-
Notifications
You must be signed in to change notification settings - Fork 5
/
toy_RL_pipeline.py
883 lines (683 loc) · 36.8 KB
/
toy_RL_pipeline.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
"""Defines and trains RL models for either CartPole or FrozenLake environments"""
import numpy as np
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.distributions import Categorical
from collections import deque
import os
import argparse
from copy import deepcopy
import config
import data
import cart_pole_env
import frozen_lake_env
import utils
import MCTS
DEVICE = config.DEVICE
EPS = config.EPS
torch.manual_seed(config.SEED)
# Define command line arguments for experiment
parser = argparse.ArgumentParser(description='Train_RL_Model')
parser.add_argument('--train_models', action='store_true', help='enable training of RL models')
parser.add_argument('--folder_name', type=str,
help='Brief description of experiment (no spaces)')
parser.add_argument('--sparse_env', type=int, choices={0, 1}, default=1,
help='sparsifies environment rewards')
parser.add_argument('--relative_rewards', type=int, choices={0, 1}, default=0,
help='use rewards relative to supervised model')
parser.add_argument('--env_name', type=str, default='CartPole',
choices=['CartPole', 'FrozenLake'], help='select RL environment')
parser.add_argument('--use_pretrained_critic', type=int, choices={0, 1}, default=1,
help='critic is initialized with pretrained model')
parser.add_argument('--pretrain_critic_n_episodes', type=int, default=0,
help='number of iterations to pretrain the critic (default: 0)')
parser.add_argument('--n_episodes', type=int, default=2500,
help='max number of iterations to train the RL model (default: 2500)')
parser.add_argument('--verbose_training', type=int, choices={0, 1}, default=0,
help='print results during training')
parser.add_argument('--init_critic', type=int, choices={0, 1}, default=1, help='initializes critic model')
parser.add_argument('--transfer_weights', type=int, choices={0, 1}, default=1,
help='transfers weights from supervised model to actor')
parser.add_argument('--use_policy_distillation', type=int, choices={0, 1}, default=0,
help='adds policy distillation error to reward function')
parser.add_argument('--MCTS_thresh', type=float, default=0,
help='Uses MCTS unless max certainty is above specified prob (default: 0)')
# Mostly for helper functions and debugging
parser.add_argument('--update_RL_models', type=int, choices={0, 1}, default=1,
help='allows the update of the rl models')
parser.add_argument('--use_MLE', type=int, choices={0, 1}, default=0,
help='can use MLE instead of sample')
parser.add_argument('--load_models', action='store_true', help='Load pretrained model from prior point')
parser.add_argument('--load_model_folder_name', type=str,
help='folder which contains the saved models to be used')
args = parser.parse_args()
args.save_models = 0
if args.train_models:
args.save_models = 1
saved_RL_model_results = data.SaveRLModelResults(args.env_name, args.folder_name)
saved_RL_model_results.check_folder_exists()
# Specify which encoder / decoder is used
args.FROZENLAKE_ENCODER = 'FrozenLake/FrozenLakeEncoder_medium.pt'
args.FROZENLAKE_DECODER = 'FrozenLake/FrozenLakeDecoder_medium.pt'
args.CARTPOLE_DECODER = 'CartPole/CartpoleDecoder.pt'
args.CP_PRETRAINED_CRITIC = 'CP_critic_model_585.pt'
args.FL_PRETRAINED_CRITIC = 'FL_critic_model_5000.pt'
#%% Manual Testing - turn train models on while keeping save models off then modify as you like
#Run test
#args.train_models = 1
#args.env_name = 'FrozenLake'
#args.verbose_training = 1
#saved_RL_model_results = data.SaveRLModelResults(args.env_name, 'Test')
#
#args.n_episodes = 3000
#args.load_models = 1
#args.load_model_folder_name = os.path.join(config.saved_RL_model_path, 'FrozenLake/Medium/Test3/Transfer_Weights10/')
#args.update_RL_models = 0
#args.use_MLE = 0
#args.MCTS_thresh = 0.65
#%% Load environment data
def load_FrozenLake_data():
"""Loads FrozenLake training data"""
input_map_df = data.load_np_data('Data/RL_Data/FL_input_map_df.npy')
visited_states_df = np.load('Data/RL_Data/FL_states_df.npy', allow_pickle=True)
selected_actions_df = np.load('Data/RL_Data/FL_selected_action_sequence_df.npy', allow_pickle=True)
train_data, val_data, test_data = data.train_test_split(
list(zip(input_map_df, visited_states_df, selected_actions_df)))
assert all([len(train_data[i][1]) == len(train_data[i][2]) for i in range(len(train_data))]), \
'Lengths of states and actions are not equal'
return train_data, val_data, test_data
def load_CartPole_data():
"""Loads CartPole training data"""
visited_states_df = np.load('Data/RL_Data/CP_states_df.npy')
selected_actions_df = np.load('Data/RL_Data/CP_selected_action_sequence_df.npy')
input_map_df = np.zeros((len(visited_states_df),1))
train_data, val_data, test_data = data.train_test_split(
list(zip(input_map_df, visited_states_df, selected_actions_df)))
assert all([len(train_data[i][1]) == len(train_data[i][2]) for i in range(len(train_data))]), \
'Lengths of states and actions are not equal'
return train_data, val_data, test_data
def load_env_data(env_name):
"""Initializes data for relevant environment"""
if env_name == 'CartPole':
train_data, val_data, test_data = load_CartPole_data()
elif env_name == 'FrozenLake':
train_data, val_data, test_data = load_FrozenLake_data()
else:
print("Please select one of the following environments: ['FrozenLake', 'CartPole']")
return train_data, val_data, test_data
train_data, val_data, test_data = load_env_data(args.env_name)
#%% Define supervised models
class MLPStateEncoder(nn.Module):
"""Uses a MLP state encoder for FrozenLake"""
def __init__(self, input_size, hidden_size, output_size):
super(MLPStateEncoder, self).__init__()
self.name = 'MLPStateEncoder'
self.hidden_size = hidden_size
self.h1 = nn.Linear(input_size, self.hidden_size)
self.dropout = nn.Dropout(p=0.6)
self.h2 = nn.Linear(self.hidden_size, output_size)
def forward(self, x):
x = torch.tensor(x, dtype=torch.float32, device=DEVICE).view(1,1,-1)
x = self.h1(x)
x = self.dropout(x)
x = F.relu(x)
action_scores = self.h2(x)
return action_scores
class CNNStateEncoder(nn.Module):
"""CNN encoder model designed for FrozenLake with a 5x5 input state space.
Would need to manually update hyperparameters for a different state space size"""
def __init__(self, output_size):
super(CNNStateEncoder, self).__init__()
self.name = 'CNNStateEncoder'
self.conv1 = nn.Conv2d(1, 16, kernel_size=3, stride=1, padding=1)
self.maxpool1 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
self.conv2 = nn.Conv2d(16, 16, kernel_size=3, stride=2, padding=1)
self.maxpool2 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
self.linear_output = nn.Linear(144, output_size)
def forward(self, x):
x = torch.tensor(x, dtype=torch.float32, device=DEVICE).view(1,1,5,5)
x = F.relu(self.conv1(x))
x = self.maxpool1(x)
x = F.relu(self.conv2(x))
x = x.view(1,1,-1)
output = self.linear_output(x)
return output
class GeneralDecoderRNN(nn.Module):
"""Vanilla decoder (WITH NO EMBEDDINGS) which decodes based on single context vector"""
def __init__(self, input_size, hidden_size, output_size):
super(GeneralDecoderRNN, self).__init__()
self.name = 'GeneralDecoderRNN'
self.hidden_size = hidden_size
