forked from dnishio/DSAC
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sqil.py
585 lines (492 loc) · 23.1 KB
/
sqil.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
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
from __future__ import unicode_literals
from __future__ import print_function
from __future__ import division
from __future__ import absolute_import
from future import standard_library
from builtins import * # NOQA
standard_library.install_aliases() # NOQA
import collections
import os
import copy
from logging import getLogger
import chainer
from chainer import cuda
import chainer.functions as F
import numpy as np
from chainerrl.agent import AttributeSavingMixin
from chainerrl.agent import BatchAgent
from chainerrl.misc.batch_states import batch_states
from chainerrl.misc.copy_param import synchronize_parameters
from chainerrl.replay_buffer import ReplayUpdater
# from chainerrl.replay_buffer import batch_experiences
from replay_buffer import batch_experiences
import sac
from replay_buffer import AbsorbReplayBuffer
def _mean_or_nan(xs):
"""Return its mean a non-empty sequence, numpy.nan for a empty one."""
return np.mean(xs) if xs else np.nan
class TemperatureHolder(chainer.Link):
"""Link that holds a temperature as a learnable value.
Args:
initial_log_temperature (float): Initial value of log(temperature).
"""
def __init__(self, initial_log_temperature=0):
super().__init__()
with self.init_scope():
self.log_temperature = chainer.Parameter(
np.array(initial_log_temperature, dtype=np.float32))
def __call__(self):
"""Return a temperature as a chainer.Variable."""
return F.exp(self.log_temperature)
class SQIL(sac.SoftActorCritic):
"""Soft Actor-Critic (SAC).
See https://arxiv.org/abs/1812.05905
Args:
policy (Policy): Policy.
q_func1 (Link): First Q-function that takes state-action pairs as input
and outputs predicted Q-values.
q_func2 (Link): Second Q-function that takes state-action pairs as
input and outputs predicted Q-values.
policy_optimizer (Optimizer): Optimizer setup with the policy
q_func1_optimizer (Optimizer): Optimizer setup with the first
Q-function.
q_func2_optimizer (Optimizer): Optimizer setup with the second
Q-function.
replay_buffer (ReplayBuffer): Replay buffer
gamma (float): Discount factor
gpu (int): GPU device id if not None nor negative.
replay_start_size (int): if the replay buffer's size is less than
replay_start_size, skip update
minibatch_size (int): Minibatch size
update_interval (int): Model update interval in step
phi (callable): Feature extractor applied to observations
soft_update_tau (float): Tau of soft target update.
logger (Logger): Logger used
batch_states (callable): method which makes a batch of observations.
default is `chainerrl.misc.batch_states.batch_states`
burnin_action_func (callable or None): If not None, this callable
object is used to select actions before the model is updated
one or more times during training.
initial_temperature (float): Initial temperature value. If
`entropy_target` is set to None, the temperature is fixed to it.
entropy_target (float or None): If set to a float, the temperature is
adjusted during training to match the policy's entropy to it.
temperature_optimizer (Optimizer or None): Optimizer used to optimize
the temperature. If set to None, Adam with default hyperparameters
is used.
act_deterministically (bool): If set to True, choose most probable
actions in the act method instead of sampling from distributions.
