forked from mihahauke/deep_rl_vizdoom
/
async_learner.py
executable file
·498 lines (422 loc) · 21 KB
/
async_learner.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
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
# -*- coding: utf-8 -*-
import numpy as np
import random
import time
from tqdm import trange
from threading import Thread
from util.coloring import red, green, blue, yellow
from time import strftime
import tensorflow as tf
import sys
import os
from vizdoom import SignalException, ViZDoomUnexpectedExitException
from util import sec_to_str, threadsafe_print
from vizdoom_wrapper import VizdoomWrapper
from util.logger import log
from util.misc import setup_vector_summaries
import networks
import logging
class A3CLearner(Thread):
def __init__(self,
thread_index,
network_type,
global_steps_counter,
scenario_tag=None,
run_id_string=None,
session=None,
tf_logdir=None,
global_network=None,
optimizer=None,
learning_rate=None,
test_only=False,
write_summaries=True,
enable_progress_bar=True,
deterministic_testing=True,
save_interval=1,
writer_max_queue=10,
writer_flush_secs=120,
**settings):
super(A3CLearner, self).__init__()
log("Creating actor-learner #{}.".format(thread_index))
self.thread_index = thread_index
self._global_steps_counter = global_steps_counter
self.write_summaries = write_summaries
self.save_interval = save_interval
self.enable_progress_bar = enable_progress_bar
self._model_savefile = None
self._train_writer = None
self._test_writer = None
self._summaries = None
self._session = session
self.deterministic_testing = deterministic_testing
self.local_steps = 0
# TODO epoch as tf variable?
self._epoch = 1
self.train_scores = []
self.local_steps_per_epoch = settings["local_steps_per_epoch"]
self._run_tests = settings["test_episodes_per_epoch"] > 0 and settings["run_tests"]
self.test_episodes_per_epoch = settings["test_episodes_per_epoch"]
self._epochs = np.float32(settings["epochs"])
self.max_remembered_steps = settings["max_remembered_steps"]
self.gamma = np.float32(settings["gamma"])
if self.write_summaries and thread_index == 0 and not test_only:
assert tf_logdir is not None
self.run_id_string = run_id_string
self.tf_models_path = settings["models_path"]
if not os.path.isdir(tf_logdir):
os.makedirs(tf_logdir)
if self.tf_models_path is not None:
if not os.path.isdir(settings["models_path"]):
os.makedirs(settings["models_path"])
self.doom_wrapper = VizdoomWrapper(**settings)
misc_len = self.doom_wrapper.misc_len
img_shape = self.doom_wrapper.img_shape
self.use_misc = self.doom_wrapper.use_misc
self.actions_num = self.doom_wrapper.actions_num
# TODO add debug log
self.local_network = eval(network_type)(actions_num=self.actions_num, img_shape=img_shape, misc_len=misc_len,
thread=thread_index, **settings)
if not test_only:
self.learning_rate = learning_rate
# TODO check gate_gradients != Optimizer.GATE_OP
grads_and_vars = optimizer.compute_gradients(self.local_network.ops.loss,
var_list=self.local_network.get_params())
grads, local_vars = zip(*grads_and_vars)
grads_and_global_vars = zip(grads, global_network.get_params())
self.train_op = optimizer.apply_gradients(grads_and_global_vars, global_step=tf.train.get_global_step())
self.global_network = global_network
self.local_network.prepare_sync_op(global_network)
if self.thread_index == 0 and not test_only:
self._model_savefile = settings["models_path"] + "/" + self.run_id_string
if self.write_summaries:
self.scores_placeholder, summaries = setup_vector_summaries(scenario_tag + "/scores")
lr_summary = tf.summary.scalar(scenario_tag + "/learning_rate", self.learning_rate)
summaries.append(lr_summary)
self._summaries = tf.summary.merge(summaries)
self._train_writer = tf.summary.FileWriter("{}/{}/{}".format(tf_logdir, self.run_id_string, "train"),
flush_secs=writer_flush_secs, max_queue=writer_max_queue)
self._test_writer = tf.summary.FileWriter("{}/{}/{}".format(tf_logdir, self.run_id_string, "test"),
flush_secs=writer_flush_secs, max_queue=writer_max_queue)
