forked from snu-larr/ibc_official
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
1154 lines (924 loc) · 67.3 KB
/
train.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
os.environ['MUJOCO_GL'] = 'egl'
os.environ['EGL_DEVICE_ID'] = '0'
import matplotlib.pyplot as plt
import seaborn as sns
import copy
import math
import pickle as pkl
import sys
import time
from queue import Queue
import numpy as np
import hydra
import torch
import utils
from logger import Logger
from replay_buffer import ReplayBuffer, HindsightExperienceReplayWrapperVer2
from video import VideoRecorder
from hgg.hgg import goal_distance
from ibc import add_noise_to_goal
torch.backends.cudnn.benchmark = True
class Workspace(object):
def __init__(self, cfg):
self.work_dir = os.getcwd()
print(f'workspace: {self.work_dir}')
self.model_dir = utils.make_dir(self.work_dir, 'model')
self.forward_buffer_dir = utils.make_dir(self.work_dir, 'forward_buffer')
self.backward_buffer_dir = utils.make_dir(self.work_dir, 'backward_buffer')
self.cfg = cfg
max_episode_timesteps_dict = {'sawyer_door' : 200 if cfg.done_on_success else 100, # to prevent collecting too many non-moving transition data at goal state
'tabletop_manipulation' : 100 if cfg.done_on_success else 50,
'fetch_reach_ergodic' : 50,
'fetch_push_ergodic' : 50,
'fetch_pickandplace_ergodic' : 50,
'point_umaze' : 100 if cfg.done_on_success else 50,
}
num_train_steps_dict = {'sawyer_door' : int(2e6),
'tabletop_manipulation' : int(2e6),
'fetch_reach_ergodic' : int(1e6),
'fetch_push_ergodic' : int(2e6),
'fetch_pickandplace_ergodic' : int(2e6),
'point_umaze' : int(1e6),
}
hgg_save_freq_dict = {'sawyer_door' : 2000,
'tabletop_manipulation' : 1000,
'fetch_reach_ergodic' : 1000,
'fetch_push_ergodic' : 1000,
'fetch_pickandplace_ergodic' : 1000,
'point_umaze' : 1000,
}
num_K_dict = {'sawyer_door' : 2,
'tabletop_manipulation' : 5,
'fetch_reach_ergodic' : 50,
'fetch_push_ergodic' : 50,
'fetch_pickandplace_ergodic' : 50,
'point_umaze' : 50,
}
cfg.max_episode_timesteps = max_episode_timesteps_dict[cfg.env]
cfg.num_train_steps = num_train_steps_dict[cfg.env]
cfg.hgg_save_freq = hgg_save_freq_dict[cfg.env]
cfg.num_K = num_K_dict[cfg.env]
self.logger = Logger(self.work_dir,
save_tb=cfg.log_save_tb,
log_frequency=cfg.log_frequency_step,
action_repeat=cfg.action_repeat,
agent='ibc')
utils.set_seed_everywhere(cfg.seed)
self.device = torch.device(cfg.device)
if cfg.env in ['tabletop_manipulation', 'sawyer_door', 'fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
if cfg.env in ['tabletop_manipulation', 'sawyer_door']:
import earl_benchmark
env_loader = earl_benchmark.EARLEnvs(cfg.env, reward_type='sparse')
env, eval_env = env_loader.get_envs()
earl_env = True
if cfg.sparse_reward_type == 'negative':
reward_offset = -1.0 # sparse reward [-1, 0]
elif cfg.sparse_reward_type == 'positive':
reward_offset = 0.0
elif cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
import env_loader
if cfg.full_state_goal:
assert cfg.env not in ['fetch_reach_ergodic', 'point_umaze']
if cfg.xml_path in ['None', 'none', None]:
cfg.xml_path = None
if 'fetch' in cfg.env:
loader = env_loader.GymEnvs(cfg.env+'2', reward_type="sparse", full_state_goal=cfg.full_state_goal, xml_path=cfg.xml_path)
else:
loader = env_loader.GymEnvs(cfg.env, reward_type="sparse", full_state_goal=cfg.full_state_goal, xml_path=cfg.xml_path)
env, eval_env = loader.get_envs()
earl_env = True
if cfg.sparse_reward_type == 'negative':
reward_offset = 0.0 # sparse reward [-1, 0]
elif cfg.sparse_reward_type == 'positive':
reward_offset = 1.0
if cfg.use_curriculum:
'''
NOTE : hgg env should be used only for obatining initial & final goal.
Do not use for rollout as its (train env) horizon is long!
