Method forward

wan/modules/multitalk_model.py:396–429

(self, audio_embeds, audio_embeds_vf)

        self.norm = nn.LayerNorm(output_dim) if norm_output_audio else nn.Identity()

    def forward(self, audio_embeds, audio_embeds_vf):
        video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1]
        B, _, _, S, C = audio_embeds.shape

        # process audio of first frame
        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
        batch_size, window_size, blocks, channels = audio_embeds.shape
        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)

        # process audio of latter frame
        audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c")
        batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape
        audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf)

        # first projection
        audio_embeds = torch.relu(self.proj1(audio_embeds)) 
        audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf)) 
        audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B)
        audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B)
        audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1) 
        batch_size_c, N_t, C_a = audio_embeds_c.shape
        audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a)

        # second projection
        audio_embeds_c = torch.relu(self.proj2(audio_embeds_c))

        context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.output_dim)

        # normalization and reshape
        with amp.autocast(dtype=torch.float32):
            context_tokens = self.norm(context_tokens)
        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)

        return context_tokens


class WanModel(ModelMixin, ConfigMixin):