Spectrum xAI is a comprehensive pipeline for explainability in artificial intelligence. It provides a suite of state-of-the-art explainability methods to analyze and interpret deep learning models. The pipeline supports evaluation metrics such as IoU, F1-Score, AUC, and execution time, along with visualizations for better understanding of model predictions.

• Explainability Methods:

  • Grad-CAM
  • Gradient SHAP
  • Integrated Gradients
  • LIME
  • Saliency Map
  • Score-CAM

• Evaluation Metrics:

  • Intersection over Union (IoU)
  • F1-Score
  • Area Under the Curve (AUC)
  • Execution Time

• Supported Model:

  • DenseNet (fully implemented)

• Visualization:

  • Heatmaps and thresholded masks for each explainability method.
  • Ground truth comparison for qualitative evaluation.

1. Clone the repository:

   git clone https://github.com/IsmailHatim/Spectrum cd Spectrum

2. Install the required dependencies:

   pip install -r requirements.txt

3. Ensure you have the DAGM dataset in the data/dataset/ directory. You can download it from the Kaggle Challenge

To train a model (e.g., DenseNet121) on the first class of the DAGM dataset:

python run_training.py --model_name densenet121 --epochs 10 --batch_size 32 --learning_rate 0.0001 --plot

To run an explainability method (e.g., Grad-CAM) on a trained model and save figures:

python run_explanation.py --method gradcam --index 0 --threshold 0.5 --model_name densenet121 --conv_layer_index -2 --save

To run the whole evluation using a specific method (e.g., Grad-CAM) on a trained model:

python run_evaluation.py --method gradcam --threshold 0.5 --model_name densenet121 --conv_layer_index -2

1. Data Loading:

   • The pipeline uses the DAGM dataset, which is divided into training and testing splits for each class.

2. Model Training:

   • Train models like DenseNet121, on the dataset.
   • Save trained models for later use.

3. Explainability Methods:

   • Apply explainability methods to generate heatmaps and thresholded masks for model predictions.

4. Evaluation:

   • Evaluate the explainability methods using IoU, F1-Score, AUC, and execution time.

5. Visualization:

   • Visualize the input image, heatmap, thresholded mask, and ground truth for qualitative analysis.

1. Train a DenseNet121 model on Class 1 of the DAGM dataset:

   python run_training.py --model_name densenet121 --epochs 10 --batch_size 32 --learning_rate 0.0001 --plot

2. Run Grad-CAM on the trained model:

   python run_explanation.py --method gradcam --index 0 --threshold 0.5 --model_name densenet121 --conv_layer_index -2 --save

3. Run evluation using Grad-CAM on the whole test set:

   python run_evaluation.py --method gradcam --threshold 0.5 --model_name densenet121 --conv_layer_index -2

4. Visualize the results:

   • Input image
   • Grad-CAM heatmap
   • Thresholded mask
   • Ground truth
   • Mean and Standard Deviation metrics

• Explainability Methods:

  • Extend the pipeline with additional explainability methods and other models.

• Dataset Support:

  • Generalize the pipeline to work with other datasets.

Contributions are welcome! Please fork the repository and submit a pull request.

• PyTorch
• Captum
• LIME
• DAGM Dataset