages/nlp/features.py in string_to_arrow(type_str)\r\n 32 if str(type_str + \"_\") not in pa.__dict__:\r\n 33 raise ValueError(\r\n---> 34 f\"Neither {type_str} nor {type_str + '_'} seems to be a pyarrow data type. \"\r\n 35 f\"Please make sure to use a correct data type, see: \"\r\n 36 f\"https://arrow.apache.org/docs/python/api/datatypes.html#factory-functions\"\r\n\r\nValueError: Neither list<item: int64> nor list<item: int64>_ seems to be a pyarrow data type. Please make sure to use a correct data type, see: https://arrow.apache.org/docs/python/api/datatypes.html#factory-functions.\r\n```\r\n\r\nIf I just create a pa- table manually like is done in the jsonloader -- it seems to work fine. Ths JSON I'm trying to load isn't overly complex - 1 integer field, the rest text fields with a nested list of objects with text fields .", "I'll close this -- It's still unclear how to go about troubleshooting the json example as I mentioned above. If I decide it's worth the trouble, I'll create another issue, or wait for a better support for using nlp for making custom data-loaders." ]

[ "I'm pretty sure you can get some inspiration from the squad_v2 script. It looks like the dataset is quite big so it will take some time for the users to generate it, but it should be reasonable.\r\n\r\nAlso you are saying that you are still making the dataset grow in size right ?\r\nIt's probably good practice to let the users do their training/evaluations with the exact same version of the dataset.\r\nWe allow for each dataset to specify a version (ex: 1.0.0) and increment this number every time there are new samples in the dataset for example. Does it look like a good solution for you ? Or would you rather have one final version with the full dataset ?", "It would also be good if there is any possibility for versioning, I think this way is much better than the dynamic way.\nIf you mean that part to put the tiles into one is the generation it would take up to 15-20 minutes on home computer hardware.\nAre there any compression or optimization algorithms while generating the dataset ?\nOtherwise the hardware limit is around 32 GB ram at the moment.\nIf everything works well we will add some more gigabytes of data in future what would make it pretty memory costly.", "15-20 minutes is fine !\r\nAlso there's no RAM limitations as we save to disk every 1000 elements while generating the dataset by default.\r\nAfter generation, the dataset is ready to use with (again) no RAM limitations as we do memory-mapping.", "Wow, that sounds pretty cool.\nActually I have the problem of running out of memory while tokenization on our local machine.\nThat wouldn't happen again, would it ?", "You can do the tokenization step using `my_tokenized_dataset = my_dataset.map(my_tokenize_function)` that writes the tokenized texts on disk as well. And then `my_tokenized_dataset` will be a memory-mapped dataset too, so you should be fine :)", "Does it have an affect to the trainings speed ?", "In your training loop, loading the tokenized texts is going to be fast and pretty much negligible compared to a forward pass. You shouldn't expect any slow down.", "Closing this one. Feel free to re-open if you have other questions" ]

