Enhancing KL Severity Grading with Focal Loss Optimization and Interpretability through Grad CAM Analysis

Project Structure

.
├── dataset
│   ├── train
│   │   ├── 0
│   │   ├── 1
│   │   ├── 2
│   │   ├── 3
│   │   └── 4
│   ├── val
│   │   ├── 0
│   │   ├── 1
│   │   ├── 2
│   │   ├── 3
│   │   └── 4
│   └── test
│       ├── 0
│       ├── 1
│       ├── 2
│       ├── 3
│       └── 4
├── models
├── custom_densenets.py
├── custom_resnets.py
├── Grad-CAM.ipynb
├── hyperval.py
├── README.md
├── requirements.txt
├── se_nets.py
├── test.ipynb
├── train.py
└── utils.py

Running the files

Dataset can be downloaded from here.

Training:

Example statement:

python3 train.py -m resnet -d dataset -b 32 -l ce -o test -e 100 --learning_rate 1e-3

Make sure your path to the dataset is correct. Other parameters can be changed. Do refer to the parser arguments for the same.

Computing accuracy on Test set, Grad-CAM

To change any of the parameters, go to the notebook and change the parameters in the parser statement

Hyperparameter Optimization:

Example statement:

python3 hyperval.py -m resnet -o test -e 100 -n 100 -s study_name

Other parameters can be changed. Do refer to the parser arguments for the same.

Results

Model weights can downloaded from here.

Architecture	Accuracy	Precision	Recall	F1
SE-ResNet18 (CE loss)	68.12	63.67	68.12	64.69
SE-DenseNet (CE loss)	69.14	67.21	69.14	67.85
SE-ResNet18 (FL)	67.75	68.56	67.75	67.83
SE-DenseNet (FL)	68.05	70.68	68.05	68.78
SE-ResNet18 (tuned FL)	69.38	67.59	69.38	67.75