'''
if cfg.env in ['tabletop_manipulation', 'sawyer_door']:
hgg_env_loader = earl_benchmark.EARLEnvs(cfg.env, reward_type='sparse')
hgg_env, hgg_eval_env = hgg_env_loader.get_envs()
elif cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
if cfg.full_state_goal:
assert cfg.env not in ['fetch_reach_ergodic', 'point_umaze']
if 'fetch' in cfg.env:
hgg_env_loader = env_loader.GymEnvs(cfg.env+'2', reward_type="sparse", full_state_goal=cfg.full_state_goal, xml_path=cfg.xml_path)
else:
hgg_env_loader = env_loader.GymEnvs(cfg.env, reward_type="sparse", full_state_goal=cfg.full_state_goal, xml_path=cfg.xml_path)
hgg_env, hgg_eval_env = hgg_env_loader.get_envs()
if cfg.no_backward_if_forward_succeed:
if cfg.env in ['tabletop_manipulation', 'sawyer_door', 'point_umaze']:
if cfg.env in ['sawyer_door']:
assert cfg.num_K==2, 'testing init and final goal'
init_g = hgg_eval_env.initial_states[0,:].copy()
final_g = hgg_eval_env.goal_states[0,:].copy()
custom_task_goal = np.stack([init_g, final_g], axis=0) # [2, dim]
elif cfg.env=='tabletop_manipulation':
initial_states = env.initial_state.copy()[None, :]
goal_states = env.goal_states.copy()
custom_task_goal = np.concatenate([initial_states, goal_states], axis =0)
elif cfg.env=='point_umaze':
if not cfg.load_pretrained:
state_1 = np.linspace([0,0], [8,0], 10)
state_2 = np.linspace([8,0], [8,8], 10)
state_3 = np.linspace([8,8], [0,8], 10)
custom_task_goal = np.concatenate([state_1, state_2, state_3], axis =0)
custom_task_goal += np.random.uniform(-1,1, size=custom_task_goal.shape)*0.1
else:
custom_task_goal = np.array([[0.0, 8.0]])
env.set_custom_task_goal(custom_task_goal)
if cfg.use_curriculum:
hgg_env.set_custom_task_goal(custom_task_goal)
# print(f'set custom task goal for env : {cfg.env}')
sawyer_velocity_info = None
if cfg.env=='sawyer_door':
env.set_velocity_info(sawyer_velocity_info)
eval_env.set_velocity_info(sawyer_velocity_info)
if cfg.use_curriculum:
hgg_env.set_velocity_info(sawyer_velocity_info)
hgg_eval_env.set_velocity_info(sawyer_velocity_info)
if cfg.backward_proprioceptive_only:
from env_utils import RewardChangeWrapperEnv
env = RewardChangeWrapperEnv(env, env_name = cfg.env)
eval_env = RewardChangeWrapperEnv(eval_env, env_name = cfg.env)
if cfg.use_curriculum:
hgg_env = RewardChangeWrapperEnv(hgg_env, env_name = cfg.env)
hgg_eval_env = RewardChangeWrapperEnv(hgg_eval_env, env_name = cfg.env)
from env_utils import RewardOffsetWrapper
env = RewardOffsetWrapper(env, reward_offset=reward_offset)
eval_env = RewardOffsetWrapper(eval_env, reward_offset=reward_offset)
if cfg.use_curriculum:
hgg_env = RewardOffsetWrapper(hgg_env, reward_offset=reward_offset)
hgg_eval_env = RewardOffsetWrapper(hgg_eval_env, reward_offset=reward_offset)
from env_utils import NonEpisodicWrapper, DoneOnSuccessWrapper, HERGoalEnvWrapper
if cfg.done_on_success:
env = DoneOnSuccessWrapper(env, reward_offset=0.0, earl_env = earl_env)
eval_env = DoneOnSuccessWrapper(eval_env, reward_offset=0.0, earl_env = earl_env)
if cfg.use_curriculum:
hgg_env = DoneOnSuccessWrapper(hgg_env, reward_offset=0.0, earl_env = earl_env)
hgg_eval_env = DoneOnSuccessWrapper(hgg_eval_env, reward_offset=0.0, earl_env = earl_env)
if cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
env = HERGoalEnvWrapper(env, env_name= cfg.env)
eval_env = HERGoalEnvWrapper(eval_env, env_name= cfg.env)
if cfg.use_curriculum:
hgg_env = HERGoalEnvWrapper(hgg_env, env_name= cfg.env)
hgg_eval_env = HERGoalEnvWrapper(hgg_eval_env, env_name= cfg.env)
env = NonEpisodicWrapper(env, cfg.env, forward_env_obs_type=cfg.forward_env_obs_type, backward_env_obs_type=cfg.backward_env_obs_type)
eval_env = NonEpisodicWrapper(eval_env, cfg.env, forward_env_obs_type=cfg.forward_env_obs_type, backward_env_obs_type=cfg.backward_env_obs_type)
if cfg.use_curriculum:
hgg_env = NonEpisodicWrapper(hgg_env, cfg.env, forward_env_obs_type=cfg.forward_env_obs_type, backward_env_obs_type=cfg.backward_env_obs_type)
hgg_eval_env = NonEpisodicWrapper(hgg_eval_env, cfg.env, forward_env_obs_type=cfg.forward_env_obs_type, backward_env_obs_type=cfg.backward_env_obs_type)
from env_utils import StateWrapper
self.env = StateWrapper(env)
self.eval_env = StateWrapper(eval_env)
if cfg.use_curriculum:
self.hgg_env = StateWrapper(hgg_env)
self.hgg_eval_env = StateWrapper(hgg_eval_env)
forward_obs_spec = self.env.observation_spec('forward')
backward_obs_spec = self.env.observation_spec('backward')
action_spec = self.env.action_spec()
cfg.agent.action_shape = action_spec.shape
cfg.forward_agent.action_shape = action_spec.shape
