fedral / SRFlow

Official SRFlow training code: Super-Resolution using Normalizing Flow in PyTorch

SRFlow

Official SRFlow training code: Super-Resolution using Normalizing Flow in PyTorch

[Paper] ECCV 2020 Spotlight

Setup: Data, Environment, PyTorch Demo

git clone https://github.com/andreas128/SRFlow.git && cd SRFlow && ./setup.sh

This oneliner will:

• Clone SRFlow
• Setup a python3 virtual env
• Install the packages from requirements.txt
• Download the pretrained models
• Download the validation data
• Run the Demo Jupyter Notebook

If you want to install it manually, read the setup.sh file. (Links to data/models, pip packages)

Demo: Try Normalizing Flow in PyTorch

./run_jupyter.sh

This notebook lets you:

• Load the pretrained models.
• Super-resolve images.
• Measure PSNR/SSIM/LPIPS.
• Infer the Normalizing Flow latent space.

Testing: Apply the included pretrained models

source myenv/bin/activate                      # Use the env you created using setup.sh
cd code
CUDA_VISIBLE_DEVICES=-1 python test.py ./confs/SRFlow_DF2K_4X.yml      # Diverse Images 4X (Dataset Included)
CUDA_VISIBLE_DEVICES=-1 python test.py ./confs/SRFlow_DF2K_8X.yml      # Diverse Images 8X (Dataset Included)
CUDA_VISIBLE_DEVICES=-1 python test.py ./confs/SRFlow_CelebA_8X.yml    # Faces 8X

For testing, we apply SRFlow to the full images on CPU.

Training: Reproduce or train on your Data

The following commands train the Super-Resolution network using Normalizing Flow in PyTorch:

source myenv/bin/activate                      # Use the env you created using setup.sh
cd code
python train.py -opt ./confs/SRFlow_DF2K_4X.yml      # Diverse Images 4X (Dataset Included)
python train.py -opt ./confs/SRFlow_DF2K_8X.yml      # Diverse Images 8X (Dataset Included)
python train.py -opt ./confs/SRFlow_CelebA_8X.yml    # Faces 8X

• To reduce the GPU memory, reduce the batch size in the yml file.
• CelebA does not allow us to host the dataset. A script will follow.

Dataset: How to train on your own data

The following command creates the pickel files that you can use in the yaml config file:

cd code
python prepare_data.py /path/to/img_dir

The precomputed DF2K dataset gets downloaded using setup.sh. You can reproduce it or prepare your own dataset.

Our paper explains

• How to train Conditional Normalizing Flow
  We designed an architecture that archives state-of-the-art super-resolution quality.
• How to train Normalizing Flow on a single GPU
  We based our network on GLOW, which uses up to 40 GPUs to train for image generation. SRFlow only needs a single GPU for training conditional image generation.
• How to use Normalizing Flow for image manipulation
  How to exploit the latent space for Normalizing Flow for controlled image manipulations
• See many Visual Results
  Compare GAN vs Normalizing Flow yourself. We've included a lot of visuals results in our [Paper].

GAN vs Normalizing Flow - Blog

• Sampling: SRFlow outputs many different images for a single input.
• Stable Training: SRFlow has much fewer hyperparameters than GAN approaches, and we did not encounter training stability issues.
• Convergence: While GANs cannot converge, conditional Normalizing Flows converge monotonic and stable.
• Higher Consistency: When downsampling the super-resolution, one obtains almost the exact input.

Get a quick introduction to Normalizing Flow in our [Blog].

Wanna help to improve the code?

If you found a bug or improved the code, please do the following:

• Fork this repo.
• Push the changes to your repo.
• Create a pull request.

Paper

[Paper] ECCV 2020 Spotlight

@inproceedings{lugmayr2020srflow,
  title={SRFlow: Learning the Super-Resolution Space with Normalizing Flow},
  author={Lugmayr, Andreas and Danelljan, Martin and Van Gool, Luc and Timofte, Radu},
  booktitle={ECCV},
  year={2020}
}