Source Code for Our Paper "When Relation Networks meet GANs: Relation GANs with Triplet Loss"
- python3
- pytorch
- torchvision
- numpy
- scipy
- tensorflow-gpu
We provide PyTorch implementations for Relation GAN and some measuring tools.
| GAN loss | name | FID(Cifar10) |
|---|---|---|
| WGAN-GP | wgangp | 63.7±0.11 |
| LS_GAN | ls_gan | 14.9±0.11 |
| vanilla GAN | sgan | 26.4±0.16 |
| Relativistic_GAN | rele | 24.1±0.19 |
| Our | relu_mean | 13.5±0.080 |
Frechet Inception Distance(https://github.com/mseitzer/pytorch-fid) Inception Score (https://github.com/google/compare_gan) Kernel Inception distance (https://github.com/google/compare_gan) Multi-scale Structural Similarity for Image Quality (https://github.com/google/compare_gan)
All gan loss function is in 'model' file folder.
In order to evaluate all model in a generally recognized method.We use both tensorflow model and pytorch model to get final result.
Our pytorch inceptionv3 model can be download here (https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth)
Our tensorflow inceptionv3 model can be download here (http://download.tensorflow.org/models/frozen_inception_v1_2015_12_05.tar.gz)
git clone
cd Final_RelationGAN\Final\
#train gan model with relation loss and 64 resolution.
python train.py --name Relation --which_loss mean_relu --dataroot path_to_img --gpu_id 1 --resize_or_crop resize_and_crop --which_step lateset --loadSize 64 --which_model_netG basic_64 --which_model_netD relation_64
#test gan model with relation loss and 64 resolution.
python test.py --name Relation --which_loss mean_relu --result_path path_to_save_reult --gpu_id 1 --which_step lateset --loadSize 64 --which_model_netG basic_64 --which_model_netD relation_64
#get FID SCORE with Inception-v3 model trained by pytorch.
python FID_Measure.py --name Relation --which_loss mean_relu --dataroot path_to_img --gpu_id 1 --resize_or_crop resize_and_crop --loadSize 64 --which_model_netG basic_64 --which_model_netD relation_64
#get IS,KID,MS_SSIM with Inception-v3 model trained by tensorflow.
python IS_Score_Tensorflow.py --name Relation --which_loss mean_relu --dataroot path_to_img --gpu_id 1 --resize_or_crop resize_and_crop --which_step lateset --loadSize 64 --which_model_netG basic_64 --which_model_netD relation_64
Our code is based on pytorch-CycleGAN-and-pix2pix(https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix)