ML Project

Overview

A multi-class scene classification system using EfficientNetV2-S with two-phase transfer learning from ImageNet pre-trained weights. The model classifies 224×224 RGB images into 6 scene categories: Buildings, Forest, Glacier, Mountain, Sea, and Street.

Training follows a two-phase strategy: Phase 1 freezes the backbone and trains only the custom classification head for 5 epochs (lr=1e-3), then Phase 2 unfreezes the entire network for fine-tuning over 15 epochs (lr=1e-4). Training uses AdamW optimizer, CosineAnnealingLR scheduling, mixed precision, and label smoothing (0.1) for regularization.

Achieves 95% top-1 accuracy with ~20.3M parameters. Grad-CAM visualizations are included to provide interpretability — highlighting the image regions the model attends to when making predictions.