This repository contains codes, scripts, and data used for team WillCode4EXP submission to AIAT Hackathon competing on classification of X-ray for Covid/Normal/Pneumonia classes. The method can be roughly described as modifying UNET+Attention model to classify visualize decision made with x-ray region.

These code are intended for educational purpose. UNET is modifid for the Hackathon from other repository and presented for people who what to study on how my team modified UNet for classification.

uNetImplement.py and uNetParts.py for UNET architecture are not entirely mine. These are codes modified from a generous repository milesial/Pytorch-UNet/ who show me how to implement UNet

visit the link below for more info https://github.com/milesial/Pytorch-UNet/tree/master/unet

For any people who learn from my codes, please also give credit to the original repository when the credit is due.

The codes were developed using Pytorch with following packages

numpy scipy pandas scikit-learn pillow flask opencv matplotlib scikit-image torchvision natsort

For installation using anaconda

conda install numpy scipy
conda install pandas scikit-learn
conda install -c pillow flask
conda install -c conda-forge opencv matplotlib scikit-image

competeRunUnet_FullAttention.py is main script for training validating and testing run. There is option to set number of epoch, load existing model, generate results .csv file, and etc.

competeRunUnet_FullAttention_getAttentionMap.py is a side script to extract attention maps and save into numpy array (.npy) for later visualization using Opencv.

uNetImplement.py and uNetParts.py are network architecture implementation files.

opencvProcess.py is script for visualization with attention map.

ForthTrialSub.csv is submission used in the competition.

compete__ are folders contain data for training, validation, and testing. The data in these folder are organized with subfolder indicating class label.

compete__All are data folders without any class label. The unorganized folders are easier to load in some scripts.

All the data has been manually crop to focus more on chest area.

Generating submission .csv results from test images (CompeteTest folder).

python competeRunUnet_FullAttention.py

competeRunUnet_FullAttention.py is automatically set to validate run be default, which mean there will be no training during the run. Many part of codes has option to run with GPU/CPU version of Pytorch. Feel free to select the modes as appropriated

This is my first repository that I able to publicly release. Feel free to email me for questions/comments/suggestions at [email protected] or [email protected]. Pull request also appreciated.