The discriminator network is also a deep convolutional network. It contains eight convolution blocks followed by two dense (fully connected) layers. Each convolution block is followed by a batch normalization layer. There are two dense layers at the end of the network, which work as a classification block. The last layer predicts the probability of an image belonging to the real dataset or to the fake dataset. The detailed configuration of the discriminator network is shown in the following table:
|
Layer name |
Hyperparameters |
Input shape |
Output shape |
|
Input layer |
None |
(256, 256, 3) |
(256, 256, 3) |
|
2D convolution layer |
filters=64, kernel_size=3, strides=1, padding='same', activation='leakyrelu' |
(256, 256, 3) |
(256, 256, 64) |
|
2D convolution layer |
filters=64, kernel_size=3, strides=2, padding='same', activation='leakyrelu' |
(256, 256, 64) |
(128, 128, 64) |
|
Batch normalization layer |
momentum=0.8 |
(128, 128, 64) |
(128, 128, 64) |
|
2D convolution layer |
filters=128, kernel_size=3, strides=1, padding='same', activation='leakyrelu' |
(128, 128, 64) |
(128, 128, 128) |
|
Batch normalization layer |
momentum=0.8 |
(128, 128, 128) |
(128, 128, 128) |
|
2D convolution layer |
filters=128, kernel_size=3, strides=2, padding='same', activation='leakyrelu' |
(128, 128, 128) |
(64, 64, 128) |
|
Batch normalization layer |
momentum=0.8 |
(64, 64, 128) |
(64, 64, 128) |
|
2D convolution layer |
filters=256, kernel_size=3, strides=1, padding='same', activation='leakyrelu' |
(64, 64, 128) |
(64, 64, 256) |
|
Batch normalization layer |
momentum=0.8 |
(64, 64, 256) |
(64, 64, 256) |
|
2D convolution layer |
filters=256, kernel_size=3, strides=2, padding='same', activation='leakyrelu' |
(64, 64, 256) |
(32, 32, 256) |
We now have a clear understanding of the architecture of both networks. In the next section, let's have a look at the objective function required to train SRGANs.