-
Notifications
You must be signed in to change notification settings - Fork 0
/
hiu_sac.py
1090 lines (941 loc) · 39.5 KB
/
hiu_sac.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
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import os
import numpy as np
import torch
import math
from models import MultiPolicyNet, MultiQNet, MultiVNet
from itertools import chain
import logger.logger as logger
import gtimer as gt
import tqdm
from utils import interaction, rollout, np_ify, torch_ify
from utils import soft_param_update_from_to, hard_buffer_update_from_to
class HIUSAC(object):
def __init__(
self,
env,
# Learning models
nets_hidden_size=64,
nets_nonlinear_op='relu',
use_q2=True,
explicit_vf=False,
# RL algorithm behavior
total_episodes=10,
train_steps=100,
eval_rollouts=10,
max_horizon=100,
fixed_horizon=True,
# Target models update
soft_target_tau=5e-3,
target_update_interval=1,
# Replay Buffer
replay_buffer_size=1e6,
batch_size=64,
# Values
discount=0.99,
# Optimization
optimization_steps=1,
optimizer='adam',
optimizer_kwargs=None,
policy_lr=3e-4,
qf_lr=3e-4,
policy_weight_decay=1.e-5,
q_weight_decay=1.e-5,
# Entropy
i_entropy_scale=1.,
u_entropy_scale=None,
auto_alpha=True,
max_alpha=10,
min_alpha=0.01,
i_tgt_entro=None,
u_tgt_entros=None,
# Multitask
multitask=True,
combination_method='convex',
# Others
norm_input_pol=False,
norm_input_vfs=False,
seed=610,
render=False,
gpu_id=-1,
):
"""Hierarchical Intentional-Unitentional Soft Actor-Critic algorithm.
Args:
env (Env): OpenAI-Gym-like environment with multigoal option.
nets_hidden_size (int): Number of units in hidden layers for all
the networks.
use_q2 (bool): Use two parameterized Q-functions.
explicit_vf (bool): Use a parameterized soft state value function.
total_episodes (int): Number of episodes (iterations) to run the
algorithm.
train_steps (int): Number of training steps per episode.
eval_rollouts (int): Number of rollouts to perform by the policy
at the end of each episode.
max_horizon (int): Maximum length of each rollout.
fixed_horizon (bool):
soft_target_tau (float): Interpolation factor for the target
networks.
target_update_interval (int): How often (gap between training
steps) the target networks are updated. Training steps
replay_buffer_size (int): Maximum length of the replay buffer.
batch_size (int): Minibatch size for SGD.
discount (float): Discount factor (between 0 and 1).
optimization_steps (int): Number of optimization steps after each
interaction.
optimizer (str): name of the (mini-batch) SGD optimizer.
Options: 'adam' or 'rmsprop'.
optimizer_kwargs (dict): Keyword arguments for the optimizer.
policy_lr (float): Policy learning rate.
qf_lr (float): State-action and state value functions learning rate
policy_weight_decay (float): Weight decay (L2 penalty) in policy
network.
q_weight_decay (float): Weight decay (L2 penalty) in value
networks.
i_entropy_scale (float): Scale value for entropy in the compound
task.
u_entropy_scale (list or tuple of float): Scale value for the
entropies in the composable tasks.
auto_alpha (int): Compute entropy regularization term automatically
max_alpha (float): Maximum entropy regularization value.
min_alpha (float): Minimum entropy regularization value.
i_tgt_entro (float): Target entropy value in the compound policy.
u_tgt_entros (list or tuple of float): Target entropy value in the
composable policy.
multitask (bool): If False a single-task process is carried out,
resulting in the SAC algorithm.
combination_method (str): Combination method of the policy.
norm_input_pol (bool): Normalize the input of the policy.
norm_input_vfs (bool): Normalize the input of the value functions.
seed (int): Seed value for the random number generators.
render (bool): Rendering the interaction.
