Differences between the paper and the code on the generation of fake image

[donydchen]

Hi Albert, thanks for sharing your code.

In the paper, the fake image is generated by using the equation
.

While in the code, I find that the fake image is generated in a different way, which is
fake_imgs_masked = fake_img_mask * self._real_img + (1 - fake_img_mask) * fake_imgs,
refers to https://github.com/albertpumarola/GANimation/blob/master/models/ganimation.py#L228 and https://github.com/albertpumarola/GANimation/blob/master/models/ganimation.py#L275, so it should be

Would you please kindly clarify this? Thanks in advance.

[albertpumarola]

Your are right, it is an error in the equation. I will update de arxiv. Thank you for pointing this out.

[ZHANG-SHI-CHANG]

Hi Albert, thanks for sharing your code.

In the paper, the fake image is generated by using the equation
.

While in the code, I find that the fake image is generated in a different way, which is
fake_imgs_masked = fake_img_mask * self._real_img + (1 - fake_img_mask) * fake_imgs,
refers to https://github.com/albertpumarola/GANimation/blob/master/models/ganimation.py#L228 and https://github.com/albertpumarola/GANimation/blob/master/models/ganimation.py#L275, so it should be

Would you please kindly clarify this? Thanks in advance.

兄弟，你复现实验的时候有没有遇到attention趋于1的问题，在我实验的过程中论文里的两个约束loss起不到作用，attention还是会趋于1

[albertpumarola]

If it is tending to 1, suposing that the dataset is correct, you can try augmenting the lambda of this constraint. Be careful tunning it, if you increase it too much A will tend to 0. When I was tunning it I was aiming for a lambda that constrained A around 0.5.