Issue in BicycleGAN paper

arnabsinha · June 18, 2020, 6:26pm

In the BicycleGAN paper ( https://arxiv.org/pdf/1711.11586.pdf ), the Pytorch implemented code has the following function implemented:

github.com

junyanz/BicycleGAN/blob/0d7459a6b1eb577e024a5165097f0ae1609d0040/models/networks.py#L623


        layers += [nl_layer()]
        layers += [convMeanpool(inplanes, outplanes)]
        self.conv = nn.Sequential(*layers)
        self.shortcut = meanpoolConv(inplanes, outplanes)

    def forward(self, x):
        out = self.conv(x) + self.shortcut(x)
        return out


class E_ResNet(nn.Module):
    def __init__(self, input_nc=3, output_nc=1, ndf=64, n_blocks=4,
                 norm_layer=None, nl_layer=None, vaeLike=False):
        super(E_ResNet, self).__init__()
        self.vaeLike = vaeLike
        max_ndf = 4
        conv_layers = [
            nn.Conv2d(input_nc, ndf, kernel_size=4, stride=2, padding=1, bias=True)]
        for n in range(1, n_blocks):
            input_ndf = ndf * min(max_ndf, n)
            output_ndf = ndf * min(max_ndf, n + 1)

Could anyone confirm with me if the forward function is returning output as a 1x1 dimensional vector representing the point estimate of the predicted mean?

If so, in the following function, is this value of mean i.e. mu broadcasted to a |z| dimensional vector using the reparameterization trick?

github.com

junyanz/BicycleGAN/blob/0d7459a6b1eb577e024a5165097f0ae1609d0040/models/bicycle_gan_model.py#L81


    self.real_B = input['B' if AtoB else 'A'].to(self.device)
    self.image_paths = input['A_paths' if AtoB else 'B_paths']

def get_z_random(self, batch_size, nz, random_type='gauss'):
    if random_type == 'uni':
        z = torch.rand(batch_size, nz) * 2.0 - 1.0
    elif random_type == 'gauss':
        z = torch.randn(batch_size, nz)
    return z.to(self.device)

def encode(self, input_image):
    mu, logvar = self.netE.forward(input_image)
    std = logvar.mul(0.5).exp_()
    eps = self.get_z_random(std.size(0), std.size(1))
    z = eps.mul(std).add_(mu)
    return z, mu, logvar

def test(self, z0=None, encode=False):
    with torch.no_grad():
        if encode:  # use encoded z
            z0, _ = self.netE(self.real_B)

ptrblck · June 19, 2020, 8:27am

You could add print statements to the methods in question and print the shape of the tensors as well as parameters.
Let me know, if I misunderstood the question.

arnabsinha · June 19, 2020, 9:26am

Yes I was thinking of cloning the repo but I am afraid my laptop might not have the required specs or might hang during the training. I am eagerly waiting for codespaces on GitHub for that. I shall try to reduce the training set size. Could you figure out from the snippet posted above about the shape of the output variable?