-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathdmc.py
197 lines (156 loc) · 6.75 KB
/
dmc.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
import numpy as np
from collections import OrderedDict, deque
import warnings
import dm_env
from dm_env import specs
from dm_control import suite
with warnings.catch_warnings():
warnings.filterwarnings('ignore', category=DeprecationWarning)
from dm_control import manipulation
from dm_control.suite.wrappers import action_scale, pixels
MANIP_PIXELS_KEY = 'front_close'
class FlattenObservationWrapper(dm_env.Environment):
def __init__(self, env):
self._env = env
self._obs_spec = OrderedDict()
wrapped_obs_spec = env.observation_spec().copy()
dim = 0
for key in wrapped_obs_spec.keys():
if key != MANIP_PIXELS_KEY:
spec = wrapped_obs_spec[key]
assert spec.dtype == np.float64
assert type(spec) == specs.Array
dim += np.prod(spec.shape)
self._obs_spec['features'] = specs.Array(shape=(dim,),
dtype=np.float32,
name='features')
if MANIP_PIXELS_KEY in wrapped_obs_spec:
spec = wrapped_obs_spec[MANIP_PIXELS_KEY]
self._obs_spec['pixels'] = specs.BoundedArray(shape=spec.shape[1:],
dtype=spec.dtype,
minimum=spec.minimum,
maximum=spec.maximum,
name='pixels')
self._obs_spec['state'] = specs.Array(
shape=self._env.physics.get_state().shape,
dtype=np.float32,
name='state')
def _transform_observation(self, time_step):
obs = OrderedDict()
features = []
for key, value in time_step.observation.items():
if key != MANIP_PIXELS_KEY:
features.append(value.ravel())
obs['features'] = np.concatenate(features, axis=0)
obs['state'] = self._env.physics.get_state().copy()
if MANIP_PIXELS_KEY in time_step.observation:
obs['pixels'] = time_step.observation[MANIP_PIXELS_KEY][0]
return time_step._replace(observation=obs)
def reset(self):
time_step = self._env.reset()
return self._transform_observation(time_step)
def step(self, action):
time_step = self._env.step(action)
return self._transform_observation(time_step)
def observation_spec(self):
return self._obs_spec
def action_spec(self):
return self._env.action_spec()
def __getattr__(self, name):
return getattr(self._env, name)
class FrameStackWrapper(dm_env.Environment):
def __init__(self, env, k):
self._env = env
self._k = k
self._frames = deque([], maxlen=k)
wrapped_obs_spec = env.observation_spec()
assert 'features' in wrapped_obs_spec
assert 'pixels' in wrapped_obs_spec
self._obs_spec = OrderedDict()
self._obs_spec['features'] = wrapped_obs_spec['features']
self._obs_spec['state'] = wrapped_obs_spec['state']
pixels_spec = wrapped_obs_spec['pixels']
self._obs_spec['pixels'] = specs.BoundedArray(shape=np.concatenate(
[[pixels_spec.shape[2] * k], pixels_spec.shape[:2]], axis=0),
dtype=pixels_spec.dtype,
minimum=0,
maximum=255,
name=pixels_spec.name)
def _transform_observation(self, time_step):
assert len(self._frames) == self._k
obs = OrderedDict()
obs['features'] = time_step.observation['features']
obs['state'] = time_step.observation['state']
obs['pixels'] = np.concatenate(list(self._frames), axis=0)
return time_step._replace(observation=obs)
def reset(self):
time_step = self._env.reset()
pixels = time_step.observation['pixels'].transpose(2, 0, 1).copy()
for _ in range(self._k):
self._frames.append(pixels.copy())
return self._transform_observation(time_step)
def step(self, action):
time_step = self._env.step(action)
self._frames.append(time_step.observation['pixels'].transpose(
2, 0, 1).copy())
return self._transform_observation(time_step)
def observation_spec(self):
return self._obs_spec
def action_spec(self):
return self._env.action_spec()
def __getattr__(self, name):
return getattr(self._env, name)
class ActionRepeatWrapper(dm_env.Environment):
def __init__(self, env, amount):
self._env = env
self._amount = amount
def step(self, action):
reward = 0.0
for i in range(self._amount):
time_step = self._env.step(action)
reward += time_step.reward or 0.0
if time_step.last():
break
return time_step._replace(reward=reward)
def observation_spec(self):
return self._env.observation_spec()
def action_spec(self):
return self._env.action_spec()
def reset(self):
return self._env.reset()
def __getattr__(self, name):
return getattr(self._env, name)
def split_env_name(env_name):
if env_name == 'ball_in_cup_catch':
return 'ball_in_cup', 'catch'
if env_name.startswith('point_mass'):
return 'point_mass', env_name.split('_')[-1]
domain = env_name.split('_')[0]
task = '_'.join(env_name.split('_')[1:])
return domain, task
def make(env_name, frame_stack, action_repeat, seed):
domain, task = split_env_name(env_name)
if domain == 'manip':
env = manipulation.load(f'{task}_vision', seed=seed)
else:
env = suite.load(domain,
task,
task_kwargs={'random': seed},
visualize_reward=False)
# apply action repeat and scaling
env = ActionRepeatWrapper(env, action_repeat)
env = action_scale.Wrapper(env, minimum=-1.0, maximum=+1.0)
# flatten features
env = FlattenObservationWrapper(env)
if domain != 'manip':
# per dreamer: https://github.com/danijar/dreamer/blob/02f0210f5991c7710826ca7881f19c64a012290c/wrappers.py#L26
camera_id = 2 if domain == 'quadruped' else 0
render_kwargs = {'height': 84, 'width': 84, 'camera_id': camera_id}
env = pixels.Wrapper(env,
pixels_only=False,
render_kwargs=render_kwargs)
env = FrameStackWrapper(env, frame_stack)
action_spec = env.action_spec()
assert np.all(action_spec.minimum >= -1.0)
assert np.all(action_spec.maximum <= +1.0)
return env