Spaces:
Running
Running
Commit History
-
fix: clean all non-ASCII chars, fix CI (lint + tests), secure API keys
Krishna Venkatesh- Remove all emojis, em/en dashes, Unicode arrows from source code, HTML, JS, docs - Fix test_llm_no_key_returns_error by mocking os.getenv - Move Roboflow API key from hardcoded to env var - Fix flake8: unused imports, bare except, f-string issues, line length - Run black + isort across entire codebase - Exclude data/training from flake8, add E402 to extend-ignore - All 29 tests passing, black/isort/flake8 cleanc81c805
-
feat: implement four-pipeline ensemble (CV + Roboflow + GPT-4o + finetuned YOLOv8)
Krishna Venkatesh- Add pipeline_roboflow.py: Roboflow pretrained YOLOv8 integration - Add pipeline_llm.py: GPT-4o Vision API integration - Add pipeline_finetuned.py: Local finetuned YOLOv8 with fallback - Add ensemble.py: Weighted voting (A:1.0, B:2.5, C:4.0, D:3.0) with agreement bonuses - Add pipeline_ensemble.py: Parallel thread orchestration with progressive updates - Add /upload_ensemble endpoint in Flask app - Prepare Angelina ICCV 2021 dataset (290 images, 85,263 annotations) - Add YOLOv8 training pipeline: convert_to_yolo.py, train_yolo.py - Create dataset.yaml for YOLOv8 training (203/43/44 split) - Start YOLOv8 training on 290 real Braille images - Update README for v2 four-pipeline architecture - Add DEVPOST_V2_GUIDE.md with 4 hackathon submission templates - All 29 tests passing (5 ensemble + 4 LLM + 20 core) - Ensemble works immediately with base model, improves when finetuned weights ready Training in progress: YOLOv8n on Apple M4 (~2-3 hours), auto-loads when complete.8df3deb