Package contains, in PyTorch implemented, neural networks with problem specific pre-structuring architectures and utils that help building and understanding models.
For HCNNCell and HcnnLstmCell the teacher_forcing output is scaled in the dropout layer if teacher_forcing < 1. Therefore, the correction of the state is wrong and the intermediate state does not contain the values of the observation.
The following paper presents three GRU variants. Variant 3 ist similar to the LSTM implementation in ECNN and HCNN. One difference is that r_t is fixed to a vector with ones in our implementation.
R. Dey and F. M. Salem, "Gate-variants of Gated Recurrent Unit (GRU) neural networks,"
2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS),
Boston, MA, USA, 2017, pp. 1597-1600, doi: 10.1109/MWSCAS.2017.8053243