gpu_id (int): GPU ID. For CPU use value -1
"""
self.seed = seed
np.random.seed(seed)
torch.cuda.manual_seed(seed)
torch.manual_seed(seed)
self.env = env
self.env.seed(seed)
if multitask:
self.num_intentions = self.env.n_subgoals
else:
self.num_intentions = 0
# Algorithm hyperparameters
self.obs_dim = np.prod(env.observation_space.shape).item()
self.action_dim = np.prod(env.action_space.shape).item()
self.total_episodes = total_episodes
self.train_steps = train_steps
self.eval_rollouts = eval_rollouts
self.max_horizon = max_horizon
self.fixed_horizon = fixed_horizon
self.render = render
self.discount = discount
self.soft_target_tau = soft_target_tau
self.target_update_interval = target_update_interval
self.norm_input_pol = norm_input_pol
self.norm_input_vfs = norm_input_vfs
# Policy Network
self.policy = MultiPolicyNet(
num_intentions=max(self.num_intentions, 1),
obs_dim=self.obs_dim,
action_dim=self.action_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear=nets_nonlinear_op,
shared_batch_norm=False,
intention_non_linear=nets_nonlinear_op,
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_pol,
combination_method=combination_method,
)
# Value Function Networks
self.qf1 = MultiQNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
action_dim=self.action_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear=nets_nonlinear_op,
shared_batch_norm=False,
intention_non_linear=nets_nonlinear_op,
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
if use_q2:
self.qf2 = MultiQNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
action_dim=self.action_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear='relu',
shared_batch_norm=False,
intention_non_linear='relu',
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
else:
self.qf2 = None
if explicit_vf:
self.vf = MultiVNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear='relu',
shared_batch_norm=False,
intention_non_linear='relu',
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
self.target_vf = MultiVNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear='relu',
shared_batch_norm=False,
intention_non_linear='relu',
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
self.target_vf.load_state_dict(self.vf.state_dict())
self.target_vf.eval()
self.target_qf1 = None
self.target_qf2 = None
else:
self.vf = None
self.target_vf = None
self.target_qf1 = MultiQNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
action_dim=self.action_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear='relu',
shared_batch_norm=False,
intention_non_linear='relu',
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
self.target_qf1.load_state_dict(self.qf1.state_dict())
self.target_qf1.eval()
if use_q2:
self.target_qf2 = MultiQNet(
num_intentions=self.num_intentions + 1,
obs_dim=self.obs_dim,
action_dim=self.action_dim,
shared_sizes=(nets_hidden_size,),
intention_sizes=(nets_hidden_size, nets_hidden_size),
shared_non_linear='relu',
shared_batch_norm=False,
intention_non_linear='relu',
intention_final_non_linear='linear',
intention_batch_norm=False,
input_normalization=norm_input_vfs,
)
self.target_qf2.load_state_dict(self.qf2.state_dict())
self.target_qf2.eval()
else:
self.target_qf2 = None
# Replay Buffer
self.replay_buffer = MultiGoalReplayBuffer(
max_size=int(replay_buffer_size),
obs_dim=self.obs_dim,
action_dim=self.action_dim,
num_intentions=self.num_intentions,
)
self.batch_size = batch_size
# Move models to GPU (if applicable)
self.torch_device = \
torch.device("cuda:" + str(gpu_id) if gpu_id >= 0 else "cpu")
for model in self.trainable_models + self.non_trainable_models:
model.to(device=self.torch_device)
# Ensure non trainable models have fixed parameters
for model in self.non_trainable_models:
model.eval()
for param in model.parameters():