cfg.agent.action_range = [
float(action_spec.low.min()),
float(action_spec.high.max())
]
cfg.forward_agent.action_range = [
float(action_spec.low.min()),
float(action_spec.high.max())
]
self.max_episode_timesteps = cfg.max_episode_timesteps
cfg.state_dim = self.env.obs_dim
if cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
cfg.state_goal_dim = self.env.obs_dim+self.env.goal_dim*2 # obs, ag, dg
else:
cfg.state_goal_dim = self.env.obs_dim+self.env.goal_dim # obs(include ag), dg
if cfg.traj_length in ['none', None, 'None']:
cfg.traj_length = None
cfg.forward_agent.obs_shape = forward_obs_spec.shape
self.forward_agent = hydra.utils.instantiate(cfg.forward_agent)
self.backward_agent = None
cfg.agent.obs_shape = backward_obs_spec.shape
self.backward_agent = hydra.utils.instantiate(cfg.agent)
self.forward_buffer = ReplayBuffer(forward_obs_spec.shape, action_spec.shape,
cfg.replay_buffer_capacity,
self.device, traj_length=cfg.traj_length, env_name=cfg.env,
)
from env_utils import WraptoGoalEnv
self.goal_env = goal_env = WraptoGoalEnv(self.env, env_name= cfg.env, sparse_reward_type = cfg.sparse_reward_type)
if self.cfg.use_curriculum:
self.hgg_goal_env = hgg_goal_env = WraptoGoalEnv(self.hgg_env, env_name= cfg.env, sparse_reward_type = cfg.sparse_reward_type)
self.hgg_goal_eval_env = hgg_goal_eval_env = WraptoGoalEnv(self.hgg_eval_env, env_name= cfg.env, sparse_reward_type = cfg.sparse_reward_type)
if cfg.use_forward_her:
assert cfg.forward_env_obs_type=='state_goal'
n_sampled_goal = 4
self.forward_buffer = HindsightExperienceReplayWrapperVer2(self.forward_buffer,
n_sampled_goal=n_sampled_goal,
wrapped_env=goal_env,
env_name = cfg.env,
)
self.backward_buffer = ReplayBuffer(backward_obs_spec.shape, action_spec.shape,
cfg.replay_buffer_capacity,
self.device, traj_length=cfg.traj_length, env_name=cfg.env,
)
if cfg.use_backward_her:
assert cfg.backward_env_obs_type=='state_goal'
self.backward_buffer = HindsightExperienceReplayWrapperVer2(self.backward_buffer,
n_sampled_goal=n_sampled_goal,
wrapped_env=goal_env,
env_name = cfg.env,
)
if cfg.use_curriculum:
from hgg.hgg import TrajectoryPool, MatchSampler
if cfg.forward_curriculum:
self.forward_curriculum_achieved_trajectory_pool = TrajectoryPool(**cfg.hgg_kwargs.trajectory_pool_kwargs)
self.forward_curriculum_sampler = MatchSampler(goal_env=self.hgg_goal_env, goal_eval_env = self.hgg_goal_eval_env, env_name=cfg.env, achieved_trajectory_pool = self.forward_curriculum_achieved_trajectory_pool,
agent_type='forward', **cfg.hgg_kwargs.match_sampler_kwargs)
self.forward_curriculum_sampler.set_networks(critic=self.forward_agent.critic, policy = self.forward_agent.actor)
if cfg.backward_curriculum:
self.backward_curriculum_achieved_trajectory_pool = TrajectoryPool(**cfg.hgg_kwargs.trajectory_pool_kwargs)
self.backward_curriculum_sampler = MatchSampler(goal_env=self.hgg_goal_env, goal_eval_env = self.hgg_goal_eval_env, env_name=cfg.env, achieved_trajectory_pool = self.backward_curriculum_achieved_trajectory_pool,
agent_type='backward', **cfg.hgg_kwargs.match_sampler_kwargs)
self.backward_curriculum_sampler.set_networks(critic=self.backward_agent.critic, policy = self.backward_agent.actor)
else:
pass
from replay_buffer import ForwardBackwardReplayBufferWrapper
self.wrapped_replay_buffer = ForwardBackwardReplayBufferWrapper(self.forward_buffer, self.backward_buffer)
self.non_episodic_video_recorder = VideoRecorder(self.work_dir+'/non_episodic' if cfg.save_video else None, dmc_env=False, env_name=cfg.env)
self.eval_video_recorder = VideoRecorder(self.work_dir if cfg.save_video else None, dmc_env=False, env_name=cfg.env)
self.step = 0
self.forward_step = 0
self.backward_step = 0
def get_agent(self, option = 'forward'):
if option=='forward':
return self.forward_agent
elif option=='backward':
return self.backward_agent
def get_buffer(self, option = 'forward'):
self.wrapped_replay_buffer.set_option(option)
if option=='forward':
if self.cfg.use_forward_her:
self.forward_buffer.replay_buffer.sample_type = None
else:
self.forward_buffer.sample_type = None
return self.forward_buffer
elif option=='backward':
return self.backward_buffer
def get_hgg_sampler(self, option='forward'):
if option=='forward':
return self.forward_curriculum_sampler
elif option=='backward':
return self.backward_curriculum_sampler
def get_hgg_achieved_trajectory_pool(self, option='forward'):
if option=='forward':
return self.forward_curriculum_achieved_trajectory_pool
elif option=='backward':
return self.backward_curriculum_achieved_trajectory_pool
def evaluate(self):
