gym-subproc-example.py

#!/usr/bin/env python3

import numpy as np
import gym
from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv

env_name = 'Pendulum-v0'
nproc = 8
T = 10

def make_env(env_id, seed):
    def _f():
        env = gym.make(env_id)
        env.seed(seed)
        return env
    return _f

envs = [make_env(env_name, seed) for seed in range(nproc)]
envs = SubprocVecEnv(envs)

xt = envs.reset()
for t in range(T):
    ut = np.stack([envs.action_space.sample() for _ in range(nproc)])
    xtp1, rt, done, info = envs.step(ut)

Hey bamos this is very nice.

However I struggle using it in my RL PPO algorithm. Cause using this I also have to parallelize my policy.act(state, memory) method... to be policy.act(states, memoys)

Here is my code snippet:

def playSteps(policy, env, steps):
    done            = 0
    xt              = envs.reset()
    memory          = [Memory()]*12
    for _ in range(100):
        # note reset is done inside the gym environment!
        ut = []
        for i in range(12):
            ut.append(policy.act([xt[i], memory[i]]))   #ut = np.stack([envs.action_space.sample() for _ in range(nproc)])
        state, rewards, done, info = envs.step(ut)

How did you solve these kind of problem?