param.requires_grad = False
# Entropy regularization coefficients (Alphas).
if u_entropy_scale is None:
u_entropy_scale = [i_entropy_scale
for _ in range(self.num_intentions)]
self.entropy_scales = torch.tensor(u_entropy_scale+[i_entropy_scale],
device=self.torch_device)
if i_tgt_entro is None:
i_tgt_entro = float(-self.action_dim)
if u_tgt_entros is None:
u_tgt_entros = [i_tgt_entro for _ in range(self.num_intentions)]
self.tgt_entros = torch.tensor(u_tgt_entros + [i_tgt_entro],
device=self.torch_device)
self._auto_alpha = auto_alpha
self.max_alpha = max_alpha
self.min_alpha = min_alpha
self.log_alphas = torch.zeros(self.num_intentions+1,
device=self.torch_device,
requires_grad=True)
# Select optimizer function and hyperparameters
self.optimization_steps = optimization_steps
if optimizer.lower() == 'adam':
optimizer_class = torch.optim.Adam
if optimizer_kwargs is None:
optimizer_kwargs = dict(
amsgrad=True,
# amsgrad=False,
)
elif optimizer.lower() == 'rmsprop':
optimizer_class = torch.optim.RMSprop
if optimizer_kwargs is None:
optimizer_kwargs = dict(
)
else:
raise ValueError('Wrong optimizer')
# Values optimizer
qvals_params = self.qf1.parameters()
if self.qf2 is not None:
qvals_params = chain(qvals_params, self.qf2.parameters())
self.qvalues_optimizer = optimizer_class(
qvals_params,
lr=qf_lr,
weight_decay=q_weight_decay,
**optimizer_kwargs
)
if self.vf is not None:
self.vvalues_optimizer = optimizer_class(
self.vf.parameters(),
lr=qf_lr,
weight_decay=q_weight_decay,
**optimizer_kwargs
)
else:
self.vvalues_optimizer = None
# Policy optimizer
self._policy_optimizer = optimizer_class(
self.policy.parameters(),
lr=policy_lr,
weight_decay=policy_weight_decay,
**optimizer_kwargs
)
# Alpha optimizers
self._alphas_optimizer = optimizer_class(
[self.log_alphas],
lr=policy_lr,
**optimizer_kwargs
)
# Internal variables
self.num_train_interactions = 0
self.num_train_steps = 0
self.num_eval_interactions = 0
self.num_episodes = 0
# Log variables
self.logging_qvalues_error = 0
self.logging_vvalues_error = 0
self.logging_policies_error = 0
self.logging_entros = torch.zeros((
self.batch_size, self.num_intentions + 1
))
self.logging_means = torch.zeros((
self.batch_size, self.num_intentions + 1, self.action_dim
))
self.logging_stds = torch.zeros((
self.batch_size, self.num_intentions + 1, self.action_dim
))
self.logging_weights = torch.zeros((
self.batch_size, self.num_intentions, self.action_dim
))
self.logging_eval_rewards = np.zeros((
self.eval_rollouts, self.num_intentions + 1
))
self.logging_eval_returns = np.zeros((
self.eval_rollouts, self.num_intentions + 1
))
@property
def trainable_models(self):
models = [
self.policy,
self.qf1
]
if self.qf2 is not None:
models.append(self.qf2)
if self.vf is not None:
models.append(self.vf)
return models
@property
def non_trainable_models(self):
models = [
self.target_qf1
]
if self.target_qf2 is not None:
models.append(self.target_qf2)
if self.target_vf is not None:
models.append(self.target_vf)
return models
def train(self, init_episode=0):
"""Train the HIU policy with HIU algorithm.
Args:
init_episode (int): Initial iteration.
Returns:
np.ndarray: Array with the expected accumulated reward obtained
with the HIU policy during the learning process.
"""
if init_episode == 0:
# Eval and log
self.eval()
self.log(write_table_header=True)
gt.reset()
gt.set_def_unique(False)
expected_accum_rewards = np.zeros(self.total_episodes)
episodes_iter = range(init_episode, self.total_episodes)
if not logger.get_log_stdout():
# Fancy iterable bar
episodes_iter = tqdm.tqdm(episodes_iter)
for it in gt.timed_for(episodes_iter, save_itrs=True):
# Put models in training mode
for model in self.trainable_models:
model.train()
obs = self.env.reset()
rollout_steps = 0
for step in range(self.train_steps):
if self.render:
self.env.render()
interaction_info = interaction(
self.env, self.policy, obs,
device=self.torch_device,
intention=None, deterministic=False,
)
self.num_train_interactions += 1
rollout_steps += 1
gt.stamp('sample')
# Add data to replay_buffer
self.replay_buffer.add_interaction(**interaction_info)
# Only train when there are enough samples from buffer
if self.replay_buffer.available_samples() > self.batch_size:
for ii in range(self.optimization_steps):
self.learn()
gt.stamp('train')
# Reset environment if it is done
if interaction_info['termination'] \
or rollout_steps > self.max_horizon:
obs = self.env.reset()
rollout_steps = 0
else:
obs = interaction_info['next_obs']
# Evaluate current policy to check performance
expected_accum_rewards[it] = self.eval()
# Log the episode data
self.log()
self.num_episodes += 1
return expected_accum_rewards
def eval(self):
"""Evaluate deterministically the HIU policy.
