fairseq/models/bert_seq2seq.py [366:383]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                )
                if self.normalize:
                    x = self.layer_norm(x)
                inner_states.append(x)
        # T x B x C -> B x T x C
        x = x.transpose(0, 1)

        if self.project_out_dim is not None:
            x = self.project_out_dim(x)

        if self.adaptive_softmax is None and self.load_softmax:
            # project back to size of vocabulary
            if self.share_input_output_embed:
                x = F.linear(x, self.embed_tokens.weight)
            else:
                x = F.linear(x, self.embed_out)

        return x, {'attn': attn, 'inner_states': inner_states, 'predicted_lengths': encoder_out['predicted_lengths']}
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



fairseq/models/disco_transformer.py [251:268]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
                )
                if self.normalize:
                    x = self.layer_norm(x)
                inner_states.append(x)
        # T x B x C -> B x T x C
        x = x.transpose(0, 1)

        if self.project_out_dim is not None:
            x = self.project_out_dim(x)

        if self.adaptive_softmax is None and self.load_softmax:
            # project back to size of vocabulary
            if self.share_input_output_embed:
                x = F.linear(x, self.embed_tokens.weight)
            else:
                x = F.linear(x, self.embed_out)

        return x, {'attn': attn, 'inner_states': inner_states, 'predicted_lengths': encoder_out['predicted_lengths']}
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -