can we train bidirectional model? about lrp_for_lstm

[ArrasL]

Hi,

Yes you can train the bi-LSTM model with whatever neural network toolbox you like (Tensorflow, Pytorch, Keras, Dynet...).

Then you just need to save the trained model's weights as numpy arrays to use the LRP code (and check the ordering of the gates in your weights to make them compatible with the LRP code, here are some hints how to do this).

[akshayuppal3]

Also what happens if I think of changing the classes to binary intstead of multiclass?

[ArrasL]

Binary classification is just a special case of multiclass classification. So there is nothing you need to care about or change in the code if you want to apply it to a binary prediction task (the number of classes is not hard coded).

Just save your weights in a python dict of numpy arrays (akin model in the __init__ method of the class LSTM_bidi). The input, hidden layer and output dimensions will then be inferred from the weights' shapes.

Typically, the only places where you need to adapt the code, will be the __init__ and set_inputmethods, where you load your trained model, and define the input sequence x (x_rev).
The other methods don't need to be changed, if your model is a standard LSTM.