Version 0.22.0

Major Features and Improvements

Crop values in natural language stats generator.
Switch to using PyBind11 instead of SWIG for wrapping C++ libraries.
CSV decoder support for multivalent columns by using tfx_bsl's decoder.
When inferring a schema entry for a feature, do not add a shape with dim = 0 when min_num_values = 0.
Add utility methods tfdv.get_slice_stats to get statistics for a slice and tfdv.compare_slices to compare statistics of two slices using Facets.
Make tfdv.load_stats_text and tfdv.write_stats_text public.
Add PTransforms tfdv.WriteStatisticsToText and tfdv.WriteStatisticsToTFRecord to write statistics proto to text and tfrecord files respectively.
Modify tfdv.load_statistics to handle reading statistics from TFRecord and text files.
Added an extra requirement group mutual-information. As a result, barebone TFDV does not require scikit-learn any more.
Added an extra requirement group visualization. As a result, barebone TFDV does not require ipython any more.
Added an extra requirement group all that specifies all the extra dependencies TFDV needs. Use pip install tensorflow-data-validation[all] to pull in those dependencies.
Depends on pyarrow>=0.16,<0.17.
Depends on apache-beam[gcp]>=2.20,<3.
Depends on `ipython>=7,<8;python_version>="3"'.
Depends on `scikit-learn>=0.18,<0.24'.
Depends on tensorflow>=1.15,!=2.0.*,<3.
Depends on tensorflow-metadata>=0.22.0,<0.23.
Depends on tensorflow-transform>=0.22,<0.23.
Depends on tfx-bsl>=0.22,<0.23.

(Known issue resolution) It is no longer necessary to use Apache Beam 2.17 when running TFDV on Windows. The current release of Apache Beam will work.

tfdv.GenerateStatistics now accepts a PCollection of pa.RecordBatch instead of pa.Table.
All the TFDV coders now output a PCollection of pa.RecordBatch instead of a PCollection of pa.Table.
tfdv.validate_instances and tfdv.api.validation_api.IdentifyAnomalousExamples now takes pa.RecordBatch as input instead of pa.Table.
The StatsGenerator interface (and all its sub-classes) now takes pa.RecordBatch as the input data instead of pa.Table.
Custom slicing functions now accepts a pa.RecordBatch instead of pa.Table as input and should output a tuple (slice_key, record_batch).