self.gru = nn.GRU(input_size, self.hidden_size)
self.out = nn.Linear(self.hidden_size, output_size)
self.softmax = nn.LogSoftmax(dim=1)
def forward(self, input, hidden):
output = torch.tensor(input, dtype=torch.float32, device=DEVICE).view(1,1,-1)
output, hidden = self.gru(output, hidden)
output = self.softmax(self.out(output[0]))
return output, hidden
def initHidden(self):
return torch.zeros(1, 1, self.hidden_size, device=DEVICE)
class TeacherRNN(nn.Module):
"""Vanilla decoder (WITH NO EMBEDDINGS) which decodes based on single context vector
Has the same architecture as the General Decoder RNN - the only difference is the addition
of temperature and softmax instead of log softmax"""
def __init__(self, input_size, hidden_size, output_size):
super(TeacherRNN, self).__init__()
self.name = 'GeneralDecoderRNN'
self.hidden_size = hidden_size
self.gru = nn.GRU(input_size, self.hidden_size)
self.out = nn.Linear(self.hidden_size, output_size)
self.softmax = nn.Softmax(dim=1)
def forward(self, input, hidden, temperature=1):
output = torch.tensor(input, dtype=torch.float32, device=DEVICE).view(1,1,-1)
output, hidden = self.gru(output, hidden)
output = self.softmax(self.out(output[0]) / temperature)
return output, hidden
def initHidden(self):
return torch.zeros(1, 1, self.hidden_size, device=DEVICE)
#%% Train and evaluate models
def train_supervised(input_map, visited_states, selected_actions, encoder, decoder,
encoder_optimizer, decoder_optimizer, criterion):
"""Train supervised models for CartPole or FrozenLake"""
if encoder != None:
encoder.zero_grad()
hidden_state = encoder(input_map)
else:
hidden_state = decoder.initHidden()
decoder.zero_grad()
target_length = len(selected_actions)
loss = 0
for i in range(target_length):
state_input = visited_states[i]
model_output, hidden_state = decoder(
state_input, hidden_state)
loss += criterion(model_output, torch.tensor([selected_actions[i]],
dtype=torch.long, device=DEVICE))
loss.backward()
if encoder != None:
encoder_optimizer.step()
decoder_optimizer.step()
return loss.item() / target_length
def trainIters(input_data, encoder, decoder,
encoder_optimizer, decoder_optimizer, n_iters=20, print_every=10):
"""Applies training loop to train models on data"""
if encoder != None:
encoder.train()
decoder.train()
start = time.time()
print_loss_total = 0 # Reset every print_every
criterion = nn.NLLLoss()
# Sample n random pairs
selected_indices = np.random.choice(len(input_data), n_iters, replace=False)
# For EACH pair train model to decrease loss
for idx, selected_idx in enumerate(selected_indices):
loss = train_supervised(input_data[selected_idx][0], input_data[selected_idx][1],
input_data[selected_idx][2], encoder, decoder,
encoder_optimizer, decoder_optimizer, criterion)
print_loss_total += loss
iter = idx+1
if iter % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print('%s (%d %d%%) %.4f' % (utils.timeSince(start, iter / n_iters),
iter, iter / n_iters * 100, print_loss_avg))
def evaluate(input_map, visited_states, selected_actions, encoder, decoder, criterion):
"""Evaluate the performance of the trained models"""
with torch.no_grad():
if encoder != None:
hidden_state = encoder(input_map)
else:
hidden_state = decoder.initHidden()
target_length = len(visited_states)
loss = 0
for i in range(target_length):
state_input = visited_states[i]
model_output, hidden_state = decoder(
state_input, hidden_state)
loss += criterion(model_output, torch.tensor([selected_actions[i]],
dtype=torch.long, device=DEVICE))
return loss.item() / target_length
def validationError(input_data, encoder, decoder, verbose=True):
"""Evalutes the error on a set of input pairs in terms of loss.