avg_episode_reward = 0
avg_episode_success_rate = 0
if self.cfg.backward_proprioceptive_only:
self.eval_env.set_proprioceptive_only(False)
assert not self.eval_env.proprioceptive_only
for episode in range(self.cfg.num_eval_episodes):
observes = []
obs = self.eval_env.reset()
observes.append(obs)
self.eval_video_recorder.init(enabled=(episode == 0))
episode_reward = 0
episode_success = 0
episode_step = 0
done = False
agent = self.get_agent(option='forward')
while not done:
with utils.eval_mode(agent):
action = agent.act(obs, goal_env=self.goal_env, sample=False)
next_obs, reward, done, info = self.eval_env.step(action)
self.eval_video_recorder.record(self.eval_env)
episode_reward += reward
episode_step += 1
obs = next_obs
observes.append(obs)
if self.eval_env.is_successful(obs):
avg_episode_success_rate+=1.0
avg_episode_reward += episode_reward
self.eval_video_recorder.save(f'{self.step}.mp4')
avg_episode_reward /= self.cfg.num_eval_episodes
avg_episode_success_rate = avg_episode_success_rate/self.cfg.num_eval_episodes
self.logger.log('eval/episode_reward', avg_episode_reward, self.step)
self.logger.log('eval/episode_success_rate', avg_episode_success_rate, self.step)
self.logger.dump(self.step, ty='eval')
def run(self):
self._run_non_episodic()
def _run_non_episodic(self):
episode, episode_reward, episode_step = 0, 0, 0
episode_observes =[]
episode_rewards = []
recent_non_episodic_episode_reward = Queue(50)
if self.cfg.use_curriculum:
recent_sampled_forward_goals = Queue(self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes)
recent_sampled_backward_goals = Queue(self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes)
recent_non_episodic_10k_steps = Queue(10000)
recent_non_episodic_50k_steps = Queue(50000)
recent_non_episodic_100k_steps = Queue(100000)
recent_non_episodic_1k_episodes = Queue(1000)
recent_non_episodic_100_episodes = Queue(100)
forward_episode, backward_episode = 0,0
episode += 1
forward_episode +=1
start_time = time.time()
done = True
option = 'forward'
init_state = None
# pre-collect initial states
initial_states = []
full_initial_states = []
if self.cfg.backward_proprioceptive_only: # to make achieved_goal in full_initial_states as gripper position
self.env.set_proprioceptive_only(True)
for i in range(1000):
obs = self.env.reset()
initial_states.append(obs[:self.env.obs_dim])
full_initial_states.append(obs)
initial_states = np.stack(initial_states, axis =0)
full_initial_states = np.stack(full_initial_states, axis =0)
if self.cfg.backward_proprioceptive_only:
self.env.set_proprioceptive_only(False)
if self.cfg.use_curriculum:
temp_obs = self.hgg_goal_eval_env.reset()
recent_sampled_forward_goals.put(self.hgg_goal_env.convert_obs_to_dict(temp_obs)['achieved_goal'].copy())
if self.cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
K = self.cfg.num_K
assert K > 0
predetermined_initial_goals = []
predetermined_desired_goals = []
for i in range(K):
temp_obs = self.hgg_goal_env.convert_obs_to_dict(self.hgg_env.reset())
goal_a = temp_obs['observation'].copy()
if self.cfg.backward_proprioceptive_only:
goal_a = add_noise_to_goal(goal_a, self.cfg.env)
goal_d = temp_obs['desired_goal'].copy()
predetermined_initial_goals.append(goal_a.copy())
predetermined_desired_goals.append(goal_d.copy())
while self.step <= self.cfg.num_train_steps:
if done: # only done when training horizon is over
self.env.set_option('forward', init_state=None)
if self.cfg.backward_proprioceptive_only:
self.env.set_proprioceptive_only(False)
obs = self.env.reset()
print('done = True at step : ', self.step)
if option=='backward':
forward_episode +=1
option = 'forward'
episode_reward = 0
episode_step = 0
episode_observes = []
episode_rewards = []
elif self.step > 0:
if last_timestep:
fps = episode_step / (time.time() - start_time)
self.logger.log('train/fps', fps, self.step)
start_time = time.time()
if recent_non_episodic_episode_reward.full():
recent_non_episodic_episode_reward.get()
recent_non_episodic_episode_reward.put(episode_reward)
self.logger.log('train/recent_non_episodic_episode_reward', np.array(recent_non_episodic_episode_reward.queue).mean(), self.step)
if self.cfg.use_curriculum :
if self.cfg.env in ['tabletop_manipulation', 'sawyer_door', 'point_umaze']: # pure_obs==ag
assert self.env.custom_task_goal is not None
if option=='forward' and self.cfg.forward_curriculum and (forward_episode % self.cfg.hgg_kwargs.hgg_sampler_update_frequency ==0 or forward_episode==5):
initial_goals = []
desired_goals = []
if self.cfg.env in ['tabletop_manipulation', 'sawyer_door']:
