Skip to content

ctde_v0.run_warmab

run_warmab

Warm-start A/B — from-scratch vs --init-from a checkpoint, side by side.

Runs at 16²/4 (small enough to fit alongside a 32² job). Warm arm inits actor+critic from --ckpt (the LPAC backbone is scale-invariant, so a 32² checkpoint loads into a 16² run). Same best architecture both arms; the only difference is the init. Reports coverage/connectivity so we can see whether warm-start converges higher/faster than current-code-from-scratch.

PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python         ctde_v0/run_warmab.py --ckpt runs/slamattn/seed0/max_slam/model.eqx --seeds 2 --jobs 1