mrzhu-cool/pix2pix-pytorch

PyTorch implementation of "Image-to-Image Translation Using Conditional Adversarial Networks".

pix2pix-pytorch's Issues

mode.eval() when test

Hi, I notice that when testing the model, you don't set the model to eval mode. Is this a bug or some special reasons?

Models' forward are incorrect

Hey just realized that the model forward/sequence is wrong.
E.g
Current code:
input is (nc) x 256 x 256

e1 = self.conv1(input)
# input is (ngf) x 128 x 128
e2 = self.batch_norm2(self.conv2(self.leaky_relu(e1)))

This should be
input is (nc) x 256 x 256

e1 = self.conv1(input)
# input is (ngf) x 128 x 128
e2 = self.conv2(self.leaky_relu(self.batch_norm2(e1)))

Hence both the discriminator and generator models are incorrect.

代码结构问题

请问，你这个代码的结构和论文里边的结构完全一样么？
我看的不是很一样啊，论文里是编码器解码器的结构啊。
麻烦说下你这里的结构，如果方便的话，告诉我下我想改成论文里边那样子要改哪些~

should I add set_requires_grad(net_d, True/False) for discriminator during training?

Hello,
Thank you for your great work. However, I think you should add set_requires_grad(net_d, True/False) for discriminator during training. Is it true?

modified code:

    # (1) Update D network
    ######################
    set_requires_grad(net_d, True) # add it here
    optimizer_d.zero_grad()        
    # train with fake
    fake_ab = torch.cat((real_a, fake_b), 1)
    pdb.set_trace()
    pred_fake = net_d.forward(fake_ab.detach())
    loss_d_fake = criterionGAN(pred_fake, False)

    # train with real
    real_ab = torch.cat((reala, real_b), 1)
    pred_real = net_d.forward(real_ab)
    loss_d_real = criterionGAN(pred_real, True)
    
    # Combined D loss
    loss_d = (loss_d_fake + loss_d_real) * 0.5

    loss_d.backward()
   
    optimizer_d.step()

    set_requires_grad(net_d, False) # add it here

    ######################

I am looking forward to hearing from you. Thank you in advance!

How many epochs ??

@mrzhu-cool
For how many iterations should I train to regenerate the results for the Facades dataset ??

What is the function of h_offset and w_offset in dataset.py?

I can't understand the offset in dataset.py, Is it set for translation invariant? May you explain it for me, Thank you.

PatchGAN part in the implementation

Hi,
I had hard time finding the PatchGAN part of code,Although you have commented before writing discriminator.Can you please point out where is it exactly there and how are you using it?
Thanks

Consider only L1 loss

I want to consider only L1 loss.Is it correct to comment out the generator loss as

#loss_g = loss_g_gan + loss_g_l1
and consider only,
loss_g = loss_g_l1

Do I need to change anything in discriminator part?

Thanks in advance

Input and output sizes of images？

@mrzhu-cool
Why do I find that the output size will be smaller in the training results？
How can I modify the code to customize the output size or the same size as the input image...

why have you implemented resnet instead of unet?

I think you should have implemented unet for Generator structure not resnet

Can I implement Wasserstein loss?

Hi!

Thank you for your good works.

I wanna apply Wasserstein loss to script

How can I do?

Thank you!

Artifacts in the generated images

Hi,

I am getting a lot of artifacts when i am applying the code to train and test on face images.
@mrzhu-cool Any suggestions on how to avoid them ??

Thanks in advance