"""
saved_attributes = (
'policy',
'q_func1',
'q_func2',
'target_q_func1',
'target_q_func2',
'policy_optimizer',
'q_func1_optimizer',
'q_func2_optimizer',
'temperature_holder',
'temperature_optimizer',
)
def __init__(
self,
policy,
q_func1,
q_func2,
reward_func,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
reward_func_optimizer,
replay_buffer,
replay_buffer_demo,
gamma,
is_discrete,
gpu=None,
replay_start_size=10000,
minibatch_size=100,
update_interval=1,
phi=lambda x: x,
soft_update_tau=5e-3,
logger=getLogger(__name__),
batch_states=batch_states,
burnin_action_func=None,
initial_temperature=1.,
entropy_target=None,
temperature_optimizer=None,
act_deterministically=True,
sample_interval=20,
lamda=1.0,
penalty_lamda=10.0,
):
self.policy = policy
self.q_func1 = q_func1
self.q_func2 = q_func2
self.reward_func = reward_func
if gpu is not None and gpu >= 0:
cuda.get_device_from_id(gpu).use()
self.policy.to_gpu(device=gpu)
self.q_func1.to_gpu(device=gpu)
self.q_func2.to_gpu(device=gpu)
if self.reward_func:
self.reward_func.to_gpu(device=gpu)
self.xp = self.policy.xp
self.replay_buffer = replay_buffer
self.replay_buffer_demo = replay_buffer_demo
self.gamma = gamma
self.gpu = gpu
self.phi = phi
self.soft_update_tau = soft_update_tau
self.sample_interval = sample_interval
self.minibatch_size = minibatch_size
self.logger = logger
self.policy_optimizer = policy_optimizer
self.q_func1_optimizer = q_func1_optimizer
self.q_func2_optimizer = q_func2_optimizer
self.reward_func_optimizer = reward_func_optimizer
self.replay_start_size = replay_start_size
self.batch_states = batch_states
self.burnin_action_func = burnin_action_func
self.initial_temperature = initial_temperature
self.entropy_target = entropy_target
self.lamda = lamda
if self.entropy_target is not None:
self.temperature_holder = TemperatureHolder(
initial_log_temperature=np.log(initial_temperature))
if temperature_optimizer is not None:
self.temperature_optimizer = temperature_optimizer
else:
self.temperature_optimizer = chainer.optimizers.Adam()
self.temperature_optimizer.setup(self.temperature_holder)
if gpu is not None and gpu >= 0:
self.temperature_holder.to_gpu(device=gpu)
else:
self.temperature_holder = None
self.temperature_optimizer = None
self.act_deterministically = act_deterministically
self.t = 0
self.prev_episode_end_t = 0
self.last_state = None
self.last_action = None
# Target model
self.target_q_func1 = copy.deepcopy(self.q_func1)
self.target_q_func2 = copy.deepcopy(self.q_func2)
# Statistics
self.q1_record = collections.deque(maxlen=1000)
self.q2_record = collections.deque(maxlen=1000)
self.entropy_record = collections.deque(maxlen=1000)
self.q_func1_loss_record = collections.deque(maxlen=100)
self.q_func2_loss_record = collections.deque(maxlen=100)
self.reward_demo_record = collections.deque(maxlen=100)
self.reward_samp_record = collections.deque(maxlen=100)
self.reward_loss_record = collections.deque(maxlen=100)
self.is_discrete = is_discrete
self.penalty_lamda = penalty_lamda
def act_and_train(self, obs, reward):
self.logger.debug('t:%s r:%s', self.t, reward)
if (self.burnin_action_func is not None
and self.policy_optimizer.t == 0):
action = self.burnin_action_func()
else:
action = self.select_greedy_action(obs, self.xp.zeros((1, 1), dtype=np.float32))
self.t += 1
if self.last_state is not None:
assert self.last_action is not None
# Add a transition to the replay buffer
self.replay_buffer.append(
state=self.last_state,
action=self.last_action,
reward=reward,
next_state=obs,
next_action=action,
is_state_terminal=False)
self.last_state = obs
self.last_action = action
# self.replay_updater.update_if_necessary(self.t)
return self.last_action
def act(self, obs):
return self.select_greedy_action(
obs, self.xp.zeros((1, 1), dtype=np.float32), deterministic=self.act_deterministically)
def batch_act(self, batch_obs):
return self.batch_select_greedy_action(
batch_obs, self.xp.zeros((len(batch_obs), 1), dtype=np.float32), deterministic=self.act_deterministically)
def batch_act_and_train(self, batch_obs):
"""Select a batch of actions for training.
Args:
batch_obs (Sequence of ~object): Observations.
Returns:
Sequence of ~object: Actions.
