-
Notifications
You must be signed in to change notification settings - Fork 5
/
models.py
220 lines (181 loc) · 8.37 KB
/
models.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
from tensorflow.keras.layers import Input, Dense, Conv2D, MaxPool2D, ZeroPadding2D, Flatten
from tensorflow import keras
import tensorflow as tf
import numpy as np
class BaseModel():
"""
Base class used to implement a model.
This class itself should not instantiated.
"""
def __init__(self, comm, controller, rank, n_acts, obs_shape, sess_config=None):
if sess_config is None:
self.sess = tf.Session()
else:
self.sess = tf.Session(config=sess_config)
### Initialize MPI variables ###
self.comm = comm
self.controller = controller
self.rank = rank
### Initialize placeholder and other network ops ###
self.sess_config = sess_config
self.obs_shape = obs_shape
self.n_acts = n_acts
self.update_policy = self.train_policy
self.create_phs(obs_shape=self.obs_shape)
self.optimizer = tf.train.AdamOptimizer(learning_rate=1e-3)
self.create_policy_ops()
self.sess.run(tf.global_variables_initializer())
# Sync the weights of models on all processes
self.sync_weights()
def create_phs(self, obs_shape):
"""
Creates placeholders (input ops) for the model.
"""
self.obs_ph = tf.placeholder(
dtype=tf.float32, shape=(None, *list(obs_shape)))
self.next_obs_ph = tf.placeholder(
dtype=tf.float32, shape=(None, *list(obs_shape)))
self.act_ph = tf.placeholder(dtype=tf.int32, shape=(None,))
self.rew_ph = tf.placeholder(dtype=tf.float32, shape=(None,))
self.gae_ph = tf.placeholder(dtype=tf.float32, shape=(None,))
def create_policy_ops(self):
"""
Creates the ops for the policy and value network.
Additionally creates the ops for updating the network.
This function should be overridden by subclasses.
"""
self.act_out = 0
self.value_op = 0
self.policy_update = 0
self.value_update = 0
def gen_actions(self, obs):
"""
Generates actions (1 to 1) for each observation passed
into the function.
"""
if type(obs) == list:
obs = np.asarray(obs)
assert (type(obs) == np.ndarray), \
"Observations must be a numpy array!"
assert (list(obs.shape)[1:] == list(self.obs_shape)), \
"Observations must have the shape, (batch_size, dimensions..., 1)!"
return self.sess.run(self.act_out, feed_dict={self.obs_ph: obs})[0]
def gen_actions_and_values(self, obs):
"""
Generates actions and values (1 to 1) for each observation
passed into the function.
"""
if type(obs) == list:
obs = np.asarray(obs)
assert (type(obs) == np.ndarray), \
"Observations must be a numpy array!"
assert (list(obs.shape)[1:] == list(self.obs_shape)), \
"Observations must have the shape, (batch_size, dimensions..., 1)!"
acts, vals = self.sess.run(
[self.act_out, self.value_op], feed_dict={self.obs_ph: obs})
return acts[0], vals[0]
def train_policy(self, states, actions, rewards, gaes):
"""
Updates the policy given training data from
environment simulation.
"""
self.sess.run([self.policy_update, self.value_update],
feed_dict={self.obs_ph: states,
self.act_ph: actions,
self.rew_ph: rewards,
self.gae_ph: gaes})
def sync_weights(self):
"""
Sync the weights between model on all processes
using MPI.
"""
if self.rank == self.controller:
self.comm.bcast(self.sess.run(
tf.trainable_variables()), self.controller)
else:
sync_vars = self.comm.bcast(None, self.controller)
t_vars = tf.trainable_variables()
for pair in zip(t_vars, sync_vars):
self.sess.run(tf.assign(pair[0], pair[1]))
class OneDimModel(BaseModel):
"""
Vanilla Policy Gradient implemented for an
environment with 1-dimensional states.
