Method score

funasr/models/language_model/rnn/decoders.py:1152–1207

Score. Args: yseq: TODO. state: TODO. x: TODO.

(self, yseq, state, x)

        )

    def score(self, yseq, state, x):
        # to support mutiple encoder asr mode, in single encoder mode,
        # convert torch.Tensor to List of torch.Tensor
        """Score.
        
            Args:
                yseq: TODO.
                state: TODO.
                x: TODO.
            """
        if self.num_encs == 1:
            x = [x]

        att_idx, z_list, c_list = state["workspace"]
        vy = yseq[-1].unsqueeze(0)
        ey = self.dropout_emb(self.embed(vy))  # utt list (1) x zdim
        if self.num_encs == 1:
            att_c, att_w = self.att[att_idx](
                x[0].unsqueeze(0),
                [x[0].size(0)],
                self.dropout_dec[0](state["z_prev"][0]),
                state["a_prev"],
            )
        else:
            att_w = [None] * (self.num_encs + 1)  # atts + han
            att_c_list = [None] * (self.num_encs)  # atts
            for idx in range(self.num_encs):
                att_c_list[idx], att_w[idx] = self.att[idx](
                    x[idx].unsqueeze(0),
                    [x[idx].size(0)],
                    self.dropout_dec[0](state["z_prev"][0]),
                    state["a_prev"][idx],
                )
            h_han = torch.stack(att_c_list, dim=1)
            att_c, att_w[self.num_encs] = self.att[self.num_encs](
                h_han,
                [self.num_encs],
                self.dropout_dec[0](state["z_prev"][0]),
                state["a_prev"][self.num_encs],
            )
        ey = torch.cat((ey, att_c), dim=1)  # utt(1) x (zdim + hdim)
        z_list, c_list = self.rnn_forward(ey, z_list, c_list, state["z_prev"], state["c_prev"])
        if self.context_residual:
            logits = self.output(torch.cat((self.dropout_dec[-1](z_list[-1]), att_c), dim=-1))
        else:
            logits = self.output(self.dropout_dec[-1](z_list[-1]))
        logp = F.log_softmax(logits, dim=1).squeeze(0)
        return (
            logp,
            dict(
                c_prev=c_list[:],
                z_prev=z_list[:],
                a_prev=att_w,
                workspace=(att_idx, z_list, c_list),
            ),
        )