-
Notifications
You must be signed in to change notification settings - Fork 310
/
gaussian_gru_policy.py
364 lines (309 loc) · 14.2 KB
/
gaussian_gru_policy.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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
"""Gaussian GRU Policy.
A policy represented by a Gaussian distribution
which is parameterized by a Gated Recurrent Unit (GRU).
"""
# pylint: disable=wrong-import-order
import akro
import numpy as np
import tensorflow as tf
from garage.experiment import deterministic
from garage.tf.models import GaussianGRUModel
from garage.tf.policies.policy import Policy
# pylint: disable=too-many-ancestors
class GaussianGRUPolicy(GaussianGRUModel, Policy):
"""Gaussian GRU Policy.
A policy represented by a Gaussian distribution
which is parameterized by a Gated Recurrent Unit (GRU).
Args:
env_spec (garage.envs.env_spec.EnvSpec): Environment specification.
name (str): Model name, also the variable scope.
hidden_dim (int): Hidden dimension for GRU cell for mean.
hidden_nonlinearity (Callable): Activation function for intermediate
dense layer(s). It should return a tf.Tensor. Set it to
None to maintain a linear activation.
hidden_w_init (Callable): Initializer function for the weight
of intermediate dense layer(s). The function should return a
tf.Tensor.
hidden_b_init (Callable): Initializer function for the bias
of intermediate dense layer(s). The function should return a
tf.Tensor.
recurrent_nonlinearity (Callable): Activation function for recurrent
layers. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
recurrent_w_init (Callable): Initializer function for the weight
of recurrent layer(s). The function should return a
tf.Tensor.
output_nonlinearity (Callable): Activation function for output dense
layer. It should return a tf.Tensor. Set it to None to
maintain a linear activation.
output_w_init (Callable): Initializer function for the weight
of output dense layer(s). The function should return a
tf.Tensor.
output_b_init (Callable): Initializer function for the bias
of output dense layer(s). The function should return a
tf.Tensor.
hidden_state_init (Callable): Initializer function for the
initial hidden state. The functino should return a tf.Tensor.
hidden_state_init_trainable (bool): Bool for whether the initial
hidden state is trainable.
learn_std (bool): Is std trainable.
std_share_network (bool): Boolean for whether mean and std share
the same network.
init_std (float): Initial value for std.
layer_normalization (bool): Bool for using layer normalization or not.
state_include_action (bool): Whether the state includes action.
If True, input dimension will be
(observation dimension + action dimension).
"""
def __init__(self,
env_spec,
hidden_dim=32,
name='GaussianGRUPolicy',
hidden_nonlinearity=tf.nn.tanh,
hidden_w_init=tf.initializers.glorot_uniform(
seed=deterministic.get_tf_seed_stream()),
hidden_b_init=tf.zeros_initializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_init=tf.initializers.glorot_uniform(
seed=deterministic.get_tf_seed_stream()),
output_nonlinearity=None,
output_w_init=tf.initializers.glorot_uniform(
seed=deterministic.get_tf_seed_stream()),
output_b_init=tf.zeros_initializer(),
hidden_state_init=tf.zeros_initializer(),
hidden_state_init_trainable=False,
learn_std=True,
std_share_network=False,
init_std=1.0,
layer_normalization=False,
state_include_action=True):
if not isinstance(env_spec.action_space, akro.Box):
raise ValueError('GaussianGRUPolicy only works with '
'akro.Box action space, but not {}'.format(
env_spec.action_space))
self._env_spec = env_spec
self._obs_dim = env_spec.observation_space.flat_dim
self._action_dim = env_spec.action_space.flat_dim
self._hidden_dim = hidden_dim
self._hidden_nonlinearity = hidden_nonlinearity
self._hidden_w_init = hidden_w_init
self._hidden_b_init = hidden_b_init
self._recurrent_nonlinearity = recurrent_nonlinearity
self._recurrent_w_init = recurrent_w_init
self._output_nonlinearity = output_nonlinearity
self._output_w_init = output_w_init
self._output_b_init = output_b_init
self._hidden_state_init = hidden_state_init
self._hidden_state_init_trainable = hidden_state_init_trainable
self._learn_std = learn_std
self._std_share_network = std_share_network
self._init_std = init_std
self._layer_normalization = layer_normalization
self._state_include_action = state_include_action
if state_include_action:
self._input_dim = self._obs_dim + self._action_dim
else:
self._input_dim = self._obs_dim
self._f_step_mean_std = None
super().__init__(
output_dim=self._action_dim,
hidden_dim=hidden_dim,
name=name,
hidden_nonlinearity=hidden_nonlinearity,
hidden_w_init=hidden_w_init,
hidden_b_init=hidden_b_init,
recurrent_nonlinearity=recurrent_nonlinearity,
recurrent_w_init=recurrent_w_init,
output_nonlinearity=output_nonlinearity,
output_w_init=output_w_init,
output_b_init=output_b_init,
hidden_state_init=hidden_state_init,
hidden_state_init_trainable=hidden_state_init_trainable,
layer_normalization=layer_normalization,
learn_std=learn_std,
std_share_network=std_share_network,
init_std=init_std)
self._prev_actions = None
self._prev_hiddens = None
self._init_hidden = None
self._initialize_policy()
def _initialize_policy(self):
"""Initialize policy."""
state_input = tf.compat.v1.placeholder(shape=(None, None,
self._input_dim),
name='state_input',
dtype=tf.float32)
step_input_var = tf.compat.v1.placeholder(shape=(None,
self._input_dim),
name='step_input',
dtype=tf.float32)
step_hidden_var = tf.compat.v1.placeholder(shape=(None,
self._hidden_dim),
name='step_hidden_input',
dtype=tf.float32)
(_, step_mean, step_log_std, step_hidden,
self._init_hidden) = super().build(state_input, step_input_var,
step_hidden_var).outputs
self._f_step_mean_std = (
tf.compat.v1.get_default_session().make_callable(
[step_mean, step_log_std, step_hidden],
feed_list=[step_input_var, step_hidden_var]))
# pylint: disable=arguments-differ
def build(self, state_input, name=None):
"""Build policy.
