Generative Adversarial Network

Goal:

• Train a generator-discriminator pair on CIFAR10 using techniques from ACGAN and Wasserstein GANs
• Compare the performance of GAN with a baseline ResNet module trained on CIFAR10
• Create synthetic images maximizing class output scores or a particular features to understand how CNN works

Results Visualization

• Real Images:
• Gradients:
• Perturbed Images:

Components

• Utilize techniques from ACGAN and Wasserstein GANs

• BN and ReLU are applied after transposed conv as well as conv in generator

• For transposed conv, use torch.nn.ConvTranspose2d. Transposed conv first blow up input region by inserting 0s between adjacent pixels. Then perform a conv operation

• 0.5 mean and 0.5 std normalization is applied on training images. This change pixel range from 0 to 1 to -1 to 1. However, the output of generator is 0 to 1. So use tanh fn to scale to -1 to 1 in order to match the input range of the discriminator

• discriminator contains fc1 and fc10 as the output:

  out1 = self.fc1(x)
  out10 = self.fc10(x)
  return out1, out10

  fc1 output a real number, that can be very positive or very negative. fc10 is cross entropy loss

• use 196 intermediate featmaps and train for 200 epochs (tutorial train for 500 epochs)

• random noise 1d vector of 100,1 is used as starting point for generator. 1st 10 numbers are the one-hot encoding for the target class of this fake image. The generator transform the 100 noise vector into a fake image

• Discriminator without Generator: just train discriminator define

• Discriminator with Generator: GAN algorithm, train both discriminator and generator

• visualization: One method is to input a noisy initialized image and calculate the gradient of a particular feature within the network with respect to the input. This gradient indicates how changes in the input image affect the magnitude of the feature. This process can be repeated with the image being adjusted slightly each time with the gradient. The final image is then a synthesized image causing a very high activation of that particular feature.