Recurrent Networks in TMVA

The initial plan was to have two layers of abstraction (inspired from Tensorflow). One layer for the Recurrent Cell, and one for the time unrolling using this cell. See this presentation for more details. However after writing the deep learning module, we concluded that writing the entire RNNLayer as a single class would be better suited. The cell technique was still kept in Forward() and Backward() functions to maintain better modularity.

At https://github.com/tmvadnn/tmva-dnn-tutorial and TestFullRNN you can find some examples using the Recurrent Networks. The repository isstill under construction, and we will be adding better examples.

Full code can be found here

Here is an example of an RNN->Reshape->DenseLayer network learning the identity function (input of dimensionality two stored as a state of size 3 in RNN, then reconstructed by DenseLayer).