def eval_time(args):
with open(args.config) as fp:
json_data = json.load(fp)
video_path = os.path.join("./videos", args.agent)
config = GameConfig.deserialize(json_data)
config.agents_config[args.agent]["save_path"] += "_vs_time_pt_check.zip"
env = DummyVecEnv(
[lambda: retro.make(config.game_name, state=config.eval_state[1])])
agent = AgentLoader.get_agent(args.agent,
config.agents_config,
env,
load=True)
env.close()
env = DummyVecEnv([
lambda: retro.make(
config.game_name, state=config.eval_state[1], record=video_path)
])
obs = env.reset()
done = False
while not done:
actions, _ = agent.agent.predict(obs)
obs, rew, done, info = env.step(actions)
# env.render()
env.close()
def main(config: str, agent: str):
with open(config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
log_dir = config.agents_config[agent]["save_path"]
# if agent == "DQN":
# env = make_atari_env(config.game_name, n_envs=1,
# seed=0, monitor_dir=log_dir)
# elif agent == "PPO":
# env = make_atari_env(config.game_name, n_envs=8,
# seed=0, monitor_dir=log_dir)
# else:
# env = make_atari_env(config.game_name, n_envs=16,
# seed=0, monitor_dir=log_dir)
env = gym_super_mario_bros.make(config.game_name)
env = JoypadSpace(env, SIMPLE_MOVEMENT)
env = Monitor(env, log_dir)
# env = VecFrameStack(env, n_stack=4)
agent = AgentLoader.get_agent(agent, config.agents_config, env)
def eval_100_trials(args):
with open(args.config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
config.agents_config[args.agent]["save_path"] += "_vs_time_pt.zip"
env = DummyVecEnv(
[lambda: retro.make(config.game_name, state=config.eval_state[1])])
agent = AgentLoader.get_agent(args.agent,
config.agents_config,
env,
load=True)
rew_list = []
trials = 100
for i in tqdm(range(trials)):
obs = env.reset()
done = False
reward = 0
while not done:
actions, _ = agent.agent.predict(obs)
obs, rew, done, info = env.step(actions)
reward += rew
rew_list.append(reward)
env.close()
count = sum(i > 0 for i in rew_list)
print("win percentage = {}%".format(count / trials * 100))
def main_vs_5(config: str):
with open(config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
config.agents_config["A2C"]["save_path"] += "_vs_5"
config.agents_config["A2C"]["tensorboard"] += "_vs_5"
env = DummyVecEnv(
[lambda: retro.make(config.game_name, state=config.train_states[0])])
agent = AgentLoader.get_agent("A2C", config.agents_config, env)
env.close()
start_time = time.time()
for st in tqdm(config.train_states, desc='Main Loop'):
print(st)
env = DummyVecEnv([
lambda: retro.make(config.game_name, state=st, scenario='scenario')
])
agent.agent.set_env(env)
agent.agent.learn(total_timesteps=10000)
agent.save()
env.close()
end_time = time.time() - start_time
print(f'\n The Training Took {end_time} seconds')
def main(config: str, agent: str):
with open(config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
log_dir = config.agents_config[agent]["save_path"]
if agent == "DQN":
env = make_atari_env(config.game_name, n_envs=1,
seed=0, monitor_dir=log_dir)
elif agent == "PPO":
env = make_atari_env(config.game_name, n_envs=8,
seed=0, monitor_dir=log_dir)
else:
env = make_atari_env(config.game_name, n_envs=16,
seed=0, monitor_dir=log_dir)
env = VecFrameStack(env, n_stack=4)
agent = AgentLoader.get_agent(agent, config.agents_config, env)
reward_callback = SaveOnBestTrainingRewardCallback(
check_freq=100, log_dir=log_dir)
start_time = time.time()
steps = 10_000_000
with ProgressBarManager_new(steps) as progress_callback:
agent.agent.learn(total_timesteps=steps, callback=[
reward_callback, progress_callback])
# agent.save()
env.close()
end_time = time.time() - start_time
print(f'\n The Training Took {end_time} seconds')
def main(config, agent):
with open(config) as fp:
json_data = json.load(fp)
video_path = os.path.join("./videos", agent, "pong")
config = GameConfig.deserialize(json_data)
config.agents_config[args.agent]["save_path"] += "best_model.zip"
# config.agents_config[args.agent]["save_path"] = "my_models/pong/pong_ppo/best_model.zip"
print(config.agents_config[args.agent]["save_path"])
# env = retro.make(config.game_name)
env = gym.make("PongNoFrameskip-v4")
agent = AgentLoader.get_agent(args.agent,
config.agents_config,
env,
load=True)
env.close()
env = gym.make("PongNoFrameskip-v4")
env = DummyVecEnv([lambda: env])
# env = retro.make(config.game_name, record=video_path)
env = VecVideoRecorder(
env,
video_path,
record_video_trigger=lambda x: x == 0,
)
obs = env.reset()
done = False
while not done:
actions, _ = agent.agent.predict(obs)
obs, rew, done, info = env.step(actions)
env.close()
def main_vs_time(config: str):
with open(config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
config.agents_config["A2C"]["save_path"] += "_vs_time_pt"
config.agents_config["A2C"]["tensorboard"] += "_vs_time"
env = DummyVecEnv(
[lambda: (retro.make(config.game_name, state=config.eval_state[0]))])
agent = AgentLoader.get_agent("A2C", config.agents_config, env)
start_time = time.time()
with ProgressBarManager_new(40000) as callback:
agent.agent.learn(total_timesteps=40000, callback=callback)
agent.save()
env.close()
end_time = time.time() - start_time
print(f'\n The Training Took {end_time} seconds')
def main_vs_time(config: str):
with open(config) as fp:
json_data = json.load(fp)
config = GameConfig.deserialize(json_data)
config.agents_config["PPO"]["save_path"] += "_vs_time_pt_check"
config.agents_config["PPO"]["tensorboard"] += "_vs_time_check"
env = DummyVecEnv(
[lambda: (retro.make(config.game_name, state=config.eval_state[0]))])
agent = AgentLoader.get_agent("PPO", config.agents_config, env)
env.close()
env = DummyVecEnv(
[lambda: (retro.make(config.game_name, state=config.eval_state[0]))])
agent.agent.set_env(env)
with ProgressBarManager_new(1000) as callback:
agent.agent.learn(1000, callback=callback)
agent.save()
env.close()