-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
1464 lines (1067 loc) · 56 KB
/
main.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 torch
from torchvision import datasets, transforms
import torch.utils.data.sampler as sampler
import torch.utils.data as data
import numpy as np
import argparse
import random
import os
from custom_datasets import *
import model
import vgg
import torchvision.models as models
from solver import Solver
from utils import *
import arguments
import torch.optim as optim
from rl.PolicyNetwork import PolicyNet
from tqdm import tqdm
from identity_sampler import IdentitySampler
def cifar_transformer():
return transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
'''
Simulates a step of the environment. Returns the new state, and the next dynamic action space.
Returns a next state: 1x10 vector
'''
def environment_step(train_dataloader, solver, task_model, num_repeats=1):
accs = torch.zeros((num_repeats,1))
class_accs_across_runs = torch.zeros((num_repeats,5))
counts_across_runs = torch.zeros((num_repeats,5))
for i in range(num_repeats):
task_model = model.FCNet(num_classes=args.num_classes) # remake a new task model each time
acc, vae, discriminator, class_accs = solver.train_without_adv_vae(train_dataloader,
task_model,
None,
None,
None, args)
counts, total = dataloader_statistics(train_dataloader, 5)
counts_across_runs[i] = counts.t()
accs[i] = acc
class_accs_across_runs[i] = class_accs
mean_class_accs = torch.mean(class_accs_across_runs, axis=0).unsqueeze(1)
mean_class_counts = torch.mean(counts_across_runs, axis=0).unsqueeze(1)
next_state_vector = torch.cat((mean_class_accs , mean_class_counts), axis=0)
return torch.mean(accs), next_state_vector.t()
'''simply gets the reward as the acc'''
def compute_reward_clean(curr_state):
curr_state = curr_state[:,0:args.num_classes].detach()
return torch.mean(curr_state)
'''
returns the delta, as well as the actual performance
'''
def compute_reward(curr_state, time_step, prev_reward,args):
curr_state = curr_state[:,0:args.num_classes].detach() #the rward should give 5 signals!
curr_reward = torch.mean(curr_state) - torch.mean(prev_reward)
prev_reward[time_step%len(prev_reward)] = torch.mean(curr_state)
return curr_reward, prev_reward
class_acc = curr_state[:,0:5].detach()
baseline = 20 + 0.8*time_step
# baseline = 1 + 2*time_step
# choices: we can try to achieve parity. Or we can try and just maximize the total acc across everything
# return torch.sum(curr_state-baseline) #equiv to acc.
# try some torch sum stuff. sum of squared differences for instance
perf = (class_acc-baseline)
print("perf is")
print(torch.mean(perf))
# return -torch.sum((class_acc - baseline)**2) # we want to achieve 20% acc in all of them...
return torch.mean(perf)
# construct a sign vector
return torch.sum(torch.clamp((class_acc - baseline),min=0, max=100) ** 2)
print("current performance is ")
print(perf)
perf_matrix = perf - perf.t() #broadcasting
return torch.mean(perf_matrix.pow(2))/2 # compute the sum across, and divide by 2
# sum along the diagonal etc. perf_matrix
torch.sum()
# trying sum of squared errors
return torch.sum((class_acc - baseline)**2) # we want to achieve 20% acc in all of them...
'''
Penalty for mode collapse.
Inputs: p and q are the distributions.
'''
def mode_collapse_penalty(p_dist,q_dist):
return torch.sum((p_dist-q_dist).pow(2))
'''
KL terms
'''
def mode_collapse_penalty_kl(p_dist,q_dist):
import torch.nn.functional as F
# p_dist += 0.05
# # smooth it
# p_dist /=torch.sum(p_dist)
p_EPSILON = 1e-10
print(p_dist, q_dist)
if len((p_dist==0).nonzero()) > 0:
print("WE HAVE A ZERO")
print(p_dist)
p_dist = p_dist + p_EPSILON
p_dist = p_dist/torch.sum(p_dist)
print("renormalized")
print(p_dist)
print("kl term is")
kl = F.kl_div(p_dist.log(), q_dist, reduction="batchmean") #reverse KL
print(kl )
return kl
'''
Returns the actual query, given an action distribution.
