def make_gym_atari_env(game, seed=1):
env = gym_atari.GymAtari(game, seed=seed)
env = gym_atari.RandomNoopsEnvironmentWrapper(env,
min_noop_steps=1,
max_noop_steps=30,
seed=seed)
return env
def environment_builder():
"""Creates Atari environment."""
env = gym_atari.GymAtari(
FLAGS.environment_name, seed=random_state.randint(1, 2**32))
return gym_atari.RandomNoopsEnvironmentWrapper(
env,
min_noop_steps=1,
max_noop_steps=30,
seed=random_state.randint(1, 2**32),
)
def num_noops_sequence(seed, num_episodes):
tape = []
environment = test_utils.DummyEnvironment(tape, episode_length=10)
wrapped_environment = gym_atari.RandomNoopsEnvironmentWrapper(
environment, max_noop_steps=8, seed=seed)
seq = []
for _ in range(num_episodes):
tape.clear()
wrapped_environment.reset()
seq.append(len(tape))
return seq
def test_specs(self):
environment = test_utils.DummyEnvironment([], episode_length=10)
wrapped_environment = gym_atari.RandomNoopsEnvironmentWrapper(
environment, max_noop_steps=5)
self.assertEqual(environment.observation_spec(),
wrapped_environment.observation_spec())
self.assertEqual(environment.action_spec(),
wrapped_environment.action_spec())
self.assertEqual(environment.reward_spec(),
wrapped_environment.reward_spec())
self.assertEqual(environment.discount_spec(),
wrapped_environment.discount_spec())
def test_basic(self, min_noop_steps, max_noop_steps):
noop_action = 3
tape = []
environment = test_utils.DummyEnvironment(tape, episode_length=10)
wrapped_environment = gym_atari.RandomNoopsEnvironmentWrapper(
environment,
min_noop_steps=min_noop_steps,
max_noop_steps=max_noop_steps,
noop_action=noop_action,
seed=42)
# Make sure noops are applied appropriate number of times (in min/max range
# and not always the same number), with correct action.
num_noop_steps = set()
for i in range(20):
# Switch between different ways of starting a new episode.
if i % 4 == 0:
tape.clear()
wrapped_environment.reset()
num_steps = len(tape)
expected_tape = (['Environment reset'] +
['Environment step (%s)' % noop_action] *
(num_steps - 1))
else:
timestep = wrapped_environment.reset()
while not timestep.last():
timestep = wrapped_environment.step(0)
tape.clear()
wrapped_environment.step(noop_action)
num_steps = len(tape)
expected_tape = (['Environment step (%s)' % noop_action] *
num_steps)
self.assertEqual(expected_tape, tape)
# +1 because of the extra initial reset() / step().
self.assertBetween(num_steps, min_noop_steps + 1,
max_noop_steps + 1)
num_noop_steps.add(num_steps)
# Do some regular steps & check pass-through of actions.
wrapped_environment.step(6)
wrapped_environment.step(7)
self.assertLen(tape, num_steps + 2)
self.assertEqual(['Environment step (6)', 'Environment step (7)'],
tape[-2:])
# Check it's not always the same number of random noop steps.
if max_noop_steps > min_noop_steps:
self.assertGreater(len(num_noop_steps), 1)
def environment_builder():
"""Creates Key-Door environment."""
env = gym_key_door.GymKeyDoor(
env_args={
constants.MAP_ASCII_PATH: FLAGS.map_ascii_path,
constants.MAP_YAML_PATH: FLAGS.map_yaml_path,
constants.REPRESENTATION: constants.PIXEL,
constants.SCALING: FLAGS.env_scaling,
constants.EPISODE_TIMEOUT: FLAGS.max_frames_per_episode,
constants.GRAYSCALE: False,
constants.BATCH_DIMENSION: False,
constants.TORCH_AXES: False,
},
env_shape=FLAGS.env_shape,
)
return gym_atari.RandomNoopsEnvironmentWrapper(
env,
min_noop_steps=1,
max_noop_steps=30,
seed=random_state.randint(1, 2**32),
)
def test_episode_end_during_noop_steps(self):
environment = test_utils.DummyEnvironment([], episode_length=5)
wrapped_environment = gym_atari.RandomNoopsEnvironmentWrapper(
environment, min_noop_steps=10, max_noop_steps=20)
with self.assertRaisesRegex(RuntimeError, 'Episode ended'):
wrapped_environment.reset()
