InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

References

•

확률론 기초 by ratsgo

•

InfoGAN by haawron

•

Generative Models Part 2: ImprovedGAN,InfoGAN,EBGAN by Taeho Kim

•

초짜 대학원생 InfoGAN [part 1], [part 2] by Jaejun Yoo

G_{Loss} = G_{GANLoss} + \lambda C_{Loss}

D_{Loss} = D_{GANLoss} + \lambda C_{Loss}

input_to_G = torch.cat((noise, continuous_code_input, discrete_code_input), 1)
fake = G(input_to_G)

input_to_D = torch.cat([images, fake])

outputs = D(input_to_D)
# outputs : 1 (real/fake) + cont_code + dist_code
Python
복사

Train D

D.zero_grad()

# Mutual Information Loss
continuous_code_output = outputs[batch_size:, 1:1+num_of_continuous_code]
discrete_code_output = outputs[batch_size:, 1+num_of_continuous_code:1+num_of_continuous_code+num_of_class_of_category]

Loss_continuous = MSE(continuous_code_output, Distribution_of_continuous_code_generator)
Loss_discrete = CrossEntropy(discrete_code_output, discrete_code_input)

Loss_D = BCE(outputs[:,0], real_and_fake_labels) + Lambda1 * Loss_continuous + Lambda2 * Loss_discrete

Loss_D.backward(retain_variables=True)
D_optimizer.step()
Python
복사

Train G

G.zero_grad()

Loss_G = BCE(outputs[batch_size:,0], real_labels) + Lambda1 * Loss_continuous + Lambda2 * Loss_discrete

Loss_G.backward()
G_optimizer.step()
Python
복사