yunjey / mnist-svhn-transfer Goto Github PK

I found two problems about the code.
First, the reconstruction loss should be L1 loss instead of L2, according to cycleGAN's paper.
Second, the generator and discriminator should update simutaneously, while in the code it firstly update discriminator and then use updated discriminator to update generator, which is wrong.

Hi,
Could you explain in more details what you mean by "minimal"? Which features miss from the original paper/implementation?

So is this code actually working? Readme says "(should be re-uploaded)", What does that mean? Some people here report that this code is just memorizing training examples but does not generalize. Is this true?

Even worse than having no code is having code that is pretending to work but will just waste your time and then you have to implement it from scratch anyway...

Hi, thanks for sharing the well-written code.
Could you please also share the testing code? I have successfully trained the model, now I just want to test it on the validation set.

Where is test code? In "main.py" line 35，I don't where is the "sample()" in solver.py.

After running model:
RuntimeError: output with shape [1, 32, 32] doesn't match the broadcast shape [3, 32, 32]

solver.py 128, 134 .. line

• Before (Now)
  d1_loss = torch.mean((out-1)**2)

• After
  d1_loss = torch.mean((out-1).clamp(max=0)**2)

when i run your code about cyclegan svhn to mnist，i find the result is very pool，maybe the sample example is good，but，the test examples is pool. for example ,it can translate 8 to 5, 2 to 0,the correct rate of the generate mnist is about 25%,when the correct rate about train mnist classify is 99%.