A summary of the generator network is as follows:
# Summary of generator network model
summary(g)
____________________________________________________________________________
Layer (type) Output Shape Param #
============================================================================
input_7 (InputLayer) [(None, 28)] 0
____________________________________________________________________________
dense_4 (Dense) (None, 6272) 181888
____________________________________________________________________________
leaky_re_lu_8 (LeakyReLU) (None, 6272) 0
____________________________________________________________________________
reshape_2 (Reshape) (None, 14, 14, 32) 0
____________________________________________________________________________
conv2d_6 (Conv2D) (None, 14, 14, 32) 25632
____________________________________________________________________________
leaky_re_lu_9 (LeakyReLU) (None, 14, 14, 32) 0
____________________________________________________________________________
conv2d_transpose_2 (Conv2DTranspose) (None, 28, 28, 32) 16416
____________________________________________________________________________
leaky_re_lu_10 (LeakyReLU) (None, 28, 28, 32) 0
____________________________________________________________________________
conv2d_7 (Conv2D) (None, 28, 28, 1) 801
============================================================================
Total params: 224,737
Trainable params: 224,737
Non-trainable params: 0
_______________________________________________________________________________________
The summary of the generator network shows the output's shape and the number of parameters for each layer. Note that the final output shape is 28 x 28 x 1. The fake images that will be generated will have these dimensions. Overall, for this network, we have 224,737 parameters.
Now that we've specified the structure of the generator network, we can develop the architecture for the discriminator network.