nota_wav2lip/models/base.py

from typing import final

import torch
from torch import nn


class Wav2LipBase(nn.Module):
    def __init__(self) -> None:
        super().__init__()

        self.audio_encoder = nn.Sequential()
        self.face_encoder_blocks = nn.ModuleList([])
        self.face_decoder_blocks = nn.ModuleList([])
        self.output_block = nn.Sequential()

    @final
    def forward(self, audio_sequences, face_sequences):
        # audio_sequences = (B, T, 1, 80, 16)
        B = audio_sequences.size(0)

        input_dim_size = len(face_sequences.size())
        if input_dim_size > 4:
            audio_sequences = torch.cat([audio_sequences[:, i] for i in range(audio_sequences.size(1))], dim=0)
            face_sequences = torch.cat([face_sequences[:, :, i] for i in range(face_sequences.size(2))], dim=0)

        audio_embedding = self.audio_encoder(audio_sequences)  # B, 512, 1, 1

        feats = []
        x = face_sequences
        for f in self.face_encoder_blocks:
            x = f(x)
            feats.append(x)

        x = audio_embedding
        for f in self.face_decoder_blocks:
            x = f(x)
            try:
                x = torch.cat((x, feats[-1]), dim=1)
            except Exception as e:
                print(x.size())
                print(feats[-1].size())
                raise e

            feats.pop()

        x = self.output_block(x)

        if input_dim_size > 4:
            x = torch.split(x, B, dim=0)  # [(B, C, H, W)]
            outputs = torch.stack(x, dim=2)  # (B, C, T, H, W)

        else:
            outputs = x

        return outputs