datasets 2.13.0

on Python PyPI

Dataset Features

• Add IterableDataset.from_spark by @maddiedawson in #5770

  • Stream the data from your Spark DataFrame directly to your training pipeline

  from datasets import IterableDataset
  from torch.utils.data import DataLoader
  
  ids = IterableDataset.from_spark(df)
  ids = ids.map(...).filter(...).with_format("torch")
  for batch in DataLoader(ids, batch_size=16, num_workers=4):
      ...

• IterableDataset formatting for PyTorch, TensorFlow, Jax, NumPy and Arrow:

  • IterableDataset Arrow formatting by @lhoestq in #5821
  • Iterable torch formatting by @lhoestq in #5852

  from datasets import load_dataset
  
  ids = load_dataset("c4", "en", split="train", streaming=True)
  ids = ids.map(...).with_format("torch")  # to get PyTorch tensors - also works with tf, np, jax etc.

• Add IterableDataset.from_file to load local dataset as iterable by @mariusz-jachimowicz-83 in #5893

  from datasets import IterableDataset
  
  ids = IterableDataset.from_file("path/to/data.arrow")

• Arrow dataset builder to be able to load and stream Arrow datasets by @mariusz-jachimowicz-83 in #5944

  from datasets import load_dataset
  
  ds = load_dataset("arrow", data_files={"train": "train.arrow", "test": "test.arrow"})

Experimental

• Add parallel module using joblib for Spark by @es94129 in #5924

General improvements and bug fixes

• Preserve stopping_strategy of shuffled interleaved dataset (random cycling case) by @mariosasko in #5816
• Fix incomplete docstring for BuilderConfig by @Laurent2916 in #5824
• [docs] Custom decoding transforms by @stevhliu in #5836
• Add accelerate as metric's test dependency to fix CI error by @mariosasko in #5848
• Add date_format param to the CSV reader by @mariosasko in #5845
• [docs] Redirects, migrated from nginx by @julien-c in #5853
• Fix infer module for uppercase extensions by @albertvillanova in #5872
• Minor tqdm optim by @lhoestq in #5860
• Always set nullable fields in the writer by @lhoestq in #5835
• Add fn_kwargs to map and filter of IterableDataset and IterableDatasetDict by @yuukicammy in #5810
• Better error message when combining dataset dicts instead of datasets by @lhoestq in #5861
• Force overwrite existing filesystem protocol by @baskrahmer in #5894
• Support working_dir in from_spark by @maddiedawson in #5826
• Raise TypeError when indexing a dataset with bool by @albertvillanova in #5859
• Fix minor typo in docs loading.mdx by @albertvillanova in #5900
• Fix FixedSizeListArray casting by @mariosasko in #5897
• Unpin responses by @mariosasko in #5916
• Validate name parameter in make_file_instructions by @albertvillanova in #5904
• Raise error in DatasetBuilder.as_dataset when file_format is not "arrow" by @mariosasko in #5915
• Refactor extensions by @albertvillanova in #5917
• Use more efficient and idiomatic way to construct list. by @ttsugriy in #5909
• Add flatten_indices to DatasetDict by @maximxlss in #5907
• Optimize IterableDataset.from_file using ArrowExamplesIterable by @lhoestq in #5920
• Make prepare_split more robust if errors in metadata dataset_info splits by @albertvillanova in #5901
• Fix streaming parquet with image feature in schema by @lhoestq in #5921
• canonicalize data dir in config ID hash by @kylrth in #5899
• Fix link to quickstart docs in README.md by @mariosasko in #5928
• Fix string-encoding, make batch_size optional, and minor improvements in Dataset.to_tf_dataset by @alvarobartt in #5883
• Use a new low-memory approach for tf dataset index shuffling by @Rocketknight1 in #5863
• [doc build] Use secrets by @mishig25 in #5932
• Fix to_numpy when None values in the sequence by @qgallouedec in #5933
• Better row group size in push_to_hub by @lhoestq in #5935
• Avoid parallel redownload in cache by @albertvillanova in #5937
• Better filenotfound for gated by @lhoestq in #5954
• Make get_from_cache use custom temp filename that is locked by @albertvillanova in #5938
• Fix ArrowExamplesIterable.shard_data_sources by @lhoestq in #5956
• Add Arrow builder docs by @lhoestq in #5952
• Fix sequence of array support for most dtype by @qgallouedec in #5948

New Contributors

• @Laurent2916 made their first contribution in #5824
• @yuukicammy made their first contribution in #5810
• @baskrahmer made their first contribution in #5894
• @ttsugriy made their first contribution in #5909
• @maximxlss made their first contribution in #5907
• @mariusz-jachimowicz-83 made their first contribution in #5893
• @kylrth made their first contribution in #5899
• @qgallouedec made their first contribution in #5933
• @es94129 made their first contribution in #5924

Full Changelog: 2.12.0...zef