res_swin's Introduction

Res-Swin: Effective Combination of ResNet and Swin Transformer for LDCT Denoising

This is my master thesis project in the Leiden University

Overview

models folder: Include all the models of experiments, among them, RED-CNN and TransUNet refer from other repositories
warmup_scheduler folder: Implement schedueler with warmup period (reference)
main.py: File with mutiple setting parameters, and can be used for training with different dataset and visualize saved results
train.py: Overall training pipline, including creating dataset, training process and testing process.
measure.py: Measurement function of PSNR and SSIM (reference)
visualize.py: Visualization of images gotten from models to see more details and measure the performance.
Running with default settings (need to change the values of path variables):

python main.py

Requirements

The file requirements imports all the required libraries, but there are also some libraries that are not related to this project. You can install them by:

pip install -r requirements.txt

Note: There are some main essential settings for this project:

CUDA version: 11.6
python: 3.7.11
torch: 1.10.2
torchvision: 0.11.3
cudatoolkit: 11.3.1
cudnn: 8.2.1
numpy: 1.21.2
albumentations: 1.1.0
segmentation-models-pytorch

Extension

Replace ResNet by EfficientNet or other more effective CNN models (for better performance)
Replace Swin by Swinv2 or other more effective transformer models (for better performance)
Used for other similar tasks, like predicting biomass of forest, semantic segmentation (still need trying)
Optimize mix block further in terms of attention mechanism and learnable parameter

Recommend Projects