[ "The glue metric for cola is a metric for classification. It expects label ids as integers as inputs.", "I want to evaluate a sentence pair whether they are semantically equivalent, so I used MRPC and it gives the same error, does that mean we have to encode the sentences and parse as input?\r\n\r\nusing BertTokenizer;\r\n```\r\nencoded_reference=tokenizer.encode(reference, add_special_tokens=False)\r\nencoded_prediction=tokenizer.encode(prediction, add_special_tokens=False)\r\n```\r\n\r\n`glue_score = glue_metric.compute(encoded_prediction, encoded_reference)`\r\n```\r\n\r\nValueError Traceback (most recent call last)\r\n<ipython-input-9-4c3a3ce7b583> in <module>()\r\n----> 1 glue_score = glue_metric.compute(encoded_prediction, encoded_reference)\r\n\r\n6 frames\r\n/usr/local/lib/python3.6/dist-packages/nlp/metric.py in compute(self, predictions, references, timeout, **metrics_kwargs)\r\n 198 predictions = self.data[\"predictions\"]\r\n 199 references = self.data[\"references\"]\r\n--> 200 output = self._compute(predictions=predictions, references=references, **metrics_kwargs)\r\n 201 return output\r\n 202 \r\n\r\n/usr/local/lib/python3.6/dist-packages/nlp/metrics/glue/27b1bc63e520833054bd0d7a8d0bc7f6aab84cc9eed1b576e98c806f9466d302/glue.py in _compute(self, predictions, references)\r\n 101 return pearson_and_spearman(predictions, references)\r\n 102 elif self.config_name in [\"mrpc\", \"qqp\"]:\r\n--> 103 return acc_and_f1(predictions, references)\r\n 104 elif self.config_name in [\"sst2\", \"mnli\", \"mnli_mismatched\", \"mnli_matched\", \"qnli\", \"rte\", \"wnli\", \"hans\"]:\r\n 105 return {\"accuracy\": simple_accuracy(predictions, references)}\r\n\r\n/usr/local/lib/python3.6/dist-packages/nlp/metrics/glue/27b1bc63e520833054bd0d7a8d0bc7f6aab84cc9eed1b576e98c806f9466d302/glue.py in acc_and_f1(preds, labels)\r\n 60 def acc_and_f1(preds, labels):\r\n 61 acc = simple_accuracy(preds, labels)\r\n---> 62 f1 = f1_score(y_true=labels, y_pred=preds)\r\n 63 return {\r\n 64 \"accuracy\": acc,\r\n\r\n/usr/local/lib/python3.6/dist-packages/sklearn/metrics/_classification.py in f1_score(y_true, y_pred, labels, pos_label, average, sample_weight, zero_division)\r\n 1097 pos_label=pos_label, average=average,\r\n 1098 sample_weight=sample_weight,\r\n-> 1099 zero_division=zero_division)\r\n 1100 \r\n 1101 \r\n\r\n/usr/local/lib/python3.6/dist-packages/sklearn/metrics/_classification.py in fbeta_score(y_true, y_pred, beta, labels, pos_label, average, sample_weight, zero_division)\r\n 1224 warn_for=('f-score',),\r\n 1225 sample_weight=sample_weight,\r\n-> 1226 zero_division=zero_division)\r\n 1227 return f\r\n 1228 \r\n\r\n/usr/local/lib/python3.6/dist-packages/sklearn/metrics/_classification.py in precision_recall_fscore_support(y_true, y_pred, beta, labels, pos_label, average, warn_for, sample_weight, zero_division)\r\n 1482 raise ValueError(\"beta should be >=0 in the F-beta score\")\r\n 1483 labels = _check_set_wise_labels(y_true, y_pred, average, labels,\r\n-> 1484 pos_label)\r\n 1485 \r\n 1486 # Calculate tp_sum, pred_sum, true_sum ###\r\n\r\n/usr/local/lib/python3.6/dist-packages/sklearn/metrics/_classification.py in _check_set_wise_labels(y_true, y_pred, average, labels, pos_label)\r\n 1314 raise ValueError(\"Target is %s but average='binary'. Please \"\r\n 1315 \"choose another average setting, one of %r.\"\r\n-> 1316 % (y_type, average_options))\r\n 1317 elif pos_label not in (None, 1):\r\n 1318 warnings.warn(\"Note that pos_label (set to %r) is ignored when \"\r\n\r\nValueError: Target is multiclass but average='binary'. Please choose another average setting, one of [None, 'micro', 'macro', 'weighted'].\r\n\r\n```", "MRPC is also a binary classification task, so its metric is a binary classification metric.\r\n\r\nTo evaluate if pairs of sentences are semantically equivalent, maybe you could take a look at models that compute if one sentence entails the other or not (typically the kinds of model that could work well on the MRPC task).", "Closing this one. Feel free to re-open if you have other questions :)" ]

