Vision-aided GAN

video (3m) | website | paper

Can the collective knowledge from a large bank of pretrained vision models be leveraged to improve GAN training? If so, with so many models to choose from, which one(s) should be selected, and in what manner are they most effective?

We find that pretrained computer vision models can significantly improve performance when used in an ensemble of discriminators. We propose an effective selection mechanism, by probing the linear separability between real and fake samples in pretrained model embeddings, choosing the most accurate model, and progressively adding it to the discriminator ensemble. Our method can improve GAN training in both limited data and large-scale settings.

Ensembling Off-the-shelf Models for GAN Training
Nupur Kumari, Richard Zhang, Eli Shechtman, Jun-Yan Zhu
arXiv 2112.09130, 2021

Quantitative Comparison

Our method outperforms recent GAN training methods by a large margin, especially in limited sample setting. For LSUN Cat, we achieve similar FID as StyleGAN2 trained on the full dataset using only $0.7%$ of the dataset. On the full dataset, our method improves FID by 1.5x to 2x on cat, church, and horse categories of LSUN.

Example Results

Below, we show visual comparisons between the baseline StyleGAN2-ADA and our model (Vision-aided GAN) for the same randomly sample latent code.

Interpolation Videos

Latent interpolation results of models trained with our method on AnimalFace Cat (160 images), Dog (389 images), and Bridge-of-Sighs (100 photos).

Requirements

• 64-bit Python 3.8 and PyTorch 1.8.0 (or later). See https://pytorch.org/ for PyTorch install instructions.
• Cuda toolkit 11.0 or later.
• python libraries: see requirements.txt
• StyleGAN2 code relies heavily on custom PyTorch extensions. For detail please refer to the repo stylegan2-ada-pytorch

Setting up Off-the-shelf Computer Vision models

CLIP(ViT): we modify the model.py function to return intermediate features of the transformer model. To set up follow these steps.

git clone https://github.com/openai/CLIP.git
cp vision-aided-gan/training/clip_model.py CLIP/clip/model.py
cd CLIP
python setup.py install

DINO(ViT): model is automatically downloaded from torch hub.

VGG-16: model is automatically downloaded.

Swin-T(MoBY): Create a pretrained-models directory and save the downloaded model there.

Swin-T(Object Detection): follow the below step for setup. Download the model here and save it in the pretrained-models directory.

git clone https://github.com/SwinTransformer/Swin-Transformer-Object-Detection
cd Swin-Transformer-Object-Detection
pip install mmcv-full==1.3.0 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8.0/index.html
python setup.py install

for more details on mmcv installation please refer here

Swin-T(Segmentation): follow the below step for setup. Download the model here and save it in the pretrained-models directory.

git clone https://github.com/SwinTransformer/Swin-Transformer-Semantic-Segmentation.git
cd Swin-Transformer-Semantic-Segmentation
python setup.py install

Face Parsing:download the model here and save in the pretrained-models directory.

Face Normals:download the model here and save in the pretrained-models directory.

Pretrained Models

Our final trained models can be downloaded at this link

To generate images:

python generate.py --outdir=out --trunc=1 --seeds=85,265,297,849 --network=<network.pkl>

The output is stored in out directory controlled by --outdir. Our generator architecture is same as styleGAN2 and can be similarly used in the Python code as described in stylegan2-ada-pytorch.

model evaluation:

python calc_metrics.py --network <network.pkl> --metrics fid50k_full --data <dataset> --clean 1

We use clean-fid library to calculate FID metric. For LSUN Church and LSUN Horse, we calclate the full real distribution statistics. For details on calculating the real distribution statistics, please refer to clean-fid. For default FID evaluation of StyleGAN2-ADA use clean=0.

Datasets

Dataset preparation is same as given in stylegan2-ada-pytorch. Example setup for LSUN Church

LSUN Church

git clone https://github.com/fyu/lsun.git
cd lsun
python3 download.py -c church_outdoor
unzip church_outdoor_train_lmdb.zip
cd ../vision-aided-gan
python dataset_tool.py --source <path-to>/church_outdoor_train_lmdb/ --dest <path-to-datasets>/church1k.zip --max-images 1000  --transform=center-crop --width=256 --height=256

datasets can be downloaded from their repsective websites:

FFHQ, LSUN Categories, AFHQ, AnimalFace Dog, AnimalFace Cat, 100-shot Bridge-of-Sighs

Training new networks

model selection: returns the computer vision model with highest linear probe accuracy for the best FID model in a folder or the given network file.

python model_selection.py --data mydataset.zip --network  <mynetworkfolder or mynetworkpklfile>

example training command for training with a single pretrained network from scratch

python train.py --outdir=training-models/ --data=mydataset.zip --gpus 2 --metrics fid50k_full --kimg 25000 --cfg paper256 --cv input-dino-output-conv_multi_level --cv-loss multilevel_s --augcv ada --ada-target-cv 0.3 --augpipecv bgc --batch 16 --mirror 1 --aug ada --augpipe bgc --snap 25 --warmup 1  

Training configuration corresponding to training with vision-aided-loss:

• --cv=input-dino-output-conv_multi_level pretrained network and its configuration.
• --warmup=0 should be enabled when training from scratch. Introduces our loss after training with 500k images.
• --cv-loss=multilevel what loss to use on pretrained model based discriminator.
• --augcv=ada performs ADA augmentation on pretrained model based discriminator.
• --augcv=diffaugment-<policy> performs DiffAugment on pretrained model based discriminator with given poilcy.
• --augpipecv=bgc ADA augmentation strategy. Note: cutout is always enabled.
• --ada-target-cv=0.3 adjusts ADA target value for pretrained model based discriminator.
• --exact-resume=0 enables exact resume along with optimizer state.

Miscellaneous configurations:

• --appendname='' additional string to append to training directory name.
• --wandb-log=0 enables wandb logging.
• --clean=0 enables FID calculation using clean-fid if the real distribution statistics are pre-calculated.

Run python train.py --help for more details and the full list of args.

References

@article{kumari2021ensembling,
  title={Ensembling Off-the-shelf Models for GAN Training},
  author={Kumari, Nupur and Zhang, Richard and Shechtman, Eli and Zhu, Jun-Yan},
  journal={arXiv preprint arXiv:2112.09130},
  year={2021}
}

Acknowledgments

We thank Muyang Li, Sheng-Yu Wang, Chonghyuk (Andrew) Song for proofreading the draft. We are also grateful to Alexei A. Efros, Sheng-Yu Wang, Taesung Park, and William Peebles for helpful comments and discussion. Our codebase is built on stylegan2-ada-pytorch and DiffAugment.