Image Classification: FGVC-Aircraft Benchmark

This directory contains a strictly controlled benchmarking suite for CNN (ResNet-50) and ViT (ViT-B/16) architectures on the FGVC-Aircraft dataset.

Setup

1. Environment: Install dependencies: torch, torchvision, timm, umap-learn, squarify, seaborn, opencv-python.
2. Data Preparation:
   • Automatic: Run any script (e.g., python3 eda.py). The torchvision library will automatically download and extract the FGVC-Aircraft dataset to the specified --data_dir (default: ./data).
   • Manual: If automatic download fails, download the dataset from the official VGG page and extract it. Ensure the final structure is:

     data/fgvc-aircraft-2013b/A
         data/
             images/
             variants.txt
             ...
     

Component Overview

• dataset.py: Data loaders with multi-level augmentation (Light vs. RandAugment).
• models.py: Architecture definitions and Layer-wise Learning Rate Decay (LLRD).
• train.py: Core training engine with tracking, early stopping, and efficiency metrics.
• eda.py: Dataset exploration (class distribution, dimensions, UMAP feature projection).
• error_analysis.py: Confusion matrix, misclassification grids, and per-class accuracy.
• gradcam.py: Interpretability using Grad-CAM (CNNs) and Attention Rollout (ViTs).
• plot_results.py: Generates comparative learning curves and final accuracy charts from logs.

Execution

• Run EDA: python3 eda.py
• Orchestrate Grid Search: bash orchestrate_experiments.sh (Runs 30 experiments across strategies/seeds).
• Architecture Showdown: bash fair_arch_compare.sh (Controlled LR and long-budget comparison).
• Analyze Errors: python3 error_analysis.py --model <model> --checkpoint <path>

All outputs and logs are saved to logs/, checkpoints/ respectively.