-
Notifications
You must be signed in to change notification settings - Fork 0
/
training_thread.py
342 lines (227 loc) · 9.25 KB
/
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
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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
# -*- coding: utf-8 -*-
import tensorflow as tf
import numpy as np
import random
import time
import sys
import pdb
from utils.accum_trainer import AccumTrainer
from scene_loader import THORDiscreteEnvironment as Environment
from network import ActorCriticFFNetwork
from constants import ACTION_SIZE
from constants import GAMMA
from constants import LOCAL_T_MAX
from constants import ENTROPY_BETA
from constants import VERBOSE
class A3CTrainingThread(object):
def __init__(self,
thread_index,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
max_global_time_step,
device,
network_scope="network",
scene_scope="scene",
task_scope="task"):
self.thread_index = thread_index
self.learning_rate_input = learning_rate_input
self.max_global_time_step = max_global_time_step
self.network_scope = network_scope
self.scene_scope = scene_scope
self.task_scope = task_scope
self.scopes = [network_scope, scene_scope, task_scope]
self.local_network = ActorCriticFFNetwork(
action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=[scene_scope])
self.local_network.prepare_loss(ENTROPY_BETA, self.scopes)
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()
accum_grad_names = [self._local_var_name(x) for x in self.trainer.get_accum_grad_list()]
global_net_vars = [x for x in global_network.get_vars() if self._get_accum_grad_name(x) in accum_grad_names]
self.apply_gradients = grad_applier.apply_gradients(
global_net_vars, self.trainer.get_accum_grad_list() )
self.sync = self.local_network.sync_from(global_network)
self.env = None
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
def _local_var_name(self, var):
return '/'.join(var.name.split('/')[1:])
def _get_accum_grad_name(self, var):
return self._local_var_name(var).replace(':','_') + '_accum_grad:0'
def _anneal_learning_rate(self, global_time_step):
time_step_to_go = max(self.max_global_time_step - global_time_step, 0.0)
learning_rate = self.initial_learning_rate * time_step_to_go / self.max_global_time_step
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, writer, summary_op, placeholders, values, global_t):
feed_dict = {}
for k in placeholders:
feed_dict[placeholders[k]] = values[k]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
if VERBOSE: sys.stdout.write('writing to summary writer at time %d\n' % (global_t))
writer.add_summary(summary_str, global_t)
# writer.flush()
def process(self, sess, global_t, summary_writer, summary_op, summary_placeholders):
if self.env is None:
# lazy evaluation
time.sleep(self.thread_index*1.0)
self.env = Environment({
'scene_name': self.scene_scope,
'terminal_state_id': int(self.task_scope)
})
states = []
actions = []
rewards = []
values = []
targets = []
rnn_inits=[]
state_representation=[]
usf=[]
reward_vector=[]
terminal_end = False
# reset accumulated gradients
sess.run( self.reset_gradients )
# copy weights from shared to local
sess.run( self.sync )
#At each episode start we set the initial state of the RNN to zero
start_local_t = self.local_t
start_lstm_state = self.local_network.lstm_state_out
# t_max times loop
for i in range(LOCAL_T_MAX):
pi_, value_ ,usf_s_g= self.local_network.run_policy_and_value(sess, self.env.s_t,self.env.target,self.scopes)
imidia_s = self.local_network.run_state(sess, self.env.s_t,self.scopes)
#usf_s_g = self.local_network.run_usf(sess, self.env.s_t, self.env.target,self.rnn_state_init[0] ,self.rnn_state_init[1] ,self.scopes)
action = self.choose_action(pi_)
states.append(self.env.s_t)
actions.append(action)
values.append(value_)
targets.append(self.env.target)
usf.append(usf_s_g)
state_representation.append(imidia_s)
if VERBOSE and (self.thread_index == 0) and (self.local_t % 1000) == 0:
sys.stdout.write("Pi = {0} V = {1}\n".format(pi_, value_))
# process game
self.env.step(action)
# receive game result
reward = self.env.reward
terminal = self.env.terminal
# ad-hoc reward for navigation
reward = 10.0 if terminal else -0.01
if self.episode_length > 5e3: terminal = True
self.episode_reward += reward
self.episode_length += 1
self.episode_max_q = max(self.episode_max_q, np.max(value_))
# clip reward
rewards.append( np.clip(reward, -1, 1) )
self.local_t += 1
# s_t1 -> s_t
self.env.update()
if i==(LOCAL_T_MAX-1)or terminal :
imidiate_state_representation_next=[]
usf_next=[]
#reward_vector_predictor_next=[]
last_state=self.env.s_t
imidia_s_next = self.local_network.run_state(sess, self.env.s_t,self.scopes)
state_representation_next = state_representation[1:] + [imidia_s_next]
if terminal:
usf_next_imi=0
else:
usf_next_imi=self.local_network.run_usf(sess, self.env.s_t, self.env.target,self.scopes)
usf_next=usf[1:] + [usf_next_imi]
if terminal:
terminal_end = True
sys.stdout.write("time %d | thread #%d | scene %s | target #%s\n%s %s episode reward = %.3f\n%s %s episode length = %d\n%s %s episode max Q = %.3f\n" % (global_t, self.thread_index, self.scene_scope, self.task_scope, self.scene_scope, self.task_scope, self.episode_reward, self.scene_scope, self.task_scope, self.episode_length, self.scene_scope, self.task_scope, self.episode_max_q))
summary_values = {
"episode_reward_input": self.episode_reward,
"episode_length_input": float(self.episode_length),
"episode_max_q_input": self.episode_max_q,
"learning_rate_input": self._anneal_learning_rate(global_t)
}
self._record_score(sess, summary_writer, summary_op, summary_placeholders,
summary_values, global_t)
self.episode_reward = 0
self.episode_length = 0
self.episode_max_q = -np.inf
self.local_network.reset_state()
self.env.reset()
break
R = 0.0
usf_R = 0.0
if not terminal_end:
R = self.local_network.run_value(sess, self.env.s_t, self.env.target,self.scopes)
usf_R = self.local_network.run_usf(sess, self.env.s_t, self.env.target ,self.scopes)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
state_representation.reverse()
state_representation_next.reverse()
usf_next.reverse()
batch_si = []
batch_a = []
batch_td = []
batch_R = []
batch_usf_R = []
batch_t = []
# compute and accmulate gradients
for(ai, ri, si, Vi, ti,state,usf_n) in zip(actions, rewards, states, values, targets,state_representation_next,usf_next):
R = ri + GAMMA * R
usf_R = state + GAMMA*usf_R
#usf_R = state + GAMMA*usf_n
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)
batch_usf_R.append(usf_R)
batch_t.append(ti)
#We need to reverse this since in the training we unroll for 5 steps unlike in the inferences
batch_si.reverse()
batch_a.reverse()
batch_td.reverse()
batch_R.reverse()
batch_usf_R.reverse()
batch_t.reverse()
sess.run( self.accum_gradients,
feed_dict = {
self.local_network.s: batch_si,
self.local_network.a: batch_a,
self.local_network.t: batch_t,
self.local_network.td: batch_td,
self.local_network.r: batch_R,
self.local_network.return_usf: batch_usf_R,
self.local_network.initial_lstm_state: start_lstm_state,
self.local_network.step_size : [len(batch_a)],
} )
cur_learning_rate = self._anneal_learning_rate(global_t)
sess.run( self.apply_gradients,
feed_dict = { self.learning_rate_input: cur_learning_rate } )
if VERBOSE and (self.thread_index == 0) and (self.local_t % 100) == 0:
sys.stdout.write("Local timestep %d\n" % self.local_t)
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t