This is a implementation of Grounded RNNs as proposed by Ankit Vani, et. al
I have modified the GRU and Dense Layers of Keras into Grounded_GRU and Grounded_Dense respectively to add the functionality mentioned in the paper.
The main modifications in these layers are as follows:
- GRU
- Multiplied the recurrent weight matrices, which correspond to the label states of the hidden vector, with an identity matrix to convert all non-diagonal elements to zero.
- WARNING: My implementation only supports GRU models with
implementationparameter set to2
- Dense
- Performs an inverse sigmoid operation on the grounded vector and then applies the affine transformation. For more details refer to
- Add sample implementation for sequence classification.