A hierarchical 3-stage diagnosis tool for Keras/TensorFlow models. Powered by SHAP explainability. Published at ICSE 2025.
ICSE 2025 · Dalhousie University
From detection to root cause in one click
Binary fault detection using training dynamics and structural features
Classify faults into 5 categories: activation, layer, loss, optimizer, hyperparameter
SHAP-based feature attribution pinpoints the exact code-level issue
Drop in your Keras model definition
Set epochs, batch size, and data options
One-click training with live monitoring
Fault detection, categorization, and SHAP explanations
Watch loss and accuracy in real-time as your model trains
Understand exactly which features drive fault predictions
Activation, layer, loss function, optimizer, and hyperparameter faults
Upload .py files or paste code directly in the Monaco editor
Save, load, and export your diagnosis sessions
Quick diagnosis with generated dummy data for rapid prototyping