def make_env(config):
config = config.copy()
if '_' not in config['name']:
config['name'] = f"_{config['name']}"
suite, task = config['name'].split('_', 1)
if suite == 'dmc':
env = DeepMindControl(task, config.get('size', (64, 64)))
env = wrappers.FrameSkip(env, config['frame_skip'])
env = wrappers.NormalizeActions(env)
config['max_episode_steps'] = 1000
if config.get('frame_stack', 1) > 1:
env = wrappers.FrameStack(env, config['frame_stack'], np_obs=False)
elif suite == 'atari':
env = Atari(task,
config['frame_skip'], (64, 64),
grayscale=False,
life_done=True,
sticky_actions=True)
config['max_episode_steps'] = 108000
elif 'procgen' in config['name'].lower():
env = procgen.make_procgen_env(config)
config['max_episode_steps'] = env.spec.max_episode_steps or 1000
else:
raise NotImplementedError(suite)
env = wrappers.post_wrap(env, config)
return env
def make_env(config):
config = config.copy()
_, name = config['name'].split('_', 1)
name = name[0].capitalize() + name[1:]
# version = 0 if config.get('sticky_actions', True) else 4
version = 4
name = f'{name}NoFrameskip-v{version}'
config['max_episode_steps'] = max_episode_steps = 4500 * 4
env = make_atari(name, max_episode_steps=max_episode_steps)
env = wrap_deepmind(env,
reward_clip=config.get('reward_clip', True),
frame_stack=config.get('frame_stack', 4))
env = post_wrap(env, config)
return env
def make_env(config):
config = config.copy()
env_name = config['name'].lower()
if env_name.startswith('smac2'):
from env import smac2
env = smac2.make_smac_env(config)
# smac2 resembles single agent environments, in which done and reward are team-based
env = wrappers.EnvStats(env)
return env
elif env_name.startswith('smac'):
from env import smac
env = smac.make_smac_env(config)
env = wrappers.MAEnvStats(env)
return env
elif env_name.startswith('atari'):
from env import atari
env = atari.make_atari_env(config)
else:
if env_name.startswith('procgen'):
from env import procgen
env = procgen.make_procgen_env(config)
elif env_name.startswith('dmc'):
from env import dmc
env = dmc.make_dmc_env(config)
else:
env = gym.make(config['name']).env
env = wrappers.DummyEnv(env) # useful for hidding unexpected frame_skip
config.setdefault('max_episode_steps', env.spec.max_episode_steps)
if config.get('reward_scale') or config.get('reward_clip'):
env = wrappers.RewardHack(env, **config)
frame_stack = config.setdefault('frame_stack', 1)
if frame_stack > 1:
np_obs = config.setdefault('np_obs', False)
env = wrappers.FrameStack(env, frame_stack, np_obs)
frame_diff = config.setdefault('frame_diff', False)
assert not (frame_diff and frame_stack > 1), f"Don't support using FrameStack and FrameDiff at the same time"
if frame_diff:
gray_scale_residual = config.setdefault('gray_scale_residual', False)
distance = config.setdefault('distance', 1)
env = wrappers.FrameDiff(env, gray_scale_residual, distance)
env = wrappers.post_wrap(env, config)
return env
def mkenv(config):
env = DummyEnv(**config)
env = wrappers.post_wrap(env, config)
return env