|
|
#!/usr/bin/env bash |
|
|
set -euo pipefail |
|
|
|
|
|
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")"/.. && pwd)" |
|
|
OUT_DIR="$ROOT_DIR/ckpts/rl" |
|
|
TIMESTEPS=${TIMESTEPS:-200000} |
|
|
ENV_ID=${ENV_ID:-CartPole-v1} |
|
|
|
|
|
mkdir -p "$OUT_DIR" |
|
|
|
|
|
python - "$OUT_DIR" "$TIMESTEPS" "$ENV_ID" <<'PY' |
|
|
import sys |
|
|
from stable_baselines3 import PPO |
|
|
import gymnasium as gym |
|
|
|
|
|
out_dir, total_steps, env_id = sys.argv[1], int(sys.argv[2]), sys.argv[3] |
|
|
|
|
|
env = gym.make(env_id) |
|
|
model = PPO("MlpPolicy", env, n_steps=2048, batch_size=64, learning_rate=3e-4, verbose=1, tensorboard_log=out_dir) |
|
|
model.learn(total_timesteps=total_steps) |
|
|
model.save(f"{out_dir}/policy") |
|
|
print("RL adapter saved") |
|
|
PY |
|
|
|
|
|
echo "RL adapter complete → $OUT_DIR/policy.zip" |
