[Paper Review] 31.Freeze the Discriminator ; a Simple Baseline for Fine-Tuning GANs

[Paper Review] 31. Freeze the Discriminator ; a Simple Baseline for Fine-Tuning GANs

Contents

1. Abstract

2. Previous Methods
3. Proposed Methods, FreezeD

1. Abstract

GAN : heavy computational cost!

• solution : transfer learning!

  but, prone to overfitting & limited to learning small distn shifts

Proposal : FreezeD

\(\rightarrow\) **simple fine-tuning ( freeze some parts of Discriminator )! **

2. Previous Methods

1. Fine tuning

   • (traditional) fine-tune both G & D

     \(\rightarrow\) but…suffer from overfitting!

2. Scale/Shift

   • since naive fine tuning is prone to overfitting…

     scale/shift suggest to update ONLY normalization layers

     \(\rightarrow\) poor result due to restriction!

3. GLO (Generative Latent Optimization)

   • fine–tune G with supervised learning
   • loss : sum of L1 loss & perceptual loss

4. MineGAN

   • fix G and modify the latent codes

3. Proposed Methods, FreezeD

outperform methods above!

1. FreezeD

   • freeze lower layers of D

     ( just fine tune upper layers )

   • simple & effective baseline!

2. L2-SP

   • effective for classifiers
   • regularizes target models, not to move far from source models

3. Feature distillation

   • distill the activations of the source & target models$$