Is intended to be used on a validation or test set to evaluate performance"""
if encoder != None:
encoder.eval()
decoder.eval()
criterion = nn.NLLLoss()
loss = 0
for selected_idx in range(len(input_data)):
pair_loss = evaluate(input_data[selected_idx][0], input_data[selected_idx][1],
input_data[selected_idx][2],encoder, decoder, criterion)
loss += pair_loss
avg_loss = loss / len(input_data)
if verbose:
print('The average validation loss is {:.3} based on {} samples'.format(avg_loss, len(input_data)))
return avg_loss
# %% Train supervised model
# Define encoder / decoder models
def train_CartPole_supervised_models():
"""Trains the CartPole supervised model"""
encoder = None
encoder_optimizer = None
decoder = GeneralDecoderRNN(input_size=4, hidden_size=128, output_size=2).to(DEVICE)
# Set optimizer
decoder_optimizer = optim.Adam(decoder.parameters())
n_iterations = 5
for i in range(n_iterations):
print('Iteration Number: {}'.format(i))
trainIters(train_data, encoder, decoder, encoder_optimizer, decoder_optimizer,
n_iters=50, print_every=10)
validationError(val_data[:10], encoder, decoder)
return encoder, decoder
def train_FrozenLake_supervised_models():
"""Trains the FrozenLake supervised model"""
encoder = CNNStateEncoder(128).to(DEVICE)
encoder_optimizer = optim.Adam(encoder.parameters())
decoder = GeneralDecoderRNN(input_size=2, hidden_size=128, output_size=4).to(DEVICE)
# Set optimizer
decoder_optimizer = optim.Adam(decoder.parameters())
n_iterations = 5
for i in range(n_iterations):
print('Iteration Number: {}'.format(i))
trainIters(train_data, encoder, decoder, encoder_optimizer, decoder_optimizer,
n_iters=2500, print_every=500)
validationError(val_data[:10], encoder, decoder)
return encoder, decoder
def supervised_model_reward(input_env, start_state, supervised_encoder, supervised_model):
"""Returns the baseline reward of running the supervised model"""
supervised_env = deepcopy(input_env)
supervised_env.state = start_state
supervised_env.done, supervised_env.ep_reward = False, 0
state = supervised_env.state
if supervised_env.name == 'FrozenLake':
masked_input_map = frozen_lake_env.mask_map(supervised_env.input_map, flatten=False)
hidden_state = supervised_encoder(masked_input_map).detach()
else:
hidden_state = supervised_model.initHidden()
for i in range(env.max_steps):
probs, hidden_state = supervised_model(state, hidden_state)
_, topi = probs.data.topk(1)
action = topi.squeeze().item()
state, env_reward, done, _ = supervised_env.step(action)
if done:
break
return supervised_env.ep_reward
#data.save_model(decoder, os.path.join(config.saved_RL_model_path, 'CartPole/CartpoleDecoder.pt'))
#data.save_model(encoder, os.path.join(config.saved_RL_model_path, 'FrozenLake/FrozenLakeEncoder_superexpert.pt'))
#data.save_model(decoder, os.path.join(config.saved_RL_model_path, 'FrozenLake/FrozenLakeDecoder_superexpert.pt'))
#%% Define reinforcement learning models
class RLActor(nn.Module):
"""Defines the RL Actor model"""
def __init__(self, input_size, hidden_size, output_size):
super(RLActor, self).__init__()
self.name = 'RL actor model'
self.hidden_size = hidden_size
self.gru = nn.GRU(input_size, self.hidden_size) # Add state space
self.out = nn.Linear(self.hidden_size, output_size)
self.softmax = nn.Softmax(dim=1)
self.gamma = 0.9999
self.saved_action_values = []
self.rewards = []
def forward(self, input_state, hidden, temperature=1):
input_state = torch.tensor(input_state, dtype=torch.float32, device=DEVICE).view(1,1,-1)
output, hidden = self.gru(input_state, hidden)
output = self.softmax(self.out(output[0]) / temperature)
return output, hidden
def initHidden(self):
return torch.zeros(1, 1, self.hidden_size, device=DEVICE)
class RLCritic(nn.Module):
"""Defines the RL Critic model"""
def __init__(self, input_size, hidden_size):
super(RLCritic, self).__init__()
self.name = 'RL critic model'
self.hidden_size = hidden_size
self.h1 = nn.Linear(input_size, self.hidden_size)