Returns:
float: Expected accumulated reward
"""
# Put models in evaluation mode
for model in self.trainable_models:
model.eval()
env_subtask = self.env.get_active_subtask()
for ii in range(-1, self.num_intentions):
for rr in range(self.eval_rollouts):
if self.num_intentions > 0:
self.env.set_active_subtask(None if ii == -1 else ii)
rollout_info = rollout(self.env, self.policy,
max_horizon=self.max_horizon,
fixed_horizon=self.fixed_horizon,
render=self.render,
return_info_dict=True,
device=self.torch_device,
deterministic=True,
intention=None if ii == -1 else ii
)
if ii == -1:
rewards = np.array(rollout_info['reward'])
else:
rewards = np.array(rollout_info['reward_vector'])[:, ii]
self.logging_eval_rewards[rr, ii] = rewards.mean()
self.logging_eval_returns[rr, ii] = rewards.sum()
self.num_eval_interactions += rewards.size
# Set environment to training subtask.
self.env.set_active_subtask(env_subtask)
gt.stamp('eval')
return self.logging_eval_returns[-1].mean().item()
def learn(self):
"""Improve the HIU policy with HIU algorithm.
The method computes a 'training step' of the algorithm.
Returns:
None
"""
# Get batch from the replay buffer
batch = self.replay_buffer.random_batch(self.batch_size,
device=self.torch_device)
# Get common data from batch
obs = batch['observations']
actions = batch['actions']
next_obs = batch['next_observations']
# Concatenate all (sub)task rewards
i_rewards = batch['rewards'].unsqueeze(-1)
if self.num_intentions > 0:
u_rewards = batch['reward_vectors'].unsqueeze(-1)
hiu_rewards = torch.cat((u_rewards, i_rewards), dim=1)
else:
hiu_rewards = i_rewards
# Concatenate all (sub)task terminations
i_terminals = batch['terminations'].unsqueeze(-1)
if self.num_intentions > 0:
u_terminals = batch['termination_vectors'].unsqueeze(-1)
hiu_terminations = torch.cat((u_terminals, i_terminals), dim=1)
else:
hiu_terminations = i_terminals
policy_prior_log_probs = 0.0 # Uniform prior # TODO: Normal prior
# Alphas
alphas = self.entropy_scales*self.log_alphas.exp()
alphas.unsqueeze_(dim=-1)
# Actions for batch observation
i_new_actions, policy_info = self.policy(
obs,
deterministic=False,
intention=None,
log_prob=True,
)
i_new_log_pi = policy_info['i_log_prob']
i_new_mean = policy_info['i_mean']
i_new_std = policy_info['i_std']
# Unintentional policy info
if self.num_intentions > 0:
u_new_actions = policy_info['u_actions']
u_new_log_pi = policy_info['u_log_probs']
u_new_means = policy_info['u_means']
u_new_stds = policy_info['u_stds']
activation_weights = policy_info['activation_weights']
else:
u_new_actions = None
u_new_log_pi = None
u_new_means = None
activation_weights = None
# Actions for batch next_observation
with torch.no_grad():
i_next_actions, policy_info = self.policy(
next_obs,
deterministic=False,
intention=None,
log_prob=True,
)
i_next_log_pi = policy_info['i_log_prob']
# i_next_mean = policy_info['i_mean']
# i_next_std = policy_info['i_std']
# Unintentional policy info
if self.num_intentions > 0:
u_next_actions = policy_info['u_actions']
u_next_log_pi = policy_info['u_log_probs']
# u_next_means = policy_info['u_means']
# u_next_stds = policy_info['u_stds']
else:
u_next_actions = None
u_next_log_pi = None