# just use the task_model as a noisy predictor. Take 10 samples, and then take the one that is predicted as most likely to
# be reported as the correct sample
return a tupel: target class (action selected) and datapoint tuple (x[best_idx], _[best_idx], idx[best_idx])
'''
def get_query(action, unlabelled_dataloader, task_model, args):
targ_class = action.sample()
# now, we just need to sample the class
datapoint = None
# from the unlabelled indices, sample an appropriate point from the class
iters = 0
lowest_score = 100
lowest_datapoint = None
# print(len(unlabelled_dataset))
for batch in unlabelled_dataloader:
x, _, idx = batch
preds = task_model(x.to(args.device))
# best_idx = torch.argmin(torch.abs(preds[:,targ_class] -0.5))
best_idx = torch.argmin((preds[:,targ_class]))
return targ_class, (x[best_idx], _[best_idx], idx[best_idx])
# break
# return targ_class, lowest_datapoint
# pass
'''
Gets query using k-means clustering. We assume a k-means clustering is passed in
Meta note: this is often done: we have a function that needs to "continue" its work: this is continuation, generator, yield etc.
args:
- kmeans: a kmeans objects
- action: the action label
- unlabelled_data: a NUMPY array of the full dataset (x,y, class and label). We will then hstack or something. And also make sure to delete the row when we sample
- should be of size 3 X ?
- in reality, should be of size (N X [sum (feature_space) +1]). i.e. we will squahs all the features across.
- unlabelled_mask: a boolean mask that specifies whether a point has already been labelled or not
- ideally, i would like to do kmeans clustering on only PART of the array. Since, i can get the x,y coord of the
- data points, but then i want to convert BACK to
- what the actual indices are!.. yes we should be able to do kmeans on a numpy array. And then have the cluster appear as another feature
of data points that
'''
def get_query_via_kmeans(action, unlabelled_data, args):
# data is [features, label, idx, pseudo_label]
print("these are some action zero statistics")
print(action.probs)
print((action.probs<0).nonzero())
rand = False
if torch.rand(size=()) < args.epsilon:
rand = True
rand_idx = torch.randint(len(unlabelled_data), size=())
datapoint = (unlabelled_data[rand_idx, 0:2], unlabelled_data[rand_idx, 2], unlabelled_data[rand_idx, 3])
unlabelled_data = np.delete(unlabelled_data, rand_idx, 0) # test to make sure this works
targ_cluster = torch.tensor(unlabelled_data[rand_idx][-1],dtype=torch.long).view(-1)
return targ_cluster, datapoint, unlabelled_data,rand
targ_cluster = action.sample()
iters = 0
datapoint = None
while iters < 100:
rand_idx = torch.randint(len(unlabelled_data), size=())
# we assume the kmeans is appended right at the very end
if unlabelled_data[rand_idx][-1] == targ_cluster:
datapoint = (unlabelled_data[rand_idx,0:2], unlabelled_data[rand_idx,2], unlabelled_data[rand_idx,3])
unlabelled_data = np.delete(unlabelled_data, rand_idx, 0 ) # test to make sure this works
break
iters+=1
return targ_cluster, datapoint, unlabelled_data, rand
# now, we just keep track of the indices
# now, we just need to sample the class
datapoint = None
# from the unlabelled indices, sample an appropriate point from the class
iters = 0
lowest_score = 100
lowest_datapoint = None
# print(len(unlabelled_dataset))
def random_baseline(args, num_iters=100):
# runs a random baseline
if args.dataset == "ring":
print("Using Ring dataset...")
test_dataloader = data.DataLoader(
Ring(args.data_path, transform=simple_data_transformer(), return_idx=False, testset=True),
batch_size=args.batch_size, drop_last=False
)
train_dataset = Ring(args.data_path, simple_data_transformer())
print(len(train_dataset))
args.num_images = 2500
args.budget = 1 #how many we can label at each round
args.initial_budget = 1
args.num_classes = 5
solver = Solver(args, test_dataloader)
all_indices = set(np.arange(args.num_images))
initial_indices = random.sample(all_indices, args.initial_budget)
sampler = data.sampler.SubsetRandomSampler(initial_indices)
current_indices = list(initial_indices)
unlabeled_indices = np.setdiff1d(list(all_indices), current_indices)
# dataset with labels available
train_dataloader = data.DataLoader(train_dataset, sampler=sampler,
batch_size=args.batch_size, drop_last=False)
accuracies = []
for i in range(num_iters):
task_model = model.FCNet(num_classes=args.num_classes) # remake a new task model each time
unlabeled_indices = np.setdiff1d(list(all_indices), current_indices)
unlabeled_sampler = data.sampler.SubsetRandomSampler(unlabeled_indices)
unlabeled_dataloader = data.DataLoader(train_dataset,
sampler=unlabeled_sampler, batch_size=args.batch_size, drop_last=False)
if args.sampling_method == "adversary" or args.sampling_method == "adversary_1c":
