Open kristosh opened 5 years ago
I think you should addition labels
You mean where? When I add the input to the generator so the z_gen = Input(shape=(self.latent_dim,)) during the training, I am concatenating some noise with the labels and pass it as input to the critic. Then, when i train the critic I am doing:
d_loss = self.critic_model.train_on_batch([imgs, noise],[valid, fake, dummy, batch_lbls])
imgs are the target images, noise the input to the generator, and batch labels the correspondent labels to the target samples.
I would like to modify the critic of the Network with the purpose of not only working as a discriminator but also as an auxiliar classifier. Therefore, I would like the last layer of the critic to output the Wasserstein distance and multi cross entropy results from the classification. I tried to modify the code accordingly. Therefore, now my init function looks like:
The critic model:
`def build_critic_iwGANs():
While the training of the critic:
d_loss = self.critic_model.train_on_batch([imgs, noise],[valid, fake, dummy, batch_lbls])
Where batch_lbls are the categorical real annotation of the real distribution. While the system to work and produce extremely nice visual results, the classification performance during the training continues to be dumb and do not learn anything. Is there any obvious bug in my code?