@staticmethod
def choose_action_index(policy, deterministic=False):
if deterministic:
return np.argmax(policy)
r = random.random()
cummulative_sum = 0.0
for i, p in enumerate(policy):
cummulative_sum += p
if r <= cummulative_sum:
return i
return len(policy) - 1
def make_training_step(self):
states_img = []
states_misc = []
actions = []
rewards_reversed = []
values_reversed = []
advantages = []
Rs = []
# TODO check how default session works
self._session.run(self.local_network.ops.sync)
initial_network_state = None
if self.local_network.has_state():
initial_network_state = self.local_network.get_current_network_state()
terminal = None
steps_performed = 0
for _ in range(self.max_remembered_steps):
steps_performed += 1
current_img, current_misc = self.doom_wrapper.get_current_state()
policy, state_value = self.local_network.get_policy_and_value(self._session, [current_img, current_misc])
action_index = A3CLearner.choose_action_index(policy)
values_reversed.insert(0, state_value)
states_img.append(current_img)
states_misc.append(current_misc)
actions.append(action_index)
reward = self.doom_wrapper.make_action(action_index)
terminal = self.doom_wrapper.is_terminal()
rewards_reversed.insert(0, reward)
self.local_steps += 1
if terminal:
if self.thread_index == 0:
self.train_scores.append(self.doom_wrapper.get_total_reward())
self.doom_wrapper.reset()
if self.local_network.has_state():
self.local_network.reset_state()
break
if terminal:
R = 0.0
else:
R = self.local_network.get_value(self._session, self.doom_wrapper.get_current_state())
for ri, Vi in zip(rewards_reversed, values_reversed):
R = ri + self.gamma * R
advantages.insert(0, R - Vi)
Rs.insert(0, R)
train_op_feed_dict = {
self.local_network.vars.state_img: states_img,
self.local_network.vars.a: actions,
self.local_network.vars.advantage: advantages,
self.local_network.vars.R: Rs
}
if self.use_misc:
train_op_feed_dict[self.local_network.vars.state_misc] = states_misc
if self.local_network.has_state():
train_op_feed_dict[self.local_network.vars.initial_network_state] = initial_network_state
train_op_feed_dict[self.local_network.vars.sequence_length] = [len(actions)]
self._session.run(self.train_op, feed_dict=train_op_feed_dict)
return steps_performed
def run_episode(self, deterministic=True):
self.doom_wrapper.reset()
if self.local_network.has_state():
self.local_network.reset_state()
while not self.doom_wrapper.is_terminal():
current_state = self.doom_wrapper.get_current_state()
action_index = self._get_best_action(self._session, current_state, deterministic=deterministic)
self.doom_wrapper.make_action(action_index)
total_reward = self.doom_wrapper.get_total_reward()
return total_reward
def test(self, episodes_num=None, deterministic=True):
if episodes_num is None:
episodes_num = self.test_episodes_per_epoch
test_start_time = time.time()
test_rewards = []
for _ in trange(episodes_num, desc="Testing", file=sys.stdout,
leave=False, disable=not self.enable_progress_bar):
total_reward = self.run_episode(deterministic=self.deterministic_testing)
test_rewards.append(total_reward)
self.doom_wrapper.reset()
if self.local_network.has_state():
self.local_network.reset_state()
test_end_time = time.time()
test_duration = test_end_time - test_start_time
min_score = np.min(test_rewards)
max_score = np.max(test_rewards)
mean_score = np.mean(test_rewards)
score_std = np.std(test_rewards)
log(
"TEST: mean: {}, min: {}, max: {}, test time: {}".format(
green("{:0.3f}±{:0.2f}".format(mean_score, score_std)),
red("{:0.3f}".format(min_score)),
blue("{:0.3f}".format(max_score)),
sec_to_str(test_duration)))
return test_rewards
def _print_train_log(self, scores, overall_start_time, last_log_time, steps):
current_time = time.time()
mean_score = np.mean(scores)
score_std = np.std(scores)
min_score = np.min(scores)
max_score = np.max(scores)
elapsed_time = time.time() - overall_start_time
global_steps = self._global_steps_counter.get()
local_steps_per_sec = steps / (current_time - last_log_time)
