You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

110 lines
4.9 KiB

from sb3_contrib import MaskablePPO
from sb3_contrib.common.maskable.callbacks import MaskableEvalCallback
from sb3_contrib.common.wrappers import ActionMasker
from stable_baselines3.common.logger import Logger, CSVOutputFormat, TensorBoardOutputFormat, HumanOutputFormat
import gymnasium as gym
from minigrid.core.actions import Actions
from minigrid.wrappers import RGBImgObsWrapper, ImgObsWrapper
import time
from utils import MiniGridShieldHandler, create_log_dir, ShieldingConfig, MiniWrapper, expname, shield_needed, shielded_evaluation
from sb3utils import MiniGridSbShieldingWrapper, parse_sb3_arguments, ImageRecorderCallback, InfoCallback
from stable_baselines3.common.callbacks import EvalCallback
import os, sys
from copy import deepcopy
GRID_TO_PRISM_BINARY=os.getenv("M2P_BINARY")
def mask_fn(env: gym.Env):
return env.create_action_mask()
def nomask_fn(env: gym.Env):
return [1.0] * 7
def main():
args = parse_sb3_arguments()
formula = args.formula
shield_value = args.shield_value
shield_comparison = args.shield_comparison
log_dir = create_log_dir(args)
#new_logger = Logger(log_dir, output_formats=[CSVOutputFormat(os.path.join(log_dir, f"progress_{expname(args)}.csv")), TensorBoardOutputFormat(log_dir)])
new_logger = Logger(log_dir, output_formats=[CSVOutputFormat(os.path.join(log_dir, f"progress_{expname(args)}.csv")), TensorBoardOutputFormat(log_dir), HumanOutputFormat(sys.stdout)])
if shield_needed(args.shielding):
shield_handler = MiniGridShieldHandler(GRID_TO_PRISM_BINARY, args.grid_file, args.prism_output_file, formula, shield_value=shield_value, shield_comparison=shield_comparison, nocleanup=args.nocleanup)
env = gym.make(args.env, render_mode="rgb_array")
env = RGBImgObsWrapper(env)
env = ImgObsWrapper(env)
env = MiniWrapper(env)
eval_env = deepcopy(env)
eval_env.disable_random_start()
if args.shielding == ShieldingConfig.Full:
env = MiniGridSbShieldingWrapper(env, shield_handler=shield_handler, create_shield_at_reset=False)
env = ActionMasker(env, mask_fn)
eval_env = MiniGridSbShieldingWrapper(eval_env, shield_handler=shield_handler, create_shield_at_reset=False)
eval_env = ActionMasker(eval_env, mask_fn)
elif args.shielding == ShieldingConfig.Training:
env = MiniGridSbShieldingWrapper(env, shield_handler=shield_handler, create_shield_at_reset=False)
env = ActionMasker(env, mask_fn)
eval_env = ActionMasker(eval_env, nomask_fn)
elif args.shielding == ShieldingConfig.Evaluation:
env = ActionMasker(env, nomask_fn)
eval_env = MiniGridSbShieldingWrapper(eval_env, shield_handler=shield_handler, create_shield_at_reset=False)
eval_env = ActionMasker(eval_env, mask_fn)
elif args.shielding == ShieldingConfig.Disabled:
env = ActionMasker(env, nomask_fn)
eval_env = ActionMasker(eval_env, nomask_fn)
else:
assert(False) # TODO Do something proper
model = MaskablePPO("CnnPolicy", env, verbose=1, tensorboard_log=log_dir, device="auto")
model.set_logger(new_logger)
steps = args.steps
# Evaluation
eval_freq=max(500, int(args.steps/30))
n_eval_episodes=5
render_freq = eval_freq
if shielded_evaluation(args.shielding):
from sb3_contrib.common.maskable.evaluation import evaluate_policy
evalCallback = MaskableEvalCallback(eval_env, best_model_save_path=log_dir,
log_path=log_dir, eval_freq=eval_freq,
deterministic=True, render=False, n_eval_episodes=n_eval_episodes)
imageAndVideoCallback = ImageRecorderCallback(eval_env, render_freq, n_eval_episodes=1, evaluation_method=evaluate_policy, log_dir=log_dir, deterministic=True, verbose=0)
else:
from stable_baselines3.common.evaluation import evaluate_policy
evalCallback = EvalCallback(eval_env, best_model_save_path=log_dir,
log_path=log_dir, eval_freq=eval_freq,
deterministic=True, render=False, n_eval_episodes=n_eval_episodes)
imageAndVideoCallback = ImageRecorderCallback(eval_env, render_freq, n_eval_episodes=1, evaluation_method=evaluate_policy, log_dir=log_dir, deterministic=True, verbose=0)
model.learn(steps,callback=[imageAndVideoCallback, InfoCallback()])
#vec_env = model.get_env()
#obs = vec_env.reset()
#terminated = truncated = False
#while not terminated and not truncated:
# action_masks = None
# action, _states = model.predict(obs, action_masks=action_masks)
# print(action)
# obs, reward, terminated, truncated, info = env.step(action)
# # action, _states = model.predict(obs, deterministic=True)
# # obs, rewards, dones, info = vec_env.step(action)
# vec_env.render("human")
# time.sleep(0.2)
if __name__ == '__main__':
main()