GradientTape compute saliency in the loss function

Question

I\'m trying to build an LSTM network to classify sentences and provide explanation for the classification using saliency. This network must learn from the t