train_ppo.py

import sys
from agent import Agent
from stable_baselines3 import PPO
from stable_baselines3.common.logger import configure
from stable_baselines3.common.env_util import make_vec_env
from snek.environment import Snek

if '--novid' in sys.argv:
    import os
    os.environ["SDL_VIDEODRIVER"] = "dummy"

def main():
    env = Snek()
    env = make_vec_env(Snek, n_envs=2)
    #agent = Agent(env)
    
    new_logger = configure('./results', ["stdout", "csv", "json", "log"])
    model = PPO("MlpPolicy", env, verbose=1)
    model.set_logger(new_logger)
    model.learn(total_timesteps=20000000, log_interval=4)
    model.save("ppo_snek")

    total_len = 0
    num_episodes = 0
    
    env = Snek(render_enabled=True)
    obs = env.reset()

    while True:
        action, _states = model.predict(obs, deterministic=False)
        obs, _, done, _ = env.step(action)
        
        env.render() # Comment out this call to train faster

        if done:
            total_len += env.player.len
            num_episodes += 1
            obs = env.reset()
            print('Average len: {:.1f}'.format(total_len / num_episodes))

if __name__ == '__main__':
    main()