Network architectures

Generator architectures

The generator is consist of the encoder-decoder architecture:

encoder:

1. Conv2D(filers = 64, strides = 2), LeakyReLu(slope = 0.2)
2. Conv2D(filers = 128, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
3. Conv2D(filers = 256, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
4. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
5. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
6. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
7. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
8. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
9. Conv2D(filers = 512, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
10. Conv2D(filers = 512, strides = 2), ReLu

decoder:

1. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, Dropout(rate = 0.5), ReLU
2. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, Dropout(rate = 0.5), ReLU
3. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, Dropout(rate = 0.5), ReLU
4. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, ReLU
5. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, ReLU
6. Conv2DTranspose(filter = 512, strides = 2), BatchNorm, ReLU
7. Conv2DTranspose(filter = 256, strides = 2), BatchNorm, ReLU
8. Conv2DTranspose(filter = 128, strides = 2), BatchNorm, ReLU
9. Conv2DTranspose(filter = 64, strides = 2), BatchNorm, ReLU
10. Conv2DTranspose(filter = 1, strides = 2), Tanh

Also, the generator has skip-connections between layers of the encoder and layers of the decoder like the U-Net architecture.

skip-connection:

• encoder 1st layer - decoder 9th layer
• encoder 2nd layer - decoder 8th layer
• encoder 3rd layer - decoder 7th layer
• encoder 4th layer - decoder 6th layer
• encoder 5th layer - decoder 5th layer
• encoder 6th layer - decoder 4th layer
• encoder 7th layer - decoder 3rd layer
• encoder 8th layer - decoder 2nd layer
• encoder 9th layer - decoder 1st layer

Discriminator architectures

The discriminator architecture is described in the following notation:

• Conv2D(filers = 64, strides = 2), LeakyReLu(slope = 0.2)
• Conv2D(filers = 128, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
• Conv2D(filers = 256, strides = 2), BatchNorm, LeakyReLu(slope = 0.2)
• Conv2D(filers = 512, strides = 1), BatchNorm, LeakyReLu(slope = 0.2)
• Conv2D(filers = 1, strides = 1), Sigmoid

The receptive field size used in our discriminator is 70 x 70.

Hyperparameter

The loss configuration of the objective functions of the Generator

• Total loss = loss of J2 + 100 * loss of J1

Batch

• Batch iteration : 500,000
• Batch size : 1

Optimizer

• Optimizer : Adam solver
• Learning rate : 0.0002
• momentum beta 1 parameter : 0.5
• momentum beta 2 parameter : 0.999

Initializer

• The Initializer of the Convolution Layers : normal distribution, mean : 0.0, stddev : 0.02
• The gamma initializer of the BatchNormalization layers : normal distribution, mean : 1.0, stddev : 0.02