-
Notifications
You must be signed in to change notification settings - Fork 1
/
architecture.py
236 lines (173 loc) · 11.2 KB
/
architecture.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
import tensorflow as tf
import numpy as np
class Actor:
def __init__(self, num_inputs, h_size, name):
"""Constructor of Actor class.
Args:
num_inputs (int): Number of the inputs of the actor
h_size (int): Size of the LSTM output.
name (str): name of the context.
"""
self.num_variables = 10
self.cell = tf.contrib.rnn.BasicLSTMCell(num_units=h_size, state_is_tuple=True)
# Input
self.inp = tf.placeholder(shape=[None, num_inputs], dtype=tf.float32)
self.initializer = tf.contrib.layers.xavier_initializer()
# LSTM
self.batch_size = tf.placeholder(dtype=tf.int32, shape=[])
self.train_length = tf.placeholder(dtype=tf.int32)
self.rnn_inp = tf.reshape(self.inp, [self.batch_size, self.train_length, num_inputs])
self.state_in = self.cell.zero_state(self.batch_size, tf.float32)
self.rnn, self.rnn_state = tf.nn.dynamic_rnn(inputs=self.rnn_inp, cell=self.cell,
dtype=tf.float32, initial_state=self.state_in, scope=name+'_rnn')
self.rnn = tf.reshape(self.rnn, shape=[-1, h_size])
# MLP
self.b1 = tf.Variable(self.initializer([1, 1000]))
self.W1 = tf.Variable(self.initializer([h_size, 1000]))
self.h1 = tf.nn.relu(tf.matmul(self.rnn, self.W1) + self.b1)
self.b2 = tf.Variable(self.initializer([1, 100]))
self.W2 = tf.Variable(self.initializer([1000, 100]))
self.h2 = tf.nn.relu(tf.matmul(self.h1, self.W2) + self.b2)
self.b3 = tf.Variable(self.initializer([1, 50]))
self.W3 = tf.Variable(self.initializer([100, 50]))
self.h3 = tf.nn.relu(tf.matmul(self.h2, self.W3) + self.b3)
self.b4 = tf.Variable(self.initializer([1, 1]))
self.W4 = tf.Variable(self.initializer([50, 1]))
self.a_unscaled = tf.nn.tanh(tf.matmul(self.h3, self.W4)+self.b4)
self.a = tf.multiply(self.a_unscaled, .05)
# Gradients
self.network_params = tf.trainable_variables()[-self.num_variables:]
self.critic_gradient = tf.placeholder(tf.float32, [None, 1])
self.unnormalized_actor_gradients = tf.gradients(self.a, self.network_params, - self.critic_gradient)
self.actor_gradients = list(map(lambda x: tf.div(x, 32), self.unnormalized_actor_gradients))
# Optimization
self.optimizer = tf.train.AdamOptimizer(1e-4)
self.upd = self.optimizer.apply_gradients(zip(self.actor_gradients, self.network_params))
self.update_parameters = []
def create_op_holder(self, params, tau):
""" Use target network op holder if needed"""
self.update_parameters = [self.network_params[i].assign(tf.multiply(params[i], tau) +
tf.multiply(self.network_params[i], 1. - tau))
for i in range(len(self.network_params))]
class Critic:
def __init__(self, num_inputs, h_size, name):
"""Constructor of Critic class.
Args:
num_inputs (int): Number of the inputs of the critic
h_size (int): Size of the LSTM output.
name (str): name of the context.
"""
self.num_variables = 10
self.cell = tf.contrib.rnn.BasicLSTMCell(num_units=h_size, state_is_tuple=True)
# Input
self.inp = tf.placeholder(shape=[None, num_inputs], dtype=tf.float32)
self.initializer = tf.contrib.layers.xavier_initializer()
# LSTM
self.batch_size = tf.placeholder(dtype=tf.int32, shape=[])
self.train_length = tf.placeholder(dtype=tf.int32)
self.rnn_inp = tf.reshape(self.inp, [self.batch_size, self.train_length, num_inputs])
self.state_in = self.cell.zero_state(self.batch_size, tf.float32)
self.rnn, self.rnn_state = tf.nn.dynamic_rnn(inputs=self.rnn_inp, cell=self.cell,
dtype=tf.float32, initial_state=self.state_in, scope=name+'_rnn')
self.rnn = tf.reshape(self.rnn, shape=[-1, h_size])
# MLP
self.b1 = tf.Variable(self.initializer([1, 1000]))
self.W1 = tf.Variable(self.initializer([h_size, 1000]))
self.h1 = tf.nn.relu(tf.matmul(self.rnn, self.W1) + self.b1)
self.b2 = tf.Variable(self.initializer([1, 100]))
self.W2 = tf.Variable(self.initializer([1000, 100]))
self.h2 = tf.nn.relu(tf.matmul(self.h1, self.W2) + self.b2)
self.b3 = tf.Variable(self.initializer([1, 50]))
self.W3 = tf.Variable(self.initializer([100, 50]))
self.h3 = tf.nn.relu(tf.matmul(self.h2, self.W3) + self.b3)
self.b4 = tf.Variable(self.initializer([1, 1]))
self.W4 = tf.Variable(self.initializer([50, 1]))
self.q = tf.matmul(self.h3, self.W4) + self.b4
# Gradients
self.network_params = tf.trainable_variables()[-self.num_variables:]
self.target_q = tf.placeholder(tf.float32, [None, 1])
# Optimization
self.loss = tf.reduce_mean(tf.square(self.target_q-self.q))
self.optimizer = tf.train.AdamOptimizer(1e-4)
self.upd = self.optimizer.minimize(self.loss)
# Gradients
self.critic_gradients = tf.gradients(self.q, self.inp)
self.update_parameters = []
def create_op_holder(self, params, tau):
""" Use target network op holder if needed"""
self.update_parameters = [self.network_params[i].assign(tf.multiply(params[i], tau) +
tf.multiply(self.network_params[i], 1. - tau))
for i in range(len(self.network_params))]
class Agent:
def __init__(self, a_dof, c_dof, h_size, name, batch_size, train_length, tau):
"""Constructor of Agent class. Each agent is composed of the main actor-critic pair and the target actor-critic
pair. Target architectures help stabilizing the training of the agents.