[ "It looks like it comes from `mwparserfromhell`.\r\n\r\nWould it be possible to get the bad `section` that causes this issue ? The `section` string is from `datasets/wikipedia.py:L548` ? You could just add a `try` statement and print the section if the line `section_text.append(section.strip_code().strip())` crashes.\r\n\r\nIt will help us know if we have to fix it on our side or if it is a `mwparserfromhell` issue.", "Hi, \r\n\r\nThank you for you answer.\r\nI have try to print the bad section using `try` and `except`, but it is a bit weird as the error seems to appear 3 times for instance, but the two first error does not print anything (as if the function did not go in the `except` part).\r\nFor the third one, I got that (I haven't display the entire text) :\r\n\r\n> error : ==== Parque nacional Cajas ====\r\n> {{AP|Parque nacional Cajas}}\r\n> [[Archivo:Ecuador cajas national park.jpg|thumb|left|300px|Laguna del Cajas]]\r\n> El parque nacional Cajas está situado en los [[Cordillera de los Andes|Andes]], al sur del [[Ecuador]], en la provincia de [[Provincia de Azuay|Azuay]], a 33\r\n> [[km]] al noroccidente de la ciudad de [[Cuenca (Ecuador)|Cuenca]]. Los accesos más comunes al parque inician todos en Cuenca: Desde allí, la vía Cuenca-Mol\r\n> leturo atraviesa en Control de [[Surocucho]] en poco más de 30 minutos de viaje; más adelante, esta misma carretera pasa a orillas de la laguna La Toreadora donde están el Centro Administrativo y de Información del parque. Siguiendo de largo hacia [[Molleturo]], por esta vía se conoce el sector norte del Cajas y se serpentea entre varias lagunas mayores y menores.\r\n> Para acceder al parque desde la costa, la vía Molleturo-Cuenca es también la mejor opción.\r\n\r\nHow can I display the link instead of the text ? I suppose it will help you more ", "The error appears several times as Apache Beam retries to process examples up to 4 times irc.\r\n\r\nI just tried to run this text into `mwparserfromhell` but it worked without the issue.\r\n\r\nI used this code (from the `wikipedia.py` script):\r\n```python\r\nimport mwparserfromhell as parser\r\nimport re\r\nimport six\r\n\r\nraw_content = r\"\"\"==== Parque nacional Cajas ====\r\n{{AP|Parque nacional Cajas}}\r\n[[Archivo:Ecuador cajas national park.jpg|thumb|left|300px|Laguna del Cajas]]\r\nEl parque nacional Cajas está situado en los [[Cordillera de los Andes|Andes]], al sur del [[Ecuador]], en la provincia de [[Provincia de Azuay|Azuay]], a 33\r\n[[km]] al noroccidente de la ciudad de [[Cuenca (Ecuador)|Cuenca]]. Los accesos más comunes al parque inician todos en Cuenca: Desde allí, la vía Cuenca-Mol\r\nleturo atraviesa en Control de [[Surocucho]] en poco más de 30 minutos de viaje; más adelante, esta misma carretera pasa a orillas de la laguna La Toreadora donde están el Centro Administrativo y de Información del parque. Siguiendo de largo hacia [[Molleturo]], por esta vía se conoce el sector norte del Cajas y se serpentea entre varias lagunas mayores y menores.\r\n\"\"\"\r\n\r\nwikicode = parser.parse(raw_content)\r\n\r\n# Filters for references, tables, and file/image links.\r\nre_rm_wikilink = re.compile(\"^(?:File|Image|Media):\", flags=re.IGNORECASE | re.UNICODE)\r\n\r\ndef rm_wikilink(obj):\r\n return bool(re_rm_wikilink.match(six.text_type(obj.title)))\r\n\r\ndef rm_tag(obj):\r\n return six.text_type(obj.tag) in {\"ref\", \"table\"}\r\n\r\ndef rm_template(obj):\r\n return obj.name.lower() in {\"reflist\", \"notelist\", \"notelist-ua\", \"notelist-lr\", \"notelist-ur\", \"notelist-lg\"}\r\n\r\ndef try_remove_obj(obj, section):\r\n try:\r\n section.remove(obj)\r\n except ValueError:\r\n # For unknown reasons, objects are sometimes not found.