Datasets:

---

ans-imran

/

github-issues

like 0

[]

[]

[ "Hi ! Thanks for reporting\r\nIf I recall correctly this has been recently fixed #1995\r\nCan you try to upgrade your local version of `datasets` ?\r\n```\r\npip install --upgrade datasets\r\n```", "Hi Ihoestq,\r\n\r\nThank you. It works after upgrading the datasets\r\n" ]

[ { "color": "d876e3", "default": true, "description": "Further information is requested", "id": 1935892912, "name": "question", "node_id": "MDU6TGFiZWwxOTM1ODkyOTEy", "url": "https://api.github.com/repos/huggingface/datasets/labels/question" } ]

[]

[ "An educated guess: does this refer to the fact that depending on the custom column names in the dataset files (csv in this case), there is a dataset loader being created? and this dataset loader - using the \"custom data configuration\" is used among all jobs running using this particular csv files? (thinking out loud here...)\r\n\r\nIf this is the case, it may be ok for my use case (have to think about it more), still a bit surprising given that datasets caching is disabled (or so I hope) by the lines I pasted above. ", "Hi ! Currently disabling the caching means that all the dataset transform like `map`, `filter` etc. ignore the cache: it doesn't write nor read processed cache files.\r\nHowever `load_dataset` reuses datasets that have already been prepared: it does reload prepared dataset files.\r\n\r\nIndeed from the documentation:\r\n> datasets.set_caching_enabled(boolean: bool)\r\n\r\n> When applying transforms on a dataset, the data are stored in cache files. The caching mechanism allows to reload an existing cache file if it’s already been computed.\r\n> Reloading a dataset is possible since the cache files are named using the dataset fingerprint, which is updated after each transform.\r\n> If disabled, the library will no longer reload cached datasets files when applying transforms to the datasets. More precisely, if the caching is disabled:\r\n> - cache files are always recreated\r\n> - cache files are written to a temporary directory that is deleted when session closes\r\n> - cache files are named using a random hash instead of the dataset fingerprint - use datasets.Dataset.save_to_disk() to save a transformed dataset or it will be deleted when session closes\r\n> - caching doesn’t affect datasets.load_dataset(). If you want to regenerate a dataset from scratch you should use the download_mode parameter in datasets.load_dataset().", "Thank you for the clarification. \r\n\r\nThis is a bit confusing. On one hand, it says that cache files are always recreated and written to a temporary directory that is removed; on the other hand the last bullet point makes me think that since the default according to the docs for `download_mode (Optional datasets.GenerateMode) – select the download/generate mode - Default to REUSE_DATASET_IF_EXISTS` => it almost sounds that it could reload prepared dataset files. Where are these files stored? I guess not in the temporary directory that is removed... \r\n\r\nI find this type of api design error-prone. When I see as a programmer `datasets.set_caching_enabled(False)` I expect no reuse of anything in the cache. ", "It would be nice if the documentation elaborated on all the possible values for `download_mode` and/or a link to `datasets.GenerateMode`. \r\nThis info here:\r\n```\r\n \"\"\"`Enum` for how to treat pre-existing downloads and data.\r\n The default mode is `REUSE_DATASET_IF_EXISTS`, which will reuse both\r\n raw downloads and the prepared dataset if they exist.\r\n The generations modes:\r\n | | Downloads | Dataset |\r\n | -----------------------------------|-----------|---------|\r\n | `REUSE_DATASET_IF_EXISTS` (default)| Reuse | Reuse |\r\n | `REUSE_CACHE_IF_EXISTS` | Reuse | Fresh |\r\n | `FORCE_REDOWNLOAD` | Fresh | Fresh |\r\n```", "I have another question. Assuming that I understood correctly and there is reuse of datasets files when caching is disabled (!), I'm guessing there is a directory that is created based on some information on the dataset file. I'm interested in the situation where I'm loading a (custom) dataset from local disk. What information is used to create the directory/filenames where the files are stored?\r\n\r\nI'm concerned about the following scenario: if I have a file, let's say `train.csv` at path `the_path`, run once, the dataset is prepared, some models are run, etc. Now let's say there is an issue and I recreate `train.csv` at the same path `the_path`. Is there enough information in the temporary name/hash to *not* reload the *old* prepared dataset (e.g., timestamp of the file)? Or is it going to reload the *old* prepared file? ", "Thanks for the feedback, we'll work in improving this aspect of the documentation.\r\n\r\n> Where are these files stored? I guess not in the temporary directory that is removed...\r\n\r\nWe're using the Arrow file format to load datasets. Therefore each time you load a dataset, it is prepared as an arrow file on your disk. By default the file is located in the ~/.cache/huggingface/datasets/<dataset_name>/<config_id>/<version> directory.\r\n\r\n> What information is used to create the directory/filenames where the files are stored?\r\n\r\nThe config_id contains a hash that takes into account:\r\n- the dataset loader used and its source code (e.g. the \"csv\" loader)\r\n- the arguments passed to the loader (e.g. the csv delimiter)\r\n- metadata of the local data files if any (e.g. their timestamps)\r\n\r\n> I'm concerned about the following scenario: if I have a file, let's say train.csv at path the_path, run once, the dataset is prepared, some models are run, etc. Now let's say there is an issue and I recreate train.csv at the same path the_path. Is there enough information in the temporary name/hash to not reload the old prepared dataset (e.g., timestamp of the file)? Or is it going to reload the old prepared file?\r\n\r\nYes the timestamp of the local csv file is taken into account. If you edit your csv file, the config_id will change and loading the dataset will create a new arrow file.", "Thank you for all your clarifications, really helpful! \r\n\r\nIf you have the bandwidth, please do revisit the api wrt cache disabling. Anywhere in the computer stack (hardware included) where you disable the cache, one assumes there is no caching that happens. ", "That makes total sense indeed !\r\nI think we can do the change", "I have another question about caching, this time in the case where FORCE_REDOWNLOAD is used to load the dataset, the datasets cache is one directory as defined by HF_HOME and there are multiple concurrent jobs running in a cluster using the same local dataset (i.e., same local files in the cluster). Does anything in the naming convention and/or file access/locking that you're using prevent race conditions between the concurrent jobs on the caching of the local dataset they all use?\r\n\r\nI noticed some errors (can provide more details if helpful) in load_dataset/prepare_split that lead to my question above. \r\n\r\nLet me know if my question is clear, I can elaborate more if needed @lhoestq Thank you!", "I got another error that convinces me there is a race condition (one of the test files had zero samples at prediction time). I think it comes down to the fact that the `config_id` above (used in the naming for the cache) has no information on who's touching the data. If I have 2 concurrent jobs, both loading the same dataset and forcing redownload, they may step on each other foot/caching of the dataset. ", "We're using a locking mechanism to prevent two processes from writing at the same time. The locking is based on the `filelock` module.\r\nAlso directories that are being written use a suffix \".incomplete\" so that reading is not possible on a dataset being written.\r\n\r\nDo you think you could provide a simple code to reproduce the race condition you experienced ?", "I can provide details about the code I'm running (it's really-really close to some official samples from the huggingface transformers examples, I can point to the exact sample file, I kept a record of that). I can also describe in which conditions this race occurs (I'm convinced it has to do with forcing the redownloading of the dataset, I've been running hundreds of experiments before and didn't have a problem before I forced the redownload). I also can provide samples of the different stack errors I get and some details about the level of concurrency of jobs I was running. I can also try to imagine how the race manifests (I'm fairly sure that it's a combo of one job cleaning up and another job being in the middle of the run).\r\n\r\nHowever, I have to cleanup all this to make sure I'm no spilling any info I shouldn't be spilling. I'll try to do it by the end of the week, if you think all this is helpful. \r\n\r\nFor now, I have a workaround. Don't use forcing redownloading. And to be ultra careful (although I don't think this is a problem), I run a series of jobs that will prepare the datasets and I know there is no concurrency wrt the dataset. Once that's done (and I believe even having multiple jobs loading the datasets at the same time doesn't create problems, as long as REUSE_DATASET_IF_EXISTS is the policy for loading the dataset, so the filelock mechanism you're using is working in that scenario), the prepared datasets will be reused, no race possible in any way. \r\n\r\nThanks for all the details you provided, it helped me understand the underlying implementation and coming up with workarounds when I ran into issues. ", "Hi! I have the same challenge with caching, where the **.cache** folder is required even though it isn't possible for me.\r\n\r\nI'd like to run transformers in Snowflake, using Snowpark for Python, this would mean I could provide configurable transformers in real-time for business users without having data leave an environment (for security reasons). With no need for data transfer,n the compute is faster. It is a large use case - is it possible to entirely disable caching in certain scenarios?\r\n@lhoestq ?\r\n", "You can try to change the location of the cache folder using the `HF_CACHE_HOME` environment variable, and set a location where you have read/write access.", "Thanks @lhoestq \r\n\r\nI wanted to do that, however, snowflake does not allow it to write at all. I'm asking around to see if they can help me out with that issue 😅" ]

