Method forward_hidden

funasr/models/mfcca/mfcca_encoder.py:409–455

Calculate forward propagation. Args: xs_pad (torch.Tensor): Input tensor (#batch, L, input_size). ilens (torch.Tensor): Input length (#batch). prev_states (torch.Tensor): Not to be used now. Returns: torch.Tensor: Output tensor (#batch,

(
        self,
        xs_pad: torch.Tensor,
        ilens: torch.Tensor,
        prev_states: torch.Tensor = None,
    )

        return xs_pad, olens, None

    def forward_hidden(
        self,
        xs_pad: torch.Tensor,
        ilens: torch.Tensor,
        prev_states: torch.Tensor = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
        """Calculate forward propagation.
        Args:
            xs_pad (torch.Tensor): Input tensor (#batch, L, input_size).
            ilens (torch.Tensor): Input length (#batch).
            prev_states (torch.Tensor): Not to be used now.
        Returns:
            torch.Tensor: Output tensor (#batch, L, output_size).
            torch.Tensor: Output length (#batch).
            torch.Tensor: Not to be used now.
        """
        masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
        if (
            isinstance(self.embed, Conv2dSubsampling)
            or isinstance(self.embed, Conv2dSubsampling6)
            or isinstance(self.embed, Conv2dSubsampling8)
        ):
            short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
            if short_status:
                raise TooShortUttError(
                    f"has {xs_pad.size(1)} frames and is too short for subsampling "
                    + f"(it needs more than {limit_size} frames), return empty results",
                    xs_pad.size(1),
                    limit_size,
                )
            xs_pad, masks = self.embed(xs_pad, masks)
        else:
            xs_pad = self.embed(xs_pad)
        num_layer = len(self.encoders)
        for idx, encoder in enumerate(self.encoders):
            xs_pad, masks = encoder(xs_pad, masks)
            if idx == num_layer // 2 - 1:
                hidden_feature = xs_pad
        if isinstance(xs_pad, tuple):
            xs_pad = xs_pad[0]
            hidden_feature = hidden_feature[0]
        if self.normalize_before:
            xs_pad = self.after_norm(xs_pad)
            self.hidden_feature = self.after_norm(hidden_feature)

        olens = masks.squeeze(1).sum(1)
        return xs_pad, olens, None