Hi Erik! Great work with the repo. I was following the CGAN code alo

Ideas on CGAN code,about eriklindernoren/keras-gan

Comments (4)

eriklindernoren commented on July 27, 2024 4

Yeah, the implementations will in some cases not exactly mirror the ones described in the papers. Here is how I interpret the architecture described in the text you posted:

def build_generator(self):

  # Noise layer
  noise = Input(shape=(self.latent_dim,))
  noise_emb = Dense(200, activation='relu')(noise)      # out_shape: (batch_size, 200)

  # Label layer
  label = Input(shape=(self.num_classes+1,))
  label_emb = Dense(1000, activation='relu')(label)     # out_shape: (batch_size, 1000)

  # Joint layers
  merged = Concatenate()([noise_emb, label_emb])        # out_shape: (batch_size, 1200)
  combined_out = Dense(1200, activation='relu')(merged) # out_shape: (batch_size, 1200)
  img_flat = Dense(np.prod(self.img_shape), activation='sigmoid') \
                                (combined_out)          # out_shape: (batch_size, 784)
  img = Reshape(self.img_shape)(img_flat)               # out_shape: (batch_size, 28, 28, 1)


  return Model([noise, label], img)

The labels will have to be one-hot encoded before fed as input to the generator. self.num_classes will have to be set to 10 in the MNIST case.

from keras-gan.

andrearama commented on July 27, 2024 2

Might I ask what exactly the main change is supposed to do, and the motivation of such choice?
I mean the line:
Flatten()(Embedding(self.num_classes, self.latent_dim)(label))

Thanks

from keras-gan.

miranthajayatilake commented on July 27, 2024

Thanks a lot for this Erik! Exactly what I wanted to figure out! Keep up the great work.

from keras-gan.