# train the models on the current data
# we also want to check which sampled_indice is best, and which one should be ideal, according to the dataset!
acc, vae, discriminator = solver.train(train_dataloader,
task_model,
vae,
discriminator,
unlabeled_dataloader)
else:
# train the models on the current data
acc,_,_,class_acc = solver.train_without_adv_vae(train_dataloader,
task_model,
None,
None,
None, args)
print('Final accuracy with {}% of data is: {:.2f}'.format(int(i), acc))
print(class_acc)
accuracies.append(acc)
sampled_indices = solver.sample_for_labeling(None, None, unlabeled_dataloader, task_model)
inquiry_sampler = data.sampler.SubsetRandomSampler(sampled_indices)
inquiry_dataloader = data.DataLoader(train_dataset, sampler=inquiry_sampler ,
batch_size=args.batch_size, drop_last=False)
new_datapoints_batch = np.zeros((0,3))
for datapoint_batch, label_batch, _ in inquiry_dataloader:
train_ex_batch = np.concatenate((datapoint_batch, np.expand_dims(label_batch, axis=1)), axis=1)
new_datapoints_batch = np.concatenate((new_datapoints_batch, train_ex_batch), axis=0) # concat the
print(sampled_indices)
current_indices = list(current_indices) + list(sampled_indices) #really they just want a set here...
sampler = data.sampler.SubsetRandomSampler(current_indices)
train_dataloader = data.DataLoader(train_dataset, sampler=sampler,
batch_size=args.batch_size, drop_last=False)
with open(os.path.join(args.out_path, "{}_current_accs.txt".format(args.sampling_method)), "a") as acc_file:
acc_file.write("{} {}\n".format(acc, class_acc))
visual_labelled_dataset = np.zeros((0, 3)) # each dimension does not require something new!
for datapoint_batch, label_batch, _ in train_dataloader: # will be tuple of n by 1
train_ex_batch = np.concatenate((datapoint_batch, np.expand_dims(label_batch, axis=1)), axis=1)
visual_labelled_dataset = np.concatenate((visual_labelled_dataset, train_ex_batch), axis=0) # concat the
# visualize_training_dataset(i, args.num_classes, visual_labelled_dataset, new_datapoints_batch, args.sampling_method)
return accuracies
'''
Computes the per-class statistics for the datapoints in the dataloader
'''
def dataloader_statistics(train_dataloader, num_classes):
per_class = torch.zeros((num_classes, 1))
for datapoint, label, idx in train_dataloader:
for dp, lb, _ in zip(datapoint, label, idx):
per_class[lb] += 1
return per_class, torch.sum(per_class)
'''
new datapoints is a numpy array, n by D+1 (D for the features, 1 for the class)
'''
def visualize_training_dataset(iteration, num_classes, prev_dataset, new_datapoints, name="rl"):
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
for cluster in range(num_classes):
# k_means_data = unlabelled_dataset[unlabelled_dataset[...,-1]==cluster]
# fig, ax = plt.subplots()
k_means_data = prev_dataset[prev_dataset[:,-1]==cluster]
ax.scatter(k_means_data [:,0], k_means_data [:,1])
# break
ax.scatter(new_datapoints[:,0],new_datapoints[:,1], s=100)
fig.savefig(os.path.join(args.out_path, "{}_viz_{}".format(name,iteration)))
fig.show()
plt.close(fig)
'''
function which computes the REINFORCE paramter updates
'''
def learn(y,y_pred, reward, criterion, optim):
'''assert the loss is indeed the CE loss'''
loss = criterion( y_pred, y)
loss = reward * loss
loss *= -1
optim.zero_grad()
loss.sum().backward()
optim.step()
def discount_rewards(r):
GAMMA = 0.9
discounted_r = torch.zeros(r.size())
running_add = 0
for t in reversed(range(len(r))):
print("Step: {}".format(t))
running_add = running_add * GAMMA + r[t] # we slowly accumulate the reward into all of the steps along the way
discounted_r[t] = running_add
# therefore, this has the effect of adding the rewards back into the original actions..
# print(r)
# print(discounted_r)
return discounted_r
'''
Processes the reward, like reward normalization and importantly,
GAMMA DISCOUNTING.
