the following are the Generator and discriminator model architectures for a (modified) WGAN model that I have come up with for downsampling rainfall data by a factor of 4 (only as a test). The input shape of the dataset is (8030, 14, 21) and the output shape I want is (8030, 28, 42).

def build_generator(input_shape):
   inputs = Input(shape=input_shape, name='generator_input')

   # Downsampling
   downsample_3 = conv_block(inputs, 128, (3, 3), 'downsample_3')
   downsample_2 = conv_block(downsample_3, 64, (3, 3), 'downsample_2')
   downsample_1 = conv_block(downsample_2, 32, (3, 3), 'downsample_1')

   # Bottleneck
   bottleneck_0 = conv_block(downsample_1, 16, (3, 3), 'bottleneck_0')
   bottleneck_00 = conv_block(bottleneck_0, 16, (3, 3), 'bottleneck_00')

   # Upsampling
   upsample_1 = deconv_block(bottleneck_00, 32, (3, 3), 'upsample_1')
   upsample_2 = deconv_block(upsample_1, 64, (3, 3), 'upsample_2')
   upsample_3 = deconv_block(upsample_2, 128, (3, 3), 'upsample_3')

   # Final Upsample to required shape
   outputs = Conv2DTranspose(filters=1, kernel_size=(3, 3), kernel_initializer=he_normal(), padding='same', 
                          strides=(2, 2), activation='relu', name='final_upsample_conv')(upsample_3)

   model = Model(inputs=inputs, outputs=outputs, name='generator')

   return model

def build_discriminator(input_shape):
   input_layer = Input(shape=input_shape, name='discriminator_input')

   x = conv_block(input_layer, filters=16, kernel_size=(3, 3), name='conv1')#, use_batch_norm=False)
   x = conv_block(x, filters=32, kernel_size=(3, 3), name='conv2')# use_batch_norm=False)
   x = conv_block(x, filters=128, kernel_size=(3, 3), name='conv3')# use_batch_norm=False)

   x = MaxPooling2D()(x)
   x = Dropout(0.25)(x)
   output_layer = Dense(1, activation='linear')(x)

   model = Model(inputs=input_layer, outputs=output_layer, name='discriminator')

   return model'''

Going by WGAN implementations in the literature I use the RMSprop optimizer instead of ADAM with a learning rate of 5e-5 and momentum of 0.5. Additionally I clip the discriminator weights to fall between -1e-2 and 1e-2. Traditionally (to the best of my knowledge), discriminator models output a single value loss conveying whether it is able to distinguish real output from fake outputs. In my case the discriminator output has the shape (None, 14, 21, 1) and it seems to work better but I don’t understand why. Does anyone have an idea why this might happen? Thanks in advance!