# This process is just copying firstly assigned custom task goal.
K = self.cfg.num_K
assert K > 0 and K == self.env.custom_task_goal.shape[0]
if self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes==K:
for i in range(self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes):
temp_obs = self.hgg_goal_env.convert_obs_to_dict(self.hgg_env.reset())
goal_a = temp_obs['achieved_goal'].copy()
if self.cfg.backward_proprioceptive_only:
goal_a = add_noise_to_goal(goal_a, self.cfg.env)
goal_d = self.env.custom_task_goal[i].copy()
initial_goals.append(goal_a.copy())
desired_goals.append(goal_d.copy())
elif self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes > K:
for i in range(self.cfg.hgg_kwargs.match_sampler_kwargs.num_episodes):
temp_obs = self.hgg_goal_env.convert_obs_to_dict(self.hgg_env.reset())
goal_a = temp_obs['achieved_goal'].copy()
if self.cfg.backward_proprioceptive_only:
goal_a = add_noise_to_goal(goal_a, self.cfg.env)
temp_idx = int(i%K)
goal_d = self.env.custom_task_goal[temp_idx].copy()
initial_goals.append(goal_a.copy())
desired_goals.append(goal_d.copy())
else:
raise NotImplementedError
elif self.cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
initial_goals = copy.deepcopy(predetermined_initial_goals)
desired_goals = copy.deepcopy(predetermined_desired_goals)
hgg_sampler = self.get_hgg_sampler(option=option)
hgg_sampler.update(initial_goals, desired_goals)
# non episodic
if option == 'forward':
self.logger.log('train/forward_episode_reward', episode_reward, self.step)
self.logger.log('train/forward_episode', forward_episode, self.step)
forward_episode +=1
# if forward was success keep option as forward
if (self.cfg.no_backward_if_forward_succeed and self.eval_env.is_successful(obs)):
option = 'forward'
self.env.set_proprioceptive_only(False)
if self.env.forward_env_obs_type=='state_goal':
if self.cfg.forward_curriculum:
hgg_sampler = self.get_hgg_sampler(option='forward')
n_iter = 0
while True:
forward_goal = hgg_sampler.sample(np.random.randint(len(hgg_sampler.pool)), backward_proprioceptive=False).copy()
# exclude already success goal
obs_for_success_check = obs.copy()
obs_for_success_check = self.env.replace_goal_in_obs(obs_for_success_check, forward_goal)
if not self.env.is_successful(obs_for_success_check, proprioceptive_only=False):
break
n_iter +=1
if n_iter==2:
self.env.reset_goal(add_noise = self.cfg.add_noise_to_forward_goal)
forward_goal = self.env.goal.copy().astype(np.float32)
break
self.env.reset_goal(goal=forward_goal, add_noise = self.cfg.add_noise_to_forward_goal) # random target goal for every episode
forward_goal = self.env.goal.copy().astype(np.float32)
if recent_sampled_forward_goals.full():
recent_sampled_forward_goals.get()
recent_sampled_forward_goals.put(forward_goal)
else:
prev_goal = self.goal_env.convert_obs_to_dict(obs)['desired_goal']
same_as_before = True
while same_as_before:
self.env.reset_goal(add_noise = self.cfg.add_noise_to_forward_goal) # random target goal for every episode
temp_goal = self.env.goal.copy().astype(np.float32)
same_as_before = np.linalg.norm(prev_goal-temp_goal, axis =-1) < 0.001
forward_goal = self.env.goal.copy().astype(np.float32)
obs_info = {'goal' : forward_goal}
else:
obs_info = None
if self.env.backward_env_obs_type=='state_goal':
obs = self.env.replace_goal_in_obs(obs, forward_goal)
else:
raise NotImplementedError
# print('option is changed from forward to forward. goal in obs : {}'.format(self.env.get_pure_goal_from_obs(obs).squeeze()))
else: # change from forward to backward
option = 'backward'
if self.cfg.backward_proprioceptive_only:
self.env.set_proprioceptive_only(True)
if self.cfg.forward_curriculum and self.cfg.backward_proprioceptive_only : # consider backward curriculum goal for backward reaching
hgg_sampler = self.get_hgg_sampler(option='forward')
n_iter = 0
while True:
backward_goal = hgg_sampler.sample(np.random.randint(len(hgg_sampler.pool)), backward_proprioceptive=True).copy()
# exclude already success goal
obs_for_success_check = obs.copy()
obs_for_success_check = self.env.replace_goal_in_obs(obs_for_success_check, backward_goal)
if not self.env.is_successful(obs_for_success_check, proprioceptive_only=True):
break
n_iter +=1
if n_iter==2:
break
if recent_sampled_backward_goals.full():
recent_sampled_backward_goals.get()
recent_sampled_backward_goals.put(backward_goal)
elif self.cfg.use_curriculum and self.cfg.backward_curriculum : # consider backward curriculum goal (previous)
raise NotImplementedError
else: # just utilize initial state
if self.cfg.env in ['fetch_push_ergodic', 'fetch_pickandplace_ergodic']:
if self.env.full_state_goal:
assert full_initial_states.shape[-1]==37
backward_goal = full_initial_states[np.random.randint(0, full_initial_states.shape[0]), -12:-6]
else:
assert full_initial_states.shape[-1]==31
backward_goal = full_initial_states[np.random.randint(0, full_initial_states.shape[0]), -6:-3]
elif self.cfg.env in ['fetch_reach_ergodic']:
assert full_initial_states.shape[-1]==16
backward_goal = full_initial_states[np.random.randint(0, full_initial_states.shape[0]), -6:-3]
elif self.cfg.env in ['point_umaze']:
assert full_initial_states.shape[-1]==11
backward_goal = full_initial_states[np.random.randint(0, full_initial_states.shape[0]), -4:-2]
elif self.cfg.env in ['sawyer_door']:
backward_goal = initial_states[np.random.randint(0, initial_states.shape[0]), :7]
else:
backward_goal = initial_states[np.random.randint(0, initial_states.shape[0])]