Args:
state_input (tf.Tensor) : State input.
name (str): Name of the policy, which is also the name scope.
Returns:
tfp.distributions.MultivariateNormalDiag: Policy distribution.
tf.Tensor: Step means, with shape :math:`(N, S^*)`.
tf.Tensor: Step log std, with shape :math:`(N, S^*)`.
tf.Tensor: Step hidden state, with shape :math:`(N, S^*)`.
tf.Tensor: Initial hidden state, with shape :math:`(S^*)`.
"""
_, step_input_var, step_hidden_var = self.inputs
return super().build(state_input,
step_input_var,
step_hidden_var,
name=name)
@property
def input_dim(self):
"""int: Dimension of the policy input."""
return self._input_dim
def reset(self, do_resets=None):
"""Reset the policy.
Note:
If `do_resets` is None, it will be by default `np.array([True])`
which implies the policy will not be "vectorized", i.e. number of
parallel environments for training data sampling = 1.
Args:
do_resets (numpy.ndarray): Bool that indicates terminal state(s).
"""
if do_resets is None:
do_resets = np.array([True])
if self._prev_actions is None or len(do_resets) != len(
self._prev_actions):
self._prev_actions = np.zeros(
(len(do_resets), self.action_space.flat_dim))
self._prev_hiddens = np.zeros((len(do_resets), self._hidden_dim))
self._prev_actions[do_resets] = 0.
self._prev_hiddens[do_resets] = self._init_hidden.eval()
def get_action(self, observation):
"""Get single action from this policy for the input observation.
Args:
observation (numpy.ndarray): Observation from environment.
Returns:
numpy.ndarray: Actions
dict: Predicted action and agent information.
Note:
It returns an action and a dict, with keys
- mean (numpy.ndarray): Mean of the distribution.
- log_std (numpy.ndarray): Log standard deviation of the
distribution.
- prev_action (numpy.ndarray): Previous action, only present if
self._state_include_action is True.
"""
actions, agent_infos = self.get_actions([observation])
return actions[0], {k: v[0] for k, v in agent_infos.items()}
def get_actions(self, observations):
"""Get multiple actions from this policy for the input observations.
Args:
observations (numpy.ndarray): Observations from environment.
Returns:
numpy.ndarray: Actions
dict: Predicted action and agent information.
Note:
It returns an action and a dict, with keys
- mean (numpy.ndarray): Means of the distribution.
- log_std (numpy.ndarray): Log standard deviations of the
distribution.
- prev_action (numpy.ndarray): Previous action, only present if
self._state_include_action is True.
"""
if not isinstance(observations[0],
np.ndarray) or len(observations[0].shape) > 1:
observations = self.observation_space.flatten_n(observations)
if self._state_include_action:
assert self._prev_actions is not None
all_input = np.concatenate([observations, self._prev_actions],
axis=-1)
else:
all_input = observations
means, log_stds, hidden_vec = self._f_step_mean_std(
all_input, self._prev_hiddens)
rnd = np.random.normal(size=means.shape)
samples = rnd * np.exp(log_stds) + means
samples = self.action_space.unflatten_n(samples)
prev_actions = self._prev_actions
self._prev_actions = samples
self._prev_hiddens = hidden_vec
agent_infos = dict(mean=means, log_std=log_stds)
if self._state_include_action:
agent_infos['prev_action'] = np.copy(prev_actions)
return samples, agent_infos
@property
def state_info_specs(self):
"""State info specifcation.
Returns:
List[str]: keys and shapes for the information related to the
policy's state when taking an action.
"""
if self._state_include_action:
return [
('prev_action', (self._action_dim, )),
]
return []
@property
def env_spec(self):
"""Policy environment specification.
Returns:
garage.EnvSpec: Environment specification.
"""
return self._env_spec
def clone(self, name):
"""Return a clone of the policy.
It copies the configuration of the primitive and also the parameters.
Args:
name (str): Name of the newly created policy. It has to be
different from source policy if cloned under the same
computational graph.
Returns:
garage.tf.policies.GaussianGRUPolicy: Newly cloned policy.
"""
new_policy = self.__class__(
name=name,
env_spec=self._env_spec,
hidden_dim=self._hidden_dim,
hidden_nonlinearity=self._hidden_nonlinearity,
hidden_w_init=self._hidden_w_init,
hidden_b_init=self._hidden_b_init,
recurrent_nonlinearity=self._recurrent_nonlinearity,
recurrent_w_init=self._recurrent_w_init,
output_nonlinearity=self._output_nonlinearity,
output_w_init=self._output_w_init,
output_b_init=self._output_b_init,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self._hidden_state_init_trainable,
learn_std=self._learn_std,
std_share_network=self._std_share_network,
init_std=self._init_std,
layer_normalization=self._layer_normalization,
state_include_action=self._state_include_action)
new_policy.parameters = self.parameters
return new_policy
def __getstate__(self):
"""Object.__getstate__.
Returns:
dict: the state to be pickled for the instance.
"""
new_dict = super().__getstate__()
del new_dict['_f_step_mean_std']
del new_dict['_init_hidden']
return new_dict
def __setstate__(self, state):
"""Object.__setstate__.
Args:
state (dict): Unpickled state.
"""
super().__setstate__(state)
self._initialize_policy()