v1.3.2

• Small improvements for attention vizualization in colab.

v1.3.1

• Improvements for TF Eager compatibility

v1.3.0

• WARNING: Checkpoints produced with old versions break with this new release due to new variable scoping
• Various changes make T2T models and problems compatible with the new TF Eager mode - we'll have more on that soon
• tpu_trainer becoming more fully featured
• Internal refactoring moving towards more flexibility in specifying the Estimator input_fn and model_fn

v1.2.9

• Quick bug fixes

v1.2.8

• Batch norm should now work in T2T - fixed the custom variable getters
• Simplified ImageModality and removal of SmallImageModality
• Simplified ClassLabelModality and removal of ClassLabel1DModality
• New modality with CTC loss
• New vanilla_gan model that's a good example of a simple GAN
• TPU advances: Xception, Resnet50, and Transformer verified to work, code path uses Experiment, usage doc for Cloud TPU alpha customers
• Various small fixes, improvements, features

v1.2.7: Merge pull request #396 from rsepassi/push

• Fixed data generators for translation tasks. Great thanks to @vince62s and @martinpopel for your PRs and reviews and all the help!
• Updated LSTM models and attention. Great thanks @kolloldas for the attention work and @epurdyf for pointing out initializer problems!
• Added some variations of the transformer model.
• Bug-fixes and cleanups.

v1.2.6

• Refactored Translate problems courtesy of @vince62s
• Fast beam search decoding for the Transformer model (set --decode_hparams='use_last_position=True' with the t2t-decoder to use it)
• Various improvements and bug fixes

v1.2.5

• Various bug fixes, improvements, and additions
• Checkpoint Breaking Note: We'd like to have good defaults as well as have immutable hparams sets and so we have an updating naming scheme we're trying, starting with Transformer hparams. transformer_base is now an alias name that points to a versioned hparams set, i.e. transformer_base now calls transformer_base_v2. The previous version of transformer_base is now transformer_base_v1 (so if you have an old checkpoint, use hparams set transfomer_base_v1). This way if you're just trying something out, you can use transformer_base and know that you have a set of up-to-date good defaults. If you want to maintain reproducibility across T2T versions, use one of the versioned names, e.g. transformer_base_v2.

v1.2.4

• Various cleanups, fixes, and feature additions - see commit history
• More robust Travis CI tests

v1.2.3

• Transformer now supports fast decoding! The decoding path used to recompute the entire sequence on each additional timestep, but now caches as it goes.
• We now support SavedModel exports
• New more thorough documentation
• Travis builds for all PRs and commits for Python 2 and 3!
• The decoding flags for t2t_decoder have all been merged into a single HParams object that can be modified with the flag --decode_hparams
• Various feature additions, bug fixes, and improvements
• Note: Parameter checkpoints for the Transformer model may be broken because of a bug with variable sharing in layer_norm