TensorFlow 2.2

Improvements

• Bumped to TensorFlow 2.2.0
• Decoder refactored to be used with default Keras fit predict pipeline
• Example updated to show usage with Keras pipeline
• Per-epoch training time on toy example decreased to 4s from 11s

Limitation

• Tested with tf.data.Dataset only
• Need to pass drop_remainder=True to batch() method of tf.data.Dataset. Cannot handle case of unknown batch_size

Improvisation to Beam Search

• Beam search is improvised to give more variance in predictions
• Bug with copy token fixed

Fixes

• Decoder correctly masks special tokens while copying now
• BeamSearch stops adding to probability after end token

Fix bug

bug: fix not copying special tokens

Add debugging ability

add reporting attentive and selective weights while calculating loss

Add features

• add option to specify length in transform

First release

• Basic workflow demonstrated in example
• Metrics: BLEU
• Encoder, Decoder generation output layer are customizable