Remove redundant transforms.

Remove PreNormTransformerDecoderLayer, use builtin module.

Some cosmetic updates to drop Python 3.6 support.

Update model zoo URLs on webpage, bump versions and deps. (#31)

Update gitignore to fix #30

Bump opencv-python from 4.1.2.30 to 4.2.0.32 (#24)

Upgrade to v1.3

Cache Lookahead params after loading checkpoint.

Change docstrings from numpy format to Google format.

Remove LmdbReader and data-efficiency experiment support.

Update checkpoint URLs to dropbox (weights unchanged).

Merge branch 'decoding' to 'master'

Update version and add CHANGELOG.

Add nucleus sampling and CaptionDecoderFactory.

Fix beam search bug, thanks @ashkamath and @alcinos