#!/usr/bin/env bash # The overnight version of next_run.sh: same zero-thought kickoff (auto-names # the run, resumes from the newest committed final.zip, same standing flags), # but sized for a ~12-hour run instead of ~2.5 hours. # # --timesteps 90000000: runs 05-07 all trained at ~7.5M steps/hour on the # box (measured from the tfevents wall times), so 90M new steps ~= 12h. # On resume SB3 treats the budget as additional steps, same as the 20M runs. # --checkpoint-every 500000: at the default 100k cadence a 90M run would # commit ~900 checkpoint zips (~190MB); 500k keeps it to ~180 (~38MB), # about the same repo growth as a normal run. # # Usage: ./next_long_run.sh [extra train.py args...] # Extras still win on conflict, e.g. ./next_long_run.sh --timesteps 60000000 # for a shorter overnight run. set -euo pipefail exec "$(dirname "$0")/next_run.sh" --timesteps 90000000 --checkpoint-every 500000 "$@"