Stanford Cars - Image Classification using Deep Learning

I am currently working on this project and best so far I got 84.2% 89% accuracy using ResNet50 with PyTorch & FastAI library. This project helped me to understand following things while experimenting around them-

(1) 'fit_one_cycle' (based on Leslie Smith's 1-cycle policy) helps in getting better results than 'fit' method.

(2) 'Differential Learning Rates' (i.e. lower for initial layers and relatively higher for final layers ) give better results than 'Constant Learning Rate'.

Update : By just training my model on realtively higher reslution images cross-validation accuracy improved by ~5% i.e. 89%;

Exposure :

• Explore Cyclical Learning Rates, Differential Learning Rates and Transfer Learning concepts.

• Tech Stack : FastAI [deep learning library built on top of Pytorch]