self.h2 = nn.Linear(self.hidden_size, 1)
self.saved_state_values = []
def forward(self, state_input, hidden):
state_input = torch.tensor(state_input, dtype=torch.float32, device=DEVICE).view(1,1,-1)
x = torch.cat((state_input[0], hidden[0]), 1)
x = self.h1(x)
x = F.relu(x)
value_score = self.h2(x)
return value_score
#optionally use dropout in the network
#self.dropout = nn.Dropout(p=0.6)
#x = self.dropout(x)
def select_action(input_state, input_hidden_state, actor_model, critic_model=None,
teacher_model=None, K=1, use_MLE=False, MCTS_thresh=0):
"""Applies the model on a given input and hidden state to make a prediction of which action to take
Can use MLE, MCTS, or sampling to select an action"""
probs, hidden_state = actor_model(input_state, input_hidden_state)
m = Categorical(probs)
# Use MLE instead of sampling distribution
if use_MLE:
_, topi = probs.data.topk(1)
action = topi.squeeze()
elif torch.max(probs).detach() < MCTS_thresh:
action, hidden_state, _ = MCTS.UCT_search(
env, input_state, input_hidden_state, actor_model, critic_model,
5, env.action_space, 1000)
action = torch.tensor(action, device=config.DEVICE)
else:
action = m.sample()
actor_model.saved_action_values.append(m.log_prob(action))
if critic_model != None:
state_value = critic_model(input_state, input_hidden_state)
critic_model.saved_state_values.append(state_value)
if teacher_model != None:
# Add policy distillation error
actor_probs, _ = actor_model(input_state, input_hidden_state, K)
supervised_probs, _ = teacher_model(input_state, input_hidden_state, K)
KL_error = utils.KL_divergence(actor_probs, supervised_probs, K)
return action.item(), hidden_state, KL_error.item()
return action.item(), hidden_state, None
def REINFORCE_update(actor_model, actor_optimizer):
"""Update the model when using REINFORCE instead of Actor-Critic"""
R = 0
policy_loss = []
returns = []
# Discount the rewards back to present
for r in actor_model.rewards[::-1]:
R = r + actor_model.gamma * R
returns.insert(0, R)
# Scale the rewards
returns = torch.tensor(returns)
###returns = (returns - returns.mean()) / (returns.std() + EPS)
# Calculate the loss
for log_prob, R in zip(actor_model.saved_action_values, returns):
policy_loss.append(-log_prob * R)
# Update network weights
actor_optimizer.zero_grad()
policy_loss = torch.cat(policy_loss).sum()
policy_loss.backward()
actor_optimizer.step()
# Clear memory
del actor_model.rewards[:]
del actor_model.saved_action_values[:]
def actor_critic_update(actor_model, actor_optimizer, critic_model, critic_optimizer,
only_update_critic=False):
"""Update the model when using Actor-Critic"""
R = 0
saved_actions = actor_model.saved_action_values
saved_states = critic_model.saved_state_values
policy_losses = [] # list to save actor (policy) loss
value_losses = [] # list to save critic (value) loss
returns = [] # list to save the true values
# calculate the true value using rewards returned from the environment
for r in actor_model.rewards[::-1]:
# calculate the discounted value
R = r + actor_model.gamma * R
returns.insert(0, R)
# Scale the rewards
returns = torch.tensor(returns)
###returns = (returns - returns.mean()) / (returns.std() + EPS) # scaling reduced performance
for log_prob, value, R in zip(saved_actions, saved_states, returns):
advantage = R - value.item()
# calculate actor (policy) loss
policy_losses.append(-log_prob * advantage)
# calculate critic (value) loss using L1 smooth loss
value_losses.append(F.smooth_l1_loss(value, torch.tensor([R], device=DEVICE)))
# reset gradients
actor_optimizer.zero_grad()
critic_optimizer.zero_grad()
if only_update_critic:
loss = torch.stack(value_losses).sum()
# perform backprop
loss.backward()
critic_optimizer.step()
else:
# sum up all the values of policy_losses and value_losses
loss = torch.stack(policy_losses).sum() + torch.stack(value_losses).sum()
# perform backprop
loss.backward()
actor_optimizer.step()