\r\n pass\r\n\r\nsection_text = []\r\nfor section in wikicode.get_sections(flat=True, include_lead=True, include_headings=True):\r\n for obj in section.ifilter_wikilinks(matches=rm_wikilink, recursive=True):\r\n try_remove_obj(obj, section)\r\n for obj in section.ifilter_templates(matches=rm_template, recursive=True):\r\n try_remove_obj(obj, section)\r\n for obj in section.ifilter_tags(matches=rm_tag, recursive=True):\r\n try_remove_obj(obj, section)\r\n\r\n section_text.append(section.strip_code().strip())\r\n```", "Not sure why we're having this issue. Maybe could you get also the file that's causing that ?", "thanks for your answer.\r\nHow can I know which file is causing the issue ? \r\nI am trying to load the spanish wikipedia data. ", "Because of the way Apache Beam works we indeed don't have access to the file name at this point in the code.\r\nWe'll have to use some tricks I think :p \r\n\r\nYou can append `filepath` to `title` in `wikipedia.py:L512` for example. [[EDIT: it's L494 my bad]]\r\nThen just do `try:...except:` on the call of `_parse_and_clean_wikicode` L500 I guess.\r\n\r\nThanks for diving into this ! I tried it myself but I run out of memory on my laptop\r\nAs soon as we have the name of the file it should be easier to find what's wrong.", "Thanks for your help.\r\n\r\nI tried to print the \"title\" of the document inside the` except (mwparserfromhell.parser.ParserError) as e`,the title displayed was : \"Campeonato Mundial de futsal de la AMF 2015\". (Wikipedia ES) Is it what you were looking for ?", "Thanks a lot @Shiro-LK !\r\n\r\nI was able to reproduce the issue. It comes from [this table on wikipedia](https://es.wikipedia.org/wiki/Campeonato_Mundial_de_futsal_de_la_AMF_2015#Clasificados) that can't be parsed.\r\n\r\nThe file in which the problem occurs comes from the wikipedia dumps, and it can be downloaded [here](https://dumps.wikimedia.org/eswiki/20200501/eswiki-20200501-pages-articles-multistream6.xml-p6424816p7924815.bz2)\r\n\r\nParsing the file this way raises the parsing issue:\r\n\r\n```python\r\nimport mwparserfromhell as parser\r\nfrom tqdm.auto import tqdm\r\nimport bz2\r\nimport six\r\nimport logging\r\nimport codecs\r\nimport xml.etree.cElementTree as etree\r\n\r\nfilepath = \"path/to/eswiki-20200501-pages-articles-multistream6.xml-p6424816p7924815.bz2\"\r\n\r\ndef _extract_content(filepath):\r\n \"\"\"Extracts article content from a single WikiMedia XML file.\"\"\"\r\n logging.info(\"generating examples from = %s\", filepath)\r\n with open(filepath, \"rb\") as f:\r\n f = bz2.BZ2File(filename=f)\r\n if six.PY3:\r\n # Workaround due to:\r\n # https://github.com/tensorflow/tensorflow/issues/33563\r\n utf_f = codecs.getreader(\"utf-8\")(f)\r\n else:\r\n utf_f = f\r\n # To clear root, to free-up more memory than just `elem.clear()`.\r\n context = etree.iterparse(utf_f, events=(\"end\",))\r\n context = iter(context)\r\n unused_event, root = next(context)\r\n for unused_event, elem in tqdm(context, total=949087):\r\n if not elem.tag.endswith(\"page\"):\r\n continue\r\n namespace = elem.tag[:-4]\r\n title = elem.find(\"./{0}title\".format(namespace)).text\r\n ns = elem.find(\"./{0}ns\".format(namespace)).text\r\n id_ = elem.find(\"./{0}id\".format(namespace)).text\r\n # Filter pages that are not in the \"main\" namespace.\r\n if ns != \"0\":\r\n root.clear()\r\n continue\r\n raw_content = elem.find(\"./{0}revision/{0}text\".format(namespace)).text\r\n root.clear()\r\n\r\n if \"Campeonato Mundial de futsal de la AMF 2015\" in title:\r\n yield (id_, title, raw_content)\r\n\r\nfor id_, title, raw_content in _extract_content(filepath):\r\n wikicode = parser.parse(raw_content)\r\n```\r\n\r\nThe copied the raw content that can't be parsed [here](https://pastebin.com/raw/ZbmevLyH).\r\n\r\nThe minimal code to reproduce is:\r\n```python\r\nimport mwparserfromhell as parser\r\nimport requests\r\n\r\nraw_content = requests.get(\"https://pastebin.com/raw/ZbmevLyH\").content.decode(\"utf-8\")\r\nwikicode = parser.parse(raw_content)\r\n\r\n```\r\n\r\nI will create an issue on mwparserfromhell's repo to see if we can fix that\r\n", "This going to be fixed in the next `mwparserfromhell` release :)", "Fixed in `mwparserfromhell` version 0.6." ]

