v1.2.2

• Easy modular dataset construction with Problem.dataset
• New ClassLabelEncoder for classification problems
• Decoding now only loads the model once for all decodes (no more reloads across batches)
• Various other fixes, improvements, including a move to tf.estimator.Estimator

v1.2.1

• New t2t-decoder binary for decoding. t2t-trainer no longer does decoding
• Various fixes and improvements to code and documentation

v1.2.0

• We've updated our TensorFlow requirement to version 1.3.0, which allowed us to move our input pipeline to the new tf.contrib.data.Dataset API
• Attention visualizations available by iPython notebook
• Multiple new layers and utilities, bug fixes, etc.

v1.1.9

• Many problems have been renamed to make the hierarchy more clear; you will need to rename input files
• Various model and layer improvements and additions

v1.1.8

• Simplified mixture of experts code - breaks checkpoints that use them
• New models and attention functions
• Other cleanups and corrections

v1.1.7

• Improved residual block structure; note that this breaks old checkpoints (sorry!)
• Multiple improvements in utilities (T2TModel, evaluation, new modalities, new metrics)

v1.1.6

Merge pull request #209 from rsepassi/push

v1.1.6

v1.1.5

Large code moves:

• common_*.py and modalities.py have moved from models/ to the new folder layers/. models/ folder is now just for T2TModel definitions.
• trainer_utils.py has been broken up into several files (devices.py, decoding.py, input_fn_builder.py, model_builder.py.

v1.1.4

Merge pull request #201 from rsepassi/push

v1.1.4

v1.1.3

Merge pull request #188 from rsepassi/push

v1.1.3