"""
if (self.burnin_action_func is not None
and self.policy_optimizer.t == 0):
batch_action = [self.burnin_action_func()
for _ in range(len(batch_obs))]
else:
batch_action = self.batch_select_greedy_action(
batch_obs, self.xp.zeros((len(batch_obs), 1), dtype=np.float32))
self.batch_last_obs = list(batch_obs)
self.batch_last_action = list(batch_action)
return batch_action
def batch_select_greedy_action(self, batch_obs, batch_abs, deterministic=False):
with chainer.using_config('train', False), chainer.no_backprop_mode():
batch_xs = self.batch_states(batch_obs, self.xp, self.phi)
if deterministic:
batch_action = self.policy(batch_xs, batch_abs).most_probable.array
else:
batch_action = self.policy(batch_xs, batch_abs).sample().array
return list(cuda.to_cpu(batch_action))
def select_greedy_action(self, obs, absorb, deterministic=False):
return self.batch_select_greedy_action(
[obs], absorb, deterministic=deterministic)[0]
def batch_observe_and_train(
self, batch_obs, batch_reward, batch_done, batch_reset):
for i in range(len(batch_obs)):
self.t += 1
if self.batch_last_obs[i] is not None:
assert self.batch_last_action[i] is not None
# Add a transition to the replay buffer
self.replay_buffer.append(
state=self.batch_last_obs[i],
action=self.batch_last_action[i],
reward=batch_reward[i],
next_state=batch_obs[i],
next_action=None,
is_state_terminal=batch_done[i],
)
if batch_reset[i] or batch_done[i]:
self.batch_last_obs[i] = None
#self.replay_updater.update_if_necessary(self.t)
def _calc_target_v(self, state, is_absorb):
action_distrib = self.policy(state, is_absorb)
actions, log_prob = \
action_distrib.sample_with_log_prob()
# Starting from the goal state we can execute only non-actions.
actions *= (1 - is_absorb)
entropy_term = self.temperature * log_prob
if self.is_discrete:
q1 = F.select_item(self.target_q_func1(state, is_absorb), actions)
q2 = F.select_item(self.target_q_func2(state, is_absorb), actions)
else:
q1 = self.target_q_func1(state, is_absorb, actions)
q2 = self.target_q_func2(state, is_absorb, actions)
entropy_term = entropy_term[..., None]
entropy_term *= (1 - is_absorb)
q = F.minimum(q1, q2)
assert q.shape == entropy_term.shape
return F.flatten(q - entropy_term)
def _compute_q_loss(self, batch):
"""B(D, r)"""
batch_reward = self.xp.concatenate([self.xp.ones_like(batch['reward'][:self.minibatch_size]),
self.xp.zeros_like(batch['reward'][self.minibatch_size:])], axis=0)
batch_state = batch['state']
batch_next_state = batch['next_state']
batch_actions = batch['action']
batch_discount = batch['discount']
batch_terminal = batch['is_state_terminal']
batch_absorb = batch['is_state_absorb']
batch_next_absorb = batch['is_next_state_absorb']
with chainer.no_backprop_mode(), chainer.using_config('train', False):
target_next_v = self._calc_target_v(batch_next_state, batch_next_absorb)
if self.reward_func:
# reward for gan
D = F.sigmoid(self.reward_func(batch_state, batch_absorb, batch_actions))
batch_reward = F.flatten(F.log(D + 1e-8) - F.log(1 - D + 1e-8)) # + 0.5 * batch_reward / self.temperature / self.lamda
batch_reward = F.flatten(batch_reward)
self.reward_demo_record.extend(cuda.to_cpu(batch_reward.array[:self.minibatch_size]))
self.reward_samp_record.extend(cuda.to_cpu(batch_reward.array[self.minibatch_size:]))
target_q = batch_reward + batch_discount * \
(1.0 - batch_terminal) * target_next_v
if self.is_discrete:
predict_q1 = F.flatten(F.select_item(self.q_func1(batch_state, batch_absorb), batch_actions))
predict_q2 = F.flatten(F.select_item(self.q_func2(batch_state, batch_absorb), batch_actions))
else:
predict_q1 = F.flatten(self.q_func1(batch_state, batch_absorb, batch_actions))
predict_q2 = F.flatten(self.q_func2(batch_state, batch_absorb, batch_actions))
# soft bellman error
loss1 = 0.5 * F.mean_squared_error(target_q, predict_q1)
loss2 = 0.5 * F.mean_squared_error(target_q, predict_q2)
self.q1_record.extend(cuda.to_cpu(predict_q1.array))
self.q2_record.extend(cuda.to_cpu(predict_q2.array))
return loss1, loss2
def _compute_pi_loss(self, batch):
"""Compute loss for actor."""