'''
#TODO: use logging instead of copious prints
def process_reward(reward_history):
print("processed reward; reward is now")
adv = discount_rewards(reward_history)
processed_reward_history = adv
processed_reward_history = (adv - adv.mean()) / (adv.std() + 1e-7)
print(reward_history)
print(processed_reward_history)
return processed_reward_history
def rl_main(args):
# a full game, where you pick 10 examples
# args.mine_episodes = 10
args.episode_length = 10
args.num_episodes = 10
args.epsilon = 0.2 # try with full policy. and try with using the full vector to compute a reward. But it really is just a multiple. Unless we specifically penalize assigning 0 counts
# probably starting with 10 or so points randomly would be very good. but would invalidate past work
with open(os.path.join(args.out_path, "args.txt"), "w") as file:
for key,val in vars(args).items():
file.write("{}:{}\n".format(key,val))
if args.dataset == "ring":
print("Using Ring dataset...")
test_dataloader = data.DataLoader(
Ring(args.data_path, transform=simple_data_transformer(), return_idx=False, testset=True),
batch_size=args.batch_size, drop_last=False
)
train_dataset = Ring(args.data_path, simple_data_transformer())
print(len(train_dataset))
args.num_images = 2500
args.budget = 1 #how many we can label at each round
args.initial_budget = 1
args.num_classes = 5
elif args.dataset == "mnist":
print("Using MNIST dataset...")
test_dataloader = data.DataLoader(
MNIST(args.data_path),
batch_size=args.batch_size, drop_last=False
)
train_dataset = MNIST(args.data_path)
print(len(train_dataset))
args.num_images = 2500
args.budget = 1 # how many we can label at each round
args.initial_budget = 1
args.num_classes = 10
random.seed(args.torch_manual_seed)
torch.manual_seed(args.torch_manual_seed)
args.cuda = args.cuda and torch.cuda.is_available()
solver = Solver(args, test_dataloader)
all_indices = set(np.arange(args.num_images))
initial_indices = random.sample(all_indices, args.initial_budget)
sampler = data.sampler.SubsetRandomSampler(initial_indices)
current_indices = list(initial_indices)
unlabeled_indices = np.setdiff1d(list(all_indices), current_indices)
unlabeled_sampler = data.sampler.SubsetRandomSampler(unlabeled_indices)
unlabeled_dataloader = data.DataLoader(train_dataset,
sampler=unlabeled_sampler, batch_size=args.batch_size, drop_last=False)
# dataset with labels available
train_dataloader = data.DataLoader(train_dataset, sampler=sampler,
batch_size=args.batch_size, drop_last=False)
# iterate the train_dataloader, and compute some statistics. And also, increment quantities.
'''
FORMULATION1: We will feed in the class_specific accuracies.
'''
ROLLING_AVG_LEN = 5
prev_reward = torch.ones((ROLLING_AVG_LEN,1))
prev_reward *=20
print("prev_reward{}".format(prev_reward))
STATE_SPACE = args.num_classes
ACTION_SPACE = args.num_classes
CLASS_DIST_SPACE = args.num_classes
pol_class_net = PolicyNet(STATE_SPACE + CLASS_DIST_SPACE, ACTION_SPACE ) # gradient, or hessian in the network..; per class accs as well
args.old_model_path = "/h/johnchen/Desktop/git_stuff/vaal_querying/testing_zero/bs_16/model.pt"
if os.path.exists(args.old_model_path) and args.use_old:
pol_class_net.load_state_dict(torch.load(args.old_model_path))
pol_optimizer = optim.Adam(pol_class_net.parameters(), lr=5e-2)
curr_state = torch.zeros((1,STATE_SPACE + CLASS_DIST_SPACE)) #only feed it in the past state directly
import copy
task_model = model.FCNet(num_classes=args.num_classes)
# inference_model = task_model
# inference_model.to(args.device)
# task_model = vgg.vgg16_bn(num_classes=args.num_classes)
accuracies = []
criterion = torch.nn.CrossEntropyLoss(reduction="none")
# feel like supporting a desparate cause; might delete later
entire_loader = DataLoader(train_dataset, batch_size=len(train_dataset), shuffle=True)
# ask on SO: multi item getting using pytorch, dataloader
features, labels, idx = next(iter(entire_loader))
features = features.numpy()[unlabeled_indices] # we should exactly not be using it as this. Actually, it is OK. we are just saing what is labelled and what is not
labels = np.expand_dims(labels.numpy()[unlabeled_indices], 1)
idx = np.expand_dims(idx.numpy()[unlabeled_indices], 1)
# X = np.hstack((features,labels ,idx )) #strange that this doesn't work
# X = np.concatenate((features.reshape(len(features),-1), labels,idx), axis=1)
X = np.concatenate((features, labels,idx), axis=1)
from sklearn.cluster import KMeans
kmeans_obj = KMeans(n_clusters=args.num_classes, random_state=0) # we can also fit one kmeans at the very start.
