DIP-Project

Directory Structure

• DRAC2022_TaskB code used for our backbone selection experiments
• new code used for our advanced experiments on ensemble learning, dataset amplification, pseudo-labeling, etc.
• new2 code with no k-fold
• new2_kfold code with k-fold
• new3 prototye code(a new structure with 2 levels of resnet)
• README.md
• environment.yml

Contents

• Clone The Repo
• Download Dataset
• Environment
• Usage

Clone The Repo

Using the following command line can clone this repo into your machine.

git clone https://github.com/3116218257/DIP-Project.git
cd DIP-Project

Then create an empty directory for the dataset.

mkdir data
cd data

Download Dataset

You need to download the Task B dataset, you can get it from https://data.mendeley.com/datasets/s8kbw25s3x/1. If you have downloaded the zipped file, use this command.

unzip 'B. Image Quality Assessment.zip'
rm -rf 'B. Image Quality Assessment.zip'

Environment

Just create a virtual environment for our project using command line.

conda env create -f environment.yml

If there still are some missing package, download manually the packages in the environment.yml.

Usage

There are two sections, the DARC2022_TaskB directory contains code for our backbone selection experiments, the new directory contains code for our advanced experiments, you can play with both of them.

Reminder

Please change the dataset directory in DRAC2022_TaskB/dataset.py and DRAC2022_TaskB/train.py before executing.

Backbone Selection

You can choose different backbones in this section. Please refer to the configuration settings in train.py and change the args.sh file to use a new backbone.
Then simply run train.py using command.

sh args.sh.

Finally run test.py to get your needed .csv file.

python test.py

Advanced Experiments

The series of new directories contains different version of code that we used throughout our whole experiment process, we suggest you use the code in new, which we used in our final experiments. The process of execution is similar with the operations above, use train.sh and test.sh to get the final .csv file. For configuration settings, please refer to config.py.