Release 0.21.0

Release 0.21.0

Major Features and Improvements

• Started depending on the CSV parsing / type inferring utilities provided
  by tfx-bsl (since tfx-bsl 0.15.2). This also brings performance improvements
  to the CSV decoder (~2x faster in decoding. Type inferring performance is not
  affected).
• Compute bytes statistics for features of BYTES type. Avoid computing topk and
  uniques for such features.
• Added LiftStatsGenerator which computes lift between one feature (typically a
  label) and all other categorical features.

Bug Fixes and Other Changes

• Exclude examples in which the entire sparse feature is missing when
  calculating sparse feature statistics.
• Validate min_examples_count dataset constraint.
• Document the schema fields, statistics fields, and detection condition for
  each anomaly type that TFDV detects.
• Handle null array in cross feature stats generator, top-k & uniques combiner
  stats generator, and sklearn mutual information generator.
• Handle infinity in basic stats generator.
• Set num_missing and num_examples correctly in the presence of sparse
  features.
• Compute weighted feature stats for all weighted features declared in schema.
• Depends on tensorflow-metadata>=0.21.0,<0.22.
• Depends on pyarrow>=0.15 (removed the upper bound as it is determined by
  tfx-bsl).
• Depends on tfx-bsl>=0.21.0,<0.22
• Depends on apache-beam>=2.17,<3

Breaking Changes

• Changed the behavior regarding to statistics over CSV data:

  • Previously, if a CSV column was mixed with integers and empty strings, FLOAT
    statistics will be collected for that column. A change was made so INT
    statistics would be collected instead.

• Removed csv_decoder.DecodeCSVToDict as Dict[str, np.ndarray] had no longer
  been the internal data representation any more since 0.14.

Deprecations