critic_optimizer.step()
# reset rewards and action buffer
del actor_model.rewards[:]
del actor_model.saved_action_values[:]
del critic_model.saved_state_values[:]
# %% Define environment, models and transfer weights
def load_FrozenLake_models():
"""Instantiate supervised model and load saved weights (CNN model)"""
supervised_encoder = CNNStateEncoder(128).to(DEVICE)
supervised_model = GeneralDecoderRNN(input_size=2, hidden_size=128, output_size=4).to(DEVICE)
data.load_model(supervised_encoder, os.path.join(
config.saved_RL_model_path, args.FROZENLAKE_ENCODER))
data.load_model(supervised_model, os.path.join(
config.saved_RL_model_path, args.FROZENLAKE_DECODER))
return supervised_encoder, supervised_model
def load_CartPole_models():
"""Instantiate supervised model and load saved weights"""
supervised_encoder = None
supervised_model = GeneralDecoderRNN(input_size=4, hidden_size=128, output_size=2).to(DEVICE)
data.load_model(supervised_model, os.path.join(config.saved_RL_model_path, args.CARTPOLE_DECODER))
return supervised_encoder, supervised_model
def create_environment(env_name):
"""Instantiates the environment with the selected hyperparameters"""
if env_name == 'FrozenLake':
input_map = frozen_lake_env.generate_random_map(5)
env = frozen_lake_env.FrozenLakeEnv(input_map, map_frozen_prob=0.75,
sparse=args.sparse_env, changing_map=True)
elif env_name == 'CartPole':
env = cart_pole_env.CartPoleEnv(args.sparse_env)
else:
print("Please select one of the following environments: ['FrozenLake', 'CartPole']")
env.seed(config.SEED)
return env
def init_actor_critic_models(state_space, action_space, init_critic=True,
transfer_weights=True, supervised_model=None):
"""Instantiates the actor and critic models as well as the optimizers"""
# Define actor and critic
actor_model = RLActor(input_size=state_space, hidden_size=128, output_size=action_space).to(DEVICE)
# Transfer weights to actor and set optimizer
if transfer_weights:
actor_model.load_state_dict(supervised_model.state_dict())
actor_optimizer = optim.Adam(actor_model.parameters())
if init_critic:
critic_model = RLCritic(input_size=(state_space + actor_model.hidden_size),
hidden_size=128).to(DEVICE)
critic_optimizer = optim.Adam(critic_model.parameters())
else:
critic_model = None
critic_optimizer = None
return actor_model, critic_model, actor_optimizer, critic_optimizer
#%% Load models and environment
def init_RL_environment(env_name, init_critic=True, transfer_weights=True):
"""Instantiates the RL environment and models through using the subfunctions"""
assert env_name in ['FrozenLake', 'CartPole'], \
"Please select one of the following environments: ['FrozenLake', 'CartPole']"
# Load supervised models
if env_name == 'FrozenLake':
supervised_encoder, supervised_model = load_FrozenLake_models()
elif env_name == 'CartPole':
supervised_encoder, supervised_model = load_CartPole_models()
# Create environment
env = create_environment(env_name)
# Create environment
actor_model, critic_model, actor_optimizer, critic_optimizer = init_actor_critic_models(
state_space=env.state_space, action_space=env.action_space,
init_critic=init_critic, transfer_weights=transfer_weights, supervised_model=supervised_model)
return supervised_encoder, supervised_model, env, \
actor_model, critic_model, actor_optimizer, critic_optimizer
class HyperParams(object):
"""Sets the experiment hyperparameters"""
def __init__(self, env_name, n_episodes):
assert env_name in ['FrozenLake', 'CartPole'], \
"Please select one of the following environments: ['FrozenLake', 'CartPole']"
if env_name == 'CartPole':
self.print_every = 5
self.t_weight = utils.EpsilonDecay(1, 0, n_episodes, 'Linear') # only Linear works
self.K = 5
self.early_stopping_reward_thresh = 295
self.early_stopping_n_mean = 3
self.beta = 0.25
self.sm_baseline_reward = 250
self.distillation_n_mean = 20
elif env_name == 'FrozenLake':
self.print_every = 20
