disA2 and disB2 seem no use in model.py ?

[LeoXing1996]

First of all, thanks for your great work !

In your code, disA2 and disB2 are trained via random attribute code, and used to update generator in backward_G_alone.

However, backward_G_alone doesn't seem to be used in training process.

What is the purpose of disA2, disB2 and backward_G_alone ?

[hytseng0509]

Thank you for your interest! We call the function backward_G_alone here during the training phase.

[LeoXing1996]

Oh, my fault! I miss this function! But I still confuse about only train disA2 and disB2 via random attribute code and only train disA1 and disB1 via attribute from image.
Why the model need additional discriminators 😕 ?

[HsinYingLee]

I think it is first used in BiCycleGAN. It's not for any sound theoretical reasons, but based on experimental results. For high-level explanation, the distribution of translation via random code and via exemplar are quite different, especially in the early stage of training. Using two separate discriminators is more stable and can facilitate the training.