Home

Awesome

Omni-GAN-DGP

This repository contains the code for the paper, Omni-GAN: On the Secrets of cGANs and Beyond. </br > In particular, it contains the code for the ImageNet and DGP experiments. </br > We recommend that you refer to another project, Omni-GAN-PyTorch, to learn Omni-GAN quickly.

My tasks

✔️ Training code for ImageNet experiments. To do by the end of this month (2021-10-19)
⬜️ The inversion code for DGP with Omni-INR-GAN experiments.

===========================================================

<p float="left"> <img src=.github/truncation_curve.png width="600" /> </p>

DGP experiments

<p float="left"> <img src=.github/colorization.png width="800" /> </p> <p float="left"> <img src=.github/SR60.png width="800" /> </p>

Envs

git clone --recursive https://github.com/PeterouZh/Omni-GAN-DGP.git
cd Omni-GAN-DGP

# Create virtual environment
conda create -y --name omnigan python=3.6.7
conda activate omnigan

pip install torch==1.8.2+cu102 torchvision==0.9.2+cu102 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.8/index.html

pip install --no-cache-dir tl2==0.0.3
pip install --no-cache-dir -r requirements.txt

Prepare dataset

export CUDA_VISIBLE_DEVICES=0
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/make_hdf5.py \
  --tl_config_file configs/make_hdf5.yaml \
  --tl_command make_hdf5_ImageNet128 \
  --tl_outdir results/make_hdf5_ImageNet128 \
  --tl_opts data_root datasets/ImageNet/train \
    index_filename datasets/ImageNet_hdf5/I128_index.npz \
    saved_hdf5_file datasets/ImageNet_hdf5/ILSVRC128.hdf5

Note: data_root: the path of ImageNet training images.

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/calculate_inception_moments.py \
  --tl_config_file configs/make_hdf5.yaml \
  --tl_command calculate_inception_moments_ImageNet128 \
  --tl_outdir results/calculate_inception_moments_ImageNet128 \
  --tl_opts data_root datasets/ImageNet_hdf5/ILSVRC128.hdf5 \
    saved_inception_file datasets/ImageNet_hdf5/I128_inception_moments.npz


export CUDA_VISIBLE_DEVICES=0
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/make_hdf5.py \
  --tl_config_file configs/make_hdf5.yaml \
  --tl_command make_hdf5_ImageNet256 \
  --tl_outdir results/make_hdf5_ImageNet256 \
  --tl_opts data_root datasets/ImageNet/train \
    index_filename datasets/ImageNet_hdf5/I256_index.npz \
    saved_hdf5_file datasets/ImageNet_hdf5/ILSVRC256.hdf5


export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/calculate_inception_moments.py \
  --tl_config_file configs/make_hdf5.yaml \
  --tl_command calculate_inception_moments_ImageNet256 \
  --tl_outdir results/calculate_inception_moments_ImageNet256 \
  --tl_opts data_root datasets/ImageNet_hdf5/ILSVRC256.hdf5 \
    saved_inception_file datasets/ImageNet_hdf5/I256_inception_moments.npz

Evaluation (Omni-GAN)

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/train.py \
  --tl_config_file configs/omnigan_imagenet128.yaml \
  --tl_command eval_ImageNet128 \
  --tl_outdir results/eval_ImageNet128 \
  --tl_opts inception_file  datasets/ImageNet_hdf5/I128_inception_moments.npz \
    evaluation.G_ema_model datasets/pretrained/omnigan_r128_G_ema.pth

Train (Omni-GAN)

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PYTHONPATH=./BigGAN_Pytorch_lib:./
python scripts/train.py \
  --tl_config_file configs/omnigan_imagenet128.yaml \
  --tl_command train_ImageNet128 \
  --tl_outdir results/train_ImageNet128 \
  --tl_opts args.data_root datasets/ImageNet_hdf5/ILSVRC128.hdf5 \
    inception_file  datasets/ImageNet_hdf5/I128_inception_moments.npz

For 256x256 (Omni-GAN)

Omni-INR-GAN

Acknowledgments