About the loss fucntion of ccgan about keras-gan HOT 4 CLOSED

I have run this code but the result seems incorrect.

from keras-gan.

Yeah, I wrote about this in the README.

"In this implementation I have also decided to combine the adversarial loss with an l2 loss which measures the generated image's similarity to the original images. These losses are weighted similar to the approach described by Pathak et al. (2016)
in their paper Context Encoders: Feature Learning by Inpainting."

Adding the l2 loss helps the model converge more quickly. It's possible to get convergence without it too of course, but it takes way longer time.

from keras-gan.

Thank you for you replying. When I trained the ccgan model, the loss of G and loss of D remain nan like this:

168 [D loss: nan, acc: 100.00%, op_acc: 34.38%] [G loss: 64382.972656, mse: 64447.417969]

and
'
label = Dense(self.num_classes + 1, activation="softmax")(features)
'
why the output is num_classse + 1？
Regards!

from keras-gan.

Ok, I have found the problem. #2

from keras-gan.