Class autoencoder_vgg6

script/feature/model.py:295–367

vgg encoder with bilinear upsampling

        return feat_out, x

class autoencoder_vgg6(nn.Module): # robust feature extractors
    ''' vgg encoder with bilinear upsampling'''
    def __init__(self):
        super(autoencoder_vgg6, self).__init__()
        self.encoder = models.vgg19(pretrained=True).features
        self.decoder = nn.Sequential(
            # (b, 512, 14, 14)
            nn.Conv2d(512, 512, 3, stride=1, padding=1),
            nn.ReLU(True),
            # (b, 512, 28, 28)
            nn.Conv2d(512, 512, 3, stride=1, padding=1),
            nn.ReLU(True),
            # (b, 256, 56, 56)
            nn.Conv2d(512, 256, 3, stride=1, padding=1),
            nn.ReLU(True),
            # (b, 128, 112, 112)
            nn.Conv2d(256, 128, 3, stride=1, padding=1),
            nn.ReLU(True),
            # (b, 64, 224, 224)
            nn.Conv2d(128, 64, 3, stride=1, padding=1),
            nn.ReLU(True),
            # nn.Conv2d(64, 3, 3, stride=1, padding=1),
            # nn.Tanh() # MSELoss
            # nn.Sigmoid() # BCELoss
        )

    def forward(self, x, upsampleH, upsampleW): #
        feat = []
        feat_out = [] # we only use high level features
        for i in range(len(self.encoder)):
            # print("layer {} encoder layer: {}".format(i, self.encoder[i]))
            x = self.encoder[i](x)
            if i == 3: # ReLU-4
                feat.append(x)
            elif i == 8: # ReLU-9
                feat.append(x)
            elif i == 17: # ReLU-18
                feat.append(x)
            elif i == 26: # ReLU-27
                feat.append(x)
            elif i == 35: # ReLU-36
                feat.append(x)

        for i in range(len(self.decoder)):
            # print("layer {} decoder layer: {}".format(i, self.decoder[i]))
            x = self.decoder[i](x)
            if i == 1:
                _, _, h, w = feat[4].shape
                x = nn.UpsamplingBilinear2d(size=(h,w))(x)
                x = x + feat[4]
            elif i == 3:
                _, _, h, w = feat[3].shape
                x = nn.UpsamplingBilinear2d(size=(h,w))(x)
                x = x + feat[3]
            elif i == 5:
                _, _, h, w = feat[2].shape
                x = nn.UpsamplingBilinear2d(size=(h,w))(x)
                x = x + feat[2]