[]

[]

[ "cc @sebastiangehrmann" ]

[]

[]

[ "Hi, one workaround would be to save the mapped(tokenized in your case) file using `save_to_disk`, and having each process load this file using `load_from_disk`. This is what I am doing, and in this case, I turn off the ability to automatically load from the cache.\r\n\r\nAlso, multiprocessing the map function seems to be slower at the moment (#1992), hope this helps you.", "Thanks @hwijeen for the workaround, feels a bit prototypical but it works! (it seems files are written twice then though)\r\n\r\n(I haven't observed slowness using multiprocessed map function but I could be wrong)", "To my understanding, files are written twice anyhow(one after load_dataset, another aftet map). It's just that you now have it at the location where you can see, whereas it was secretlely saved at caching folder(.cache/huggingface/datasets by default)! Correct me if I'm wrong!", "Slowness in multiprocessing has been observed in certain environments but not others. We're investigating ;)", "So to answer my initial question, I was just doing something stupid as I was not re-giving the `preprocessing_num_workers` arguments when launching the distributed training (and it was then set to `None`). I initially thought the hash was computed only with the `tokenize_function` but it's all arguments. Thanks @lhoestq for clarifying!", "This cache process isn't really consistent. I just changed `per_device_train_batch_size` of training script and now it rebuilding the dataset cache!!!! Why?", "Hi ! A `map` function is recomputed if the code changes or if any of the variables it uses changes. Can you check that your function doesn't use `per_device_train_batch_size` or any variable that contains `per_device_train_batch_size` ?", "My code is actually a transformer's example for training t5, I modified a bit:\r\n\r\nhttps://github.com/puraminy/transformers/blob/4b40877132eedb566043f83de8f1d29a84d71430/examples/flax/language-modeling/run_t5_mlm_flax.py#L614\r\n\r\nNo, it doesn't use `per_device_train_batch_size`. I remember it worked for several times and then for no reason or various reasons like the above it started to build the cache again, as if it had an expiration date (maybe), or maybe I had changed the code! \r\n\r\nSo, to get rid of these problems I saved cache with a name (was forced to not use multiple_processes, because otherwise it generates multiple files) and then I load it from this cache file. " ]

[]

[]

[ "Made the required changes @lhoestq , sorry it took so much time!" ]

[]

[]

[]

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

[]

[ "I ping @krandiash to keep him up to date.", "TODO:\r\n- [x] Add a section in the docs about this.\r\n- ~Add a warning if someone tries to specify `cache_file_name=` in `map`, `filter` etc. on a dataset that is in memory, since the computation is not going to be cached in this case.~", "@lhoestq I have a question, regarding:\r\n> Also maybe we should add a warning if someone tries to specify cache_file_name= in map, filter etc. on a dataset that is in memory, since the computation is not going to be cached in this case.\r\n\r\n- It might be the case that the user has an in-memory dataset and might want to use `map` and cache it, by passing `cache_file_name=`\r\n- This is indeed allowed by the library and works as expected: the dataset is cached.\r\n\r\nWhy adding a warning?", "Yes right, I meant if `load_from_cache_file` is set to True and `cache_file_name ` is None. my bad :p" ]

[]

[]