[ "I wonder if this means downloading failed? That corpus has a really slow server.", "This dataset seems to have a decoding problem that results in inconsistencies in the number of generated examples.\r\nSee #215.\r\nThat's why we end up with a `NonMatchingSplitsSizesError `." ]

[ "Oh yes, this is a backward compatibility feature with tensorflow_dataset in which a `Sequence` or `dict` is converted in a `dict` of `lists`, unfortunately it is not very intuitive, see here: https://github.com/huggingface/nlp/blob/master/src/nlp/features.py#L409\r\n\r\nTo avoid this behavior, you can just define the list in the feature with a simple list or a tuple (which is also simpler to write).\r\nIn your case, the features could be as follow:\r\n``` python\r\n...\r\nfeatures=nlp.Features({\r\n \"title\": nlp.Value(\"string\"),\r\n \"vertexSet\": [[{\r\n \"name\": nlp.Value(\"string\"),\r\n \"sent_id\": nlp.Value(\"int32\"),\r\n \"pos\": nlp.features.Sequence(nlp.Value(\"int32\")),\r\n \"type\": nlp.Value(\"string\"),\r\n }]],\r\n ...\r\n }),\r\n...\r\n```" ]

[ "For one dataset you can have different configurations that each have their own `nlp.Features`.\r\nWe imagine having one configuration per subtask for example.\r\nThey are loaded with `nlp.load_dataset(\"my_dataset\", \"my_config\")`.\r\n\r\nFor example the `glue` dataset has many configurations. It is a bit different from your case though because each configuration is a dataset by itself (sst2, mnli).\r\nAnother example is `wikipedia` that has one configuration per language." ]

