Aitoolbox

AIToolbox training setup with automatic experiment tracking.

With this release, TrainLoop supports in addition to the single GPU and DataParallel training also the DistributedDataParallel multi-GPU setting.
Furthermore this release also fully enables mixed-precision training via Nvidia APEX for both single GPU as well as multi-GPU training settings.

Implemented necessary building blocks to enable machine translation experiments. This includes more flexible train loop making seq2seq modelling easier, MT result package and underlying MT BLEU performance metrics.

torchtrain TrainLoops have been extended to enable the use of train / validation / test sets. Previously only train and validation sets were possible.

Implemented a very basic kerastrain port of torchtrain. As the name suggests this is a TrainLoop oriented experiment management system for Keras. However, currently, this is still work in progress and is not on the level of torchtrain which still has more functionality and enables greater flexibility.