global_steps_per_sec = global_steps / elapsed_time
global_mil_steps_per_hour = global_steps_per_sec * 3600 / 1000000.0
log(
"TRAIN: {}(GlobalSteps), mean: {}, min: {}, max: {}, "
" LocalSpd: {:.0f} STEPS/s GlobalSpd: "
"{} STEPS/s, {:.2f}M STEPS/hour, total elapsed time: {}".format(
global_steps,
green("{:0.3f}±{:0.2f}".format(mean_score, score_std)),
red("{:0.3f}".format(min_score)),
blue("{:0.3f}".format(max_score)),
local_steps_per_sec,
blue("{:.0f}".format(
global_steps_per_sec)),
global_mil_steps_per_hour,
sec_to_str(elapsed_time)
))
def run(self):
# TODO this method is ugly, make it nicer
try:
overall_start_time = time.time()
last_log_time = overall_start_time
local_steps_for_log = 0
while self._epoch <= self._epochs:
steps = self.make_training_step()
local_steps_for_log += steps
global_steps = self._global_steps_counter.inc(steps)
# Logs & tests
if self.local_steps_per_epoch * self._epoch <= self.local_steps:
self._epoch += 1
if self.thread_index == 0:
self._print_train_log(self.train_scores, overall_start_time, last_log_time, local_steps_for_log)
if self._run_tests:
test_scores = self.test(deterministic=self.deterministic_testing)
if self.write_summaries:
train_summary = self._session.run(self._summaries,
{self.scores_placeholder: self.train_scores})
self._train_writer.add_summary(train_summary, global_steps)
if self._run_tests:
test_summary = self._session.run(self._summaries,
{self.scores_placeholder: test_scores})
self._test_writer.add_summary(test_summary, global_steps)
last_log_time = time.time()
local_steps_for_log = 0
log("Learning rate: {}".format(self._session.run(self.learning_rate)))
# Saves model
if self._epoch % self.save_interval == 0:
self.save_model()
log("")
self.train_scores = []
except (SignalException, ViZDoomUnexpectedExitException):
threadsafe_print(red("Thread #{} aborting(ViZDoom killed).".format(self.thread_index)))
def run_training(self, session):
self._session = session
self.start()
def save_model(self):
savedir = os.path.dirname(self._model_savefile)
if not os.path.exists(savedir):
log("Creating directory: {}".format(savedir))
os.makedirs(savedir)
log("Saving model to: {}".format(self._model_savefile))
saver = tf.train.Saver(self.local_network.get_params())
saver.save(self._session, self._model_savefile)
def load_model(self, session, savefile):
saver = tf.train.Saver(self.local_network.get_params())
log("Loading model from: {}".format(savefile))
saver.restore(session, savefile)
log("Loaded model.")
def _get_best_action(self, sess, state, deterministic=True):
policy = self.local_network.get_policy(sess, state)
action_index = self.choose_action_index(policy, deterministic=deterministic)
return action_index
class ADQNLearner(A3CLearner):
def __init__(self,
global_target_network=None,
unfreeze_thread=False,
frozen_global_steps=40000,
initial_epsilon=1.0,
final_epsilon=0.1,
epsilon_decay_steps=10e06,
epsilon_decay_start_step=0,
**args):
super(ADQNLearner, self).__init__(**args)
self.global_target_network = global_target_network
self.unfreeze_thread = unfreeze_thread
if unfreeze_thread:
self.frozen_global_steps = frozen_global_steps
else:
self.frozen_global_steps = None
# Epsilon
# TODO randomize epsilon somehow
self.epsilon_decay_rate = (initial_epsilon - final_epsilon) / epsilon_decay_steps
self.epsilon_decay_start_step = epsilon_decay_start_step
self.initial_epsilon = initial_epsilon
self.final_epsilon = final_epsilon
def get_current_epsilon(self):
eps = self.initial_epsilon - (self.local_steps - self.epsilon_decay_start_step) * self.epsilon_decay_rate
return np.clip(eps, self.final_epsilon, 1.0)
def make_training_step(self):
states_img = []
states_misc = []
actions = []
rewards_reversed = []
target_qs = []
self._session.run(self.local_network.ops.sync)
initial_network_state = None
if self.local_network.has_state():
initial_network_state = self.local_network.get_current_network_state()
terminal = None
steps_performed = 0
for _ in range(self.max_remembered_steps):
