Why can't my gan's implementation generate images like real ones?

vision
1 
    self.meta_optim.zero_grad()
    loss_q.backward(retain_graph=True)
    # optimize the discriminator
    self.meta_optim.step()

    for i in range(tasks_per_batch):
        qry_gen = self.generator(x_qry[i], vars=None, bn_training=True)
        for p in fast_weights[i]:
            p.detach()
        dis_feat = self.discriminator(x_qry[i], fast_weights[i], bn_training=True, feat=True)
        gen_feat = self.discriminator(qry_gen, fast_weights[i], bn_training=True, feat=True)
        fm_loss = torch.mean(torch.abs(torch.mean(gen_feat, 0) - torch.mean(dis_feat, 0)))

        nsample = gen_feat.size(0)
        gen_feat_norm = gen_feat / gen_feat.norm(p=2, dim=1).reshape(-1, 1).expand_as(gen_feat)
        cosine = torch.mm(gen_feat_norm, gen_feat_norm.t())
        mask = Variable((torch.ones(cosine.size()) - torch.diag(torch.ones(nsample))).cuda())
        pt_loss = 0.8 * torch.sum((cosine * mask) ** 2) / (nsample * (nsample - 1))

        loss_gen = fm_loss + pt_loss
        with torch.no_grad():
            corrects["gen_discrim"][0] += fm_loss

    self.gen_optim.zero_grad()
    loss_gen.backward()
    # optimize the generator
    self.gen_optim.step()

images_step6000

The left is real images, and the right is generated images. Is my backward procedure mistaked? Or maybe other problems like model, loss function, etc.?

2

I’m not familiar with your code, but some minor issues you should check:

  • p.detach() won’t do anything, as it’s not the inplace version (p.detach_() would be the inplace method)
  • Variables are deprecated since PyTorch 0.4 so you can use tensors in newer versions
  • Are you sure you need retain_graph=True? Sometimes this line is added as a workaround for valid errors, which will not fix the actual error.
4

Thanks vey much for your reply. I will try to fix some problems you mentioned.
If not using retain_graph=True, the program will error.

image
image1205×145 7.34 KB