cluster_preds = kmeans_obj.fit_predict(X[:,0:2])
oracle_clusters = True
if oracle_clusters:
unlabelled_dataset = np.concatenate((X, labels), axis=1)
else:
# we can also just predict (should be fast) again on new datapoints, using the trained classifier. But why not just memorize
unlabelled_dataset = np.concatenate((X, np.expand_dims(cluster_preds,axis=1)), axis=1)
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
# try and predict directly in the data space?
# try and graph where in the dataset it does it as well.
# in this case, we would again need some fix of the policy gradient.
# we can no longer just do an easy cross entropy
# instead, we would be operating more in the regime of .
# this is a nice analysis on problems of this type!
# just about rotating and sculpting to your particular area you want
# ultra few shot learning with fixed dimension, horizon!
# contributions: to the field of meta/few shot learning using an active learning with reinforcement learning approach
# keep on layering intersections, until you get the particular area you want.
# even the approach of doing few shot learning, using active learning is pretty novel IMO
# we would be trying to essentially do q learning on the choice of datapoint. but make sure you pick in the data space (not action, but continuous choice of the datapoint)
# the key is really then, trying to do X
# we could literally do an entire course of lin alg during the break!
# really, digging into the problems of policy gradient
# now let's graph the unlabelled dataset
for cluster in range(args.num_classes):
# k_means_data = unlabelled_dataset[unlabelled_dataset[...,-1]==cluster]
# fig, ax = plt.subplots()
k_means_data = unlabelled_dataset[unlabelled_dataset[:,-1]==cluster]
ax.scatter(k_means_data [:,0], k_means_data [:,1])
ax.scatter(kmeans_obj.cluster_centers_[cluster][0], kmeans_obj.cluster_centers_[cluster][1], s=100)
fig.savefig(os.path.join(args.out_path, "cluster_{}".format(cluster)))
# break
fig.show()
# gradient_accum = torch.zeros((args.rl_batch_steps, 1), requires_grad=False) # accumulate all the losses
# try making it an empty thing
gradient_accum = torch.zeros((args.episode_length), requires_grad=False) # accumulate all the losses
# loss.backward(0 => doesn't actually execute an update of the weights. we could probably call loss.backward individually
batched_accs = []
# try combining it with the state. and also, just try doing an epsilon greedy policy
import torch.nn.functional as F
for i in tqdm(range(args.num_episodes)):
# need y, y_pred and the reward
action_history = torch.FloatTensor([])
reward_history = torch.FloatTensor([])
taken_action_history = torch.LongTensor([])
for j in range(args.episode_length):
pol_optimizer.zero_grad()
# here we need a fake label, in order to back prop the loss. And don't backprop immediately, instead, get the gradient,
# hold it, wait for the reward, and then backprop on that quantity
action_vector = pol_class_net (curr_state )
# torch.nn.functional.log_softmax(action_vector)
# action_dist = torch.zeros((1))
# the huge bug is as follows:
# make the network output a softmax
# then, the loss is the cross entropy:
print(F.softmax(action_vector))
if (len(F.softmax(action_vector)[F.softmax(action_vector)<0]) > 0 ):
print("huh 0 probs!!")