# u_next_means = None
# u_next_stds = None
# Intention Mask
intention_mask = torch.eye(self.num_intentions + 1,
device=self.torch_device,
).unsqueeze(-1)
hiu_obs = obs.unsqueeze(-2).expand(
self.batch_size, self.num_intentions + 1, self.obs_dim
)
hiu_actions = actions.unsqueeze(-2).expand(
self.batch_size, self.num_intentions + 1, self.action_dim
)
hiu_next_obs = next_obs.unsqueeze(-2).expand(
self.batch_size, self.num_intentions + 1, self.obs_dim
)
if u_new_actions is None:
hiu_new_actions = i_new_actions.unsqueeze(-2)
else:
hiu_new_actions = torch.cat(
(u_new_actions, i_new_actions.unsqueeze(-2)),
dim=-2
)
if u_new_log_pi is None:
hiu_new_log_pi = i_new_log_pi.unsqueeze(-2)
else:
hiu_new_log_pi = torch.cat(
(u_new_log_pi, i_new_log_pi.unsqueeze(-2)),
dim=-2
)
if u_next_actions is None:
hiu_next_actions = i_next_actions.unsqueeze(-2)
else:
hiu_next_actions = torch.cat(
(u_next_actions, i_next_actions.unsqueeze(-2)),
dim=-2
)
if u_next_log_pi is None:
hiu_next_log_pi = i_next_log_pi.unsqueeze(-2)
else:
hiu_next_log_pi = torch.cat(
(u_next_log_pi, i_next_log_pi.unsqueeze(-2)),
dim=-2
)
# ###################### #
# Policy Evaluation Step #
# ###################### #
if self.target_vf is None:
with torch.no_grad():
# Estimate from target Q-value(s)
# Q1_target(s', a')
hiu_next_q1 = self.target_qf1(hiu_next_obs, hiu_next_actions)
if self.target_qf2 is not None:
# Q2_target(s', a')
hiu_next_q2 = self.target_qf2(hiu_next_obs, hiu_next_actions)
# Minimum Unintentional Double-Q
hiu_next_q = torch.min(hiu_next_q1, hiu_next_q2)
else:
hiu_next_q = hiu_next_q1
# Get only the corresponding intentional values
next_q_intention_mask = intention_mask.expand_as(hiu_next_q)
hiu_next_q = torch.sum(hiu_next_q*next_q_intention_mask, dim=-2)
# Vtarget(s')
hiu_next_v = hiu_next_q - alphas*hiu_next_log_pi
else:
with torch.no_grad():
# Vtarget(s')
hiu_next_v = self.target_vf(hiu_next_obs)
# Get only the corresponding intentional values
next_v_intention_mask = intention_mask.expand_as(hiu_next_v)
hiu_next_v = torch.sum(hiu_next_v*next_v_intention_mask, dim=-2)
# Calculate Bellman Backup for Q-values
hiu_q_backup = hiu_rewards + (1. - hiu_terminations) * self.discount * hiu_next_v
# Predictions Q(s,a)
hiu_q1_pred = self.qf1(obs, actions, intention=None)
# Critic loss: Mean Squared Bellman Error (MSBE)
hiu_qf1_loss = \
0.5*torch.mean((hiu_q1_pred - hiu_q_backup)**2, dim=0).squeeze(-1)
hiu_qf1_loss = torch.sum(hiu_qf1_loss)
if self.qf2 is not None:
hiu_q2_pred = self.qf2(obs, actions, intention=None)
# Critic loss: Mean Squared Bellman Error (MSBE)
hiu_qf2_loss = \
0.5*torch.mean((hiu_q2_pred - hiu_q_backup)**2, dim=0).squeeze(-1)
hiu_qf2_loss = torch.sum(hiu_qf2_loss)
else:
hiu_qf2_loss = 0
self.qvalues_optimizer.zero_grad()
qvalues_loss = (hiu_qf1_loss + hiu_qf2_loss)
qvalues_loss.backward()
self.qvalues_optimizer.step()
# ############################## #
# Policy Update/Improvement Step #
# ############################## #
# TODO: Decide if use the minimum btw q1 and q2. Using new_q1 for now
hiu_new_q1 = self.qf1(hiu_obs, hiu_new_actions)
hiu_new_q = hiu_new_q1
next_q_intention_mask = intention_mask.expand_as(hiu_new_q)
hiu_new_q = torch.sum(hiu_new_q*next_q_intention_mask, dim=-2)
