How can I access the model's attribution created during forward pass when using dataparallel?

1

For example, I have such a forward function():


def forward(self, x):
        mgdata = self.encoder(x)
        self.imp_mask_sigmoid = self.impmap_sigmoid(mgdata)
        self.imp_mask = self.impmap_expand(self.imp_mask_sigmoid)
        enc_data = mgdata * self.imp_mask
        dec_data = self.decoder(enc_data)
        return dec_data

When I using dataparallel, I always using model.module to access the wrapped original model.However, I found I can’t access model.module.imp_mask_sigmoid, it will complain about AttributeError: 'ContentWeightedCNN' object has no attribute 'imp_mask_sigmoid'.
How to deal with it?

Thanks very much.

2

I have solved it from github issues, plz refer to here.