[ "Hi ! Can you try to increase the block size ? For example\r\n```python\r\nblock_size_10MB = 10<<20\r\nload_dataset(\"json\", ..., block_size=block_size_10MB)\r\n```\r\nThe block size corresponds to how much bytes to process at a time from the input stream.\r\nThis will determine multi-threading granularity as well as the size of individual chunks in the dataset.\r\n\r\nYou can also try with bigger block sizes if needed", "Hi @lhoestq! Thank you for your prompt reply.\r\nI have experimented with (10<<20, 10<<28, 10<<30, 10<<33, 10<<34), since my machine has 192G of memory, but it's either the above-mentioned error or processed killed because of OOM.\r\n\r\nCould you give me a bit of background on why block size needs to be exactly calibrated?\r\nTo my understanding, small block sized should run just fine despite its slowness..\r\n\r\n\r\n", "We're using the JSON loader of pyarrow. It parses the file chunk by chunk to load the dataset.\r\nThis issue happens when there's no delimiter in one chunk of data. For json line, the delimiter is the end of line.\r\nSo with a big value for chunk_size this should have worked unless you have one extremely long line in your file.\r\n\r\nAlso what version of pyarrow are you using ?\r\n\r\nFInally I wonder if it could be an issue on pyarrow's side when using big json files. (I haven't tested big json files like yours)", "I'm using `pyarrow==3.0.0` with `datasets==1.5.0`.\r\n\r\nYour point totally makes sense. I will check if my jsonl file contains an extremely long file and let you know. \r\n\r\nHere are some different error messages that I got when tweaking `block_size`. I also suspect that this is related to the pyarrow... but I guess it would be wonderful if datasesets could give a clear guide on how to play with large datasets! (I am suddenly experiencing various issue when working with large datasets.. e.g. #1992 )\r\n```python\r\n return paj.ReadOptions(use_threads=self.use_threads, block_size=self.block_size)\r\n File \"pyarrow/_json.pyx\", line 56, in pyarrow._json.ReadOptions.__init__\r\n File \"pyarrow/_json.pyx\", line 81, in pyarrow._json.ReadOptions.block_size.__set__\r\nOverflowError: value too large to convert to int32_t\r\n```\r\n\r\n```python\r\n\r\nline 83, in _generate_tables\r\n parse_options=self.config.pa_parse_options,\r\n File \"pyarrow/_json.pyx\", line 247, in pyarrow._json.read_json\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 84, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: Exceeded maximum rows\r\n```", "I am getting the same error. When I tweak the block_size, I also find:\r\n`OverflowError: value too large to convert to int32_t`\r\nand \r\n`pyarrow.lib.ArrowInvalid: Exceeded maximum rows`\r\n", "I made more tests. I used a smaller dataset and I was getting the same error, which means that it was not necessarily linked to the dataset size. To make both my smaller and larger datasets work, I got rid of lists with the json file. I had the following data format:\r\n```python\r\n[\r\n {'key': \"a\", 'value': ['one', 'two', 'three']},\r\n {'key': \"b\", 'value': ['four', 'five', 'six']}\r\n]\r\n```\r\nI changed to:\r\n\r\n```python\r\n {'key': \"a\", 'value': 'one\\ntwo\\nthree'},\r\n {'key': \"b\", 'value': 'four\\nfive\\nsix']}\r\n```\r\nand that worked!\r\n\r\nI used the following to reformat my json file:\r\n```python\r\nwith open(file_name, \"w\", encoding=\"utf-8\") as f:\r\n for item in list_:\r\n f.write(json.dumps(item) + \"\\n\")\r\n```\r\nThis works with `block_size_10MB = 10 << 20` or without specifying `block_size`.", "Thanks @hwijeen for reporting and thanks @jpilaul for pointing this out.\r\n\r\nIndeed, those are different JSON-like formats:\r\n- the first one is the **standard JSON** format: all the file content is JSON-valid, thus all content is either a JSON object (between curly brackets `{...}`) or a JSON array (between square brackets `[...]`)\r\n- the second one is called **JSON Lines**: the entire file content is not JSON-valid, but only every line (newline-delimited) is JSON-valid\r\n\r\nCurrently PyArrow only supports **JSON Lines** format: \r\n- https://arrow.apache.org/docs/python/generated/pyarrow.json.read_json.html\r\n > Currently only the line-delimited JSON format is supported.\r\n- https://arrow.apache.org/docs/python/json.html\r\n > Arrow supports reading columnar data from line-delimited JSON files.", "Thanks @albertvillanova for your explanation, it is helpful to know (maybe add to docs?)!\r\nHowever, the problem I described above happened when I was dealing with jsonl files 😿\r\nAlthough I did not thoroughly inspect, I suspect the cause was the one extremely long document in my case.", "I see... I guess there is another problem going one then, related to the size." ]

[]

[]

[]

[ { "color": "a2eeef", "default": true, "description": "New feature or request", "id": 1935892871, "name": "enhancement", "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement" }, { "color": "c5def5", "default": false, "description": "Generic discussion on the library", "id": 2067400324, "name": "generic discussion", "node_id": "MDU6TGFiZWwyMDY3NDAwMzI0", "url": "https://api.github.com/repos/huggingface/datasets/labels/generic%20discussion" } ]

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

[]

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

[]

[ "I addressed your comments about the docstrings and the output validation :)", "I updated the bleurt mocking method and bleurt test is passing now.\r\nI also ran the slow tests and they are passing for bleurt.", "Thanks @lhoestq and @albertvillanova !" ]

[]

[]

[]

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

[]

[ "Hi @nelson-liu!\r\nHere is what I do to convert a string to class label:\r\n\r\n```python\r\nfrom datasets import load_dataset, features\r\n\r\n\r\ndset = load_dataset(...)\r\ncol_name = \"the string column name\"\r\n\r\nclass_names = dset.unique(col_name)\r\nclass_feature = features.ClassLabel(names=sorted(class_names))\r\ndset = dset.map(lambda str_value: {col_name: class_feature.str2int(str_value)}, input_columns=col_name)\r\n\r\ndset = dset.cast(features.Features({\r\n ...\r\n col_name: class_feature\r\n})\r\n```\r\n", "Hi! You can use `Dataset.class_encode_column` for this. And in the next release of `datasets` (this feature is only available on `master`), you'll also be able to use `cast` to do the conversion. \r\n\r\nAn example of conversion via `cast`: \r\n```python\r\nfrom datasets import Dataset, Features, ClassLabel\r\nd = Dataset.from_dict({\"a\": [\"no\", \"yes\", \"no\"]})\r\nd = d.cast(Features({\"a\": ClassLabel(names=[\"yes\", \"no\"])}))\r\n```" ]

[]

[]

[ "Actually, I found the answer [here](https://github.com/facebookresearch/faiss/wiki/FAQ#what-does-it-mean-when-a-search-returns--1-ids). \r\n\r\nSo we have to do some modifications to the code for instances where the index doesn't retrieve any IDs.", "@lhoestq @patrickvonplaten \r\n\r\nI also found another short bug in the retrieval part. Especially, when retrieving documents. If Faiss returns the -1 as the index, the retriever will always use the last element in the dataset.\r\n\r\nplease check [def get_doc_dicts function](https://github.com/huggingface/transformers/blob/master/src/transformers/models/rag/retrieval_rag.py#L222)\r\n\r\n\r\nDoes the use of the HNSW guarantee to retrieve valid indexes always? \r\n\r\n", "Hi !\r\nNo it happens sometimes to return -1, especially if your dataset is small.\r\nIf your dataset is big enough it shouldn't happen in my experience.\r\n\r\nIdeally we should ignore all the -1 that are returned. It should be possible to change that in RAG's code ", "I also checked with some indexes it returns more -1s. Specially with IVF\nwhen nprobr is very low. It doesn't happen when using HNSW though. But at\nthe moment if it happens, dataset will always return the last element.\nMaybe we should change it to repeat the most last valid retrieved doc id.\nWhat do you think?\n\nOn Wed, Apr 7, 2021, 21:09 Quentin Lhoest ***@***.***> wrote:\n\n> Hi !\n> No it happens sometimes to return -1, especially if your dataset is small.\n> If your dataset is big enough it shouldn't happen.\n>\n> Ideally we should ignore all the -1 that are returned. It should be\n> possible to change that in RAG's code\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2175#issuecomment-814746509>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGTENOTLBEZTXEO2RS3THQOMPANCNFSM42PRVYDA>\n> .\n>\n", "That would be an easy way to workaround this issue. Feel free to open a PR on `transformers` and ping me ! :)", "Sure. Will push everything together with RAG end to end. :) thanks a lot.\n\nOn Wed, Apr 7, 2021, 21:16 Quentin Lhoest ***@***.***> wrote:\n\n> That would be an easy way to workaround this issue. Feel free to open a PR\n> on transformers and ping me ! :)\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2175#issuecomment-814752589>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGWLROCGARKN7WOJYSTTHQPH5ANCNFSM42PRVYDA>\n> .\n>\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Also you can ignore the CI failing on `docs`, this has been fixed on master :)", "@lhoestq I need to update other stuff on GEM later today too, so will merge this one and remove columns in the next PR!", "Ok !" ]