action_dist = torch.distributions.Categorical(probs=F.softmax(action_vector)) #the diff between Softmax and softmax
action_history = torch.cat([action_history, action_dist.probs])
# we probably need logsoftmax here too
print("action dist{}\n, dist probs{}\n, self f.softmax {}\n, self.log softmax{}\n".format(action_vector,action_dist.probs,
F.softmax(action_vector, dim=1),
F.log_softmax(action_vector,
dim=1))) #intelligent to take the softmax over the right dimension
# print() #recall logsoftmax and such
# if torch.rand() < args.epsilon:
# pass
# else:
# correct_label1, action1 = get_query(action_dist, unlabeled_dataloader, inference_model, args)
correct_label, action, unlabelled_dataset, rand = get_query_via_kmeans(action_dist, unlabelled_dataset, args)
taken_action_history = torch.cat([taken_action_history, torch.LongTensor([(correct_label)])])
if not rand : #still compute the losses to avoid policy collpase
# print(rand)
pred_vector = F.softmax(action_vector.view(1,-1))
correct_label = correct_label # just a k-size list
loss = criterion(pred_vector, correct_label)
print("loss stats")
print(pred_vector, correct_label)
# labelled updates
current_indices = list(current_indices) + [int(action[2].item())] # really they just want a set here...
sampler = data.sampler.SubsetRandomSampler(current_indices)
train_dataloader = data.DataLoader(train_dataset, sampler=sampler,
batch_size=args.batch_size, drop_last=False)
# unlabelled updates
unlabeled_indices = np.setdiff1d(list(all_indices), current_indices)
unlabeled_sampler = data.sampler.SubsetRandomSampler(unlabeled_indices)
unlabeled_dataloader = data.DataLoader(train_dataset,
sampler=unlabeled_sampler, batch_size=args.batch_size, drop_last=False)
class_counts, total = dataloader_statistics(train_dataloader, args.num_classes)
print("class counts {}".format(class_counts))
#data loader not subscriptable => we should deal with the indices.
# we could also combine, and get the uncertainties, but WEIGHTED BY CLASS
# lets just try the dataloader, but it will be challenging when we have the batch size...
# print(correct_label)
print("this is the action taken by the sampler")
print(action)
acc, curr_state = environment_step(train_dataloader, solver, task_model) #might need to write a bit coupled code. This is OK for now
reward = compute_reward_clean(curr_state) #move the reward to the env. step.
reward_history = torch.cat([reward_history, reward.unsqueeze(0)])
accuracies.append(acc)
# end of episode; compute the reward and go forth!
if True:
processed_reward_history = process_reward(reward_history)
learn(taken_action_history, action_history , processed_reward_history, criterion, pol_optimizer)
with open(os.path.join(args.out_path, "accs.txt"), "a") as acc_file:
acc_file.write("{};{}\n".format(acc, curr_state))
print(curr_state)
print(acc)
# with open(os.path.join(args.out_path, "rl_current_accs.txt"), "a") as acc_file:
# acc_file.write("{} {}\n".format(acc, class_accs))
# inference_model = task_model
# inference_model.to(args.device)
task_model = model.FCNet(num_classes=args.num_classes) # remake a new task model each time
# task_model = vgg.vgg16_bn(num_classes=args.num_classes)
# graph the train dataloader at each iteration
# for cluster in range(args.num_classes):
# # k_means_data = unlabelled_dataset[unlabelled_dataset[...,-1]==cluster]
# # fig, ax = plt.subplots()
#
# k_means_data = unlabelled_dataset[unlabelled_dataset[:, -1] == cluster]
#
# ax.scatter(k_means_data[:, 0], k_means_data[:, 1])
# ax.scatter(kmeans_obj.cluster_centers_[cluster][0], kmeans_obj.cluster_centers_[cluster][1], s=100)
# fig.savefig(os.path.join(args.out_path, "cluster_{}".format(cluster)))
visual_labelled_dataset = np.zeros((0,3)) #each dimension does not require something new!
new_datapoints= np.reshape(np.asarray(action[0]), newshape=(-1,2))
for datapoint_batch, label_batch, _ in train_dataloader: #will be tuple of n by 1
train_ex_batch = np.concatenate((datapoint_batch, np.expand_dims(label_batch,axis=1)), axis=1)
visual_labelled_dataset = np.concatenate((visual_labelled_dataset, train_ex_batch), axis=0 ) #concat the
