Generative Adversarial Network (5)

Posted on In

Vanila Montage DCGAN

Results of vanila montage DCGAN according to activation function
results-of-vanila-montage-dcgan-according-to-activation-function

  • 생성기를 통해 생성된 이미지의 수준 저조
  • 모드 붕괴 발생

Increase Epoch & Seed

Results of reinforced montage DCGAN by changing epoch and seed
results-of-reinforced-montage-dcgan-by-changing-epoch-and-seed

  • 생성 이미지의 질 개선 완료
  • 모드 붕괴 개선 필요

Strong Discriminator

Loss of DCGAN training process to identify the cause of mode collapse
loss-of-dcgan-training-process-to-identify-the-cause-of-mode-collapse

  • 모드 붕괴의 원인을 생성기에 비해 상대적으로 너무 강력한 판별기로 선정
  • 6th Trial에서 판별기와 생성기의 균형 관측

Results of reinforced montage DCGAN by editing activation function and convolution layer of discriminator
results-of-reinforced-montage-dcgan-by-editing-activation-function-and-convolution-layer-of-discriminator

  • 판별기와 생성기의 loss 균형 != 양질의 생성기

Weak Generator

Results of reinforced montage DCGAN by editing convolution layer of generator
results-of-reinforced-montage-dcgan-by-editing-convolution-layer-of-generator

Results of reinforced montage DCGAN by changing learning rate of generator
results-of-reinforced-montage-dcgan-by-changing-learning-rate-of-generator

Loss of DCGAN training process according to trial
loss-of-dcgan-training-process-according-to-trial

  • 8th Trial에서 모드 붕괴 개선
  • 최적화를 위해 생성기의 learning rate, lr 조절

Strong Discriminator

Results of reinforced montage DCGAN by editing activation function of discriminator
results-of-reinforced-montage-dcgan-by-editing-activation-function-of-discriminator

  • Activation function으로 GELU가 판별기에 부정적 영향을 주어 강력한 판별기를 약화시키기 위해 사용
  • 하지만 너무 큰 성능 저하로 인해 생성기 발산

Weak Generator

Results of reinforced montage DCGAN by editing learning rate and kernel size of generator
results-of-reinforced-montage-dcgan-by-editing-learning-rate-and-kernel-size-of-generator

  • Activation function of discriminator: GELU
  • Activation function of generator: GELU $\rightarrow$ LeakyReLU
  • 모드 붕괴 개선 불가