NewStructure Classifier for Image Deep Learning

This is image classifier using PyTorch. NewStructure Classifier for Image Deep Learning can use image deep learning.

Features

• Image dataloader is image folder structure in torchvision

• image transforms use torchvision

• Early stopping control using only validation loss

• batch balanced each classes control

• Result metrics:

• Overall Accuracy

• Class wise Accuracy

• AUC Score / ROC Curve graph and saving image file

• Optimal threshold for Maximum Sensitivity and Maximum 1-Specificity and saving confusion matrix

• Confusion Matrix

• Delong Test(pROC) for two learning weight models

• Learning log: tensorboard, text file(learning log and result log in Result folder)

Main information

• PyTorch version: 1.3 or above(except for default_classifier_transformer.py: 1.7)
• Torchvision: 0.4.2
• Python: 3.6.9
• This repository is separated binary and multiclass test result code

Folder Structure

+-- root
|   +-- train
|       +-- class1
|           +-- img1.jpg
|           +-- img2.jpg
|           +-- img3.jpg
|       +-- class2
|       +-- class3
|   +-- test
|       +-- class1
|       +-- class2
|       +-- class3
|   +-- val
|       +-- class1
|       +-- class2
|       +-- class3

Reference

• Image transform: https://github.com/lucidrains/vit-pytorch