v1.2.3

Reformer

• Reversible Transformer model for machine translation and other encoder-decoder tasks
• Add code for beam search, sampling, and greedy decoding (see trax.models.beam_search.Search)
• Memory-efficient attention classes have been re-written to use even less memory and to support faster decoding (see the new SelfAttention, LSHSelfAttention and EncDecAttention classes)

RL

• Implemented the Advantage-Weighted Regression algorithm, a simple off-policy reinforcement learning algorithm.
• Extracted out a PolicyBasedTrainer, so ppo_trainer.PPO and awr_trainer.AwrTrainer now both inherit from it.
• Refactoring of the serialization code in the RL part, thanks to @koz4k !

Framework

• A lot of code cleanup and refactoring of the core abstractions by Jonni, thanks Jonni!

TF Numpy

• More ops added by @wangpengmit !