gan
gan copied to clipboard
bojone
some demo of GANs
你好，我看你是通过add_loss添加梯度损失到计算图中实现GP的，但是新版的keras好像已经不支持这样添加带梯度的loss了，请问还有别的方法实现这个操作吗？我的版本是2.4.3
您好，请问wgan在计算wasserstein distance loss的时候，怎么能确保dfake大于dreal？我得到的dfake-dreal始终为负值，谢谢您
In [wgan_div_celeba.py#L129](https://github.com/bojone/gan/blob/master/keras/wgan_div_celeba.py#L129), there is `d_train_model.add_loss(d_loss + grad_loss)` and `d_loss` is defined by `K.mean(x_real_score - x_fake_score)`. Should it be `d_train_model.add_loss(d_loss - grad_loss)` ?
Can you please explain to me what these codes below do for? Thank you very much! d_train_model.metrics_names.append('w_dist') d_train_model.metrics_tensors.append(w_dist)
The code seems to be trained on celeba hq for tens of thousands times, how long does it take for the training of the model?
