Method forward

espnet2/asr/decoder/rnn_decoder.py:171–253

(self, hs_pad, hlens, ys_in_pad, ys_in_lens, strm_idx=0)

        return z_list, c_list

    def forward(self, hs_pad, hlens, ys_in_pad, ys_in_lens, strm_idx=0):
        # to support mutiple encoder asr mode, in single encoder mode,
        # convert torch.Tensor to List of torch.Tensor
        if self.num_encs == 1:
            hs_pad = [hs_pad]
            hlens = [hlens]

        # attention index for the attention module
        # in SPA (speaker parallel attention),
        # att_idx is used to select attention module. In other cases, it is 0.
        att_idx = min(strm_idx, len(self.att_list) - 1)

        # hlens should be list of list of integer
        hlens = [list(map(int, hlens[idx])) for idx in range(self.num_encs)]

        # get dim, length info
        olength = ys_in_pad.size(1)

        # initialization
        c_list = [self.zero_state(hs_pad[0])]
        z_list = [self.zero_state(hs_pad[0])]
        for _ in range(1, self.dlayers):
            c_list.append(self.zero_state(hs_pad[0]))
            z_list.append(self.zero_state(hs_pad[0]))
        z_all = []
        if self.num_encs == 1:
            att_w = None
            self.att_list[att_idx].reset()  # reset pre-computation of h
        else:
            att_w_list = [None] * (self.num_encs + 1)  # atts + han
            att_c_list = [None] * self.num_encs  # atts
            for idx in range(self.num_encs + 1):
                # reset pre-computation of h in atts and han
                self.att_list[idx].reset()

        # pre-computation of embedding
        eys = self.dropout_emb(self.embed(ys_in_pad))  # utt x olen x zdim

        # loop for an output sequence
        for i in range(olength):
            if self.num_encs == 1:
                att_c, att_w = self.att_list[att_idx](
                    hs_pad[0], hlens[0], self.dropout_dec[0](z_list[0]), att_w
                )
            else:
                for idx in range(self.num_encs):
                    att_c_list[idx], att_w_list[idx] = self.att_list[idx](
                        hs_pad[idx],
                        hlens[idx],
                        self.dropout_dec[0](z_list[0]),
                        att_w_list[idx],
                    )
                hs_pad_han = torch.stack(att_c_list, dim=1)
                hlens_han = [self.num_encs] * len(ys_in_pad)
                att_c, att_w_list[self.num_encs] = self.att_list[self.num_encs](
                    hs_pad_han,
                    hlens_han,
                    self.dropout_dec[0](z_list[0]),