[]

[]

[ "It seems that this can be fixed from user's end by including a `date` argument, like this:\r\n\r\n`dataset = datasets.load_dataset('wikipedia', '20200501.en', date='20210420')`\r\n\r\nYou can get available dates from [here](https://dumps.wikimedia.org/enwiki/).\r\n\r\nThis is not a proper fix however as all the files will still have '20200501' in their file names." ]

[]

[]

[ "Hi ! Thanks for suggesting this fix \r\nUnfortunately it looks like it's already been fixed by #2111 \r\n\r\nFeel free to share your thoughts about this PR !\r\n\r\nI'm closing this one if you don't mind." ]

[]

[]

[ "Thanks for diving into this !\r\n\r\nBefore going further, I just want to make sure if using `eval` is the right solution\r\nPersonally I'm not a big fan of `eval` since it has many security concerns. Also storing string representations of python objects in the json files is not ideal either IMO, so maybe it's possible to change this aspect instead.\r\n\r\nMaybe it would be better to convert the `_RelativeInstruction` to a string (or \"specs\") ?\r\nIt looks like `ReadInstruction.from_spec` already exists, but not the other way around.\r\nThe specs are the string representation of instructions. For example: `train+validation[:50%]`.\r\n\r\nLet me know what you think ! And thanks again, this issue has been here for a while now ^^", "@lhoestq Yes, before going with `eval`, I thought about this approach with the \"spec\". The only issue with this approach is that we have to come up with a represenation for the `rounding` arg.\r\n\r\nWhat do you think about this (maybe too verbose)?\r\n```python\r\n>>> print(ReadInstruction(\"train\", rounding=\"pct1_dropremainder\", from_=10, to=30).to_spec())\r\ntrain[10:30](pct1_dropremainder)", "Good idea !\r\n\r\nFirst we must note that the rounding is only used for percentage instructions.\r\nFor absolute instructions there's no rounding ambiguity.\r\n\r\nBy default the rounding is set to `closest`. For example if you have a train set of 999 examples and if you provide an instruction spec `\"train[:1%]\"`, you're going to get the first ten examples (while the `pct1_dropremainder ` rounding would return 9 examples).\r\n\r\nCurrently there's no way to get an instruction with a `pct1_dropremainder` rounding strategy from an instruction spec.\r\nSo we can either drop the support of `pct1_dropremainder` or define a way to use this strategy from a spec.\r\nI don't think dropping `pct1_dropremainder` would be a good idea since it allows to load each percent to all have the same number of examples (even the last one). Therefore I think your suggestion makes total sense and we should add a representation of this rounding strategy.\r\n\r\nI like what you suggested `train[10%:30%](pct1_dropremainder)` is fine, and it seems compatible with the regex that parses the instructions specs.", "@lhoestq I've made the changes as you suggested. Ready for the review.", "@lhoestq I've added a test and addressed the comments.\r\n\r\nAdditionally, `ReadInstruction` is converted to its spec form in `builder.py` to avoid a circular import that would happen if this logic was in `arrow_reader.py`. If you think it's better to have this logic in `arrow_reader.py`, the import can be delayed by putting it inside a function." ]

[]

[]

[]

[ { "color": "72f99f", "default": false, "description": "Discussions on the datasets", "id": 2067401494, "name": "Dataset discussion", "node_id": "MDU6TGFiZWwyMDY3NDAxNDk0", "url": "https://api.github.com/repos/huggingface/datasets/labels/Dataset%20discussion" } ]

[]

[ "Hi @vyraun ! The test references for CommonGen are not publicly available: you can reach out to the original dataset authors if you would like to ask for them, but we will not be releasing them as part of GEM (March 31st was the release date for the test set inputs, references are incidentally released for some of the test sets but shouldn't really be used for benchmark submissions)\r\n\r\ncc @sebastiangehrmann", "Oh okay, thanks @yjernite ! " ]

[]

[]

[ "Hi,\r\n\r\na HF dataset can be converted to a Torch Dataset with a simple wrapper as follows:\r\n```python\r\nfrom torch.utils.data import Dataset\r\n \r\nclass HFDataset(Dataset):\r\n def __init__(self, dset):\r\n self.dset = dset\r\n\r\n def __getitem__(self, idx):\r\n return self.dset[idx]\r\n\r\n def __len__(self):\r\n return len(self.dset)\r\n\r\ntrain_ds = HFDataset(train_ds)\r\n```\r\n@lhoestq Since the Arrow Dataset already provides `__getitem__` and `__len__`, I think we could use the [virtual subclass](https://docs.python.org/3/library/abc.html#abc.ABCMeta.register) mechanism from the `abc` module to elegantly solve this issue. This mechanism would allow the Arrow Dataset to be used in place of the Torch Dataset because the `isinstance(instance of Arrow Dataset, TorchDataset)` check would return True (DeepSpeed has this check [here](https://github.com/microsoft/DeepSpeed/blob/ab5534fc4c0f8ca21ada321f9730d723aa31288b/deepspeed/runtime/engine.py#L823)).\r\n\r\nAnd it requires a minimal change in the `arrow_dataset.py` file:\r\n```python\r\nif config.TORCH_AVAILABLE:\r\n from torch.utils.data import Dataset as TorchDataset\r\n TorchDataset.register(Dataset)\r\n```", "Interesting ! Thanks for sharing this @mariosasko . I like the idea\r\nThis looks like something we should add IMO", "@mariosasko \r\nThx for your code!\r\nIt perfectly works with a small modification for HF NLP dataset:\r\n```\r\noriginal_ds = nlp.load_dataset('scientific_papers', 'arxiv')\r\ntrain_ds = HFDataset(train_ds['train']) # needs splitting\r\n```", "@lhoestq Sadly, from Python 3.7 onwards `torch.utils.data.Dataset` doesn't support the virtual subclass mechanism due to `typing.Generic` type no longer having `abc.ABCMeta` as its metaclass.\r\n\r\nWith that in mind, another option is to remove a direct type check (`isinstance(dataset, torch.utils.data.Dataset)`) in `deepspeed.initalize` and to rewrite the checks in a manner similar to `torch.utils.data.DataLoader` ([link](https://github.com/pytorch/pytorch/blob/b80c6f863f2327c712c478f67c248b94d66b65ac/torch/utils/data/dataloader.py#L197-L239)). This is exactly why the `DataLoader` works with arbitrary objects that provide `__getitem__` and `__len__` (and in our case, the `ArrowDataset`). By doing so, their code wouldn't be any stricter in comparison to the `DataLoader`.\r\n\r\nSo if you agree, I can open an issue in their repo and fix this if they like the idea.", "That makes sense ! Feel free to open an issue on their repo and discuss this idea", "@y-rokutan Hi, now if you install `deepspeed` from master (this feature will be available in the next official release), the code should work without subclassing. Let us know if you still have any issues.", "Worth mentioning that any function that expects a `torch..Dataset` (like `torch..DataLoader`) will fail a mypy-esque typecheck if a `datasets.Dataset` is passed, even though it implements the interface correctly (I think). The virtual subclass idea was a good one- I wonder if there's another workaround given the Generic issue. What we're really talking about is something similar to the structural subtyping semantics that `typing.Protocol` defines. If `torch..DataLoader` accepted anything that supports `__getitem__` and `__len__` methods this would be much easier. Not sure if there's a way to do this without the wrapper from the perspective of `datasets`." ]

