about Discriminator

[budui]

Hi, I am porting PG2 to pytorch. I am confused about the design of Discriminator: In your code, the last layer of Discriminator is a linear whose output maybe not in [0, 1]. https://github.com/charliememory/Pose-Guided-Person-Image-Generation/blob/fef3e7a3b313c23501e5c2c8178f9ea6bac8ea41/wgan_gp.py#L434

so, I add a sigmoid to output:

        self.fc = nn.Linear(8*4*8*base_channels, 1)
        self.sigmoid = nn.Sigmoid()

        self.main.apply(weights_init)

    def forward(self, x):
        x = self.main(x)
        x = x.view(-1, 8*4*8*self.base_channels)
        return self.sigmoid(self.fc(x))

copy from here

This leads to poor output

if I only change the input layer's stride to (4, 2) of DCGAN's Discriminator.

            nn.Conv2d(in_channels=nc, out_channels=nf, kernel_size=4, stride=(4, 2), padding=1, bias=False),
            nn.LeakyReLU(0.2, inplace=True),

I get a better output.

but it's still far from your result in paper.

[budui]

My implement using PyTorch1.0 is here: https://github.com/budui/Human-Pose-Transfer-For-ReID/blob/master/models/PG2.py

Anyone interested in using pytorch to implement PG2? I need your help!

[charliememory]

Hi @budui, many thanks to your effort in the PyTorch version! The discriminator architecture is borrowed from the improved_wgan_training at that moment as follows https://github.com/igul222/improved_wgan_training/blob/fa66c574a54c4916d27c55441d33753dcc78f6bc/gan_64x64.py#L428 Currently, there are new discriminator architectures like PatchGAN.

As to the artifacts, I think it may be caused by the pixel value out of range. Adding a value clamp operation in the output to keep the pixel value between [-1,1] may help to address. Otherwise, you may need to trade off the loss weights between L1 loss and adversarial loss. And adding perceptual loss will further improve the results.

[budui]

Thanks for your reply! I have tried torch.clamp yesterday, artifacts disappeared, but result is still not good.

I am wondering that how to make the last linear's ouput is in [0,1]?

Can I email you in Chinese to describe the problem I have encountered?

[charliememory]

Sure, my email is liqian.ma@esat.kuleuven.be