最最简单的超级马里奥训练过程
from nes_py.wrappers import JoypadSpace
import gym_super_mario_bros
from gym_super_mario_bros.actions import SIMPLE_MOVEMENT
import time
from matplotlib import pyplot as plt
from stable_baselines3 import PPO
env = gym_super_mario_bros.make('SuperMarioBros-v0')
env = JoypadSpace(env, SIMPLE_MOVEMENT)
tensorboard_log = r'./tensorboard_log/'
model = PPO("CnnPolicy", env, verbose=1,
tensorboard_log = tensorboard_log)
model.learn(total_timesteps=25000)
model.save("mario_model")
Using cuda device
Wrapping the env with a `Monitor` wrapper
Wrapping the env in a DummyVecEnv.
Wrapping the env in a VecTransposeImage.
Logging to ./tensorboard_log/PPO_1
D:\software\e_anaconda\envs\pytorch\lib\site-packages\gym_super_mario_bros\smb_env.py:148: RuntimeWarning: overflow encountered in ubyte_scalars
return (self.ram[0x86] - self.ram[0x071c]) % 256
-----------------------------
| time/ | |
| fps | 116 |
| iterations | 1 |
| time_elapsed | 17 |
| total_timesteps | 2048 |
-----------------------------
-----------------------------------------
| time/ | |
| fps | 81 |
| iterations | 2 |
| time_elapsed | 50 |
| total_timesteps | 4096 |
| train/ | |
| approx_kl | 0.025405666 |
| clip_fraction | 0.274 |
| clip_range | 0.2 |
| entropy_loss | -1.92 |
| explained_variance | 0.00504 |
| learning_rate | 0.0003 |
| loss | 0.621 |
| n_updates | 10 |
| policy_gradient_loss | 0.0109 |
| value_loss | 17.4 |
-----------------------------------------
-----------------------------------------
| time/ | |
| fps | 73 |
| iterations | 3 |
| time_elapsed | 83 |
| total_timesteps | 6144 |
| train/ | |
| approx_kl | 0.010906073 |
| clip_fraction | 0.109 |
| clip_range | 0.2 |
| entropy_loss | -1.92 |
| explained_variance | 0.0211 |
| learning_rate | 0.0003 |
| loss | 0.101 |
| n_updates | 20 |
| policy_gradient_loss | -0.00392 |
| value_loss | 0.187 |
-----------------------------------------
-----------------------------------------
| time/ | |
| fps | 69 |
| iterations | 4 |
| time_elapsed | 117 |
| total_timesteps | 8192 |
| train/ | |
| approx_kl | 0.009882288 |
| clip_fraction | 0.0681 |
| clip_range | 0.2 |
| entropy_loss | -1.9 |
| explained_variance | 0.101 |
| learning_rate | 0.0003 |
| loss | 0.0738 |
| n_updates | 30 |
| policy_gradient_loss | -0.00502 |
| value_loss | 0.13 |
-----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 65 |
| iterations | 5 |
| time_elapsed | 156 |
| total_timesteps | 10240 |
| train/ | |
| approx_kl | 0.008186281 |
| clip_fraction | 0.105 |
| clip_range | 0.2 |
| entropy_loss | -1.87 |
| explained_variance | 0.0161 |
| learning_rate | 0.0003 |
| loss | 0.28 |
| n_updates | 40 |
| policy_gradient_loss | -0.00649 |
| value_loss | 0.811 |
-----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 64 |
| iterations | 6 |
| time_elapsed | 190 |
| total_timesteps | 12288 |
| train/ | |
| approx_kl | 0.024062362 |
| clip_fraction | 0.246 |
| clip_range | 0.2 |
| entropy_loss | -1.9 |
| explained_variance | 0.269 |
| learning_rate | 0.0003 |
| loss | 0.54 |
| n_updates | 50 |
| policy_gradient_loss | 0.0362 |
| value_loss | 10.8 |
-----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 63 |
| iterations | 7 |
| time_elapsed | 225 |
| total_timesteps | 14336 |
| train/ | |
| approx_kl | 0.024466533 |
| clip_fraction | 0.211 |
| clip_range | 0.2 |
| entropy_loss | -1.89 |
| explained_variance | 0.839 |