[ "Update: I think I've found a solution.\r\n\r\n```python\r\noutput_types = {\"input_ids\": tf.int64, \"token_type_ids\": tf.int64, \"attention_mask\": tf.int64}\r\ndef train_dataset_gen():\r\n for i in range(len(train_dataset)):\r\n yield train_dataset[i]\r\ntf_dataset = tf.data.Dataset.from_generator(train_dataset_gen, output_types=output_types)\r\n```\r\n\r\nloads WikiText-2 in 20 ms, and WikiText-103 in 20 ms. It appears to be lazily loading via indexing train_dataset.", "Yes this is the current best solution. We should probably show it in the tutorial notebook.\r\n\r\nNote that this solution unfortunately doesn't allow to train on TPUs (yet). See #193 ", "This approach still seems quite slow. When using TFRecords with a similar training loop, I get ~3.0-3.5 it/s on multi-node, multi-GPU training. I notice a pretty severe performance regression when scaling, with observed performance numbers. Since the allreduce step takes less than 100ms/it and I've achieved 80% scaling efficiency up to 64 GPUs, it must be the data pipeline.\r\n\r\n| Nodes | GPUs | Iterations/Second |\r\n| --- | --- | --- |\r\n| 1 | 2 | 2.01 |\r\n| 1 | 8 | 0.81 |\r\n| 2 | 16 | 0.37 |\r\n\r\nHere are performance metrics over 10k steps. The iteration speed appears to follow some sort of caching pattern. I would love to use `nlp` in my project, but a slowdown from 3.0 it/s to 0.3 it/s is too great to stomach.\r\n\r\n<img width=\"1361\" alt=\"Screen Shot 2020-07-02 at 8 29 22 AM\" src=\"https://user-images.githubusercontent.com/4564897/86378156-2f8d3900-bc3e-11ea-918b-c395c3df5377.png\">\r\n", "An interesting alternative to investigate here would be to use the tf.io library which has some support for Arrow to TF conversion: https://www.tensorflow.org/io/api_docs/python/tfio/arrow/ArrowDataset\r\n\r\nThere are quite a few types supported, including lists so if the unsupported columns are dropped then we could maybe have a zero-copy mapping from Arrow to TensorFlow, including tokenized inputs and 1D tensors like the ones we mostly use in NLP: https://github.com/tensorflow/io/blob/322b3170c43ecac5c6af9e39dbd18fd747913e5a/tensorflow_io/arrow/python/ops/arrow_dataset_ops.py#L44-L72\r\n\r\nHere is an introduction on Arrow to TF using tf.io: https://medium.com/tensorflow/tensorflow-with-apache-arrow-datasets-cdbcfe80a59f", "Interesting. There's no support for strings, but it does enable int and floats so that would work for tokenized inputs. \r\n\r\nArrowStreamDataset requires loading from a \"record batch iterator\", which can be instantiated from in-memory arrays as described here: https://arrow.apache.org/docs/python/ipc.html. \r\n\r\nBut the nlp.Dataset stores its data as a `pyarrow.lib.Table`, and the underlying features are `pyarrow.lib.ChunkedArray`. I can't find any documentation about lazily creating a record batch iterator from a ChunkedArray or a Table. Have you had any success?\r\n\r\nI can't find [any uses](https://grep.app/search?q=ArrowDataset&filter[lang][0]=Python) of tfio.arrow.ArrowDataset on GitHub.", "You can use `to_batches` maybe?\r\nhttps://arrow.apache.org/docs/python/generated/pyarrow.Table.html#pyarrow.Table.to_batches", "Also note that since #322 it is now possible to do\r\n```python\r\nids = [1, 10, 42, 100]\r\nbatch = dataset[ids]\r\n```\r\nFrom my experience it is quite fast but it can take lots of memory for large batches (haven't played that much with it).\r\nLet me know if you think there could be a better way to implement it. (current code is [here](https://github.com/huggingface/nlp/blob/78628649962671b4aaa31a6b24e7275533416845/src/nlp/arrow_dataset.py#L463))", "Thanks @lhoestq! That format is much better to work with.\r\n\r\nI put together a benchmarking script. This doesn't measure the CPU-to-GPU efficiency, nor how it scales with multi-GPU multi-node training where many processes are making the same demands on the same dataset. But it does show some interesting results:\r\n\r\n```python\r\nimport nlp\r\nimport numpy as np\r\nimport tensorflow as tf\r\nimport time\r\n\r\ndset = nlp.load_dataset(\"wikitext\", \"wikitext-2-raw-v1\", split=\"train\")\r\ndset = dset.filter(lambda ex: len(ex[\"text\"]) > 0)\r\nbsz = 1024\r\nn_batches = 100\r\n\r\ndef single_item_gen():\r\n for i in range(len(dset)):\r\n yield dset[i]\r\n\r\ndef sequential_batch_gen():\r\n for i in range(0, len(dset), bsz):\r\n yield dset[i:i+bsz]\r\n\r\ndef random_batch_gen():\r\n for i in range(len(dset)):\r\n indices = list(np.random.randint(len(dset), size=(bsz,)))\r\n yield dset[indices]\r\n\r\noutput_types = {\"text\": tf.string}\r\nsingle_item = tf.data.Dataset.from_generator(single_item_gen, output_types=output_types).batch(bsz)\r\ninterleaved = tf.data.Dataset.range(10).interleave(\r\n lambda idx: tf.data.Dataset.from_generator(single_item_gen, output_types=output_types),\r\n cycle_length=10,\r\n)\r\nsequential_batch = tf.data.Dataset.from_generator(sequential_batch_gen, output_types=output_types)\r\nrandom_batch = tf.data.Dataset.from_generator(random_batch_gen, output_types=output_types)\r\n\r\ndef iterate(tf_dset):\r\n start = time.perf_counter()\r\n for i, batch in enumerate(tf_dset.take(n_batches)):\r\n pass\r\n elapsed = time.perf_counter() - start\r\n print(f\"{tf_dset} took {elapsed:.3f} secs\")\r\n\r\niterate(single_item)\r\niterate(interleaved)\r\niterate(sequential_batch)\r\niterate(random_batch)\r\n```\r\n\r\nResults:\r\n```\r\n<BatchDataset shapes: {text: <unknown>}, types: {text: tf.string}> took 23.005 secs\r\n<InterleaveDataset shapes: {text: <unknown>}, types: {text: tf.string}> took 0.135 secs\r\n<FlatMapDataset shapes: {text: <unknown>}, types: {text: tf.string}> took 0.074 secs\r\n<FlatMapDataset shapes: {text: <unknown>}, types: {text: tf.string}> took 0.550 secs\r\n```\r\n\r\n- Batching a generator which fetches a single item is terrible.\r\n- Interleaving performs well on a single process, but doesn't scale well to multi-GPU training. I believe the bottleneck here is in Arrow dataset locking or something similar. The numbers from the table above are with interleaving.\r\n- The sequential access dominates the random access (7x faster). Is there any way to bring random access times closer to sequential access? Maybe re-indexing the dataset after shuffling each pass over the data.", "Hey @jarednielsen \r\n\r\nThanks for this very interesting analysis!! IMHO to read text data one should use `tf.data.TextLineDataset`. It would be interesting to compare what you have done with simply load with a `TextLineDataset` and see if there is a difference.\r\n\r\nA good example can be found here https://www.tensorflow.org/tutorials/load_data/text", "Thanks! I'm not actually loading in raw text data, that was just the synthetic data I created for this benchmark. A more realistic use case would be a dataset of tokenized examples, which would be a dict of lists of integers. TensorFlow's TextLineDataset greedily loads the dataset into the graph itself, which can lead to out-of-memory errors - one of the main reason I'm so drawn to the `nlp` library is its zero-copy no-RAM approach to dataset loading and mapping. \r\n\r\nIt's quite helpful for running a preprocessing pipeline - a sample ELECTRA pipeline I've built is here: https://github.com/jarednielsen/deep-learning-models/blob/nlp/models/nlp/common/preprocess.py.", "Sorry, I think I badly expressed myself, my bad. What I suggested is to compare with the usual loading textual data in pure TF with `TextLineDataset` with `nlp`. I know it is not recommended with very large datasets to use it, but I was curious to see how it behaves compared to a processing with `nlp` on smaller datasets.\r\n\r\nBTW your script looks very interesting, thanks for sharing!!" ]

