Open source single image super-resolution toolbox containing various functionality for training a diverse number of state-of-the-art super-resolution models. Also acts as the companion code for the IEEE signal processing letters paper titled 'Improving Super-Resolution Performance using Meta-Attention Layers’.

Overview

Deep-FIR Codebase - Super Resolution Meta Attention Networks macOS Linux Windows License: GPL v3

About

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:

training_system

Installation

Python and Virtual Environments

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+.

Local Installation

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).

Guidelines for Generating SR Data

Setting up CelebA Dataset

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).

Setting up CelebA-HQ Dataset

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

Setting up DIV2K/Flickr2K Datasets

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

Setting up SR testing Datasets

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.

Additional Options

Further detail on generating LR data provided in Documentation/data_prep.md.

Training/Evaluating Models

Training

To train models, prepare a configuration file (details in Documentation/model_training.md) and run:

train_sisr --parameters *path_to_config_file*

Evaluation

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):

  1. SRCNN
  2. VDSR
  3. EDSR
  4. RCAN
  5. SPARNet
  6. SFTMD
  7. SRMD
  8. SAN
  9. HAN

Custom models available (all meta-models are marked with a Q-):

  1. Q-RCAN (meta-RCAN)
  2. Q-EDSR
  3. Q-SAN
  4. Q-HAN
  5. Q-SPARNet
  6. Various SFTMD variants (check SFTMD architectures file for options)

IEEE SPL Pre-Trained Model Weights

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_loc parameter 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).

Replicating SPL Results from Scratch

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.

Additional/Advanced Setup

Setting up JM (for compressing images)

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.

Setting up VGGFace (Pytorch)

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.

Setting up lightCNN

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.

Creating Custom Models

Information on how to develop and train your own models is available in Documentation/framework_development.md.

Full List of Commands Available

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.

Uninstall

Simply run:

pip uninstall Deep-FIR-SR

from any directory, with the relevant virtual environment activated.

Citation

Paper currently still in early-access, will update once fully published.

@ARTICLE{Meta-Attention,
author={Aquilina, Matthew and Galea, Christian and Abela, John and Camilleri, Kenneth P. and Farrugia, Reuben},
journal={IEEE Signal Processing Letters},
title={Improving Super-Resolution Performance using Meta-Attention Layers},
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/LSP.2021.3116518}}

License/Further Development

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
PyTorch code for our ECCV 2020 paper "Single Image Super-Resolution via a Holistic Attention Network"

HAN PyTorch code for our ECCV 2020 paper "Single Image Super-Resolution via a Holistic Attention Network" This repository is for HAN introduced in the

五维空间 114 Jan 14, 2022
PyTorch code for our paper "Attention in Attention Network for Image Super-Resolution"

Under construction... Attention in Attention Network for Image Super-Resolution (A2N) This repository is an PyTorch implementation of the paper "Atten

Haoyu Chen 51 Jan 4, 2022
Meta Language-Specific Layers in Multilingual Language Models

Meta Language-Specific Layers in Multilingual Language Models This repo contains the source codes for our paper On Negative Interference in Multilingu

Zirui Wang 19 May 5, 2021
Compare outputs between layers written in Tensorflow and layers written in Pytorch

Compare outputs of Wasserstein GANs between TensorFlow vs Pytorch This is our testing module for the implementation of improved WGAN in Pytorch Prereq

Hung Nguyen 70 Dec 17, 2021
Quickly comparing your image classification models with the state-of-the-art models (such as DenseNet, ResNet, ...)

Image Classification Project Killer in PyTorch This repo is designed for those who want to start their experiments two days before the deadline and ki

null 333 Jan 3, 2022
deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

deep-table implements various state-of-the-art deep learning and self-supervised learning algorithms for tabular data using PyTorch.

null 44 Dec 22, 2021
tsai is an open-source deep learning package built on top of Pytorch & fastai focused on state-of-the-art techniques for time series classification, regression and forecasting.

Time series Timeseries Deep Learning Pytorch fastai - State-of-the-art Deep Learning with Time Series and Sequences in Pytorch / fastai

timeseriesAI 1.6k Jan 12, 2022
Propose a principled and practically effective framework for unsupervised accuracy estimation and error detection tasks with theoretical analysis and state-of-the-art performance.

Detecting Errors and Estimating Accuracy on Unlabeled Data with Self-training Ensembles This project is for the paper: Detecting Errors and Estimating

Jiefeng Chen 4 Dec 11, 2021
PyTorch code for our ECCV 2018 paper "Image Super-Resolution Using Very Deep Residual Channel Attention Networks"

PyTorch code for our ECCV 2018 paper "Image Super-Resolution Using Very Deep Residual Channel Attention Networks"

Yulun Zhang 1k Jan 12, 2022
null 133 Jan 15, 2022
Pre-trained BERT Models for Ancient and Medieval Greek, and associated code for LaTeCH 2021 paper titled - "A Pilot Study for BERT Language Modelling and Morphological Analysis for Ancient and Medieval Greek"

Ancient Greek BERT The first and only available Ancient Greek sub-word BERT model! State-of-the-art post fine-tuning on Part-of-Speech Tagging and Mor

Pranaydeep Singh 13 Dec 17, 2021
🤗 Transformers: State-of-the-art Natural Language Processing for Pytorch, TensorFlow, and JAX.

English | 简体中文 | 繁體中文 State-of-the-art Natural Language Processing for Jax, PyTorch and TensorFlow ?? Transformers provides thousands of pretrained mo

Hugging Face 56.8k Jan 14, 2022
git《Pseudo-ISP: Learning Pseudo In-camera Signal Processing Pipeline from A Color Image Denoiser》(2021) GitHub: [fig5]

Pseudo-ISP: Learning Pseudo In-camera Signal Processing Pipeline from A Color Image Denoiser Abstract The success of deep denoisers on real-world colo

Yue Cao 40 Dec 28, 2021
[CVPR2021] Invertible Image Signal Processing

Invertible Image Signal Processing This repository includes official codes for "Invertible Image Signal Processing (CVPR2021)". Figure: Our framework

Yazhou XING 227 Jan 12, 2022
Implementation of NÜWA, state of the art attention network for text to video synthesis, in Pytorch

NÜWA - Pytorch (wip) Implementation of NÜWA, state of the art attention network for text to video synthesis, in Pytorch. This repository will be popul

Phil Wang 179 Jan 17, 2022
Code for reproducing our analysis in the paper titled: Image Cropping on Twitter: Fairness Metrics, their Limitations, and the Importance of Representation, Design, and Agency

Image Crop Analysis This is a repo for the code used for reproducing our Image Crop Analysis paper as shared on our blog post. If you plan to use this

Twitter Research 201 Jan 12, 2022
Improving Deep Network Debuggability via Sparse Decision Layers

Improving Deep Network Debuggability via Sparse Decision Layers This repository contains the code for our paper: Leveraging Sparse Linear Layers for D

Madry Lab 26 Dec 20, 2021
PyTorch code for our paper "Image Super-Resolution with Non-Local Sparse Attention" (CVPR2021).

Image Super-Resolution with Non-Local Sparse Attention This repository is for NLSN introduced in the following paper "Image Super-Resolution with Non-

null 89 Jan 11, 2022