mirror of
https://github.com/jcreek/CosmicClash.git
synced 2026-09-10 16:04:04 +00:00
feat(*): Add long training run script
This commit is contained in:
Executable
+17
@@ -0,0 +1,17 @@
|
|||||||
|
#!/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 "$0")/next_run.sh" --timesteps 90000000 --checkpoint-every 500000 "$@"
|
||||||
Reference in New Issue
Block a user