A-AIO-process-optimization-and-training
Testing/prototyping repository for AOI/AI concepts, focused on reproducible pipelines and practical model iteration.
Quick Links
- Main project:
Simple-Sim - Documentation index:
Simple-Sim/docs/INDEX.md - Latest arena report:
Simple-Sim/ARENA_REPORT.md - Training logs index:
Simple-Sim/docs/training_logs/INDEX.md - Sample gallery:
Simple-Sim/SAMPLE_GALLERY.md
Contents
Current Status
Last updated: 2026-02-20
| Topic | Status | Link |
|---|---|---|
| Arena leaderboard | Current top ranking based on latest QFN crawler fine-tune boost | Simple-Sim/ARENA_REPORT.md |
| Active research | Extended image filter robustness + QFN-focused fine-tuning | Simple-Sim/docs/training_logs/INDEX.md |
| Production recommendation | Prefer single cross-profile models for robustness/simplicity | Final Verdict |
Projects
| Project | Purpose | Start Here |
|---|---|---|
Simple-Sim |
Sandbox for synthetic data generation, training, evaluation, and model selection via arena workflows | Simple-Sim/README.md |
Arena
- Latest stats:
Simple-Sim/ARENA_REPORT.md - Historical arena reports:
Simple-Sim/docs/arena-reports-historie/INDEX.md
Research Updates
Current Focus (2026-02-20)
QFN crawler fine-tune boost with additional clean no-image-filter samples (QFN32-2D +50, QFN-3D +50) reached the current top arena ranking.
Log: 2026-02-20-Training-QFN-Crawler-NoImageFilter-Boost.md
Recent Logs
- Fine-tuned comparison (default-filter vs QFN-focused):
2026-02-20-Training-Comparison-FineTuned-vs-QFN-FineTuned.md - Extended filter restart (living log):
2026-02-17-Training-Restart-Extended-Image-Filters.md - 90° random orientation + SOIC16 restart:
2026-02-16-Training-Restart-90deg-SOIC16.md - Filter settings reference:
FILTER_SETTINGS.md
Model Strategy
Current conclusions: Final Verdict, Training Conclusions
Bundled Models (Multi-Profile)
- Require an upstream component/object classifier for routing.
- Higher per-component peak accuracy.
- Larger footprint (multiple sub-models).
Single Models (Cross-Profile)
- Self-contained (no external classifier).
- Trained across all profiles and defect types.
- Smaller footprint.
- Slightly lower per-component peak accuracy, but stronger on randomized mixed/unknown component recognition.
Recommendation: use single cross-profile models for deployment when robustness and simplicity are prioritized.
License and Usage
This repository is proprietary and not open-source.
- Full license:
LICENSE - Third-party dependency licenses:
Simple-Sim/THIRD_PARTY_LICENSES.md
Allowed
- Read code and documentation.
- Discuss concepts and provide feedback.
- Link to this repository.
Not Allowed (without prior written permission)
- Use this code (whole or partial) in projects/products/services.
- Extract or reuse modules, files, snippets, or components.
- Copy, modify, distribute, sublicense, or republish.
- Use for commercial or production deployments.
Note: public visibility on GitHub does not grant reuse rights beyond viewing the repository.
Permission requests: arn-c0de@protonmail.com
See also: CONTRIBUTING.md