Before starting,

download the dataset-->https://www.kaggle.com/datasets/egeyigiterbil/cervicalcancer/data Extract the files to a Directory named CervicalCancer the file structure should look like:

-->CervicalCancer
-------> Herlev
-------> Mendeley
-------> sipakmed
--> train.py
--> dataloader.py
--> ....

To show how we handle data visually:

python dataloader.py --dataset Herlev --classification binary --visualize

To show with augmentations:

python dataloader.py --dataset Herlev --classification binary --augmentation 2 --visualize

We can change classification type too:

python dataloader.py --dataset Herlev --classification multiclass --augmentation 0 --visualize

To train:

python train.py --model_name model1-2-3-4-5-6 --dataset Herlev-Mendeley-sipakmed --classification binary or multiclass --augmentation 0-1-2 --epochs 100 The train script will put the model into a new folder called checkpoints For no augmentations, I use 0 , and for augmentations, use 2

To test with all metrics:

python test.py --model_name model1 --checkpoint ./checkpoints/model1_best.pth --dataset Herlev --classification binary --num_visualize 15 --save_dir ./evaluation_results_binary

model_name gets the model architecture, --checkpoint gets the file name, dataset and classification are same as above, num visualize will show visualized results, and the results will be saved in new folder

multiclass:

python test.py --model_name model1 --checkpoint ./checkpoints/model1_best.pth --dataset Herlev --classification multiclass--num_visualize 15 --save_dir ./evaluation_results_multiclass

To Ensemble:

python test_ensemble.py --models model1 model2 model3 --checkpoints ./checkpoints/model1_best.pth ./checkpoints/model2_best.pth ./checkpoints/model3_best.pth --ensemble_method avg_prob --dataset Herlev --classification binary --num_visualize 15 --save_dir ./evaluation_results_ensemble --num_visualize 10

--models gets the architectures, --checkpoints are same as above, but they have to be as same order as --models, we have 3 ensembling techniques, avg_prob, max_prob, majority_vote rest are same