def make_env(process_idx, test):
# Use different random seeds for train and test envs
process_seed = process_seeds[process_idx]
env_seed = 2**31 - 1 - process_seed if test else process_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_env():
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=None),
episode_life=False,
clip_rewards=False)
env.seed(int(args.seed))
# Randomize actions like epsilon-greedy
env = chainerrl.wrappers.RandomizeAction(env, 0.01)
if args.monitor:
env = chainerrl.wrappers.Monitor(env,
args.outdir,
mode='evaluation')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(args.env),
episode_life=not test,
clip_rewards=not test)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def __init__(self,
host='localhost',
port=8001,
model='atari',
env_name='BreakoutNoFrameskip-v4',
n_stack_frames=4,
rendering_scale=None,
sync=False):
self._running = False
self._display_surf = None
self._env = chainerrl.wrappers.RandomizeAction(
atari_wrappers.make_atari(env_name, max_frames=None),
0.05)
_obs_shape = self._env.observation_space.shape
self._screen_img = rotate_and_flip_screen(
self._env.reset())
self._done = False
if rendering_scale is not None:
self.width = int(rendering_scale * _obs_shape[1])
self.height = int(rendering_scale * _obs_shape[0])
self.size = self.width, self.height
else:
self.width, self.height = _obs_shape[1], _obs_shape[0]
self.size = self.width, self.height
self._rendering_scale = rendering_scale
self.n_channels = _obs_shape[2]
self._sync = sync
if self._sync:
self._agent = SyncAgent(
host=host,
port=port,
model=model,
n_stack_frames=n_stack_frames)
else:
self._agent = AsyncAgent(
host=host,
port=port,
model=model,
observation_shape=_obs_shape,
n_stack_frames=n_stack_frames)
self._action = 0
def make_env(idx, test):
# Use different random seeds for train and test envs
process_seed = int(process_seeds[idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test)
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = chainerrl.wrappers.RandomizeAction(env, args.eval_epsilon)
env.seed(env_seed)
if args.monitor:
env = gym.wrappers.Monitor(
env, args.outdir, mode='evaluation' if test else 'training')
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_env():
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(env_name),
episode_life=False,
clip_rewards=False)
env.seed(seed)
return env
