def test_filter_stacking():
# test that filter stacking works fine by taking as input a transition with:
# - an observation of shape 210x160,
# - a reward of 100
# filtering it by:
# - rescaling the observation to 110x84
# - cropping the observation to 84x84
# - clipping the reward to 1
# - stacking 4 observations to get 84x84x4
env_response = EnvResponse({'observation': np.ones([210, 160])}, reward=100, game_over=False)
filter1 = ObservationRescaleToSizeFilter(
output_observation_space=ImageObservationSpace(np.array([110, 84]), high=255),
)
filter2 = ObservationCropFilter(
crop_low=np.array([16, 0]),
crop_high=np.array([100, 84])
)
filter3 = RewardClippingFilter(
clipping_low=-1,
clipping_high=1
)
output_filter = ObservationStackingFilter(
stack_size=4,
stacking_axis=-1
)
input_filter = InputFilter(
observation_filters={
"observation": OrderedDict([
("filter1", filter1),
("filter2", filter2),
("output_filter", output_filter)
])},
reward_filters=OrderedDict([
("filter3", filter3)
])
)
result = input_filter.filter(env_response)[0]
observation = np.array(result.next_state['observation'])
assert observation.shape == (84, 84, 4)
assert np.all(observation == np.ones([84, 84, 4]))
assert result.reward == 1
for e in gym_mujoco_envs
}
mujoco_v2['walker2d'] = 'Walker2d-v2'
# Fetch
gym_fetch_envs = ['reach', 'slide', 'push', 'pick_and_place']
fetch_v1 = {
e: "{}".format('Fetch' + lower_under_to_upper(e) + '-v1')
for e in gym_fetch_envs
}
"""
Atari Environment Components
"""
AtariInputFilter = InputFilter(is_a_reference_filter=True)
AtariInputFilter.add_reward_filter('clipping', RewardClippingFilter(-1.0, 1.0))
AtariInputFilter.add_observation_filter(
'observation',
'rescaling',
ObservationRescaleToSizeFilter(
ImageObservationSpace(
np.array([84, 84, 3]), #np.array([224, 224, 3]),
high=255)))
#AtariInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter())
#AtariInputFilter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255))
#AtariInputFilter.add_observation_filter('observation', 'stacking', ObservationStackingFilter(4))
AtariOutputFilter = NoOutputFilter()
class Atari(GymEnvironmentParameters):
def __init__(self, level=None):
def clip_filter():
return InputFilter(
reward_filters=OrderedDict([('clip', RewardClippingFilter(2, 10))]))