-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathcritic_network.py
146 lines (118 loc) · 5.92 KB
/
critic_network.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
import tensorflow as tf
import numpy as np
import math
LAYER1_SIZE = 300
LAYER2_SIZE = 600
LEARNING_RATE = 1e-3
TAU = 0.001
L2 = 0.0001
class CriticNetwork:
"""docstring for CriticNetwork"""
def __init__(self, sess, state_dim, action_dim, alextmodel_critic):
self.time_step = 0
self.sess = sess
# create q network
self.state_input, \
self.action_input, \
self.image_input, \
self.q_value_output, \
self.net = self.create_q_network(state_dim, action_dim, alextmodel_critic)
# create target q network (the same structure with q network)
self.target_state_input, \
self.target_action_input, \
self.target_image_input, \
self.target_q_value_output, \
self.target_update = self.create_target_q_network(state_dim, action_dim, self.net, alextmodel_critic)
self.create_training_method()
# initialization
self.sess.run(tf.initialize_all_variables())
self.update_target()
def create_training_method(self):
# Define training optimizer
self.y_input = tf.placeholder("float", [None, 1])
weight_decay = tf.add_n([L2 * tf.nn.l2_loss(var) for var in self.net])
self.cost = tf.reduce_mean(tf.square(self.y_input - self.q_value_output)) + weight_decay
self.optimizer = tf.train.AdamOptimizer(LEARNING_RATE).minimize(self.cost)
'''
self.optimizer = tf.train.AdamOptimizer(LEARNING_RATE)
self.parameters_gradients = self.optimizer.compute_gradients(self.cost)
for i, (grad,var) in enumerate(self.parameters_gradients):
if grad is not None:
self.parameters_gradients[i] = (tf.clip_by_value(grad, -5.0,5.0),var)
self.train_op = self.optimizer.apply_gradients(self.parameters_gradients)
'''
self.action_gradients = tf.gradients(self.q_value_output, self.action_input)
def create_q_network(self, state_dim, action_dim, alextmodel_critic):
# the layer size could be changed
layer1_size = LAYER1_SIZE
layer2_size = LAYER2_SIZE
state_input = tf.placeholder("float", [None, state_dim])
action_input = tf.placeholder("float", [None, action_dim])
image_input = tf.placeholder(tf.float32, shape=[None, 64, 64, 3])
cnn_output = alextmodel_critic(image_input)
W1 = self.variable([state_dim, layer1_size], state_dim)
b1 = self.variable([layer1_size], state_dim)
W2 = self.variable([layer1_size, layer2_size], layer1_size + action_dim)
W2_action = self.variable([action_dim, layer2_size], layer1_size + action_dim)
b2 = self.variable([layer2_size], layer1_size + action_dim)
W3 = tf.Variable(tf.random_uniform([layer2_size + 512, 1], -3e-3, 3e-3))
b3 = tf.Variable(tf.random_uniform([1], -3e-3, 3e-3))
layer1 = tf.nn.relu(tf.matmul(state_input, W1) + b1)
layer2 = tf.nn.relu(tf.matmul(layer1, W2) + tf.matmul(action_input, W2_action) + b2)
layer2 = tf.concat([layer2, cnn_output], 1)
q_value_output = tf.identity(tf.matmul(layer2, W3) + b3)
return state_input, action_input, image_input, q_value_output, [W1, b1, W2, W2_action, b2, W3, b3]
def create_target_q_network(self, state_dim, action_dim, net, alextmodel_critic):
state_input = tf.placeholder("float", [None, state_dim])
action_input = tf.placeholder("float", [None, action_dim])
image_input = tf.placeholder(tf.float32, shape=[None, 64, 64, 3])
cnn_output = alextmodel_critic(image_input)
a_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='critic')
net.extend(a_params)
ema = tf.train.ExponentialMovingAverage(decay=1 - TAU)
target_update = ema.apply(net)
target_net = [ema.average(x) for x in net]
layer1 = tf.nn.relu(tf.matmul(state_input, target_net[0]) + target_net[1])
layer2 = tf.nn.relu(tf.matmul(layer1, target_net[2]) + tf.matmul(action_input, target_net[3]) + target_net[4])
layer2 = tf.concat([layer2, cnn_output], 1)
q_value_output = tf.identity(tf.matmul(layer2, target_net[5]) + target_net[6])
return state_input, action_input,image_input, q_value_output, target_update
def update_target(self):
self.sess.run(self.target_update)
def train(self, y_batch, state_batch, action_batch):
self.time_step += 1
self.sess.run(self.optimizer, feed_dict={
self.y_input: y_batch,
self.state_input: state_batch,
self.action_input: action_batch
})
def gradients(self, state_batch, action_batch):
return self.sess.run(self.action_gradients, feed_dict={
self.state_input: state_batch,
self.action_input: action_batch
})[0]
def target_q(self, state_batch, action_batch):
return self.sess.run(self.target_q_value_output, feed_dict={
self.target_state_input: state_batch,
self.target_action_input: action_batch
})
def q_value(self, state_batch, action_batch):
return self.sess.run(self.q_value_output, feed_dict={
self.state_input: state_batch,
self.action_input: action_batch})
# f fan-in size
def variable(self, shape, f):
return tf.Variable(tf.random_uniform(shape, -1 / math.sqrt(f), 1 / math.sqrt(f)))
'''
def load_network(self):
self.saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state("saved_critic_networks")
if checkpoint and checkpoint.model_checkpoint_path:
self.saver.restore(self.sess, checkpoint.model_checkpoint_path)
print "Successfully loaded:", checkpoint.model_checkpoint_path
else:
print "Could not find old network weights"
def save_network(self,time_step):
print 'save critic-network...',time_step
self.saver.save(self.sess, 'saved_critic_networks/' + 'critic-network', global_step = time_step)
'''