self.t_weight = utils.EpsilonDecay(1, 0, n_episodes, 'Linear') # only Linear works
self.K = 5
self.early_stopping_reward_thresh = 11
self.early_stopping_n_mean = 50
self.beta = 1
self.sm_baseline_reward = 3.5
self.distillation_n_mean = 500
#%% Training loop
def train_RL_models(actor_model, critic_model, actor_optimizer, critic_optimizer,
supervised_encoder, supervised_model, teacher_model,
use_policy_distillation, update_RL_models, only_update_critic, use_MLE, MCTS_thresh, n_episodes):
"""Main training loop to train the actor and critic models"""
for i_episode in range(1, n_episodes+1):
# Reset environment at beginning of episode
state, ep_reward, done = env.reset(), 0, False
ep_env_reward = 0
ep_KL_penalty = 0
start_state = deepcopy(state)
if env.name == 'FrozenLake':
masked_input_map = frozen_lake_env.mask_map(env.input_map, flatten=False)
hidden_state = supervised_encoder(masked_input_map).detach()
else:
hidden_state = actor_model.initHidden()
# Play environment until maximum number of steps or environment terminates
for step_i in range(1, env.max_steps+1):
# Select action
action, hidden_state, KL_error = select_action(input_state=state, input_hidden_state=hidden_state,
actor_model=actor_model, critic_model=critic_model,
teacher_model=teacher_model, K=hp.K, use_MLE=use_MLE,
MCTS_thresh=MCTS_thresh)
# Apply action to environment to transition to next step
state, env_reward, done, _ = env.step(action)
# Can optionally use relative rewards to supervised model as reward
if args.sparse_env and args.relative_rewards and done:
env_reward = env_reward - supervised_model_reward(
env, start_state, supervised_encoder, supervised_model)
# Can optionally use policy distillation
if use_policy_distillation:
avg_RL_reward = np.mean(saved_RL_model_results.env_rewards[-hp.distillation_n_mean:]) \
if len(saved_RL_model_results.env_rewards) > hp.distillation_n_mean else 0
lambda_value= utils.lambda_value(beta=hp.beta, sm_baseline_reward=hp.sm_baseline_reward,
avg_rewards=avg_RL_reward)
KL_penalty = lambda_value * -KL_error
reward = env_reward + KL_penalty
ep_env_reward += env_reward
ep_KL_penalty += KL_penalty
else:
reward = env_reward
ep_reward += reward
actor_model.rewards.append(reward)
if done:
break
# Update models based on reward performance
if update_RL_models:
if critic_model != None:
actor_critic_update(actor_model, actor_optimizer, critic_model, critic_optimizer,
only_update_critic=only_update_critic)
else:
REINFORCE_update(actor_model, actor_optimizer)
if use_policy_distillation:
saved_RL_model_results.env_rewards.append(ep_env_reward)
saved_RL_model_results.KL_penalty.append(ep_KL_penalty)
if args.verbose_training and (i_episode % hp.print_every == 0):
avg_env_reward = np.mean(saved_RL_model_results.env_rewards[-hp.print_every:])
avg_KL_penalty = np.mean(saved_RL_model_results.KL_penalty[-hp.print_every:])
print('Episode {} | Avg env reward: {:.2f} | Avg KL penalty: {:.2f} | Lambda value: {:.2f}'.format(
i_episode, avg_env_reward, avg_KL_penalty, lambda_value))
early_stopping_value = np.mean(saved_RL_model_results.env_rewards[-hp.early_stopping_n_mean:]) \
if len(saved_RL_model_results.env_rewards) > hp.early_stopping_n_mean else 0
if (early_stopping_value >= hp.early_stopping_reward_thresh) or (i_episode == n_episodes-1):
if args.save_models:
saved_RL_model_results.save_top_models(actor_model, 'actor_{:.1f}.pt'.format(early_stopping_value))
if args.init_critic:
saved_RL_model_results.save_top_models(critic_model, 'critic_{:.1f}.pt'.format(early_stopping_value))
saved_RL_model_results.export_rewards('model_performance.txt')
break
else:
saved_RL_model_results.env_rewards.append(ep_reward)
if args.verbose_training and (i_episode % hp.print_every == 0):
avg_env_reward = np.mean(saved_RL_model_results.env_rewards[-hp.print_every:])
print('Episode {} | Average reward: {:.2f}'.format(i_episode, avg_env_reward))
early_stopping_value = np.mean(saved_RL_model_results.env_rewards[-hp.early_stopping_n_mean:]) \
