PaTz β€” Physics-Aware Transformer with eXplainability
● API Offlinepatz_env

Live Training Dashboard

RECONNECTING

Real-time training metrics and logs via WebSocket /ws/training. All metrics parsed from the actual loguru training output.

Status
IDLE
Epoch
β€”
Best val/Fmax
β€”
Last Total Loss
β€”
GPU Memory
β€”
ETA
β€”
Training Control
Runs train_patz.py via conda patz_env
Hyperparameters
Effective batch = 8 Γ— 4 = 32
Total Loss per Epoch
No training data yet
Component Losses (score / func / phys)
No training data yet
Validation Metrics (Fmax & AUROC)
No training data yet
Physics Energies (E_bond / E_angle / E_clash)
No training data yet
Training Log0 lines
Connecting to WebSocket…
Loss Formula
L(ΞΈ) = L_score + Ξ»_phys Β· E_phys + Ξ»_func Β· L_func
E_phys = w_bondΒ·E_bond + w_angleΒ·E_angle + w_clashΒ·E_clash
Score matching loss L_score = E[||s_ΞΈ(x_t,t,c) + Ξ΅/Οƒ(t)||Β²] Β· Functional loss L_func = BCE over 354 GO terms (weighted)