-
Notifications
You must be signed in to change notification settings - Fork 26
/
A3C_advantage_async_actor_critic.py
168 lines (151 loc) · 6.94 KB
/
A3C_advantage_async_actor_critic.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
#!/usr/bin/python
# -*- coding: utf-8 -*-
# author: [email protected]
import threading
import numpy as np
from common import *
class A3CModel(object):
"""
Advantage async actor-critic model.
ref:
paper "asynchronous methods for deep reinforcement learning"
"""
def __init__(self, actions=225, train_dir="./a3c_models", gpu_id=0):
self.action_num = actions
self.learn_rate = 1e-4
self.train_dir = train_dir
self.gpu_id = gpu_id
if not os.path.isdir(self.train_dir):
os.mkdir(self.train_dir)
self.sess, self.saver, self.graph_ops = self.build_graph()
def policy_model(self, _input, shape, board_size=15):
fc = full_connect(_input, (shape[0], board_size * board_size), "fc_p", activate=None)
softmax_linear = tf.nn.softmax(fc)
return softmax_linear
def value_model(self, _input, shape):
fc = full_connect(_input, (shape[0], 1), "fc_v", activate=None)
return fc
def inference(self, _input):
# first conv1
conv1 = conv2d(_input, (5, 5, 3, 32), "conv_1", stride=1)
# norm1
norm1 = tf.nn.lrn(conv1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm_1')
# conv2 ~ conv_k
pre_layer = norm1
for i in xrange(5):
conv_k = conv2d(pre_layer, (3, 3, 32, 32), "conv_%d" % (i + 2), stride=1)
norm2 = tf.nn.lrn(conv_k, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm_%d' % (i + 2))
pre_layer = norm2
# last layer
conv_n = conv2d(pre_layer, (1, 1, 32, 32), "conv_n", stride=1)
norm_n = tf.nn.lrn(conv_n, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm_n')
reshape = tf.reshape(norm_n, [-1, 225 * 32])
# dim = reshape.get_shape()[1].value
logits = full_connect(reshape, (225 * 32, 1024), "fc_1")
return logits
def build_graph(self):
with tf.Graph().as_default(), tf.device('/gpu:%d' % self.gpu_id):
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0), trainable=False)
# init placeholder
state_ph = tf.placeholder(tf.float32, [None, 15, 15, 3])
action_ph = tf.placeholder(tf.float32, shape=[None, self.action_num])
target_ph = tf.placeholder(tf.float32, shape=[None])
# create model
shared_dim = 1024
shared = self.inference(state_ph, 3, out_dim=shared_dim)
policy_out = self.policy_model(shared, [shared_dim])
value_out = self.value_model(shared, [shared_dim])
# calculate loss
policy_loss = tf.reduce_mean(
-tf.log(tf.reduce_mean(tf.mul(policy_out, action_ph), reduction_indices=1)) * tf.square(
target_ph - value_out))
value_loss = tf.reduce_mean(tf.square(target_ph - value_out))
l2_loss = tf.add_n(tf.get_collection('losses'), name='l2_loss')
total_loss = policy_loss + value_loss + l2_loss
# optimizer
optimizer = tf.train.AdamOptimizer(self.learn_rate).minimize(total_loss, global_step=global_step)
# optimizer = tf.train.GradientDescentOptimizer(learn_rate).minimize(total_loss, global_step=global_step)
saver = tf.train.Saver()
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=False)
)
sess.run(init)
# restore model
restore_model(sess, self.train_dir, saver)
graph_ops = state_ph, action_ph, target_ph, optimizer, policy_out, value_out, total_loss, global_step
return sess, saver, graph_ops
def train(self, thread_num=3):
envs = [Environment() for _ in xrange(thread_num)]
actor_learner_threads = []
t_max, gamma = 64, 0.99
constants = self.action_num, t_max, gamma, self.train_dir
for idx in xrange(envs):
_thread = threading.Thread(target=actor_learner_thread,
args=(idx, envs[idx], self.sess, self.saver, self.graph_ops, constants))
_thread.start()
actor_learner_threads.append(_thread)
for _thread in actor_learner_threads:
_thread.join()
def sample_action(probs):
probs = probs - np.finfo(np.float32).epsneg
histogram = np.random.multinomial(1, probs)
action_index = int(np.nonzero(histogram)[0])
return action_index
def actor_learner_thread(thread_id, env, sess, saver, graph_ops, constants):
state_ph, action_ph, target_ph, optimizer, policy_out, value_out, total_loss, global_step = graph_ops
action_num, t_max, gamma, train_dir = constants
terminal = False
episode_reward = 0
episode_step = 0
episode_loss = []
episode_count = 0
state = env.get_state()
start_time = time.time()
while True:
states, actions, rewards = [], [], []
t_start = episode_step
while not terminal and ((episode_step - t_start) < t_max):
probs = sess.run([policy_out], feed_dict={state_ph: [state]})[0][0]
action = sample_action(probs)
state_n, reward_n, terminal = env.step_forward(action)
one_hot_action = np.zeros(action_num)
one_hot_action[action] = 1
states.append(state)
actions.append(one_hot_action)
rewards.append(reward_n)
state = state_n
episode_step += 1
episode_reward += reward_n
rewards_R = np.zeros(len(rewards))
if not terminal:
R = sess.run([value_out], feed_dict={state_ph: [state]})[0][0]
else:
R = 0
for t_idx in xrange(episode_step - t_start - 1, -1, -1):
R = rewards[t_idx] + gamma * R
rewards_R[t_idx] = R
_, loss, gt = sess.run([optimizer, total_loss, global_step], feed_dict={state_ph: states, action_ph: actions,
target_ph: rewards_R})
episode_loss.append(float(loss))
gt = int(gt)
if gt % 500 == 0:
save_model(sess, train_dir, saver, "policy_rl_a3c", global_step=gt)
if terminal:
elapsed_time = int(time.time() - start_time)
env.reset()
state = env.get_state()
terminal = False
episode_loss = sum(episode_loss) / len(episode_loss)
episode_count += 1
logger.info("thread=%d, T=%d, episode(count=%d, step=%d, loss=%.5f, reward=%d), time=%d(s)" %
(thread_id, gt, episode_count, episode_step, episode_loss, episode_reward, elapsed_time))
start_time = time.time()
episode_reward = 0
episode_step = 0
episode_loss = []
if __name__ == "__main__":
model = A3CModel()
model.train(thread_num=3)