if len(saved_RL_model_results.env_rewards) > hp.early_stopping_n_mean else 0
if (early_stopping_value >= hp.early_stopping_reward_thresh) or (i_episode == n_episodes-1):
if args.save_models:
saved_RL_model_results.save_top_models(actor_model, 'actor_{:.1f}.pt'.format(early_stopping_value))
if args.init_critic:
saved_RL_model_results.save_top_models(critic_model, 'critic_{:.1f}.pt'.format(early_stopping_value))
saved_RL_model_results.export_rewards('model_performance.txt')
break
def load_RL_models(folder_name, actor_file_name='best', critic_file_name='best'):
"""Instantiate RL models and load trained weights"""
actor_model = RLActor(input_size=env.state_space, hidden_size=128, output_size=env.action_space).to(DEVICE)
if actor_file_name == 'best':
actor_file_name = 'actor_{:.1f}.pt'.format(
data.get_top_n_models(
os.path.join(config.saved_RL_model_path, args.env_name, folder_name), 'actor', n=1)[0])
data.load_model(actor_model, os.path.join(config.saved_RL_model_path, args.env_name,
folder_name, actor_file_name))
if args.init_critic:
critic_model = RLCritic(input_size=(env.state_space + actor_model.hidden_size),
hidden_size=128).to(DEVICE)
if critic_file_name == 'best':
critic_file_name = 'critic_{:.1f}.pt'.format(
data.get_top_n_models(
os.path.join(config.saved_RL_model_path, args.env_name, folder_name), 'critic', n=1)[0])
data.load_model(critic_model, os.path.join(config.saved_RL_model_path, args.env_name,
folder_name, critic_file_name))
return actor_model, critic_model
else:
return actor_model, None
def load_pretrained_critic(env_name):
"""Intantial RL critic and load trained weights"""
if env_name =='CartPole':
data.load_model(critic_model, os.path.join(config.saved_RL_model_path, env_name,
args.CP_PRETRAINED_CRITIC))
elif env_name =='FrozenLake':
data.load_model(critic_model, os.path.join(config.saved_RL_model_path, env_name,
args.FL_PRETRAINED_CRITIC))
else:
print("Please select one of the following environments: ['FrozenLake', 'CartPole']")
#%% Train and Evaluate Model
if args.train_models:
"""Instantiates models and environment, trains and evaluates the model"""
# Instantiate models and environment
supervised_encoder, supervised_model, env, \
actor_model, critic_model, actor_optimizer, critic_optimizer = init_RL_environment(
env_name=args.env_name, init_critic=args.init_critic, transfer_weights=args.transfer_weights)
# Create folder if saving models
if args.save_models:
saved_RL_model_results.init_folder(args, actor_model, critic_model)
# Optionally load trained models
if args.load_models:
actor_model, critic_model = load_RL_models(args.load_model_folder_name)
# Instantiate teacher model if using policy distillation
if args.use_policy_distillation:
teacher_model = TeacherRNN(input_size=env.state_space, hidden_size=128,
output_size=env.action_space).to(DEVICE)
teacher_model.load_state_dict(supervised_model.state_dict())
else:
teacher_model = None
# Load pretrained critic
if args.use_pretrained_critic and critic_model is not None:
load_pretrained_critic(args.env_name)
# Optionally pretrain critic
if args.pretrain_critic_n_episodes > 0:
hp = HyperParams(args.env_name, args.pretrain_critic_n_episodes)
train_RL_models(actor_model, critic_model, actor_optimizer, critic_optimizer,
supervised_encoder, supervised_model, teacher_model,
use_policy_distillation=False, update_RL_models=True,
only_update_critic=True, use_MLE=False, MCTS_thresh=0,
n_episodes=args.pretrain_critic_n_episodes)
# Instantiate hyperparameters
hp = HyperParams(args.env_name, args.n_episodes)
# Train models
train_RL_models(actor_model, critic_model, actor_optimizer, critic_optimizer,
supervised_encoder, supervised_model, teacher_model,
use_policy_distillation=args.use_policy_distillation,
update_RL_models=args.update_RL_models,
only_update_critic=False, use_MLE=args.use_MLE,
MCTS_thresh=args.MCTS_thresh, n_episodes=args.n_episodes)