[]

[]

[]

[]

[]

[ "Hi @mariosasko,\r\nJust came across this PR and I was wondering if we can use\r\n`description = \"\\n\\n\".join(OrderedDict.fromkeys([info.description for info in dataset_infos]))`\r\n\r\nThis will obviate the need for `unique` and is almost as fast as `set`. We could have used `dict` inplace of `OrderedDict` but it's available 3.7+ onwards", "Hi,\r\n\r\nlet's see what @lhoestq thinks. Although my approach adds more code, it's more readable IMO.", "Yeah, that's true. Your approach is more readable." ]

[ { "color": "94203D", "default": false, "description": "", "id": 2107841032, "name": "nlp-viewer", "node_id": "MDU6TGFiZWwyMTA3ODQxMDMy", "url": "https://api.github.com/repos/huggingface/datasets/labels/nlp-viewer" } ]

[]

[ "This looks like an issue with the cc100 dataset itself but not sure\r\nDid you try loading cc100 on your machine ?", "Hi\nloading works fine, but the viewer only is broken\nthanks\n\nOn Wed, Apr 7, 2021 at 12:17 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> This looks like an issue with the cc100 dataset itself but not sure\n> Did you try loading cc100 on your machine ?\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2162#issuecomment-814793809>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMRUO33JSOYGT6RETWLTHQWNLANCNFSM42IUOR6Q>\n> .\n>\n", "Hi! This visualization tool is deprecated now. The viewer at https://huggingface.co/datasets/cc100 works fine, so I'm closing this issue." ]

[]

[]

[ "Not yet but it’s on the short/mid-term roadmap (requested by many indeed).", "oh, great, really awesome feature to have, thank you very much for the great, fabulous work", "We'll work on dataset streaming soon. This should allow you to only load the examples you need ;)", "thanks a lot Quentin, this would be really really a great feature to have\n\nOn Wed, Apr 7, 2021 at 12:14 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> We'll work on dataset streaming soon. This should allow you to only load\n> the examples you need ;)\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2161#issuecomment-814791922>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMROD62QAKIJMAKWISTTHQWBVANCNFSM42IUI5JQ>\n> .\n>\n", "Is streaming completed? On the 1.8.0 docs it is mentioned (https://huggingface.co/docs/datasets/dataset_streaming.html), but when following the example I get the following error:\r\n\r\n```\r\n>>> dataset2 = load_dataset(\"amazon_us_reviews\", \"Pet_Products_v1_00\", split='train', streaming=True)\r\n\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n<ipython-input-21-1eedab26cff1> in <module>()\r\n----> 1 en_dataset = load_dataset('oscar', \"unshuffled_deduplicated_en\", split='train', streaming=True)\r\n\r\n3 frames\r\n/usr/local/lib/python3.7/dist-packages/datasets/builder.py in _create_builder_config(self, name, custom_features, **config_kwargs)\r\n 339 if value is not None:\r\n 340 if not hasattr(builder_config, key):\r\n--> 341 raise ValueError(f\"BuilderConfig {builder_config} doesn't have a '{key}' key.\")\r\n 342 setattr(builder_config, key, value)\r\n 343 \r\n\r\nValueError: BuilderConfig OscarConfig(name='unshuffled_deduplicated_en', version=1.0.0, data_dir=None, data_files=None, description='Unshuffled and deduplicated, English OSCAR dataset') doesn't have a 'streaming' key.\r\n```\r\n\r\nUPDATE: Managed to get streaming working by building from source and installing the additional `datasets[streaming]` package:\r\n\r\n```\r\n!pip install git+https://github.com/huggingface/datasets.git\r\n!pip install datasets[streaming]\r\n```", "Hi ! Streaming is available on `master` only right now. We'll make a new release 1.9.0 on Monday :)" ]

[]

[]

[ "Hi.\r\nI cannot always reproduce this issue, and on later runs I did not see it so far. Sometimes also I set 8 processes but I see less being showed, is this normal, here only 5 are shown for 8 being set, thanks\r\n\r\n```\r\n#3: 11%|███████████████▊ | 172/1583 [00:46<06:21, 3.70ba/s]\r\n#4: 9%|█████████████▏ | 143/1583 [00:46<07:46, 3.09ba/s]\r\n#7: 6%|█████████ | 98/1583 [00:45<11:34, 2.14ba/s]\r\n#5: 8%|███████████▍ | 124/1583 [00:46<09:03, 2.68ba/s]\r\n#6: 7%|██████████▏ \r\n```", "closing since I cannot reproduce it again, thanks " ]

[ { "color": "e99695", "default": false, "description": "Requesting to add a new dataset", "id": 2067376369, "name": "dataset request", "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request" } ]

[]

[ "closing since I think this is cc100, just the name has been changed. thanks " ]

[ { "color": "94203D", "default": false, "description": "", "id": 2107841032, "name": "nlp-viewer", "node_id": "MDU6TGFiZWwyMTA3ODQxMDMy", "url": "https://api.github.com/repos/huggingface/datasets/labels/nlp-viewer" } ]

[]

[ "Thanks for reporting !\r\nThe viewer doesn't have all the dependencies of the datasets. We may add openpyxl to be able to show this dataset properly", "This viewer tool is deprecated now and the new viewer at https://huggingface.co/datasets/fake_news_english works fine, so I'm closing this issue" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Just note that docstrings injected from PyArrow do not follow the same convention for formatting types in `Args` or `Returns` as we do... Not a big problem, anyway! 😄 " ]

[]

[]

[ "Awesome!" ]

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

[ { "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "gravatar_id": "", "html_url": "https://github.com/lhoestq", "id": 42851186, "login": "lhoestq", "node_id": "MDQ6VXNlcjQyODUxMTg2", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "repos_url": "https://api.github.com/users/lhoestq/repos", "site_admin": false, "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "type": "User", "url": "https://api.github.com/users/lhoestq" } ]

[ "Hi ! Thanks for reporting. I opened a PR to fix this issue: #2201", "Nice question which helped me a lot! I have wasted a lot of time to the `DatasetDict` creation from a csv file. Hope the document of this module add some simple examples.", "Hi :) We're indeed working on tutorials that we will add to the docs !" ]

[]

[]

[]

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

[]

