Stage I Report:
Planning & Requirements:
- Defined problem: lack of holistic and interpretable testing for image classifiers
- Identified key components:
- Model & dataset understanding
- Multiple testing strategies (random, stress, feature-based, AI-guided)
- Finalized initial tech stack:
- React + Vite (UI), Node.js (API), Python (ML execution), LLM integration
- Selected baseline models (e.g., ResNet, VGG)
Stage II Report:
Prototype Design & Initial Implementation:
- Designed system architecture and workflow
- Implemented:
- Model summary module (LLM-based)
- Basic UI for model and dataset input
- Developed Random Testing module:
- Input label + dataset
- Generated pass/fail results and metrics
- Initial integration between frontend, backend, and Python inference
Stage III Report:
Stress Testing & Analysis:
- Implemented Stress Testing module:
- Background changes
- Environment variations
- Occlusion and multi-object scenarios
- Enabled before vs after comparison of predictions
- Observations:
- Significant drop in performance under transformations
- Confidence scores not reliable for robustness
- Added basic reporting for test results and failure counts
Stage IV Report:
Current Progress & Ongoing Work:
- Currently working on Feature based Testing:
- Generating visual features using LLM
- Designing targeted test cases based on features
- Improving:
- Failure analysis using LLM
- Final report generation (summary + insights)
- Next steps:
- Complete feature testing pipeline
- Integrate all modules into unified reporting