[ "Hi Jared,\r\nInteresting, thanks for raising this question. You can also do that after loading with `dataset.select()` or `dataset.filter()` which let you keep only a specific subset of rows in a dataset.\r\nWhat is your use-case for sharding?", "Thanks for the pointer to those functions! It's still a little more verbose since you have to manually calculate which ids each rank would keep, but definitely works.\r\n\r\nMy use case is multi-node, multi-GPU training and avoiding global batches of duplicate elements. I'm using horovod. You can shuffle indices, or set random seeds, but explicitly sharding the dataset up front is the safest and clearest way I've found to do so." ]

[]

[]

[]

[ "Even more complicating - \r\n\r\nAs I see it, datasets can have \"addons\".\r\nFor example, the WebNLG dataset is a dataset for data-to-text. However, a work of mine and other works enriched this dataset with text plans / underlying text structures. In that case, I see a need to load the dataset \"WebNLG\" with \"plans\" addon.\r\n\r\nSame for sign language - if there is a dataset of videos, one addon can be to run OpenPose, another to run ARKit4 pose estimation, and another to run PoseNet, or even just a video embedding addon. (which are expensive to run individually for everyone who wants to use these data)\r\n\r\nThis is something I dabbled with my own implementation to a [research datasets library](https://github.com/AmitMY/meta-scholar/) and I love to get the discussion going on these topics.", "This is a really cool idea !\r\nThe example for data objects you gave for the RWTH-PHOENIX-Weather 2014 T dataset can totally fit inside the library.\r\n\r\nFor your point about formats like `ilex`, `eaf`, or `srt`, it is possible to use any library in your dataset script.\r\nHowever most user probably won't need these libraries, as most datasets don't need them, and therefore it's unlikely that we will have them in the minimum requirements to use `nlp` (we want to keep it as light-weight as possible). If a user wants to load your dataset and doesn't have the libraries you need, an error is raised asking the user to install them.\r\n\r\nMore generally, we plan to have something like a `requirements.txt` per dataset. This could also be a place for addons as you said. What do you think ?", "Thanks, Quentin, I think a `requirements.txt` per dataset will be a good thing.\r\nI will work on adding this dataset next week, and once we sort all of the kinks, I'll add more." ]

[ "Whoops didn't mean to close this one.\r\nI did some changes, could you try to run it from the master branch ?", "Now it works, thanks!" ]