[ "@lhoestq I am going to implement the consolidation step you mentioned in #1870.", "@lhoestq I was thinking that the order of the TableBlocks is not relevant, isn't it?\r\n\r\nI mean, in order to consolidate _consecutive_ in-memory table blocks, in this case:\r\n```\r\nblocks = [in_memory_1, memory_mapped, in_memory_2]\r\n```\r\nI could reorder the list:\r\n```\r\nblocks = [in_memory_1, in_memory_2, memory_mapped]\r\n```\r\nso that the first 2 can be consolidated into a single one:\r\n```\r\nblocks = [in_memory_3, memory_mapped]\r\n```", "I think the order is important, users won't expect the dataset to be \"shuffled\" when they add a new item", "> I think the order is important, users won't expect the dataset to be \"shuffled\" when they add a new item\r\n\r\nOK, therefore I leave `_consolidate_blocks` as it is, which currently keeps the order of the blocks (no shuffling).", "Thank you guys for implementing this. Minor thing I noticed in the [documentation](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.concatenate_datasets): it says \"Converts a list of Dataset with **the same schema** into a single Dataset\". With the addition of the axis parameter, perhaps this should be reworded, no?" ]

[]

[]

[]

[]

[]

[ "Good catch ! Thanks for reporting\r\n\r\nI just opened #2180 to fix this", "Fixed in #2180" ]

[]

[]

[ "Hi @marrodion. \r\n\r\nThanks for pointing this out. It would be great to incorporate this metric-specific enhancement.\r\n\r\nAnother possibility would be to require the user to input the scheme as a string `mode=\"strict\", scheme=\"IOB2\"` and then dynamically import the corresponding module using Python `importlib`:\r\n```python\r\nif scheme:\r\n scheme = importlib.import_module(f\"seqeval.scheme.{scheme}\")\r\n```\r\n\r\nFeel free to create a Pull Request to make this contribution." ]

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

[]

[]

[]

[]

[ "Hi ! In the arrow file we store all the integers as uint8.\r\nSo your arrow file should weigh around `height x width x n_channels x n_images` bytes.\r\n\r\nWhat feature type do your TFDS dataset have ?\r\n\r\nIf it uses a `tfds.features.Image` type, then what is stored is the encoded data (as png or jpg for example). Since these encodings are made for compression, the resulting tfrecord is smaller that the arrow file.\r\n\r\nWe are working on adding a similar feature in `datasets`: the ability to store the encoded data instead of the raw integers for images, but also for audio data. This way, arrow files will have similar sizes as tfrecords for images.", "Thanks for the prompt response. You're right about the encoding, I have the `tfds.features.Image` feature type you mentioned.\r\nHowever, as described in the `dataset_info.json`, my dataset is made of 1479 (224x224x3) images. 1479 x 224 x 224 x 3 = 222630912 bytes which is far from the actual size 520803408 bytes. \r\n\r\nAnyway I look forward to the Image feature type in `datasets`. ", "@lhoestq I changed the data structure so I have a 2D Array feature type instead of a 3D Array by grouping the two last dimensions ( a 224x672 2D Array instead of a 224x224x3 3D Array). The file size is now 223973964 bytes, nearly half the previous size! Which is around of what I would expect.\r\nI found similar behavior in existing `datasets` collection, when comparing black and white vs color image, for example MNIST vs CIFAR. ", "Interesting !\r\nThis may be because of the offsets that are stored with the array data.\r\n\r\nCurrently the offsets are stored even if the `shape` of the arrays is fixed. This was needed because of some issues with pyarrow a few months ago. I think these issues have been addressed now, so we can probably try to remove them to make the file lighter.\r\n\r\nIdeally in your case the floats data should be 220 MB for both Array2D and Array3D", "Yeah for sure, can you be a bit more specific about where the offset is stored in the code base ? And any reference to pyarrow issues if you have some. I would be very interested in contributing to `datasets` by trying to fix this issue. ", "Pyarrow has two types of lists: variable length lists and fixed size lists.\r\nCurrently we store the ArrayXD data as variable length lists. They take more disk space because they must store both actual data and offsets.\r\nIn the `datasets` code this is done here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/dbac87c8a083f806467f5afc4ec9b401a7e4c15c/src/datasets/features.py#L346-L352\r\n\r\nTo use a fixed length list, one should use the `list_size` argument of `pyarrow.list_()`.\r\nI believe this would work directly modulo some changes in the numpy conversion here:\r\n\r\nhttps://github.com/huggingface/nlp/blob/dbac87c8a083f806467f5afc4ec9b401a7e4c15c/src/datasets/features.py#L381-L395" ]

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

[]

[ "#2274 has been merged. You can now merge master into this branch and use `assert_arrow_metadata_are_synced_with_dataset_features(dset)` to make sure that the metadata are good :)" ]

[]

[]

[ "That's how I loaded the dataset\r\n```python\r\nfrom datasets import load_dataset\r\nds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache')\r\n```", "Hi ! It looks like the arrow file in the folder\r\n`/usr/local/workspace/NAS_NLP/cache/wikipedia/20200501.en/1.0.0/50aa706aa417bb77d910ad61211cc672c0ef3e0f224225a5e0a18277ade8b931` is corrupted.\r\n\r\nCan you take a look and check that it's 18.3GB ?\r\n\r\nIf not, then maybe you need to redownload it:\r\n```python\r\nfrom datasets import load_dataset\r\nds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache', download_mode=\"force_redownload\")\r\n```", "> Hi ! It looks like the arrow file in the folder\r\n> `/usr/local/workspace/NAS_NLP/cache/wikipedia/20200501.en/1.0.0/50aa706aa417bb77d910ad61211cc672c0ef3e0f224225a5e0a18277ade8b931` is corrupted.\r\n> \r\n> Can you take a look and check that it's 18.3GB ?\r\n> \r\n> If not, then maybe you need to redownload it:\r\n> \r\n> ```python\r\n> from datasets import load_dataset\r\n> ds = load_dataset('wikipedia', '20200501.en', cache_dir='/usr/local/workspace/NAS_NLP/cache', download_mode=\"force_redownload\")\r\n> ```\r\n\r\nHi Ihoestq, thanks for the reply! Actually i think my issue is i couldn't download the dataset beyond 10.7G. It feels like the whole dataset is split into different volumes and after the first one was downloaded it crashed before proceeding to the next one. I did try 'force_redownload' mode but still got the same issue.", "I just tried on my side and got no issues.\r\nWhen downloading the dataset again, did it crash at 10.7GB as well ?", "> I just tried on my side and got no issues.\r\n> When downloading the dataset again, did it crash at 10.7GB as well ?\r\n\r\nYes i have tried it multiple times on different machines. I am wondering if you could share the screenshot of your dependency versions and i will try to make them the same as yours?", "I tried using `datasets` from `master` on macos with python 3.7.2\r\nI also have `requests==2.23.0` and `tqdm==4.45.0`." ]

[]