if self.cfg.backward_proprioceptive_only:
backward_goal = add_noise_to_goal(backward_goal, self.cfg.env)
self.env.reset_goal(goal = backward_goal) # random target goal for every episode
goal = self.env.goal.copy().astype(np.float32)
assert (backward_goal==goal).all()
obs = self.env.replace_goal_in_obs(obs, goal)
obs = self.env.convert_obs(obs, option)
# print('option is changed from forward to backward. goal in obs : {}'.format(self.env.get_pure_goal_from_obs(obs).squeeze()))
elif option == 'backward': # change from backward to forward
self.logger.log('train/backward_episode_reward', episode_reward, self.step)
self.logger.log('train/backward_episode', backward_episode, self.step)
backward_episode +=1
option = 'forward'
if self.cfg.backward_proprioceptive_only:
self.env.set_proprioceptive_only(False)
if self.env.forward_env_obs_type=='state_goal':
if self.cfg.forward_curriculum and self.cfg.backward_proprioceptive_only:
hgg_sampler = self.get_hgg_sampler(option='forward')
n_iter = 0
while True:
forward_goal = hgg_sampler.sample(np.random.randint(len(hgg_sampler.pool))).copy()
# exclude already success goal
obs_for_success_check = obs.copy()
obs_for_success_check = self.env.replace_goal_in_obs(obs_for_success_check, forward_goal)
if not self.env.is_successful(obs_for_success_check, proprioceptive_only=False):
break
n_iter +=1
if n_iter==2:
# sample task goal from env
self.env.reset_goal(add_noise = self.cfg.add_noise_to_forward_goal)
forward_goal = self.env.goal.copy().astype(np.float32)
break
if recent_sampled_forward_goals.full():
recent_sampled_forward_goals.get()
recent_sampled_forward_goals.put(forward_goal)
self.env.reset_goal(goal = forward_goal, add_noise = self.cfg.add_noise_to_forward_goal)
if not self.cfg.add_noise_to_forward_goal:
assert (self.env.goal.copy()==forward_goal).all()
elif self.cfg.forward_curriculum : # consider backward curriculum goal (previous)
raise NotImplementedError
else:
self.env.reset_goal(add_noise = self.cfg.add_noise_to_forward_goal) # random target goal for every episode
forward_goal = self.env.goal.copy().astype(np.float32)
obs_info = {'goal' : forward_goal}
else:
obs_info = None
if self.env.backward_env_obs_type=='state_goal':
obs = self.env.replace_goal_in_obs(obs, forward_goal)
else:
obs = self.env.convert_obs(obs, option, obs_info)
# print('option is changed from backward to forward. goal in obs : {}'.format(self.env.get_pure_goal_from_obs(obs).squeeze()))
episode_reward = 0
episode_step = 0
episode_observes = []
episode_rewards = []
episode += 1
self.env.set_option(option=option, init_state=init_state)
agent = self.get_agent(option=option)
replay_buffer = self.get_buffer(option=option)
# evaluate agent periodically
if self.step % self.cfg.eval_frequency == 0:
print('eval started...')
self.logger.log('eval/episode', episode - 1, self.step)
self.evaluate()
# save agent periodically
if self.cfg.save_model and self.step % self.cfg.save_frequency == 0:
utils.save(
self.forward_agent,
os.path.join(self.model_dir, f'forward_agent_{self.step}.pt'))
utils.save(
self.backward_agent,
os.path.join(self.model_dir, f'backward_agent_{self.step}.pt'))
if self.cfg.save_buffer and self.step % self.cfg.save_frequency == 0:
utils.save(self.forward_buffer.sample_all_data(), os.path.join(self.forward_buffer_dir, f'forward_buffer.pt'))
utils.save(self.backward_buffer.sample_all_data(), os.path.join(self.backward_buffer_dir, f'backward_buffer.pt'))
if self.step % self.cfg.non_episodic_video_save_frequency == 0 or self.step in [4000, 8000, 12000, 16000]:
self.non_episodic_video_recorder.init(enabled=True)
non_episodic_obses = []
non_episodic_actions = []
non_episodic_rewards = []
non_episodic_dones = []
non_episodic_next_obses = []
non_episodic_backward_indices_of_record = []
non_episodic_idx_for_record = 0
# sample action for data collection
if self.step < self.cfg.num_random_steps:
spec = self.env.action_spec()
action = np.random.uniform(spec.low, spec.high,
spec.shape)
else:
with utils.eval_mode(agent):
action = agent.act(obs, goal_env=self.goal_env, sample=True)
logging_dict = agent.update(self.wrapped_replay_buffer, self.step, self.goal_env)
if self.forward_step % self.cfg.logging_frequency== 0:
if logging_dict is not None: # when step = 0
if option=='forward':
for key, val in logging_dict.items():
self.logger.log('train/forward/'+key, val, self.step)
# just for debug
self.logger.log('train/forward_step', self.forward_step, self.step)
elif self.backward_step % self.cfg.logging_frequency== 0:
if logging_dict is not None: # when step = 0
if option=='backward':
for key, val in logging_dict.items():
self.logger.log('train/backward/'+key, val, self.step)
# just for debug
self.logger.log('train/backward_step', self.backward_step, self.step)
if self.step>0 and (self.step % self.cfg.logging_frequency == 0):
# just for logging
self.logger.log('train/recent_non_episodic_10k_forward_ratio', np.array(recent_non_episodic_10k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_50k_forward_ratio', np.array(recent_non_episodic_50k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_100k_forward_ratio', np.array(recent_non_episodic_100k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_10k_backward_ratio', 1-np.array(recent_non_episodic_10k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_50k_backward_ratio', 1-np.array(recent_non_episodic_50k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_100k_backward_ratio', 1-np.array(recent_non_episodic_100k_steps.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_100_forward_episode_ratio', np.array(recent_non_episodic_100_episodes.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_1k_forward_episode_ratio', np.array(recent_non_episodic_1k_episodes.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_100_backward_episode_ratio', 1-np.array(recent_non_episodic_100_episodes.queue).mean(), self.step)