# visualize_training_dataset(i, args.num_classes, visual_labelled_dataset, new_datapoints)
# stack all of them!
# and furthermore, we need to do a group by on the label.
# TODO: reset the batched points, and then readd everything back in
# i.e.: we need to reset the state, and then do another 10-point learning
# now, check the visual labelled dataset
# let's graph the vector, as we see it come
# graph the new point on the map, then graph the old collection of data as regular
# current_indices
# save the trained model
model_params = pol_class_net.state_dict()
torch.save(model_params, os.path.join(args.out_path, "model.pt"))
#
fig, ax = acc_plot(accuracies, args, label="policy gradient", name="policy gradient only")
spaced_x = list(range(len(batched_accs)))
spaced_x = [x*10 for x in spaced_x]
ax.plot(spaced_x, batched_accs, marker="x", c="purple", label="batched policy updates")
ax.legend()
fig.show()
fig.savefig(os.path.join(args.out_path, "comparison_batched_acc_plot_{}_queries".format(len(accuracies))))
print(pol_class_net)
import copy
uncertain_args = copy.deepcopy(args)
uncertain_args.sampling_method = "uncertainty"
uncertain_accs = random_baseline(uncertain_args, args.num_episodes)
random_args = copy.deepcopy(args)
random_args.sampling_method = "random"
random_accs = random_baseline(random_args, args.num_episodes)
fig, ax = acc_plot(accuracies, args, label="policy gradient")
ax.plot(range(0, len(random_accs)), random_accs, marker="x", c="orange", label="random")
ax.plot(range(0, len(uncertain_accs)), uncertain_accs, marker="^", c="green", label="uncertain")
ax.legend()
fig.show()
fig.savefig(os.path.join(args.out_path, "comparison_acc_plot_{}_queries".format(len(accuracies))))
# try comparing vs a regularly trained network (random sampling)
# now that the policy network is trained, we can use it to actually do the inference
# # ever batch size episodes, make the gradient update
# # compute a rollout, or at least one step
# # need an environment step
#
#
# print(len(train_dataloader))
#
#
# import math
# splits = range(int(math.ceil(100 / args.budget)))
#
# # # splits = [0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4]
# # splits = [args.initial_budget/float(args.num_images),
# # (args.initial_budget+args.budget)/float(args.num_images),
# # (args.initial_budget+args.budget*2)/float(args.num_images),
# # (args.initial_budget+args.budget*3)/float(args.num_images),
# # (args.initial_budget+args.budget*4)/float(args.num_images),
# # (args.initial_budget+args.budget*5)/float(args.num_images), ]
#
# current_indices = list(initial_indices)
# accuracies = []
#
# best_data_point = None
# total_optimal = 0
#
# for split in splits:
# task_model = model.FCNet(num_classes=args.num_classes) # remake a new task model each time
# # need to retrain all the models on the new images
# # re initialize and retrain the models
# # task_model = vgg.vgg16_bn(num_classes=args.num_classes)
# if args.dataset == "mnist":
# vae = model.VAE(args.latent_dim, nc=1)
# elif args.dataset == "ring":
# vae = model.VAE(args.latent_dim, nc=2)
# else:
# vae = model.VAE(args.latent_dim)
# discriminator = model.Discriminator(args.latent_dim)
#
# unlabeled_indices = np.setdiff1d(list(all_indices), current_indices)
#
# unlabeled_sampler = data.sampler.SubsetRandomSampler(unlabeled_indices)
# unlabeled_dataloader = data.DataLoader(train_dataset,
# sampler=unlabeled_sampler, batch_size=args.batch_size, drop_last=False)
# print(len(unlabeled_dataloader))
#
# if args.sampling_method == "adversary" or args.sampling_method == "adversary_1c":
# # train the models on the current data
# # we also want to check which sampled_indice is best, and which one should be ideal, according to the dataset!
# acc, vae, discriminator = solver.train(train_dataloader,
# task_model,
# vae,
# discriminator,
# unlabeled_dataloader)
# else:
# # train the models on the current data
# acc, vae, discriminator = solver.train_without_adv_vae(train_dataloader,
# task_model,
# vae,
# discriminator,
# unlabeled_dataloader)
#
# print('Final accuracy with {}% of data is: {:.2f}'.format(int(split * 100), acc))
# accuracies.append(acc)
#
# sampled_indices = solver.sample_for_labeling(vae, discriminator, unlabeled_dataloader, task_model)
#
# args.oracle_impute = False
# print("ORacle impute: {}".format(args.oracle_impute))
#
#
#
# if args.oracle_impute:
# print("main")
# print(sampled_indices)
# # compute a pass over all points
# uncertainties = []
# index_order = [None for _ in range(args.num_images)]
#
# # torch.ones((args.num_images))
#
# # [1 for i in range(args.num_images)] # we have so many points to label, in total. we should investigate that if we select a number than, the length is unchanged?
