Explainability Methods Comparison on CUB-200-2011

This project compares different explainability approaches from the Captum library for a CNN classifier trained on the CUB-200-2011 bird species dataset. The focus is on measuring explanation plausibility by comparing model explanations with ground-truth part annotations.

Project Goals

1. Train a CNN classifier on CUB-200-2011 (200 bird species)
2. Apply multiple explainability methods: Integrated Gradients, Saliency, Input×Gradient, LIME, Kernel SHAP, Occlusion, Noise Tunnel (wrapper)
3. Evaluate explanation quality (plausibility) using ground-truth part annotations
4. Analyze: Are the results due to the model or to the explanation method?

Project Structure

Explainability-project/
├── src/                       # Python package
│   ├── __init__.py
│   ├── config.py              # Configuration and hyperparameters
│   ├── data_loader.py         # Dataset loading and preprocessing
│   ├── model.py               # CNN model definition (ResNet-50)
│   ├── explainability.py      # Explainability methods (Captum)
│   ├── evaluation.py          # Plausibility metrics
│   └── visualizations_utils.py # Visualization utilities
├── train.py                   # Training script
├── model_evaluation.py        # Model evaluation script
├── explainability_analysis.ipynb  # Interactive Jupyter notebook
├── requirements.txt           # Python dependencies
├── README.md                  # This file
├── models/                    # Saved model checkpoints
├── results/                   # Evaluation results and metrics
├── visualizations/            # Generated visualizations
├──CUB_200_2011/
  └── CUB_200_2011/
      └── ...

Quick Start

1. Create Virtual Environment

python -m venv .venv
source .venv/bin/activate  # macOS/Linux
# .venv\Scripts\activate   # Windows

2. Install Dependencies

pip install -r requirements.txt

3. Prepare the Dataset

The CUB-200-2011 dataset should be in the CUB_200_2011/ directory with the following structure:

• images/ - Bird images organized by species
• parts/ - Part location annotations
• attributes/ - Attribute annotations
• Various metadata files (images.txt, classes.txt, etc.)

Download from: https://www.vision.caltech.edu/datasets/cub_200_2011

4. Download Pre-trained Model

The trained ResNet-50 model trained by us is available via Git LFS (Large File Storage). To download it:

# Install Git LFS if not already installed
# macOS
brew install git-lfs

# Ubuntu/Debian
sudo apt-get install git-lfs

# Windows (with Git for Windows)
# Git LFS is included by default

# Initialize Git LFS
git lfs install

# Pull the model file
git lfs pull

The model will be downloaded to models/best_resnet50_cub200.pth.

Note: If you cloned the repository without Git LFS installed, the model file will be a pointer file. Run git lfs pull after installing Git LFS to download the actual model.

5. Train the Model (Optional)

If you want to train the model from scratch instead of using the pre-trained one:

python train.py --epochs 30 --batch_size 32 --lr 0.001

Options:

Check 'train.py'

6. Run Model Evaluation

python model_evaluation.py

This script evaluates the trained model on the test set and generates:

• Classification metrics (accuracy, precision, recall, F1-score)
• Confusion matrix
• Per-class performance analysis

Results are saved in the results/ directory.

7. Interactive Analysis (MAIN NOTEBOOK)

Open the Jupyter notebook for interactive exploration:

jupyter notebook explainability_analysis.ipynb

Technical Details

Hardware

The training and intensive explainability computations were performed using:

• GPU: NVIDIA L4
• Cloud Platform: Lightning.ai

References

• Captum: https://captum.ai - PyTorch model interpretability library
• CUB-200-2011: https://www.vision.caltech.edu/datasets/cub_200_2011 - Fine-grained bird classification dataset

Papers:

[1] J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, "Sanity Checks for Saliency Maps," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2018.

[2] S. M. Lundberg and S.-I. Lee, "A Unified Approach to Interpreting Model Predictions," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017.

[3] M. T. Ribeiro, S. Singh, and C. Guestrin, "'Why Should I Trust You?': Explaining the Predictions of Any Classifier," in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discovery and Data Mining (KDD), 2016.

[4] K. Simonyan, A. Vedaldi, and A. Zisserman, "Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps," in Proc. ICLR Workshop, 2014.

[5] M. Sundararajan, A. Taly, and Q. Yan, "Axiomatic Attribution for Deep Networks," in Proc. 34th Int. Conf. Machine Learning (ICML), 2017.

[6] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, "The Caltech-UCSD Birds-200-2011 Dataset," California Institute of Technology, Tech. Rep. CNS-TR-2011-001, 2011.

[7] J. Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2022.

Authors

This project was developed by:

• @filippostanghellini
• @samueleviola