Gym-MiniGrid/basicrl/envs.py

import os
import numpy
import gym

from gym.spaces.box import Box

from baselines import bench
from baselines.common.atari_wrappers import make_atari, wrap_deepmind

try:
    import pybullet_envs
except ImportError:
    pass

try:
    import gym_minigrid
except:
    pass

class ScaleActions(gym.ActionWrapper):
    def __init__(self, env=None):
        super(ScaleActions, self).__init__(env)

    def _step(self, action):
        action = (numpy.tanh(action) + 1) / 2 * (self.action_space.high - self.action_space.low) + self.action_space.low
        return self.env.step(action)

def make_env(env_id, seed, rank, log_dir):
    def _thunk():
        env = gym.make(env_id)
        is_atari = hasattr(gym.envs, 'atari') and isinstance(env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
        if is_atari:
            env = make_atari(env_id)
        env.seed(seed + rank)
        if log_dir is not None:
            env = bench.Monitor(env, os.path.join(log_dir, str(rank)))
        if is_atari:
            env = wrap_deepmind(env)
        # If the input has shape (W,H,3), wrap for PyTorch convolutions
        obs_shape = env.observation_space.shape
        if len(obs_shape) == 3 and obs_shape[2] == 3:
            env = WrapPyTorch(env)

        #env = ScaleActions(env)

        return env

    return _thunk


class WrapPyTorch(gym.ObservationWrapper):
    def __init__(self, env=None):
        super(WrapPyTorch, self).__init__(env)
        obs_shape = self.observation_space.shape
        self.observation_space = Box(
            self.observation_space.low[0,0,0],
            self.observation_space.high[0,0,0],
            [obs_shape[2], obs_shape[1], obs_shape[0]]
        )

    def _observation(self, observation):
        return observation.transpose(2, 0, 1)