#
# # get the length. But also get the maximum element. Most likely, the length of the unlabelled dataloader does not change.
#
# with torch.no_grad():
# for i, pt in enumerate(unlabeled_dataloader):
# pred = task_model(pt[0].to(args.device))
# uncertainties.append(1 - pred.max().item())
# index_order[pt[2].item()] = i
#
# # uncertainties[pt[2].item()] = 1 - pred.max().item() # we need to compute the loss! (or we can just take the max uncertainty..)
#
# # uncertainties can range from 0 to 1. We want to take the maximum value
# # it might be the case that model is perfectly confident. this means that we will have 0. that is ok. it doesn't make sense if we have 1
#
# # HOW IS IT POSSIBLE WE HAVE 0 AS THE MAX
# # AND HOW IS IT POSSIBLE WE DONT EVEN HAVE THE SAME SAMPLED INDICE AS WHAT THE OTHER RETURNS
# # uncertainties = [elt for elt in uncertainties if elt is not None]
# # ensure that the uncertainties return here are indeed accurate
# # they wont actually line up unfortunately...
# print("uncertainty vs sampled index")
# print(uncertainties.index(max(uncertainties)), max(uncertainties))
# print(index_order[sampled_indices[0]], uncertainties[index_order[sampled_indices[
# 0]]]) # but it might be possible, that this quantity is not computed...no. it MUST be computed, since it is unlabelled
#
# # we only multiply by -1 at the end (to select the elemnts which are furtherst)
# # for i in range(len(uncertainties)):
# # uncertainties[i] = 1 - uncertainties[i]
#
# # print(sampled_indices)
# # unlabeled_dataloader.dataset[sampled_indices]
# best_data_point, max_acc, accs = oracle_best_point(unlabeled_dataloader, current_indices.copy(),
# train_dataset, solver, args, sampled_indices,
# index_order,
# uncertainties) # since we might have an elt with index being 2.5k, then it would mess it all up.
# # hence, a hash based approach IS best!
# #
# #
# print(sampled_indices, accs[index_order[sampled_indices.item()]])
# print(best_data_point, max_acc)
# #
# if best_data_point == sampled_indices[0]:
# total_optimal += 1
# print(max_acc) # this should be optimal
# # print()
# torch.save(accs, os.path.join(args.out_path, "accs_{}".format(split) + ".txt"))
# torch.save(uncertainties, os.path.join(args.out_path, "uncertainties_{}".format(split) + ".txt"))
# uncertainty_acc_plot(uncertainties, accs, args, split, sampled_indices, index_order)
#
# with open(os.path.join(args.out_path, "current_accs.txt"), "a") as acc_file:
# acc_file.write("{}\n".format(acc))
#
# #
# query_analysis(sampled_indices, unlabeled_dataloader, args, split)
#
# # current_indices = list(current_indices) + [best_data_point] #really they just want a set here...
# current_indices = list(current_indices) + list(sampled_indices) # really they just want a set here...
#
# sampler = data.sampler.SubsetRandomSampler(current_indices)
# train_dataloader = data.DataLoader(train_dataset, sampler=sampler,
# batch_size=args.batch_size, drop_last=False)
#
# # break
#
# acc_plot(accuracies, args)
# print("In total, we had {} out of 100 optimal".format(total_optimal))
#
# torch.save(accuracies, os.path.join(args.out_path, args.log_name + ".txt"))
def main(args):
if args.dataset == "ring":
print("Using Ring dataset...")
test_dataloader = data.DataLoader(
Ring(args.data_path, transform=simple_data_transformer(), return_idx=False, testset=True),
batch_size=args.batch_size, drop_last=False
)
train_dataset = Ring(args.data_path, simple_data_transformer())
print(len(train_dataset))
args.num_images = 2500
args.budget = 1 #how many we can label at each round
args.initial_budget = 1
args.num_classes = 5
elif args.dataset == 'mnist':
test_dataloader = data.DataLoader(
datasets.MNIST(args.path, download=True, transform=mnist_transformer(), train=False),
batch_size=args.batch_size, drop_last=False)
train_dataset = MNIST(args.data_path)
print(len(train_dataset))