Method get_att_weight

seq2seq_models.py:124–137

(self, hidden, encoder_hiddens)

        return output, hidden, attn_weights

    def get_att_weight(self, hidden, encoder_hiddens):
        seq_len = len(encoder_hiddens)

        # Create variable to store attention energies
        attn_scores = cuda_variable(torch.zeros(seq_len))  # B x 1 x S

        # Calculate energies for each encoder hidden
        for i in range(seq_len):
            attn_scores[i] = self.get_att_score(hidden, encoder_hiddens[i])

        # Normalize scores to weights in range 0 to 1,
        # resize to 1 x 1 x seq_len
        # print("att_scores", attn_scores.size())
        return F.softmax(attn_scores).view(1, 1, -1)

    # score = h^T W h^e = h dot (W h^e)
    # TODO: We need to implement different score models