self.logger.log('train/recent_non_episodic_1k_backward_episode_ratio', 1-np.array(recent_non_episodic_1k_episodes.queue).mean(), self.step)
if self.step >=500 and ((self.step % self.cfg.hgg_save_freq == 0) or self.step in np.arange(0, 500*20+1, 500)):
if self.cfg.forward_curriculum:
hgg_sampler = self.get_hgg_sampler(option='forward')
if hgg_sampler.total_cost is not None:
self.logger.log('train/forward_curriculum_total_cost_mean', hgg_sampler.total_cost.mean(), self.step)
if hgg_sampler.total_forward_cost is not None:
self.logger.log('train/forward_curriculum_total_forward_cost_mean', hgg_sampler.total_forward_cost.mean(), self.step)
if hgg_sampler.total_backward_cost is not None:
self.logger.log('train/forward_curriculum_total_backward_cost_mean', hgg_sampler.total_backward_cost.mean(), self.step)
import matplotlib.pyplot as plt
import seaborn as sns
sampled_goals_for_vis = np.array(recent_sampled_forward_goals.queue)
if self.cfg.env in ['fetch_reach_ergodic', 'fetch_pickandplace_ergodic']:
pass
else:
# plot and save
fig = plt.figure()
sns.set_style("darkgrid")
ax1 = fig.add_subplot(1,1,1)
if self.cfg.env in ['point_umaze']:
ax1.scatter(sampled_goals_for_vis[:, 0], sampled_goals_for_vis[:, 1], c='red')
plt.xlim(-2,10)
plt.ylim(-2,10)
elif self.cfg.env == 'tabletop_manipulation':
ax1.scatter(sampled_goals_for_vis[:, 2], sampled_goals_for_vis[:, 3], c='red')
ax1.scatter(sampled_goals_for_vis[:, 0], sampled_goals_for_vis[:, 1], c='blue')
plt.xlim(-2.8,2.8)
plt.ylim(-2.8,2.8)
elif self.cfg.env == 'sawyer_door':
ax1.scatter(sampled_goals_for_vis[:, 4], sampled_goals_for_vis[:, 5], c='red')
plt.xlim(-0.2,0.4)
plt.ylim(0.3,0.9)
elif self.cfg.env == 'fetch_push_ergodic':
ax1.scatter(sampled_goals_for_vis[:, 0], sampled_goals_for_vis[:, 1], c='red')
plt.xlim(0.8,1.8)
plt.ylim(0.2,1.2)
else:
raise NotImplementedError
plt.savefig(self.non_episodic_video_recorder.save_dir+'/train_hgg_forward_goals_step_'+str(self.step)+'.jpg')
plt.close()
with open(self.non_episodic_video_recorder.save_dir+'/train_hgg_forward_goals_step_'+str(self.step)+'.pkl', 'wb') as f:
pkl.dump(sampled_goals_for_vis, f)
if self.cfg.backward_proprioceptive_only:
import matplotlib.pyplot as plt
import seaborn as sns
sampled_goals_for_vis = np.array(recent_sampled_backward_goals.queue)
if self.cfg.env in ['fetch_reach_ergodic', 'fetch_pickandplace_ergodic']:
pass
elif self.cfg.env in ['sawyer_door']:
pass
elif self.cfg.env == 'fetch_push_ergodic':
pass
else:
# plot and save
fig = plt.figure()
sns.set_style("darkgrid")
ax1 = fig.add_subplot(1,1,1)
if self.cfg.env in ['point_umaze']:
ax1.scatter(sampled_goals_for_vis[:, 0], sampled_goals_for_vis[:, 1], c='blue')
plt.xlim(-2,10)
plt.ylim(-2,10)
elif self.cfg.env == 'tabletop_manipulation':
ax1.scatter(sampled_goals_for_vis[:, 0], sampled_goals_for_vis[:, 1], c='blue')
plt.xlim(-2.8,2.8)
plt.ylim(-2.8,2.8)
else:
raise NotImplementedError
plt.savefig(self.non_episodic_video_recorder.save_dir+'/train_hgg_backward_goals_step_'+str(self.step)+'.jpg')
plt.close()
with open(self.non_episodic_video_recorder.save_dir+'/train_hgg_backward_goals_step_'+str(self.step)+'.pkl', 'wb') as f:
pkl.dump(sampled_goals_for_vis, f)
next_obs, reward, done, info = self.env.step(action)
episode_reward += reward
last_timestep = True if (episode_step+1) % self.max_episode_timesteps == 0 or done else False
if last_timestep:
# NOTE: It is meaningless when consider_done_true_in_critic is False
if self.cfg.done_on_success: # earl done or success or max episode length
earl_done = info['earl_done'] # last_timestep not always means earl_done=True
if earl_done: # earl_done & (success or max episode timestep) simultaneously rarely happen -> ignore
done = False
else: # success or max episode length
# sparse reward
if done : # success
pass
else: # max episode step
done = False
else: # earl done or max episode length
if done: # earl done = True regardless of steps (assume done from env is False)
earl_done=True
else: # max episode step
earl_done=False
done = False # done = False as done_on_success=False
episode_observes.append(obs)
episode_rewards.append(reward)
if self.cfg.use_forward_her or self.cfg.use_backward_her:
if option=='forward' :
assert self.cfg.use_forward_her
replay_buffer.add(obs, action, reward, next_obs, done, last_timestep)
elif option=='backward' :
assert self.cfg.use_backward_her
replay_buffer.add(obs, action, reward, next_obs, done, last_timestep)
else: # no her
# TODO: you should consider forward, backward
replay_buffer.add(obs, action, reward, next_obs, done)
if last_timestep:
if self.cfg.use_curriculum:
if option=='forward' and (not self.cfg.forward_curriculum):
pass
elif option=='backward' and (not self.cfg.backward_curriculum) and (not self.cfg.backward_proprioceptive_only):
pass
else:
if option=='forward': # only insert traj in forward hgg pool
temp_episode_observes = copy.deepcopy(episode_observes)
temp_episode_ag = []
if self.cfg.env in ['tabletop_manipulation', 'sawyer_door']:
temp_episode_init = self.goal_env.convert_obs_to_dict(temp_episode_observes[0])['observation'] # NOTE : should it be [obs, ag] ?