[ { "avatar_url": "https://avatars.githubusercontent.com/u/17948980?v=4", "events_url": "https://api.github.com/users/theo-m/events{/privacy}", "followers_url": "https://api.github.com/users/theo-m/followers", "following_url": "https://api.github.com/users/theo-m/following{/other_user}", "gists_url": "https://api.github.com/users/theo-m/gists{/gist_id}", "gravatar_id": "", "html_url": "https://github.com/theo-m", "id": 17948980, "login": "theo-m", "node_id": "MDQ6VXNlcjE3OTQ4OTgw", "organizations_url": "https://api.github.com/users/theo-m/orgs", "received_events_url": "https://api.github.com/users/theo-m/received_events", "repos_url": "https://api.github.com/users/theo-m/repos", "site_admin": false, "starred_url": "https://api.github.com/users/theo-m/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/theo-m/subscriptions", "type": "User", "url": "https://api.github.com/users/theo-m" }, { "avatar_url": "https://avatars.githubusercontent.com/u/33657802?v=4", "events_url": "https://api.github.com/users/SBrandeis/events{/privacy}", "followers_url": "https://api.github.com/users/SBrandeis/followers", "following_url": "https://api.github.com/users/SBrandeis/following{/other_user}", "gists_url": "https://api.github.com/users/SBrandeis/gists{/gist_id}", "gravatar_id": "", "html_url": "https://github.com/SBrandeis", "id": 33657802, "login": "SBrandeis", "node_id": "MDQ6VXNlcjMzNjU3ODAy", "organizations_url": "https://api.github.com/users/SBrandeis/orgs", "received_events_url": "https://api.github.com/users/SBrandeis/received_events", "repos_url": "https://api.github.com/users/SBrandeis/repos", "site_admin": false, "starred_url": "https://api.github.com/users/SBrandeis/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/SBrandeis/subscriptions", "type": "User", "url": "https://api.github.com/users/SBrandeis" } ]

[]

[]

[]

[]

[]

[]

[ "Hi @lhoestq \r\nThanks a lot for taking time checking it. I update \"dataset_infos.json\", I added description to the function of _generate_samples in wikiann.py but I was not sure about the format to write in README. thanks. ", "Thanks !\r\n\r\nFor the fields description in the dataset card, something like this does the job:\r\n```\r\n- `tokens`: a `list` of `string` features.\r\n- `langs`: a `list` of `string` features that correspond to the language of each token.\r\n- `ner_tags`: a `list` of classification labels, with possible values including `O` (0), `B-PER` (1), `I-PER` (2), `B-ORG` (3), `I-ORG` (4), `B-LOC` (5), `I-LOC` (6).\r\n- `spans`: a `list` of `string` features, that is the list of named entities in the input text formatted as ``<TAG>: <mention>``\r\n```\r\n\r\nAlso for information, I think the trailer of rick and morty season 5 is out now :)", "Hi @lhoestq \r\nthank you! This is updated now, please feel free to let me know if I need to modify something :) thanks " ]

[]

[]

[ "@lhoestq I updated files" ]

[]

[]

[ "Hi !\r\nI think this has been fixed recently on `master`.\r\nCan you try again by installing `datasets` from `master` ?\r\n```\r\npip install git+https://github.com/huggingface/datasets.git\r\n```", "Hi!\r\n\r\nUsing that version of the code solves the issue. Thanks!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi ! Indeed only the languages of the `translate-train` data are included...\r\nI can't find a link to download the english train set on https://github.com/facebookresearch/MLQA though, do you know where we can download it ?", "Hi @lhoestq \r\nthank you very much for coming back to me, now I see, you are right, in the link you sent I see split of {split}-context-{context_language}-question-{question_language}.json with context_language=question_language=en, TFDS most probably has extracted english ones from these files as en language files, but translate-train/test do not have en indeed. thanks a lot for the great explanations", "I close the ticket, since I do not see any en existing, they have trained on \"SQuAD V1.1\" instead. Thanks. " ]

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

[ { "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "gravatar_id": "", "html_url": "https://github.com/lhoestq", "id": 42851186, "login": "lhoestq", "node_id": "MDQ6VXNlcjQyODUxMTg2", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "repos_url": "https://api.github.com/users/lhoestq/repos", "site_admin": false, "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "type": "User", "url": "https://api.github.com/users/lhoestq" } ]

[ "Hi !\r\nIndeed `save_to_disk` doesn't call pickle anymore. Though the `OverflowError` can still appear for in-memory datasets bigger than 4GB. This happens when doing this for example:\r\n```python\r\nimport pyarrow as pa\r\nimport pickle\r\n\r\narr = pa.array([0] * ((4 * 8 << 30) // 64))\r\ntable = pa.Table.from_arrays([a], names=[\"foo\"])\r\npickle.dumps(table) # fails with an OverflowError\r\npickle.dumps(table, 4) # works !\r\n```\r\nWe'll do the change to use `protocol=4`.\r\n\r\nMoreover I've also seen other users complain about this error\r\n```\r\nstruct.error: 'I' format requires 0 <= number <= 4294967295\r\n```\r\n\r\nIt looks like something related to the 4GB limit as well but I'm not able to reproduce on my side.\r\nDo you think you can provide a script that reproduces the issue ?\r\nHow big is your dataset ? (number of bytes, number of rows)\r\n\r\n", "Hi!\r\nSo I've managed to created a minimum working (well technically crashing) example for the multiprocessing case, I create a huge list of zeros, like in your example, and then I try to .map(None, num_proc=2) over it, which then crashes, here's the code:\r\n\r\n```python\r\nfrom datasets import Dataset\r\n\r\nif __name__ == '__main__':\r\n ton_of_zeroes = [0] * ((12 * 8 << 30) // 64)\r\n large_dataset = Dataset.from_dict({'col': ton_of_zeroes})\r\n print(\"Start\")\r\n large_dataset.map(function=None, num_proc=2)\r\n print(\"Done - should not print\")\r\n```\r\n\r\nThe amount of zeros could probably be reduced, I haven't tried to minimize it to find the breaking point, I just increased it from your code (which by quick glance I assumed tried to allocate over 4 GiB)\r\n\r\nRunning this results in the following traceback:\r\n\r\n```\r\nParameter 'indices'=[ 0 1 2 ... 805306365 805306366 805306367] of the transform datasets.arrow_dataset.Dataset.select couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.\r\nTraceback (most recent call last):\r\n File \"./crash_multiproc_pickle.py\", line 7, in <module>\r\n large_dataset.map(function=None, num_proc=2)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1485, in map\r\n transformed_shards = [r.get() for r in results]\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1485, in <listcomp>\r\n transformed_shards = [r.get() for r in results]\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/pool.py\", line 657, in get\r\n raise self._value\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/pool.py\", line 431, in _handle_tasks\r\n put(task)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/connection.py\", line 209, in send\r\n self._send_bytes(_ForkingPickler.dumps(obj))\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/multiprocess/reduction.py\", line 54, in dumps\r\n cls(buf, protocol, *args, **kwds).dump(obj)\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 454, in dump\r\n StockPickler.dump(self, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 437, in dump\r\n self.save(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 789, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 941, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 859, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 885, in _batch_setitems\r\n save(v)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 662, in save_reduce\r\n save(state)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/home/cernypro/dev/envs/huggingface_gpu/lib/python3.7/site-packages/dill/_dill.py\", line 941, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 859, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 885, in _batch_setitems\r\n save(v)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 846, in _batch_appends\r\n save(tmp[0])\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 846, in _batch_appends\r\n save(tmp[0])\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 789, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 819, in save_list\r\n self._batch_appends(obj)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 843, in _batch_appends\r\n save(x)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 549, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 638, in save_reduce\r\n save(args)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 774, in save_tuple\r\n save(element)\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 504, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/mnt/appl/software/Python/3.7.4-GCCcore-8.3.0/lib/python3.7/pickle.py\", line 732, in save_bytes\r\n self._write_large_bytes(BINBYTES + pack(\"<I\", n), obj)\r\nstruct.error: 'I' format requires 0 <= number <= 4294967295\r\n```\r\n\r\nMy datasets usually have hundreds of thousands to low millions of rows, with each row containing a list of 10 strings and list of vectors of different length (the strings tokenized), which in the worst case have 10\\*512\\*8 = 40960 bytes (but usually it is much smaller, as the vectors tend to be shorter. I need these groups of text lines to create training data for the Inverse Cloze Task.\r\n\r\nAnyway I don't think my particular dataset is relevant, as the tiny script I created also manages to crash.\r\nBut I think the issue is the same as the save_to_disk, from the traceback it seems that in multiprocessing, it tries to use dill to return the result of the map workers, which tries to pickle the data and can't do it, probably because it's again using the older pickle protocol. That's my guess anyway.", "I just merged a fix #2150 that allows to pickle tables bigger than 4GiB\r\nFeel free to try it on the `master` branch !", "awesome! I started getting this error as well when I tried to tokenize with a longer sequence length", "@prokopCerny does this fix work for you? I found that with the latest master, my container with 500GB RAM starts crashing when I try to map a large dataset using `num_proc`.\r\n\r\n@lhoestq would it be possible to implement some logic to keep the individual cache files small (say below 100mb)? I find this helps with loading large datasets, but the \"hack\" I was using (increasing `num_proc` to a large number) doesn't work anymore with the latest master; my container crashes even with `num_proc=200` now", "Closing since the original issue was fixed in #2150 \r\nFeel free to reopen if you are still experiencing it.\r\nFor the other problems, please open separate issues" ]

