flow-hpc-shocktrack

GPU-accelerated, fault-tolerant Schlieren/PIV shock tracking with interactive ROI, 1-px edges, and resumable training.

End-to-end pipeline to extract shock angle vs time and dominant oscillation frequency from Schlieren frames (PIV-friendly too). Designed for small datasets and Spyder. PyTorch is optional.

Features

• Interactive ROI (OpenCV; Matplotlib fallback)
• 1‑px thin edges (skeletonization) + sub‑pixel cv2.fitLine angle
• (Optional) tiny U‑Net: synthetic pretrain + fine‑tune
• Resumable training (epoch checkpoints, PAUSE file or Ctrl+C)
• Outputs: overlays, time series, FFT/Strouhal, JSON summary

Quickstart

1. pip install -r requirements.txt
2. Unzip demo_schlieren_dataset.zip → set CONFIG["data_dir"] to demo_schlieren/images/ (and mask_dir to demo_schlieren/masks/).
3. Run python shocktrack_roi.py (or in Spyder).
   ROI keys: click points → c close → y confirm (u=undo, r=reset, q=cancel).
4. Results in runs/exp_roi/ → pred/ overlays, analysis/*.png, analysis/summary.json.

Pause / Resume

• Create empty PAUSE file in the training dir (e.g., runs/exp_roi/finetune/PAUSE) to stop after current epoch.
• Or press Ctrl+C; a checkpoint is saved.
• Resume: set CONFIG["resume_finetune_dir"] (or pretrain) to that folder and run again.

Files

• shocktrack_roi.py — main script (Spyder-friendly)
• requirements.txt — core deps (no PyTorch)
• requirements-torch.txt — optional torch deps
• demo_schlieren_dataset.zip — tiny synthetic dataset (30 frames with masks)
• LICENSE (MIT), .gitignore, .gitattributes