| learning_rate | 0.0003 |
| loss | 0.435 |
| n_updates | 60 |
| policy_gradient_loss | 0.023 |
| value_loss | 3.06 |
-----------------------------------------
----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 63 |
| iterations | 8 |
| time_elapsed | 259 |
| total_timesteps | 16384 |
| train/ | |
| approx_kl | 0.01970315 |
| clip_fraction | 0.242 |
| clip_range | 0.2 |
| entropy_loss | -1.9 |
| explained_variance | 0.486 |
| learning_rate | 0.0003 |
| loss | 0.526 |
| n_updates | 70 |
| policy_gradient_loss | 0.00486 |
| value_loss | 1.57 |
----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 62 |
| iterations | 9 |
| time_elapsed | 293 |
| total_timesteps | 18432 |
| train/ | |
| approx_kl | 0.012460884 |
| clip_fraction | 0.217 |
| clip_range | 0.2 |
| entropy_loss | -1.87 |
| explained_variance | 0.74 |
| learning_rate | 0.0003 |
| loss | 0.139 |
| n_updates | 80 |
| policy_gradient_loss | -0.000311 |
| value_loss | 0.734 |
-----------------------------------------
----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 62 |
| iterations | 10 |
| time_elapsed | 327 |
| total_timesteps | 20480 |
| train/ | |
| approx_kl | 0.02535792 |
| clip_fraction | 0.298 |
| clip_range | 0.2 |
| entropy_loss | -1.88 |
| explained_variance | 0.405 |
| learning_rate | 0.0003 |
| loss | 1.17 |
| n_updates | 90 |
| policy_gradient_loss | 0.0205 |
| value_loss | 6.6 |
----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.01e+04 |
| ep_rew_mean | 891 |
| time/ | |
| fps | 62 |
| iterations | 11 |
| time_elapsed | 361 |
| total_timesteps | 22528 |
| train/ | |
| approx_kl | 0.019694094 |
| clip_fraction | 0.243 |
| clip_range | 0.2 |
| entropy_loss | -1.91 |
| explained_variance | 0.952 |
| learning_rate | 0.0003 |
| loss | 0.39 |
| n_updates | 100 |
| policy_gradient_loss | -0.00434 |
| value_loss | 1.31 |
-----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.19e+04 |
| ep_rew_mean | 884 |
| time/ | |
| fps | 61 |
| iterations | 12 |
| time_elapsed | 398 |
| total_timesteps | 24576 |
| train/ | |
| approx_kl | 0.013096321 |
| clip_fraction | 0.227 |
| clip_range | 0.2 |
| entropy_loss | -1.91 |
| explained_variance | 0.0132 |
| learning_rate | 0.0003 |
| loss | 0.669 |
| n_updates | 110 |
| policy_gradient_loss | -0.000837 |
| value_loss | 1.42 |
-----------------------------------------
-----------------------------------------
| rollout/ | |
| ep_len_mean | 1.19e+04 |
| ep_rew_mean | 884 |
| time/ | |
| fps | 61 |
| iterations | 13 |
| time_elapsed | 432 |
| total_timesteps | 26624 |
| train/ | |
| approx_kl | 0.014833134 |
| clip_fraction | 0.239 |
| clip_range | 0.2 |
| entropy_loss | -1.9 |
| explained_variance | 0.452 |
| learning_rate | 0.0003 |
| loss | 18.1 |
| n_updates | 120 |
| policy_gradient_loss | -7.3e-05 |
| value_loss | 26.3 |
-----------------------------------------
测试代码
from nes_py.wrappers import JoypadSpace
import gym_super_mario_bros
from gym_super_mario_bros.actions import SIMPLE_MOVEMENT
import time
from matplotlib import pyplot as plt
from stable_baselines3 import PPO
env = gym_super_mario_bros.make('SuperMarioBros-v0')
env = JoypadSpace(env, SIMPLE_MOVEMENT)
model = PPO.load("mario_model")
obs = env.reset()
obs=obs.copy()
done = True
while True:
if done:
state = env.reset()
action, _states = model.predict(obs)
obs, rewards, done, info = env.step(action)
obs=obs.copy()
env.render()