[ "Thanks for reporting these errors :)\r\n\r\nI can actually see two issues here.\r\n\r\nFirst, datasets like `natural_questions` require apache_beam to be processed. Right now the import is not at the right place so we have this error message. However, even the imports are fixed, the nlp viewer doesn't actually have the resources to process NQ right now so we'll have to wait until we have a version that we've already processed on our google storage (that's what we've done for wikipedia for example).\r\n\r\nSecond, datasets like `newsroom` require manual downloads as we're not allowed to redistribute the data ourselves (if I'm not wrong). An error message should be displayed saying that we're not allowed to show the dataset.\r\n\r\nI can fix the first issue with the imports but for the second one I think we'll have to see with @srush to show a message for datasets that require manual downloads (it can be checked whether a dataset requires manual downloads if `dataset_builder_instance.manual_download_instructions is not None`).\r\n\r\n", "I added apache-beam to the viewer. We can think about how to add newsroom. ", "We don't plan to host the source files of newsroom ourselves for now.\r\nYou can still get the dataset if you follow the download instructions given by `dataset = load_dataset('newsroom')` though.\r\nThe viewer also shows the instructions now.\r\n\r\nClosing this one. If you have other questions, feel free to re-open :)" ]

[ "@jxmorris12 , we use `-1` to label examples for which `gold label` is missing (`gold label = -` in the original dataset). ", "Thanks @mariamabarham! so the original dataset is missing some labels? That is weird. Is standard practice just to discard those examples training/eval?", "Yes the original dataset is missing some labels maybe @sleepinyourhat , @gangeli can correct me if I'm wrong \r\nFor my personal opinion at least if you want your model to learn to predict no answer (-1) you can leave it their but otherwise you can discard them. ", "thanks @mariamabarham :)" ]

[]

[ "I couldn't replicate this issue on my macbook :/\r\nCould you try to play with different encodings in `with open(path, encoding=...) as f` in scientific_papers.py:L108 ?", "I was able to track down the file causing the problem by adding the following to `scientific_papers.py` (starting at line 116):\r\n\r\n```python\r\n from json import JSONDecodeError\r\n try:\r\n d = json.loads(line)\r\n summary = \"\\n\".join(d[\"abstract_text\"])\r\n except JSONDecodeError:\r\n print(path, line)\r\n```\r\n\r\n\r\n\r\nFor me it was at: `/Users/johngiorgi/.cache/huggingface/datasets/f87fd498c5003cbe253a2af422caa1e58f87a4fd74cb3e67350c635c8903b259/arxiv-dataset/train.txt` with `\"article_id\": \"1407.3051\"`.\r\n\r\nNot really 100% sure at the moment, but it looks like this specific substring from `\"article_text\"` may be causing the problem?\r\n\r\n```\r\n\"after the missing - mass scale adjustment , the validity of the corrections was tested in the @xmath85 productions at 1.69 gev/@xmath1 . in fig . [\", \"fig : calibrations ] ( a ) , we show the missing - mass spectrum in the @xmath86 region in the @xmath87 reaction at 1.69 gev/@xmath1 . a fitting result with a lorentzian function for the @xmath86 ( dashed line ) and the three - body phas\r\n```\r\n\r\nperhaps because it appears to be truncated. I (think) I can recreate the problem by doing the following:\r\n\r\n```python\r\nimport json\r\n\r\n# A minimal example of the json file that causes the error\r\ninvalid_json = '{\"article_id\": \"1407.3051\", \"article_text\": [\"the missing - mass resolution was obtained to be 2.8 @xmath3 0.1 mev/@xmath4 ( fwhm ) , which corresponds to the missing - mass resolution of 3.2 @xmath3 0.2 mev/@xmath4 ( fwhm ) at the @xmath6 cusp region in the @xmath0 reaction .\", \"this resolution is at least by a factor of 2 better than the previous measurement with the same reaction ( [email protected] mev/@xmath4 in @xmath84 ) @xcite .\", \"after the missing - mass scale adjustment , the validity of the corrections was tested in the @xmath85 productions at 1.69 gev/@xmath1 . in fig . [\", \"fig : calibrations ] ( a ) , we show the missing - mass spectrum in the @xmath86 region in the @xmath87 reaction at 1.69 gev/@xmath1 . a fitting result with a lorentzian function for the @xmath86 ( dashed line ) and the three - body phas' \r\n# The line of code from `scientific_papers.py` which appears to cause the error\r\njson.loads(invalid_json)\r\n```\r\n\r\nThis is as far as I get before I am stumped.", "I just checked inside `train.txt` and this line isn't truncated for me (line 163577).\r\nCould you try to clear your cache and re-download the dataset ?", "Ah the turn-it-off-turn-it-on again solution! That did it, thanks a lot :) " ]