# Policy KL loss: - (E_a[Q(s, a) + H(.)])
policy_kl_loss = -torch.mean(
hiu_new_q - alphas*hiu_new_log_pi
+ policy_prior_log_probs,
dim=0,
)
policy_loss = torch.sum(policy_kl_loss)
# Update both Intentional and Unintentional Policies at the same time
self._policy_optimizer.zero_grad()
policy_loss.backward()
self._policy_optimizer.step()
# ################################# #
# (Optional) V-fcn improvement step #
# ################################# #
if self.vf is not None:
hiu_v_pred = self.vf(hiu_obs)
# Calculate Bellman Backup for Q-values
hiu_v_backup = hiu_new_q - alphas*hiu_new_log_pi + policy_prior_log_probs
hiu_v_backup.detach_()
# Critic loss: Mean Squared Bellman Error (MSBE)
hiu_vf_loss = \
0.5*torch.mean((hiu_v_pred - hiu_v_backup)**2, dim=0).squeeze(-1)
hiu_vf_loss = torch.sum(hiu_vf_loss)
# ####################### #
# Entropy Adjustment Step #
# ####################### #
if self._auto_alpha:
# NOTE: In SAC formula is alphas and not log_alphas
alphas_loss = - (self.log_alphas *
(hiu_new_log_pi.squeeze(-1) + self.tgt_entros
).mean(dim=0).detach()
)
hiu_alphas_loss = alphas_loss.sum()
self._alphas_optimizer.zero_grad()
hiu_alphas_loss.backward()
self._alphas_optimizer.step()
self.log_alphas.data.clamp_(min=math.log(self.min_alpha),
max=math.log(self.max_alpha))
# ########################### #
# Target Networks Update Step #
# ########################### #
if self.num_train_steps % self.target_update_interval == 0:
if self.target_vf is None:
soft_param_update_from_to(
source=self.qf1,
target=self.target_qf1,
tau=self.soft_target_tau
)
if self.target_qf2 is not None:
soft_param_update_from_to(
source=self.qf2,
target=self.target_qf2,
tau=self.soft_target_tau
)
else:
soft_param_update_from_to(
source=self.vf,
target=self.target_vf,
tau=self.soft_target_tau
)
# Always hard_update of input normalizer (if active)
if self.norm_input_vfs:
if self.target_vf is None:
hard_buffer_update_from_to(
source=self.qf1,
target=self.target_qf1,
)
if self.target_qf2 is not None:
hard_buffer_update_from_to(
source=self.qf2,
target=self.target_qf2,
)
else:
hard_buffer_update_from_to(
source=self.vf,
target=self.target_vf,
)
# Increase internal counter
self.num_train_steps += 1
# ######## #
# Log data #
# ######## #
self.logging_policies_error = policy_loss.item()
self.logging_qvalues_error = qvalues_loss.item()
self.logging_vvalues_error = hiu_vf_loss.item() \
if self.target_vf is not None else 0.
self.logging_entros.data.copy_(-hiu_new_log_pi.squeeze(dim=-1).data)
self.logging_means.data[:, -1].copy_(i_new_mean.data)
self.logging_stds.data[:, -1].copy_(i_new_std.data)
if self.num_intentions > 0:
self.logging_means.data[:, :self.num_intentions].copy_(u_new_means.data)
self.logging_stds.data[:, :self.num_intentions].copy_(u_new_stds.data)
self.logging_weights.data.copy_(activation_weights.data)
def save_training_state(self):
"""Save models
Returns:
None
"""
models_dict = {
'policy': self.policy,
'qf1': self.qf1,
'qf2': self.qf2,
'target_qf1': self.target_qf1,
'target_qf2': self.target_qf2,
'vf': self.vf,
}
replaceable_models_dict = {
'replay_buffer', self.replay_buffer,
}
logger.save_torch_models(self.num_episodes, models_dict,
replaceable_models_dict)
def load_training_state(self):
pass
def log(self, write_table_header=False):
logger.log("Logging data in directory: %s" % logger.get_snapshot_dir())
logger.record_tabular("Episode", self.num_episodes)
logger.record_tabular("Accumulated Training Steps",
self.num_train_interactions)
logger.record_tabular("Policy Error", self.logging_policies_error)
logger.record_tabular("Q-Value Error", self.logging_qvalues_error)
logger.record_tabular("V-Value Error", self.logging_vvalues_error)
for intention in range(self.num_intentions):