Args:
h_size (int): Size of the LSTM output.
name (str): name of the context.
"""
self.actor = Actor(a_dof, h_size, name+'_actor')
self.critic = Critic(c_dof, h_size, name+'_critic')
self.actor_target = Actor(a_dof, h_size, name+'_actor_target')
self.critic_target = Critic(c_dof, h_size, name+'_critic_target')
self.batch_size = batch_size
self.train_length = train_length
self.state_train = (np.zeros([self.batch_size, h_size]), np.zeros([self.batch_size, h_size]))
self.actor_target.create_op_holder(self.actor.network_params, tau)
self.critic_target.create_op_holder(self.critic.network_params, tau)
def a_actor_operation(self, session, inp, state):
return session.run([self.actor.a, self.actor.rnn_state],
feed_dict={self.actor.inp: inp,
self.actor.state_in: state,
self.actor.batch_size: 1,
self.actor.train_length: 1})
def a_actor_training(self, session, inp):
return session.run(self.actor.a, feed_dict={self.actor.inp: inp,
self.actor.state_in: self.state_train,
self.actor.batch_size: self.batch_size,
self.actor.train_length: self.train_length})
def a_target_actor_training(self, session, inp):
return session.run(self.actor_target.a, feed_dict={self.actor_target.inp: inp,
self.actor_target.state_in: self.state_train,
self.actor_target.batch_size: self.batch_size,
self.actor_target.train_length: self.train_length})
def q_target_critic(self, session, inp):
return session.run(self.critic_target.q, feed_dict={self.critic_target.inp: inp,
self.critic_target.train_length: self.train_length,
self.critic_target.batch_size: self.batch_size,
self.critic_target.state_in: self.state_train})
def gradients_critic(self, session, inp):
return session.run(self.critic.critic_gradients, feed_dict={self.critic.inp: inp,
self.critic.train_length: self.train_length,
self.critic.batch_size: self.batch_size,
self.critic.state_in: self.state_train})
def update_critic(self, session, inp, q):
session.run(self.critic.upd, feed_dict={self.critic.inp: inp,
self.critic.target_q: q,
self.critic.train_length: self.train_length,
self.critic.batch_size: self.batch_size,
self.critic.state_in: self.state_train})
def update_actor(self, session, inp, gradients):
session.run(self.actor.upd, feed_dict={self.actor.inp: inp,
self.actor.state_in: self.state_train,
self.actor.critic_gradient: gradients,
self.actor.batch_size: self.batch_size,
self.actor.train_length: self.train_length})
def update_targets(self, session):
session.run(self.actor_target.update_parameters)
session.run(self.critic_target.update_parameters)
def initialize_targets(self, session):
session.run([self.actor_target.network_params[i].assign(self.actor.network_params[i])
for i in range(len(self.actor.network_params))])
session.run([self.critic_target.network_params[i].assign(self.critic.network_params[i])
for i in range(len(self.critic.network_params))])
def importance(self, session, inp_q, inp_q_target, r, gamma, state, new_state):
act_q = session.run(self.critic_target.q, feed_dict={self.critic_target.inp: inp_q,
self.critic_target.train_length: 1,
self.critic_target.batch_size: 1,
self.critic_target.state_in: state})[0, 0]
tar_q = session.run(self.critic_target.q, feed_dict={self.critic_target.inp: inp_q_target,
self.critic_target.train_length: 1,
self.critic_target.batch_size: 1,
self.critic_target.state_in: new_state})[0, 0]
return np.abs(r + gamma*tar_q - act_q)