batch_absorb = batch['is_state_absorb']
# Don't update the actor for absorbing states.
# And skip update if all states are absorbing.
if all(batch_absorb):
return 0
batch_state = batch['state']
action_distrib = self.policy(batch_state, batch_absorb)
if self.is_discrete: # for discrete actions
prob = action_distrib.all_prob
log_prob = action_distrib.all_log_prob
q1 = self.q_func1(batch_state, batch_absorb)
q2 = self.q_func2(batch_state, batch_absorb)
else:
actions, log_prob = action_distrib.sample_with_log_prob()
q1 = self.q_func1(batch_state, batch_absorb, actions)
q2 = self.q_func2(batch_state, batch_absorb, actions)
prob = 1
log_prob = log_prob[...,None]
q = F.minimum(q1, q2)
entropy_term = self.temperature * log_prob
assert q.shape == entropy_term.shape
loss = F.average((1 - batch_absorb) * prob * (entropy_term - q))
if self.entropy_target is not None:
self.update_temperature(log_prob.array)
# Record entropy
with chainer.no_backprop_mode():
try:
self.entropy_record.extend(
cuda.to_cpu(action_distrib.entropy.array))
except NotImplementedError:
# Record - log p(x) instead
self.entropy_record.extend(
cuda.to_cpu(-log_prob.array))
return loss
def update_reward_func(self, batch_demo, batch_samp):
batch_demo_state = batch_demo['state']
batch_demo_absorb = batch_demo['is_state_absorb']
batch_demo_actions = batch_demo['action']
batch_samp_state = batch_samp['state']
batch_samp_absorb = batch_samp['is_state_absorb']
batch_samp_actions = batch_samp['action']
# -log(D_demo(x)) for GAN
y_demo = self.reward_func(batch_demo_state, batch_demo_absorb, batch_demo_actions)
loss_demo = - F.average(F.log(F.sigmoid(y_demo) + 1e-8))
# -log(1 - D_samp(x)) for GAN
y_samp = self.reward_func(batch_samp_state, batch_samp_absorb, batch_samp_actions)
loss_samp = - F.average(F.log(1 - F.sigmoid(y_samp) + 1e-8))
# grad penalty
eps = self.xp.random.uniform(0, 1, size=len(batch_demo_state)).astype("f")
if len(batch_samp_state.shape) == 2:
eps = eps[:, None]
elif len(batch_samp_state.shape) == 4:
eps = eps[:, None, None, None]
s_mid = eps * batch_demo_state + (1.0 - eps) * batch_samp_state
abs_mid = eps * batch_demo_absorb + (1.0 - eps) * batch_samp_absorb
act_mid = eps * batch_demo_actions + (1.0 - eps) * batch_samp_actions
x_mid = np.concatenate((s_mid, abs_mid, act_mid), axis=-1)
x_mid = chainer.Variable(x_mid)
y_mid = self.reward_func.forward(x_mid)
grad, = chainer.grad([y_mid], [x_mid], enable_double_backprop=True)
penalty = 0.5 * self.penalty_lamda * F.mean(F.batch_l2_norm_squared(grad))
# penalty = self.penalty_lamda * F.mean_squared_error(grad, self.xp.zeros_like(grad.array))
loss = loss_demo + loss_samp + penalty
self.reward_loss_record.append(float(loss.array))
# self.penalty_record.append(float(penalty.array))
self.reward_func_optimizer.update(lambda: loss)
def _compute_bc_loss(self, batch):
# behavioral cloning
batch_state = batch['state']
batch_actions = batch['action']
batch_absorb = batch['is_state_absorb']
action_distrib = self.policy(batch_state, batch_absorb)
action = action_distrib.sample()
loss = F.mean_squared_error(F.flatten(batch_actions), F.flatten(action))
return loss
def pretrain(self, num_iters):
"""Behavioral Cloning"""
for iter in range(num_iters):
batch_demo = batch_experiences(
self.replay_buffer_demo.sample(self.minibatch_size),
xp=self.xp, phi=self.phi, gamma=self.gamma)
# actor
loss_pi = self._compute_bc_loss(batch_demo)
self.policy_optimizer.update(lambda: loss_pi)
if iter % 1000 == 0:
print(f'iteration: {iter}, loss pi: {loss_pi.array}')
def _calc_absorb_value(self, batch_demo):
absorb_state = self.xp.zeros((1, batch_demo['state'].shape[1]), dtype=batch_demo['state'].dtype)