steps_performed += 1
current_img, current_misc = self.doom_wrapper.get_current_state()
if random.random() <= self.get_current_epsilon():
action_index = random.randint(0, self.actions_num - 1)
if self.local_network.has_state():
self.local_network.update_network_state(self._session, [current_img, current_misc])
else:
q_values = self.local_network.get_q_values(self._session, [current_img, current_misc]).flatten()
action_index = q_values.argmax()
states_img.append(current_img)
states_misc.append(current_misc)
actions.append(action_index)
reward = self.doom_wrapper.make_action(action_index)
terminal = self.doom_wrapper.is_terminal()
rewards_reversed.insert(0, reward)
self.local_steps += 1
if terminal:
if self.thread_index == 0:
self.train_scores.append(self.doom_wrapper.get_total_reward())
self.doom_wrapper.reset()
if self.local_network.has_state():
self.local_network.reset_state()
break
if terminal:
target_q = 0.0
else:
if self.global_network.has_state():
q2 = self.global_target_network.get_q_values(self._session,
self.doom_wrapper.get_current_state(),
False,
self.local_network.get_current_network_state())
else:
q2 = self.global_target_network.get_q_values(self._session,
self.doom_wrapper.get_current_state())
target_q = q2.max()
for ri in rewards_reversed:
target_q = ri + self.gamma * target_q
target_qs.insert(0, target_q)
# TODO delegate this to the network as train_batch(session, ...), maybe?
train_op_feed_dict = {
self.local_network.vars.state_img: states_img,
self.local_network.vars.a: actions,
self.local_network.vars.target_q: target_qs
}
if self.use_misc:
train_op_feed_dict[self.local_network.vars.state_misc] = states_misc
if self.local_network.has_state():
train_op_feed_dict[self.local_network.vars.initial_network_state] = initial_network_state
train_op_feed_dict[self.local_network.vars.sequence_length] = [len(actions)]
self._session.run(self.train_op, feed_dict=train_op_feed_dict)
return steps_performed
def run(self):
# TODO this method is ugly, make it nicer ...and it's the same as above.... really TODO!!
# Basically code copied from base class with unfreezing
try:
overall_start_time = time.time()
last_log_time = overall_start_time
local_steps_for_log = 0
next_target_update = self.frozen_global_steps
while self._epoch <= self._epochs:
steps = self.make_training_step()
local_steps_for_log += steps
global_steps = self._global_steps_counter.inc(steps)
# Updating target network:
if self.unfreeze_thread:
# TODO this check is dangerous
if global_steps >= next_target_update:
next_target_update += self.frozen_global_steps
if next_target_update <= global_steps:
# TODO use warn from the logger
logging.warning(yellow("Global steps ({}) <= next target update ({}).".format(
global_steps, next_target_update)))
self._session.run(self.global_network.ops.unfreeze)
# Logs & tests
if self.local_steps_per_epoch * self._epoch <= self.local_steps:
self._epoch += 1
if self.thread_index == 0:
self._print_train_log(self.train_scores, overall_start_time, last_log_time, local_steps_for_log)
if self._run_tests:
test_scores = self.test(deterministic=self.deterministic_testing)
if self.write_summaries:
train_summary = self._session.run(self._summaries,
{self.scores_placeholder: self.train_scores})
self._train_writer.add_summary(train_summary, global_steps)
if self._run_tests:
test_summary = self._session.run(self._summaries,
{self.scores_placeholder: test_scores})
self._test_writer.add_summary(test_summary, global_steps)
last_log_time = time.time()
local_steps_for_log = 0
log("Learning rate: {}".format(self._session.run(self.learning_rate)))
# Saves model
if self._epoch % self.save_interval == 0:
self.save_model()
log("")
self.train_scores = []
except (SignalException, ViZDoomUnexpectedExitException):
threadsafe_print(red("Thread #{} aborting(ViZDoom killed).".format(self.thread_index)))
def _get_best_action(self, sess, state, deterministic=True):
q = self.local_network.get_q_values(sess, state).flatten()
action_index = q.argmax()
return action_index