logger.record_tabular("Alpha [U-%02d]" % intention,
np_ify(self.log_alphas[intention].exp()).item())
logger.record_tabular("Alpha", np_ify(self.log_alphas[-1].exp()).item())
for intention in range(self.num_intentions):
logger.record_tabular(
"Entropy [U-%02d]" % intention,
np_ify(self.logging_entros[intention].mean(dim=0))
)
logger.record_tabular("Entropy",
np_ify(self.logging_entros[-1].mean(dim=0)))
act_means = np_ify(self.logging_means.mean(dim=0))
act_stds = np_ify(self.logging_stds.mean(dim=0))
for aa in range(self.action_dim):
for intention in range(self.num_intentions):
logger.record_tabular(
"Mean Action %02d [U-%02d]" % (aa, intention),
act_means[intention, aa]
)
logger.record_tabular(
"Std Action %02d [U-%02d]" % (aa, intention),
act_stds[intention, aa]
)
logger.record_tabular("Mean Action %02d" % aa, act_means[-1, aa])
logger.record_tabular("Std Action %02d" % aa, act_stds[-1, aa])
for aa in range(self.action_dim):
for intention in range(self.num_intentions):
logger.record_tabular(
"Activation Weight Action %02d [U-%02d]" % (aa, intention),
np_ify(self.logging_weights.mean(dim=(0,))[intention, aa])
)
# Evaluation Stats to plot
for ii in range(-1, self.num_intentions):
if ii > -1:
uu_str = ' [%02d]' % ii
else:
uu_str = ''
logger.record_tabular(
"Test Rewards Mean"+uu_str,
np_ify(self.logging_eval_rewards[:, ii].mean())
)
logger.record_tabular(
"Test Rewards Std"+uu_str,
self.logging_eval_rewards[:, ii].std()
)
logger.record_tabular(
"Test Returns Mean"+uu_str,
self.logging_eval_returns[:, ii].mean()
)
logger.record_tabular(
"Test Returns Std"+uu_str,
self.logging_eval_returns[:, ii].std()
)
# Add the previous times to the logger
times_itrs = gt.get_times().stamps.itrs
train_time = times_itrs.get('train', [0])[-1]
sample_time = times_itrs.get('sample', [0])[-1]
eval_time = times_itrs.get('eval', [0])[-1]
epoch_time = train_time + sample_time + eval_time
total_time = gt.get_times().total
logger.record_tabular('Train Time (s)', train_time)
logger.record_tabular('(Previous) Eval Time (s)', eval_time)
logger.record_tabular('Sample Time (s)', sample_time)
logger.record_tabular('Epoch Time (s)', epoch_time)
logger.record_tabular('Total Train Time (s)', total_time)
# Dump the logger data
logger.dump_tabular(with_prefix=False, with_timestamp=False,
write_header=write_table_header)
# Save pytorch models
self.save_training_state()
logger.log("----")
class MultiGoalReplayBuffer(object):
"""Multigoal Replay Buffer
"""
def __init__(self, max_size, obs_dim, action_dim, num_intentions):
"""
Args:
max_size (int): Maximum buffersize.
obs_dim (int): Observation space dimension.
action_dim (int): Action space dimension.
num_intentions (int):
"""
if not max_size > 1:
raise ValueError("Invalid Maximum Replay Buffer Size: {}".format(
max_size)
)
if not num_intentions >= 0:
raise ValueError("Invalid Num Intentions Size: {}".format(
num_intentions)
)
max_size = int(max_size)
num_intentions = int(num_intentions)
self.obs_buffer = torch.zeros((max_size, obs_dim))
self.acts_buffer = torch.zeros((max_size, action_dim))
self.rewards_buffer = torch.zeros((max_size, 1))
self.termination_buffer = torch.zeros((max_size, 1))
self.next_obs_buffer = torch.zeros((max_size, obs_dim))
# Update reward vector and terminal vector buffers if applicable
if num_intentions > 0:
self.rew_vects_buffer = torch.zeros((max_size, num_intentions))
self.term_vects_buffer = torch.zeros((max_size, num_intentions))
else:
self.rew_vects_buffer = None
self.term_vects_buffer = None
# self.to(device=device)
self.obs_dim = obs_dim
self.action_dim = action_dim
self._max_size = max_size
self._top = 0
self._size = 0
def add_interaction(self, obs, action, reward, termination, next_obs,
reward_vector=None, termination_vector=None):