** UPDATE ** - A newer version of this framework with more models and functionality is now available here: https://github.com/um-dsrg/RUMpy
This repository contains the main coding framework accompanying our work on meta-attention in Single Image Super-Resolution (SISR), which has been published in the IEEE Signal Processing Letters (SPL) here. A sample of the results obtained by our metadata-enhanced models is provided below:
If installing from scratch, it is first recommended to set up a new Python virtual environment prior to installing this code. With Conda, this can be achieved through the following:
conda create -n *environment_name* python=3.7 (Python 3.7 recommended but not essential).
conda activate *environment_name*
Code testing was conducted in Python 3.7, but the code should work fine with Python 3.6+.
Run the following commands from the repo base directory to fully install the package and all requirements:
cd Code
If using CPU only: conda install --file requirements.txt --channel pytorch --channel conda-forge
If using CPU + GPU: First install Pytorch and Cudatoolkit for your specific configuration using instructions here. Then, install requirements as above.
If using Aim for metrics logging, install via pip install aim. The Aim GUI does not work on Windows, but metrics should still be logged in the .aim folder.
Finally:
pip install -e . This installs all the command-line functions from Code/setup.py.
All functionality has been tested on Linux (CPU & GPU), Mac OS (CPU) and Windows (CPU & GPU).
Requirements installation is only meant as a guide and all requirements can be installed using alternative means (e.g. using pip).
Create a folder 'celeba' in the Data directory. In here, download all files from the celeba source.
Unpack all archives in this location. Run image_manipulate to generate LR images and corresponding metadata (check details in Documentation/data_prep.md for more info on how to do this).
CelebA-HQ files can be easily downloaded from here. To generate LR images, check Documentation/data_prep.md as with CelebA. For our IEEE SPL paper (super-resolving by 4x), we generated images using the following two commands:
To generate 512x512 HR images: image_manipulate --source_dir *path_to_original_images* --output_dir *path_to_new_folder* --pipeline downscale --scale 2
To generate 128x128 LR images: image_manipulate --source_dir *path_to_512x512_images* --output_dir *path_to_new_folder* --pipeline blur-downscale --scale 4
To generate pre-upscaled 512x512 LR images for SPARNet: image_manipulate --source_dir *path_to_128x128_images* --output_dir *path_to_new_folder* --pipeline upscale --scale 4
DIV2K training/validation downloadable from here.
Flickr2K dataset downloadable from here.
Similar to CelebA-HQ, for our IEEE SPL paper (super-resolving by 4x), we generated LR images using the following command:
image_manipulate --source_dir *path_to_original_HR_images* --output_dir *path_to_new_folder* --pipeline blur-downscale --scale 4
For blurred & compressed images, we used the following command (make sure to first install JM to be able to compress the images, as detailed here):
image_manipulate --source_dir *path_to_original_HR_images* --output_dir *path_to_new_folder* --pipeline blur-downscale-jm_compress --scale 4 --random_compression
All SR testing datasets are available for download from the LapSRN main page here. Generate LR versions of each image using the same commands as used for the DIV2K/Flickr2K datasets.
Further detail on generating LR data provided in Documentation/data_prep.md.
To train models, prepare a configuration file (details in Documentation/model_training.md) and run:
train_sisr --parameters *path_to_config_file*
Similarly, for evaluation, prepare an eval config file (details in Documentation/model_eval.md) and run:
eval_sisr --config *path_to_config_file*
Standard SISR models available (code for each adapted from their official repository - linked within source code):
- Q-RCAN (meta-RCAN)
- Q-EDSR
- Q-SAN
- Q-HAN
- Q-SPARNet
- Various SFTMD variants (check SFTMD architectures file for options)
All weights for the models presented in our paper are available for download here. The models are split into three folders:
- Models trained on blurry general images: These models were all trained on DIV2K/Flickr2K blurred/downsampled images. These include:
- SRMD
- SFTMD
- RCAN
- EDSR
- SAN
- HAN
- Meta-RCAN
- Meta-EDSR
- Meta-SAN
- Meta-HAN
- Models trained on blurry and compressed general images: These models were all trained on DIV2K/Flickr2K blurred/downsampled/compressed images. These include:
- RCAN
- Meta-RCAN (accepting blur kernel data only)
- Meta-RCAN (accepting compression QPI data only)
- Meta-RCAN (accepting both blur kernels and compression QPI)
- Models trained on blurry face images: These models were all trained on CelebA-HQ blurred/downsampled images. These include:
- RCAN
- SPARNet (note that SPARNET only accepts pre-upsampled images)
- Meta-RCAN
- Meta-SPARNet
- Testing config files for all of these models are available in Documentation/SPL_testing_files. To use these, you need to first download and prepare the relevant datasets as shown here. Place the downloaded model folders in ./Results to use the config files as is, or adjust the
model_locparameter to point towards the directory containing the models.
Once downloaded, these models can be used directly with the eval command (```eval_sisr``) on any other input dataset as discussed in the evaluation documentation (Documentation/model_eval.md).
All training config files for models presented in our SPL paper are provided in Documentation/sample_config_files. These configurations assume that your training/eval data is stored in the relevant directory within ./Data, so please check that you have downloaded and prepared your datasets (as detailed above) before training.
Download the reference software from here. Place the software in the directory ./JM. cd into this directory and compile the software using the commands . unixprep.sh and make. Some changes might be required for different OS versions.
To compress images, simply add the jm_compress argument when specifying image_manipulate's pipeline.
Download pre-trained weights for the VGGFace model from here (scroll to VGGFace). Place the weights file in the directory ./external_packages/VGGFace/. The weights file should be called vgg_face_dag.pth.
Download pre-trained weights for the lightCNN model from here (LightCNN-29 v1). Place the weights file in the directory ./external_packages/LightCNN/. The weights file should be called LightCNN_29Layers_checkpoint.pth.tar.
Information on how to develop and train your own models is available in Documentation/framework_development.md.
The entire list of commands available with this repository is:
- train_sisr - main model training function.
- eval_sisr - main model evaluation function.
- image_manipulate - main bulk image converter.
- images_to_video - Helper function to convert a folder of images into a video.
- extract_best_model - Helper function to extract model config and best model checkpoint from a folder to a target location.
- clean_models - Helper function to remove unnecessary model checkpoints.
- model_report - Helper function to report on models available in specified directory.
Each command can be run with the --help parameter, which will print out the available options and docstrings.
Simply run:
pip uninstall Deep-FIR-SR
from any directory, with the relevant virtual environment activated.
@ARTICLE{Meta-Attention,
author={Aquilina, Matthew and Galea, Christian and Abela, John and Camilleri, Kenneth P. and Farrugia, Reuben A.},
journal={IEEE Signal Processing Letters},
title={Improving Super-Resolution Performance Using Meta-Attention Layers},
year={2021},
volume={28},
number={},
pages={2082-2086},
doi={10.1109/LSP.2021.3116518}}
This code has been released via the GNU GPLv3 open-source license. However, this code can also be made available via an alternative closed, permissive license. Third-parties interested in this form of licensing should contact us separately.
Usages of code from other repositories is properly referenced within the code itself.
We are working on a number of different research tasks in super-resolution, we'll be updating this repo as we make further advancements!
Short-term upgrades planned:
- CI automated testing (alongside Pytest)
- Release of packaged version
- Other upgrades TBA