elif self.cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
temp_episode_init_obs = self.goal_env.convert_obs_to_dict(temp_episode_observes[0])['observation']
temp_episode_init_ag = self.goal_env.convert_obs_to_dict(temp_episode_observes[0])['achieved_goal']
temp_episode_init = np.concatenate([temp_episode_init_obs, temp_episode_init_ag], axis =-1)
else:
raise NotImplementedError
for k in range(len(temp_episode_observes)):
if self.cfg.env in ['tabletop_manipulation', 'sawyer_door']:
temp_episode_ag.append(self.goal_env.convert_obs_to_dict(temp_episode_observes[k])['achieved_goal']) # pure_obs == ag
elif self.cfg.env in ['fetch_reach_ergodic', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic', 'point_umaze']:
temp_episode_ag.append(self.goal_env.convert_obs_to_dict(temp_episode_observes[k])['observation']) # pure_obs != ag
else:
raise NotImplementedError
# NOTE : currently, add episodewise, thus list has only 1 element
achieved_trajectories = [np.array(temp_episode_ag)] # list of [ts, dim]
achieved_init_states = [temp_episode_init] # list of [ts(1), dim]
selection_trajectory_idx = {}
for i in range(len(achieved_trajectories)): # 1
# if there is a difference btw first and last timestep, then add it.
if self.cfg.backward_proprioceptive_only:
if self.cfg.env in ['sawyer_door', 'tabletop_manipulation', 'fetch_push_ergodic', 'fetch_pickandplace_ergodic']:
if 'sawyer' in self.cfg.env: # consider gripper, object related states
if goal_distance(achieved_trajectories[i][0][4:7], achieved_trajectories[i][-1][4:7])>0.01 or goal_distance(achieved_trajectories[i][0][:3], achieved_trajectories[i][-1][:3])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env=='tabletop_manipulation': # only consider object related states
if goal_distance(achieved_trajectories[i][0][2:4], achieved_trajectories[i][-1][2:4])>0.01 or goal_distance(achieved_trajectories[i][0][:2], achieved_trajectories[i][-1][:2])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['fetch_push_ergodic', 'fetch_pickandplace_ergodic']:
if goal_distance(achieved_trajectories[i][0][3:6], achieved_trajectories[i][-1][3:6])>0.01 or goal_distance(achieved_trajectories[i][0][:3], achieved_trajectories[i][-1][:3])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['fetch_reach_ergodic']:
if goal_distance(achieved_trajectories[i][0][:3], achieved_trajectories[i][-1][:3])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['point_umaze']:
if goal_distance(achieved_trajectories[i][0][:2], achieved_trajectories[i][-1][:2])>0.1:
selection_trajectory_idx[i] = True
else: # full state achieved_goal
if goal_distance(achieved_trajectories[i][0], achieved_trajectories[i][-1])>0.01:
selection_trajectory_idx[i] = True
else:
if 'sawyer' in self.cfg.env: # only consider object related states
if goal_distance(achieved_trajectories[i][0][4:7], achieved_trajectories[i][-1][4:7])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env=='tabletop_manipulation': # only consider object related states
if goal_distance(achieved_trajectories[i][0][2:4], achieved_trajectories[i][-1][2:4])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['fetch_push_ergodic', 'fetch_pickandplace_ergodic']:
if goal_distance(achieved_trajectories[i][0][3:6], achieved_trajectories[i][-1][3:6])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['fetch_reach_ergodic']:
if goal_distance(achieved_trajectories[i][0][:3], achieved_trajectories[i][-1][:3])>0.01:
selection_trajectory_idx[i] = True
elif self.cfg.env in ['point_umaze']:
if goal_distance(achieved_trajectories[i][0][:2], achieved_trajectories[i][-1][:2])>0.1:
selection_trajectory_idx[i] = True
else: # full state achieved_goal
if goal_distance(achieved_trajectories[i][0], achieved_trajectories[i][-1])>0.01: