PPO playing impala-PongNoFrameskip-v4 from https://github.com/sgoodfriend/rl-algo-impls/tree/e47a44c4d891f48885af0b1605b30d19fc67b5af
be9c115
import gym | |
import numpy as np | |
from collections import deque | |
from stable_baselines3.common.vec_env.base_vec_env import ( | |
VecEnvStepReturn, | |
VecEnvObs, | |
) | |
from torch.utils.tensorboard.writer import SummaryWriter | |
from typing import Any, Dict, List | |
from shared.stats import Episode, EpisodesStats | |
class EpisodeStatsWriter(gym.Wrapper): | |
def __init__( | |
self, | |
env, | |
tb_writer: SummaryWriter, | |
training: bool = True, | |
rolling_length=100, | |
): | |
super().__init__(env) | |
self.training = training | |
self.tb_writer = tb_writer | |
self.rolling_length = rolling_length | |
self.episodes = deque(maxlen=rolling_length) | |
self.total_steps = 0 | |
self.episode_cnt = 0 | |
self.last_episode_cnt_print = 0 | |
def step(self, actions: np.ndarray) -> VecEnvStepReturn: | |
obs, rews, dones, infos = self.env.step(actions) | |
self._record_stats(infos) | |
return obs, rews, dones, infos | |
# Support for stable_baselines3.common.vec_env.VecEnvWrapper | |
def step_wait(self) -> VecEnvStepReturn: | |
obs, rews, dones, infos = self.env.step_wait() | |
self._record_stats(infos) | |
return obs, rews, dones, infos | |
def _record_stats(self, infos: List[Dict[str, Any]]) -> None: | |
self.total_steps += getattr(self.env, "num_envs", 1) | |
step_episodes = [] | |
for info in infos: | |
ep_info = info.get("episode") | |
if ep_info: | |
episode = Episode(ep_info["r"], ep_info["l"]) | |
step_episodes.append(episode) | |
self.episodes.append(episode) | |
if step_episodes: | |
tag = "train" if self.training else "eval" | |
step_stats = EpisodesStats(step_episodes, simple=True) | |
step_stats.write_to_tensorboard(self.tb_writer, tag, self.total_steps) | |
rolling_stats = EpisodesStats(self.episodes) | |
rolling_stats.write_to_tensorboard( | |
self.tb_writer, f"{tag}_rolling", self.total_steps | |
) | |
self.episode_cnt += len(step_episodes) | |
if self.episode_cnt >= self.last_episode_cnt_print + self.rolling_length: | |
print( | |
f"Episode: {self.episode_cnt} | " | |
f"Steps: {self.total_steps} | " | |
f"{rolling_stats}" | |
) | |
self.last_episode_cnt_print += self.rolling_length | |
def reset(self) -> VecEnvObs: | |
return self.env.reset() | |