forked from miyosuda/async_deep_reinforce
-
Notifications
You must be signed in to change notification settings - Fork 0
/
a3c_training_thread.py
184 lines (137 loc) · 4.92 KB
/
a3c_training_thread.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
# -*- coding: utf-8 -*-
import tensorflow as tf
import numpy as np
import random
from accum_trainer import AccumTrainer
from game_state import GameState
from game_state import ACTION_SIZE
from game_ac_network import GameACNetwork
from constants import GAMMA
from constants import LOCAL_T_MAX
from constants import ENTROPY_BETA
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.local_network = GameACNetwork(ACTION_SIZE, device)
self.local_network.prepare_loss(ENTROPY_BETA)
# TODO: don't need accum trainer anymore with batch
self.trainer = AccumTrainer(device)
self.trainer.prepare_minimize( self.local_network.total_loss,
self.local_network.get_vars() )
self.accum_gradients = self.trainer.accumulate_gradients()
self.reset_gradients = self.trainer.reset_gradients()
self.apply_gradients = grad_applier.apply_gradients(
global_network.get_vars(),
self.trainer.get_accum_grad_list() )
self.sync = self.local_network.sync_from(global_network)
self.game_state = GameState(113 * thread_index)
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
def _anneal_learning_rate(self, global_time_step):
learning_rate = self.initial_learning_rate * (self.max_global_time_step - global_time_step) / self.max_global_time_step
if learning_rate < 0.0:
learning_rate = 0.0
return learning_rate
def choose_action(self, pi_values):
values = []
sum = 0.0
for rate in pi_values:
sum = sum + rate
value = sum
values.append(value)
r = random.random() * sum
for i in range(len(values)):
if values[i] >= r:
return i;
#fail safe
return len(values)-1
def _record_score(self, sess, summary_writer, summary_op, score_input, score, global_t):
summary_str = sess.run(summary_op, feed_dict={
score_input: score
})
summary_writer.add_summary(summary_str, global_t)
def process(self, sess, global_t, summary_writer, summary_op, score_input):
states = []
actions = []
rewards = []
values = []
terminal_end = False
# reset accumulated gradients
sess.run( self.reset_gradients )
# copy weights from shared to local
sess.run( self.sync )
start_local_t = self.local_t
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ = self.local_network.run_policy_and_value(sess, self.game_state.s_t)
action = self.choose_action(pi_)
states.append(self.game_state.s_t)
actions.append(action)
values.append(value_)
if (self.thread_index == 0) and (self.local_t % 100) == 0:
print "pi=", pi_
print " V=", value_
# process game
self.game_state.process(action)
# receive game result
reward = self.game_state.reward
terminal = self.game_state.terminal
self.episode_reward += reward
# clip reward
rewards.append( np.clip(reward, -1, 1) )
self.local_t += 1
# s_t1 -> s_t
self.game_state.update()
if terminal:
terminal_end = True
print "score=", self.episode_reward
self._record_score(sess, summary_writer, summary_op, score_input,
self.episode_reward, global_t)
self.episode_reward = 0
self.game_state.reset()
break
R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.game_state.s_t)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
# compute and accmulate gradients
for(ai, ri, si, Vi) in zip(actions, rewards, states, values):
R = ri + GAMMA * R
td = R - Vi
a = np.zeros([ACTION_SIZE])
a[ai] = 1
batch_si.append(si)
batch_a.append(a)
batch_td.append(td)
batch_R.append(R)
sess.run( self.accum_gradients,
feed_dict = {
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.td: batch_td,
self.local_network.r: batch_R } )
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run( self.apply_gradients,
feed_dict = { self.learning_rate_input: cur_learning_rate } )
if (self.thread_index == 0) and (self.local_t % 100) == 0:
print "TIMESTEP", self.local_t
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t