is_absorb = self.xp.ones((1, 1), dtype=self.xp.float32)
absorb_action = self.xp.zeros((1, 1), dtype=self.xp.float32)
# absorb_reward = self.reward_func(absorb_state, is_absorb, absorb_action)
absorb_q1 = self.q_func1(absorb_state, is_absorb, absorb_action)
absorb_q2 = self.q_func2(absorb_state, is_absorb, absorb_action)
# assert absorb_q1.shape == absorb_reward.shape
# return F.flatten(absorb_reward + self.gamma * F.minimum(absorb_q1, absorb_q2))
return F.flatten(absorb_q1), F.flatten(absorb_q2)
def concat_demo(self, batch_demo, batch_sample):
batch = {}
for key in batch_demo.keys():
batch[key] = self.xp.concatenate((batch_demo[key], batch_sample[key]), axis=0)
return batch
def train_from_demo(self, episode_len):
# update reward function
if self.reward_func:
for epoch in range(episode_len):
batch_demo = batch_experiences(
self.replay_buffer_demo.sample(self.minibatch_size),
xp=self.xp, phi=self.phi, gamma=self.gamma)
batch_sample = batch_experiences(
self.replay_buffer.sample(self.minibatch_size),
xp=self.xp, phi=self.phi, gamma=self.gamma)
self.update_reward_func(batch_demo, batch_sample)
# update actor critic
for epoch in range(episode_len):
batch_demo = batch_experiences(
self.replay_buffer_demo.sample(self.minibatch_size),
xp=self.xp, phi=self.phi, gamma=self.gamma)
batch_sample = batch_experiences(
self.replay_buffer.sample(self.minibatch_size),
xp=self.xp, phi=self.phi, gamma=self.gamma)
batch = self.concat_demo(batch_demo, batch_sample)
loss_q1, loss_q2 = self._compute_q_loss(batch)
self.q_func1_optimizer.update(lambda: loss_q1)
self.q_func2_optimizer.update(lambda: loss_q2)
loss_pi = self._compute_pi_loss(batch_sample)
# loss_pi = self._compute_pi_loss(batch)
self.policy_optimizer.update(lambda: loss_pi)
# # decay learning rate
if self.policy_optimizer.t % 100000 == 0:
if self.reward_func_optimizer:
self.reward_func_optimizer.alpha *= 0.5
self.q_func1_optimizer.alpha *= 0.5
self.q_func2_optimizer.alpha *= 0.5
self.policy_optimizer.alpha *= 0.5
# Update stats
self.q_func1_loss_record.append(float(loss_q1.array))
self.q_func2_loss_record.append(float(loss_q2.array))
self.sync_target_network()
def stop_episode_and_train(self, state, reward, done=False):
assert self.last_state is not None
assert self.last_action is not None
# Add a transition to the replay buffer
self.replay_buffer.append(
state=self.last_state,
action=self.last_action,
reward=reward,
next_state=state,
next_action=self.last_action,
is_state_terminal=done)
episode_size = self.t - self.prev_episode_end_t
self.stop_episode()
if self.t >= self.minibatch_size and self.t >= self.replay_start_size:
self.train_from_demo(episode_size)
def stop_episode(self):
self.last_state = None
self.last_action = None
self.prev_episode_end_t = self.t
self.replay_buffer.stop_current_episode()
def get_statistics(self):
return [
('average_q1', _mean_or_nan(self.q1_record)),
('average_q2', _mean_or_nan(self.q2_record)),
('average_q_func1_loss', _mean_or_nan(self.q_func1_loss_record)),
('average_q_func2_loss', _mean_or_nan(self.q_func2_loss_record)),
('n_updates', self.policy_optimizer.t),
('average_entropy', _mean_or_nan(self.entropy_record)),
('temperature', self.temperature),
('reward_demo', _mean_or_nan(self.reward_demo_record)),
('reward_sample', _mean_or_nan(self.reward_samp_record)),
# ('reward_loss', _mean_or_nan(self.reward_loss_record)),
]
def save(self, dirname, save_replay=True):
super().save(dirname)
if save_replay:
self.replay_buffer.save(os.path.join(dirname, 'replay'))
def load(self, dirname, load_replay=True):
super().load(dirname)
if load_replay:
self.replay_buffer.load(os.path.join(dirname, 'replay'))