PSGAN 初步搭建與訓練

PSGAN 初步搭建與訓練

在閱讀了妝容遷移的相關論文，比如PSGAN、BeautyGAN、SPADE、StarGAN v2等之後，我開始使用PyTorch初步復現PSGAN網絡結構。

網絡結構部分代碼如下：

# Makeup Apply Network(MANet)
class Generator(nn.Module):
    """Generator. Encoder-Decoder Architecture."""

    def __init__(self, conv_dim=64, repeat_num=6):
        super(Generator, self).__init__()

        encoder_layers = []
        encoder_layers.append(nn.Conv2d(3, conv_dim, kernel_size=7, stride=1, padding=3, bias=False))
        # MANet設置沒有affine
        encoder_layers.append(nn.InstanceNorm2d(conv_dim, affine=False))
        encoder_layers.append(nn.ReLU(inplace=True))

        # Down-Sampling
        curr_dim = conv_dim
        for i in range(2):
            encoder_layers.append(nn.Conv2d(curr_dim, curr_dim * 2, kernel_size=4, stride=2, padding=1, bias=False))
            encoder_layers.append(nn.InstanceNorm2d(curr_dim * 2, affine=False))
            encoder_layers.append(nn.ReLU(inplace=True))
            curr_dim = curr_dim * 2

        # Bottleneck
        for i in range(3):
            encoder_layers.append(ResidualBlock(dim_in=curr_dim, dim_out=curr_dim))

        decoder_layers = []
        for i in range(3):
            decoder_layers.append(ResidualBlock(dim_in=curr_dim, dim_out=curr_dim))

        # Up-Sampling
        for i in range(2):
            decoder_layers.append(
                nn.ConvTranspose2d(curr_dim, curr_dim // 2, kernel_size=4, stride=2, padding=1, bias=False))
            decoder_layers.append(nn.InstanceNorm2d(curr_dim // 2, affine=True))
            decoder_layers.append(nn.ReLU(inplace=True))
            curr_dim = curr_dim // 2

        decoder_layers.append(nn.Conv2d(curr_dim, 3, kernel_size=7, stride=1, padding=3, bias=False))
        decoder_layers.append(nn.Tanh())

        self.encoder = nn.Sequential(*encoder_layers)
        self.decoder = nn.Sequential(*decoder_layers)
        self.MDNet = MDNet()
        self.AMM = AMM()

    def forward(self, source_image, reference_image):
        fm_source = self.encoder(source_image)
        fm_reference = self.MDNet(reference_image)
        morphed_fm = self.AMM(fm_source, fm_reference)
        result = self.decoder(morphed_fm)
        return result


class MDNet(nn.Module):
    """Generator. Encoder-Decoder Architecture."""

    # MDNet is similar to the encoder of StarGAN
    def __init__(self, conv_dim=64, repeat_num=3):
        super(MDNet, self).__init__()

        layers = []
        layers.append(nn.Conv2d(3, conv_dim, kernel_size=7, stride=1, padding=3, bias=False))
        layers.append(nn.InstanceNorm2d(conv_dim, affine=True))
        layers.append(nn.ReLU(inplace=True))

        # Down-Sampling
        curr_dim = conv_dim
        for i in range(2):
            layers.append(nn.Conv2d(curr_dim, curr_dim * 2, kernel_size=4, stride=2, padding=1, bias=False))
            layers.append(nn.InstanceNorm2d(curr_dim * 2, affine=True))
            layers.append(nn.ReLU(inplace=True))
            curr_dim = curr_dim * 2

        # Bottleneck
        for i in range(repeat_num):
            layers.append(ResidualBlock(dim_in=curr_dim, dim_out=curr_dim))
        self.main = nn.Sequential(*layers)

    def forward(self, reference_image):
        fm_reference = self.main(reference_image)
        return fm_reference


# AMM暫時先不用landmark detector，先試試一般的attention效果如何
# feature map也先不乘以visual_feature_weight
# 這裏attention部分的計算也先存疑，因爲論文中提到x和y需要是同一個區域，但結構圖中是softmax
class AMM(nn.Module):
    """Attentive Makeup Morphing module"""

    def __init__(self):
        super(AMM, self).__init__()
        self.visual_feature_weight = 0.01
        self.lambda_matrix_conv = nn.Conv2d(in_channels=256, out_channels=1, kernel_size=1)
        self.beta_matrix_conv = nn.Conv2d(in_channels=256, out_channels=1, kernel_size=1)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, fm_source, fm_reference):
        batch_size, channels, width, height = fm_reference.size()
        old_lambda_matrix = self.lambda_matrix_conv(fm_reference).view(batch_size, -1, width * height)
        old_beta_matrix = self.beta_matrix_conv(fm_reference).view(batch_size, -1, width * height)

        # reshape後fm的形狀是C*(H*W)
        temp_fm_reference = fm_reference.view(batch_size, -1, height * width)
        # print('temp_fm_reference shape: ', temp_fm_reference.shape)
        # fm_source 在reshape後需要transpose成(H*W)*C
        temp_fm_source = fm_source.view(batch_size, -1, height * width).permute(0, 2, 1)
        # print('temp_fm_source shape: ', temp_fm_source.shape)
        # energy的形狀應該是N*N，N=H*W
        energy = torch.bmm(temp_fm_source, temp_fm_reference)
        attention_map = self.softmax(energy)

        new_lambda_matrix = torch.bmm(old_lambda_matrix, attention_map.permute(0, 2, 1))
        new_beta_matrix = torch.bmm(old_beta_matrix, attention_map.permute(0, 2, 1))
        new_lambda_matrix = new_lambda_matrix.view(batch_size, 1, width, height)
        new_beta_matrix = new_beta_matrix.view(batch_size, 1, width, height)

        # 對feature_map_source進行修改
        lambda_tensor = new_lambda_matrix.expand(batch_size, 256, width, height)
        beta_tensor = new_beta_matrix.expand(batch_size, 256, width, height)
        morphed_fm_source = torch.mul(lambda_tensor, fm_source)
        morphed_fm_source = torch.add(morphed_fm_source, beta_tensor)

        return morphed_fm_source

訓練數據讀取代碼如下:

def get_loader(data_config, config, mode="train"):
    dataset_name = data_config.name
    transform = transforms.Compose([
        transforms.Resize(config.img_size),
        transforms.ToTensor(),
        transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
    transform_mask = transforms.Compose([
        transforms.Resize(config.img_size, interpolation=PIL.Image.NEAREST),
        ToTensor])
    print(config.data_path)
    # """
    if mode == "train":
        dataset_train = eval(dataset_name)(data_config.dataset_path, transform=transform, mode="train", \
                                           transform_mask=transform_mask, cls_list=config.cls_list)
        dataset_test = eval(dataset_name)(data_config.dataset_path, transform=transform, mode="test", \
                                          transform_mask=transform_mask, cls_list=config.cls_list)

        # """
        data_loader_train = DataLoader(dataset=dataset_train,
                                       batch_size=config.batch_size,
                                       shuffle=True)

    if mode == "test":
        data_loader_train = None
        dataset_test = eval(dataset_name)(data_config.dataset_path, transform=transform, mode="test", \
                                          transform_mask=transform_mask, cls_list=config.cls_list)

    data_loader_test = DataLoader(dataset=dataset_test,
                                  batch_size=1,
                                  shuffle=False)

    return [data_loader_train, data_loader_test]

訓練尚未完成，但有最新迭代的妝容遷移效果，訓練效果如下：

第50輪迭代效果：

第49輪迭代效果：