PyTorch Implementation of (n)BRC

PyTorch implementation of the bistable recurrent cell (BRC) and recurrently neuromodulated bistable recurrent cell (nBRC).

The available classes, BRCLayer, nBRCLayer, BRC and nBRC, are documented in brc.py.

Download

git clone https://github.com/glambrechts/bistable-recurrent-cell
cd brc/

Example usage

See main.py for a copy-first-input benchmark with the BRC cell.

python3 main.py

Notes

The implementation is similar to that of torch.nn.GRU, such that the output of the RNN is its hidden state. A small wrapper is proposed in main.py to add a linear layer on top of the recurrent cell.

Also note that the parameter train_h0 allows to make the initial hidden state a trainable parameter of the recurrent cell.