Skip to content

andreas128/SRFlow

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SRFlow

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


News: Unified Image Super-Resolution and Rescaling [code]

SRFlow


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 DIV2K 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.

How to prepare CelebA?

1. Get HD-CelebA-Cropper

git clone https://github.com/LynnHo/HD-CelebA-Cropper

2. Download the dataset

img_celeba.7z and annotations.zip as desribed in the Readme.

3. Run the crop align

python3 align.py --img_dir ./data/data --crop_size_h 640 --crop_size_w 640 --order 3 --face_factor 0.6 --n_worker 8

4. Downsample for GT

Use the matlablike kernel to downscale to 160x160 for the GT images.

5. Downsample for LR

Downscale the GT using the Matlab kernel to the LR size (40x40 or 20x20)

6. Train/Validation

For training and validation, we use the corresponding sets defined by CelebA (Train: 000001-162770, Validation: 162771-182637)

7. Pack to pickle for training

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



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}
}