Skip to content

Added Unroll-related features and layers

Christian Fuß requested to merge develop into master
  • Added support for RNNs, which are implemented by calling a network multiple times with the respective previous output.
  • Initialization and invocation of networks was extended to make use of RNN states.
  • Support for Beamsearch was added, currently only works with batches of size 1. Patch to follow.
  • SoftmaxCrossEntropyIgnoreIndices loss and BLEU eval_metric were introduced. "use_teacher_forcing" and "save_attention_image" flags can be used in training.
  • Added possibility to extract attention layer from network and save it as image.
  • Added backend implementation of new layers from CNNArch.
  • Added tests and datasets (or scripts to generate datasets too large for git) for recurrent networks. Updated dependencies.

Merge request reports