Initial Release V0.1.0

Release Notes — v0.1.0

🟢 Initial Release — brassflows algorithm

This release marks the first functional version of the project — an end-to-end pipeline for collecting real-time visual key input, training CNNs, and running live inference using trained models.

---

✅ Included in This Release

Model Architectures

• ModelV1.1: Baseline convolutional architecture. Served as the starting point.
• ModelV1.5: Enhanced and optimized version with deeper layers, better accuracy, and improved training stability.
• ModelV2: An experimental model with a different structure. Currently underperforming and not recommended for use.

🖼️ Data Collection Pipeline

• Real-time screenshot capture using bounding boxes.
• Keypress detection (j, k, l) to automatically label and save training data.
• Organized into per-class folders for easy training.

Inference Mode

• Live classification using the trained CNN.

• Controlled via keyboard:

  • Enter to activate
  • Esc to stop

• Outputs prediction probabilities and label mappings.

• Release model weights available as latest-stable_v010.pth located in saves.

• Stable release architecture: models/modelv1.py

📊 Model Performance (Across Trials)

Trial	ModelV1.1	ModelV1.5	ModelV2
Train/Test 1	76.63%	97.98%	14.58%
Train/Test 2	86.21%	95.76%	51.23%
Train/Test 3	94.53%	100.0%	43.12%
Latest	❌	✅	❌

---

🚧 Known Limitations

• ModelV2 needs architectural refinement — too unstable to use reliably.
• Bounding box is hardcoded — ensure your screen layout matches or modify the code.
• No GUI — command-line driven only.

---

🔍 Goals of This Release

This version serves as a foundation to:

• Explore CNN behavior across varying data and architectures.
• Build a real-world, interactive AI pipeline.
• Lay the groundwork for improving accuracy, generalization, and future expansion.

---