junting
commented
7 years ago Hello @wangxiao5791509 ,
You can probably change the weighting hyperparameter alpha for a smaller value. But it is normal that the drop of adversarial cost is slow to maintain the stability of the adversarial training.
class ModelSALGAN(Model): def __init__(self, w, h, batch_size=32, G_lr=3e-4, D_lr=3e-4, alpha=1/20.):
Best,
Junting
wangxiao5791509
Dear authors: The following screen shot is my training log, I think the decreasing of loss of discriminator is slow, how can I speed it up ? I first training the network using bce loss, then load pretrained model and continue the training with adversal loss. Do I made any mistakes ? Looking forward for your replay. Thanks very much !!!