One suggestion may be useful for learning about began-pytorch HOT 4 OPEN

Thanks for sharing your code. Did you tried to train the model with your code? If so, how does the samples look like?

from began-pytorch.

@carpedm20 Actually, here is another PyTorch implementation for BEGAN: https://github.com/sunshineatnoon/Paper-Implementations/tree/master/BEGAN. Although that code lacks two tanh layer at the ends of G and D networks, I found it works well after adding two tanh. After coarsely comparing that code with this one, I found that code strictly obeys the basic training scheme of GANs while this code does not. Maybe some senior tricks are used in this code that I can't undetstand, but you can try this:

1. Use d_loss and g_loss to separately and successively update D and G instead of adding them up and updating the whole network once. (line 166 and 167 in trainer.py)
2. Use g_loss = l1(AE_G_g, sample_z_G.detach()) on line 164

from began-pytorch.

Here is some samples I generated. 500_G,5000_G,8000_G & 17500_G. I change the lr from 0.00008 to 0.00004 at 5000 iters and to 0.00002 at 10000iters.
I'm sure that I'm using non-optimal parameters and super-parameters. And I don't consider the measure when I update the lr.

from began-pytorch.

@nian-liu I've seen the code you refered. Maybe he has changed his code, the code to compute loss is similiar to mine.

And I guess the tanh() is helpful for solving pytorch's bad initialization. I'm not sure, just for communication.

from began-pytorch.