Get many many noises on C and D

[phelogges]

Thanks for your elegant work!
I'm implementing your paper by TF, but there comes many noises when I sample from outputs and it's getting worse...
I'm not sure if it happens to model collapse or something else...
I trained model on a subset of celebA with 10000 images.
The number on images is iterations with batch size 2.
Hope for your advice, thanks!

[Prinsphield]

Could you please share me a private repo, so that I can check your code?

[phelogges]

Here is my code
Thanks!

[Prinsphield]

Hello, good implementation!

There are several differences or mistakes in your code.

• [difference] The multi-scale discriminators share parameters. They are of identical network structure, but with different input image scales. However, in your implementation, they have separate parameters.
• [mistake] The strategy of training ELEGANT is iterative training: for each iteration, we train ELEGANT with respect to one attribute. For example, if training with respect to the smiling attribute, the input data should be a batch of smiling images (A), and another batch of non-smiling images (B). Either A or B cannot be a random batch images. I believe you misunderstand this part. You can refer to dataset.py and here for knowing more about how to feed the input data as well as the iterative training strategy.

[phelogges]

That will be a big mistake_(:з」∠)_

In that case, I should create graphs during training phase rather than initialization phase because of dynamic z_C and z_D...

I’ll modify my code...
Thanks!