Cakechat Versions Save

Major update:

• Migrated to Keras on top of Tensorflow + Python 3.5
• Supported multi-GPU training with horovod
• Supported cudnnGRU for quicker inference
• Optimized training model architecture (speed-up by 42%)
• Increased hidden state dimension to 768

• fixed requirements installation in Dockerfile.gpu
• tensorflow and tensorboard -> 1.12.2

• Updated 'requests' dependency to get rid of security alert

Two functions updated:

• save_model preserves parameters' names while saving their values to a dump
• load_weights restores saved weight separately for each model's parameter

Thus you can initialize all or some of your model's parameters with weights that could come from a model with different architecture. Your model's parameter will be initialized with a saved value if the parameter's name and shape are the same as for the saved parameter. Otherwise it will keep its default initialization weights.

• Code refactoring
• train.py accepts --init_weights argument that specifies the file for model's initialization
• Naming of NN models changed: model's name always includes prefix defined in config's NN_MODEL_PREFIX const, which can be general ('cakechat' for example) or specific for the given experiment.

• Python 2+3 compatibility (thanks to @Oxylibrium)

Changes:

• Fixed Tensorboard dependencies
• Fixed Setuptools broken version
• Provided pre-built Docker images

The initial release of CakeChat: Emotional Generative Dialog System.