"""
def __init__(self, comm, controller, rank, n_acts, obs_shape, sess_config=None):
super().__init__(comm, controller, rank, n_acts, obs_shape, sess_config)
self.optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
def create_policy_ops(self):
"""
Creates the ops for the policy and value network.
Additionally creates the ops for updating the network.
"""
with tf.variable_scope('policy'):
# Creating a fully connected net for the policy and value estimator
dense_1 = Dense(64, activation='relu')(self.obs_ph)
act_dense_1 = Dense(64, activation='relu')(dense_1)
act_dense_2 = Dense(64, activation='relu')(act_dense_1)
val_dense_1 = Dense(64, activation='relu')(dense_1)
val_dense_2 = Dense(64, activation='relu')(val_dense_1)
# Output probability distribution over possible actions
self.act_probs_op = Dense(
self.n_acts, activation='softmax', name='act_probs')(act_dense_2)
self.act_out = tf.random.categorical(tf.log(self.act_probs_op), 1)
# Output value of observed state
self.value_op = Dense(1, activation='linear')(val_dense_2)
self.act_masks = tf.one_hot(
self.act_ph, self.n_acts, dtype=tf.float32)
self.log_probs = tf.log(self.act_probs_op)
self.resp_acts = tf.reduce_sum(
self.act_masks * self.log_probs, axis=1)
self.policy_loss = \
-tf.reduce_mean(self.resp_acts * self.gae_ph)
self.policy_update = self.optimizer.minimize(self.policy_loss)
# Add dependency on policy update to make sure the value network
# only gets updated after the policy
with tf.control_dependencies([self.policy_update]):
self.value_loss = tf.reduce_mean(
tf.square(self.rew_ph - tf.squeeze(self.value_op)))
self.value_update = self.optimizer.minimize(self.value_loss)
class TwoDimModel(BaseModel):
"""
Vanilla Policy Gradient implemented for an
environment with 2-dimensional states.
"""
def __init__(self, comm, controller, rank, n_acts, obs_shape, sess_config=None):
super().__init__(comm, controller, rank, n_acts, obs_shape, sess_config)
def create_policy_ops(self):
"""
Creates the ops for the policy and value network.
Additionally creates the ops for updating the network.
"""
with tf.variable_scope('policy'):
# Creating a conv net for the policy and value estimator
conv_1 = Conv2D(16, 5, 3, activation='relu')(self.obs_ph)
pooling_1 = MaxPool2D(2)(conv_1)
act_conv_1 = Conv2D(32, 3, 2, activation='relu')(conv_1)
act_pool_1 = MaxPool2D(2)(act_conv_1)
act_flat = Flatten()(act_pool_1)
val_conv_1 = Conv2D(32, 3, 2, activation='relu')(conv_1)
val_pool_1 = MaxPool2D(2)(val_conv_1)
val_flat = Flatten()(val_pool_1)
# Output probability distribution over possible actions
self.act_probs_op = Dense(
self.n_acts, activation='softmax', name='act_probs')(act_flat)
self.act_out = tf.random.categorical(tf.log(self.act_probs_op), 1)
# Output value of observed state
self.value_op = Dense(1, activation='linear')(val_flat)
self.act_masks = tf.one_hot(
self.act_ph, self.n_acts, dtype=tf.float32)
self.log_probs = tf.log(self.act_probs_op)
self.resp_acts = tf.reduce_sum(
self.act_masks * self.log_probs, axis=1)
self.policy_loss = \
-tf.reduce_mean(self.resp_acts * self.gae_ph)
self.policy_update = self.optimizer.minimize(self.policy_loss)
# Add dependency on policy update to make sure the value network
# only gets updated after the policy
with tf.control_dependencies([self.policy_update]):
self.value_loss = tf.reduce_mean(
tf.square(self.rew_ph - tf.squeeze(self.value_op)))
self.value_update = self.optimizer.minimize(self.value_loss)