Attention GAN

This repository provides GAN with various attention modules. All attention modules are applied to 32x32 feature map in both generator and discriminator(SA-GAN paper demonstrated when attention module apply to 32x32 feature map have the best fid result). Used wgan-hinge loss, spectral normalization and Adam optimizer for train, for discriminator and generator have different learning rate(4e-4 for discriminator, 1e-4 for generator) to satisfy TTUR(Two Time-scale Update Rule). Because we only use one class dataset(CelebA), replaced Conditional Batch Normalization with Batch Normalization.

Provided Attention Module

SA(Self Attention)

• "Self-Attention Generative Adversarial Networks."

CCA(Criss Cross Attention)

• "CCNet: Criss-Cross Attention for Semantic Segmentation."

YLA(Your Local Attention)

• "Your Local GAN: Designing Two Dimensional Local Attention Mechanisms for Generative Models."

ESA for YLA

Number of Attention Head

Attention Module	Num of Head
Self Attention (SA)	1
Criss Cross Attention (CCA)	2
Your Local Attention (YLA)	8

Results (Trained for 100K)

Attention Module	FID
Self Attention (SA)	32.516
Criss Cross Attention (CCA)	30.016
Your Local Attention (YLA)	32.691

Generated Images (Trained for 100K)

Generated from SA(Self Attention)-GAN

Generated from CCA(Criss Cross Attention)-GAN

Generated from YLA(Your Local Attention)-GAN

Requirements

• numpy>=1.20.3
• tqdm>=4.61.0
• pillow>=7.1.2
• torch>=1.7.1
• torchvision>=0.8.2

Train

run with default config

python train.py

run with CCA example

python train.py --batch_size 64 --dataset CelebA --version CCA

Generate Images

run with default config

python generate.py --load 100

run with CCA example

python generate.py --version CCA --load 100

Adopted from

• Spetral Noramlization
• FID Calculate