Frank14f
|
b3ee72e144
|
feat(train): V5 parameterized training pipeline — Karman + Illusion verified
Calibration-driven, no_bias only, 2000x600 grid. All cases share unified
env/train/calibrate pattern. Multi-GPU server deployment ready.
Core additions:
- calibrate.py: Phase 0 calibration (karman/illusion), produces
calibration.json with rounded FORCE_SCALE, SENS_SCALE, SIM_BP/VAL
- env_karman.py: parameterized Karman cloak env (calibration + config_path)
- env_illusion.py: illusion env with FFT harmonics target (S_DIM=14)
- env_vortex.py: vortex cloaking env (lamb/taylor, MAX_STEPS=150)
- train_karman.py, train_illusion.py: parameterized training scripts
- launch_multi.sh: sequential multi-GPU launcher (7-min staggered)
- SERVER_DEPLOY.md: complete server setup, calibration, training guide
- calibrations/re100/ & calibrations/illusion_1L/: pre-run calibrations
Fixes:
- SIM_VAL[-1] 0.95 -> 1.0 (r_sim maps to full [0,1] range)
- Cross-Re configs: re50/200/400 (viscosity-only variants)
Verified end-to-end on GPU0+GPU1:
- Karman V5 20-ep: best reward 0.459 at Ep16 (monotonic rise)
- Illusion 20-ep: best reward 0.224 at Ep19 (harmonics, DTW learning)
Co-authored-by: Cursor <cursoragent@cursor.com>
|
2026-07-01 20:10:27 +08:00 |
|
Frank14f
|
2ae248421d
|
fix(oid): confirm FIFO bias bug has no structural impact
- Fixed bias_arr[4] (front 0), bias_arr[5] (bottom -4U0), bias_arr[6] (top +4U0)
- Re-ran full karman pipeline: force-sig overlap unchanged (-0.034)
- Force-OID still beats POD (0.295 vs 0.068, was 0.750 vs 0.418)
- Absolute R2 shifted because corrected FIFO changed PPO trajectory start
- Structural conclusion (force-sig near-orthogonal) is robust
Co-authored-by: Cursor <cursoragent@cursor.com>
|
2026-06-30 16:26:54 +08:00 |
|