dominic1021/LatentSync

1# https://github.com/joonson/syncnet_python/blob/master/SyncNetModel.py2 3import torch4import torch.nn as nn5 6 7def save(model, filename):8    with open(filename, "wb") as f:9        torch.save(model, f)10        print("%s saved." % filename)11 12 13def load(filename):14    net = torch.load(filename)15    return net16 17 18class S(nn.Module):19    def __init__(self, num_layers_in_fc_layers=1024):20        super(S, self).__init__()21 22        self.__nFeatures__ = 2423        self.__nChs__ = 3224        self.__midChs__ = 3225 26        self.netcnnaud = nn.Sequential(27            nn.Conv2d(1, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),28            nn.BatchNorm2d(64),29            nn.ReLU(inplace=True),30            nn.MaxPool2d(kernel_size=(1, 1), stride=(1, 1)),31            nn.Conv2d(64, 192, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),32            nn.BatchNorm2d(192),33            nn.ReLU(inplace=True),34            nn.MaxPool2d(kernel_size=(3, 3), stride=(1, 2)),35            nn.Conv2d(192, 384, kernel_size=(3, 3), padding=(1, 1)),36            nn.BatchNorm2d(384),37            nn.ReLU(inplace=True),38            nn.Conv2d(384, 256, kernel_size=(3, 3), padding=(1, 1)),39            nn.BatchNorm2d(256),40            nn.ReLU(inplace=True),41            nn.Conv2d(256, 256, kernel_size=(3, 3), padding=(1, 1)),42            nn.BatchNorm2d(256),43            nn.ReLU(inplace=True),44            nn.MaxPool2d(kernel_size=(3, 3), stride=(2, 2)),45            nn.Conv2d(256, 512, kernel_size=(5, 4), padding=(0, 0)),46            nn.BatchNorm2d(512),47            nn.ReLU(),48        )49 50        self.netfcaud = nn.Sequential(51            nn.Linear(512, 512),52            nn.BatchNorm1d(512),53            nn.ReLU(),54            nn.Linear(512, num_layers_in_fc_layers),55        )56 57        self.netfclip = nn.Sequential(58            nn.Linear(512, 512),59            nn.BatchNorm1d(512),60            nn.ReLU(),61            nn.Linear(512, num_layers_in_fc_layers),62        )63 64        self.netcnnlip = nn.Sequential(65            nn.Conv3d(3, 96, kernel_size=(5, 7, 7), stride=(1, 2, 2), padding=0),66            nn.BatchNorm3d(96),67            nn.ReLU(inplace=True),68            nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2)),69            nn.Conv3d(96, 256, kernel_size=(1, 5, 5), stride=(1, 2, 2), padding=(0, 1, 1)),70            nn.BatchNorm3d(256),71            nn.ReLU(inplace=True),72            nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1)),73            nn.Conv3d(256, 256, kernel_size=(1, 3, 3), padding=(0, 1, 1)),74            nn.BatchNorm3d(256),75            nn.ReLU(inplace=True),76            nn.Conv3d(256, 256, kernel_size=(1, 3, 3), padding=(0, 1, 1)),77            nn.BatchNorm3d(256),78            nn.ReLU(inplace=True),79            nn.Conv3d(256, 256, kernel_size=(1, 3, 3), padding=(0, 1, 1)),80            nn.BatchNorm3d(256),81            nn.ReLU(inplace=True),82            nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2)),83            nn.Conv3d(256, 512, kernel_size=(1, 6, 6), padding=0),84            nn.BatchNorm3d(512),85            nn.ReLU(inplace=True),86        )87 88    def forward_aud(self, x):89 90        mid = self.netcnnaud(x)91        # N x ch x 24 x M92        mid = mid.view((mid.size()[0], -1))93        # N x (ch x 24)94        out = self.netfcaud(mid)95 96        return out97 98    def forward_lip(self, x):99 100        mid = self.netcnnlip(x)101        mid = mid.view((mid.size()[0], -1))102        # N x (ch x 24)103        out = self.netfclip(mid)104 105        return out106 107    def forward_lipfeat(self, x):108 109        mid = self.netcnnlip(x)110        out = mid.view((mid.size()[0], -1))111        # N x (ch x 24)112 113        return out114