Vehicle Make and Model Recognition with Fine-Tuned AlexNet

This project explores the effectiveness of fine-tuning AlexNet for vehicle make and model recognition using the Stanford Cars dataset. All training and evaluations were performed using project_colab.ipynb on Google Colab with GPU acceleration (A100). The final report is included in report.pdf.

📁 Project Structure

dataset/
  train/                         # Training images pre-split
  valid/                         # Validation images pre-split
models/
  best_alexnet.pkl              # Exported best-performing AlexNet model
visualizations/
  confusion_matrix.png                   # Raw confusion matrix (unnormalized)
  confusion_matrix_normalized.png        # Normalized confusion matrix
  lr_find_plot.png                       # Learning rate finder curve
  lr_find_plot_labeled.png               # LR curve with valley marked
  sample_batch.png                       # Sample training batch shown before training
  sample_predictions.png                 # Grid of sample predictions (correct/incorrect)
  top_losses.png                         # Top high-loss predictions
  top_losses_examples.png                # Broader view of most confused predictions
  top_losses_per_class.png               # Misclassifications per vehicle class
  train_loss.png                         # Plot of training loss only
  train_vs_val_loss.png                  # Training vs validation loss curves
  val_accuracy_over_epochs.png           # Validation accuracy over training epochs
classification_report.csv              # Precision, recall, F1-score per class
classwise_accuracy.csv                 # Validation accuracy per class
confusion_matrix_data.csv              # Numerical confusion matrix values
metrics_over_epochs.csv                # Epoch-wise training/validation loss and accuracy

project.ipynb                  # Clean version for final evaluation, uses relative paths
project_colab.ipynb            # Google Colab training notebook, uses Drive paths
README.md                      # Project documentation (this file)
requirements.txt               # Package dependencies
report.pdf                     # Final 8-page project report (IEEE TPAMI format)

🧠 Notebooks

• project.ipynb: Final version for grading and code review. Uses relative paths so it works in both local and Colab environments (if files are uploaded manually).
• project_colab.ipynb: Primary notebook used for training with Google Drive. Requires the dataset to be located at drive/MyDrive/VehicleModelProject/dataset/.

📊 Evaluation Artifacts

All training diagnostics, loss curves, confusion matrices, and prediction analyses are saved in the visualizations/ folder. These are referenced throughout the report and provide insight into the model’s learning behavior, error patterns, and classification performance.

📄 Report

• report.pdf: Full technical report formatted in IEEE TPAMI style. Includes background, methodology, results, analysis, ablation studies, limitations, and future directions.