This repository was archived by the owner on Jul 7, 2023. It is now read-only.

v1.3.1#451

Merged

lukaszkaiser merged 11 commits intotensorflow:masterfrom

Dec 1, 2017

Commits on Dec 1, 2017

Discrete autoencoder with VQ-VAE as in https://arxiv.org/abs/1711.00937.

T2T Team
authored and
Ryan Sepassi
committed
Packed datasets - combine examples to constant length for efficient TPU training. Modify transformer to keep the packed-together examples from attending to one another.

nshazeer
authored and
Ryan Sepassi
committed
Make Parallelism object use reuse=True by default. Solves tpu checkpoint compatibility bug.

nshazeer
authored and
Ryan Sepassi
committed
Remove TranslateEndeWmtBpe32kPacked. We have TranslateEndeWmt32kPacked.

nshazeer
authored and
Ryan Sepassi
committed
Clean up transformer_vae and add refining.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
T2T depends on TF 1.4+, daisy_chain_getter bug fix, some Eager-mode improvements/fixes
Ryan Sepassi
committed
v1.3.1
Ryan Sepassi
committed
New BLEU cleanup and small correction to VAE.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Enable Transformer fast decoding in eager mode

Niki Parmar
authored and
Ryan Sepassi
committed
Fix decoding and training issues in external colab.

Niki Parmar
authored and
Ryan Sepassi
committed
TF Eager improvements for T2TModel
Ryan Sepassi
committed