Optimizing CNN Models for COVID-19 Detection: A Comparative Analysis of Optimizers and Loss Functions
Course project for CSC413: Neural Networks and Deep Learning
Our report can be found in report.pdf
On around 28000 training data, performed train-validation split of 80%-20%
Train the model on train set, use the model with highest validation accuracy among all epochs on test set to avoid overfitting
All evaluation metrics reported based on an unseen test dataset: 200 negative cases, 200 positive cases
Best Optimizer and Loss function: RMSprop + Cross Entropy
Accuracy: 98.0%, Sensitivity: 96.0%, Specificity: 100.0%, F-1 Score: 96.6%
Run the model by executing the file resnet152/resnet152.ipynb
Best Optimizer and Loss function: AdamW + Cross Entropy
Accuracy: 89.5%, Sensitivity: 82.5%, Specificity: 96.5%
Run the model by executing the file small_covid_net.ipynb, preferably on Kaggle
Best Optimizer and Loss function: SGD (Stochastic Gradient Descent) + Cross Entropy
Accuracy: 89.0%, Sensitivity: 82.0%, Specificity: 96.0%
Run the model by executing the files shallowCNN/train_ce_nll.py and shallowCNN/train_mse.py