[]

[]

[ "If you print those questions, you get readable texts:\r\n```python\r\n>>> questions = [\r\n... \"\\u0645\\u062a\\u0649 \\u0628\\u062f\\u0627\\u062a \\u0627\\u0644\\u0645\\u062c\\u0644\\u0629 \\u0627\\u0644\\u0645\\u062f\\u0631\\u0633\\u064a\\u0629 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645 \\u0628\\u0627\\u0644\\u0646\\u0634\\u0631?\",\r\n... \"\\u0643\\u0645 \\u0645\\u0631\\u0629 \\u064a\\u062a\\u0645 \\u0646\\u0634\\u0631\\u0647\\u0627 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0645\\u0627 \\u0647\\u064a \\u0627\\u0644\\u0648\\u0631\\u0642\\u0629 \\u0627\\u0644\\u064a\\u0648\\u0645\\u064a\\u0629 \\u0644\\u0644\\u0637\\u0644\\u0627\\u0628 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0643\\u0645 \\u0639\\u062f\\u062f \\u0627\\u0644\\u0627\\u0648\\u0631\\u0627\\u0642 \\u0627\\u0644\\u0627\\u062e\\u0628\\u0627\\u0631\\u064a\\u0629 \\u0644\\u0644\\u0637\\u0644\\u0627\\u0628 \\u0627\\u0644\\u062a\\u064a \\u0648\\u062c\\u062f\\u062a \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\",\r\n... \"\\u0641\\u064a \\u0627\\u064a \\u0633\\u0646\\u0629 \\u0628\\u062f\\u0627\\u062a \\u0648\\u0631\\u0642\\u0629 \\u0627\\u0644\\u0637\\u0627\\u0644\\u0628 \\u0627\\u0644\\u062d\\u0633 \\u0627\\u0644\\u0633\\u0644\\u064a\\u0645 \\u0628\\u0627\\u0644\\u0646\\u0634\\u0631 \\u0641\\u064a \\u0646\\u0648\\u062a\\u0631\\u062f\\u0627\\u0645?\"\r\n... ]\r\n>>> print(questions)\r\n['متى بدات المجلة المدرسية في نوتردام بالنشر?', 'كم مرة يتم نشرها في نوتردام?', 'ما هي الورقة اليومية للطلاب في نوتردام?', 'كم عدد الاوراق الاخبارية للطلاب التي وجدت في نوتردام?', 'في اي سنة بدات ورقة الطالب الحس السليم بالنشر في نوتردام?']\r\n```\r\nI don't think we can change this", "Hi @dorost1234.\r\n\r\nIn Python 3, strings are sequences of Unicode _code points_. Unicode is a specification that maps all characters (and emoji symbols) with its unique representation in terms of code points. That is what you see: Unicode code points (represented by a \\u escaped sequence of 16-bit hex values).\r\n\r\nCharacters are usually represented (on screen and papers) with a graphical element called _glyph_. That is what you would like to see: glyphs. But Python does not care about glyphs: that is the job of the GUI or the terminal; glyphs are what you get with the `print` function (if your terminal is properly configured to display those glyphs).\r\n\r\nYou have more detailed information about Unicode in the Python documentation: https://docs.python.org/3/howto/unicode.html", "thank you so much for the insightful comments. " ]

[]

[]

[ "You can filter the languages this way:\r\n```python\r\ntydiqa_en = tydiqa_dataset.filter(lambda x: x[\"language\"] == \"english\")\r\n```\r\n\r\nOtherwise maybe we can have one configuration per language ?\r\nWhat do you think of this for example ?\r\n\r\n```python\r\nload_dataset(\"tydiqa\", \"primary_task.en\")\r\n```", "Hi\nthank you very much for the great response, this will be really wonderful\nto have one configuration per language, as one need the dataset in majority\nof case per language for cross-lingual evaluations.\nThis becomes also then more close to TFDS format, which is separated per\nlanguage https://www.tensorflow.org/datasets/catalog/tydi_qa which will be\nreally awesome to have.\nthanks\n\nOn Mon, Mar 29, 2021 at 6:17 PM Quentin Lhoest ***@***.***>\nwrote:\n\n> You can filter the languages this way:\n>\n> tydiqa_en = tydiqa_dataset.filter(lambda x: x[\"language\"] == \"english\")\n>\n> Otherwise maybe we can have one configuration per language ?\n> What do you think of this for example ?\n>\n> load_dataset(\"tydiqa\", \"primary_task.en\")\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/issues/2132#issuecomment-809516799>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMXPW2PWSQ2RHG73O7TTGCY4LANCNFSM4Z7ER7IA>\n> .\n>\n", "@lhoestq I greatly appreciate any updates on this. thanks a lot" ]

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