-
Notifications
You must be signed in to change notification settings - Fork 0
/
Playground_tmp.py
1281 lines (910 loc) · 36.7 KB
/
Playground_tmp.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
# -*- coding: utf-8 -*-
"""
Created on Mar Jan 15 11:21:12 2019
@author: marcorax, pedro
This script serve as a test for different implementations, and parameter exploration
of HOTS under the name gordoNN to be tested with multiple speech recognition tests.
HOTS (The type of neural network implemented in this project) is a machine learning
method totally unsupervised.
To test if the algorithm is learning features from the dataset, a simple classification
task is accomplished with the use multiple classifiers, taking the activity
of the last layer.
If by looking at the activities of the last layer we can sort out the class of the
input data, it means that the network has managed to separate the classes.
"""
# General Porpouse Libraries
import numpy as np
import matplotlib.pyplot as plt
from matplotlib import offsetbox
import seaborn as sns
import os, gc, pickle, copy
from tensorflow.keras.callbacks import EarlyStopping
from Libs.Solid_HOTS._General_Func import create_mlp
from joblib import Parallel, delayed
from sklearn import svm
import pandas as pd
from tensorflow.keras.layers import Input, Dense, BatchNormalization,\
Dropout, Activation
from tensorflow.keras.models import Model
from tensorflow.keras import backend as K
from tensorflow.keras import optimizers, regularizers
# To use CPU for training
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # see issue #152
# os.environ["CUDA_VISIBLE_DEVICES"] = ""
# To allow more memory to be allocated on gpus incrementally
# os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"
# Data loading Libraries
from Libs.Data_loading.dataset_load import on_off_load, data_load
# 3D Dimensional HOTS or Solid HOTS
from Libs.Solid_HOTS.Network import Solid_HOTS_Net
# from Libs.Solid_HOTS._General_Func import first_layer_sampling_plot, other_layers_sampling_plot, first_layer_sampling, recording_local_surface_generator
# To avoid MKL inefficient multithreading (helped for some AMD processors before)
# now disabled
# os.environ['MKL_NUM_THREADS'] = '1'
from Libs.GORDONN.Layers.Local_Layer import Local_Layer
from Libs.GORDONN.Layers.Cross_Layer import Cross_Layer
from Libs.GORDONN.Layers.Pool_Layer import Pool_Layer
from Libs.GORDONN.Layers.MIG_Layer import MIG_Layer
from Libs.GORDONN.Network import GORDONN
# Plotting settings
sns.set(style="white")
plt.style.use("dark_background")
### Selecting the dataset
shuffle_seed = 25 # seed used for dataset shuffling if set to 0 the process will be totally random
# shuffle_seed = 0
#%% Google Commands Dataset
# 30 class of recordings are used. They consists of simple keywords
# =============================================================================
# number_files_dataset : the number of files to be loaded for each class
# train_test_ratio: ratio between the amount of files used to train
# and test the algorithm, 0.5 will mean that the
# half of the files wiil be used for training.
# use_all_addr : if False all off events will be dropped, and the total addresses
# number will correspond to the number of channels of the cochlea
# =============================================================================
number_files_dataset = 1000
# number_files_dataset = 10
train_test_ratio = 0.80
use_all_addr = False
number_of_labels = 8
dataset_folder ='Data/Google_Command'
spacing=1
# classes=['off','on']
classes=['stop', 'left', 'no', 'go', 'yes', 'down', 'right', 'up']
[dataset_train, dataset_test, labels_train, labels_test, filenames_train, filenames_test, classes] = data_load(dataset_folder, number_files_dataset, train_test_ratio, shuffle_seed, use_all_addr, spacing, class_names=classes)
#%% Network local layer
input_channels = 32 + 32*use_all_addr
channel_taus=1/np.array([2. , 2.09818492, 2.19739597, 2.30845776, 2.42444764,
2.54123647, 2.67160239, 2.80743423, 2.94395277, 3.09176148,
3.24720257, 3.41043852, 3.57786533, 3.75780207, 3.94362851,
4.14023903, 4.34601214, 4.56200522, 4.78905638, 5.02749048,
5.27770753, 5.53978449, 5.81554546, 6.10468077, 6.40818066,
6.72679013, 7.06129558, 7.41233318, 7.78086536, 8.16764435,
8.57370814, 9. ][::-1])
channel_taus = channel_taus/channel_taus[0]
tau_K = 10000
taus = tau_K*channel_taus
# Create the network
local = Local_Layer(n_features=20, local_tv_length=20,
n_input_channels=input_channels, taus=taus,
n_batch_files=None, dataset_runs=1, n_threads=22,
verbose=True)
# Learn the features
local.learn(dataset_train)
tmp = local.features
# Predict the features
local_response = local.predict(dataset_test)
hists, norm_hist = local.gen_histograms(local_response)
sign, norm_sign= local.gen_signatures(hists, norm_hist, classes, labels_test)
file_i = 0
local.response_plot(local_response, f_index=file_i, class_name=classes[labels_test[file_i]])
#%% Network local layer
input_channels = 32 + 32*use_all_addr
df = pd.read_csv (r'whitenoise.csv')
mean_rate = np.asarray(df.columns[:], dtype=float)
channel_taus = 1/mean_rate
channel_taus = channel_taus/channel_taus[0]
tau_K = 5000
taus = tau_K*channel_taus
# Create the network
local = Local_Layer(n_features=20, local_tv_length=10,
n_input_channels=input_channels, taus=taus,
n_batch_files=None, dataset_runs=1, n_threads=22,
verbose=True)
# Learn the features
local.learn(dataset_train)
tmp = local.features
# Predict the features
local_response = local.predict(dataset_test)
hists, norm_hist = local.gen_histograms(local_response)
sign, norm_sign= local.gen_signatures(hists, norm_hist, classes, labels_test)
file_i = 0
local.response_plot(local_response, f_index=file_i, class_name=classes[labels_test[file_i]])
local.features_plot()
#%% Network cross layer stack
# Create the network
cross = Cross_Layer(n_features=20, cross_tv_width=32,
n_input_channels=input_channels, taus=50e3,
n_input_features=20, n_batch_files=None,
dataset_runs=1, n_threads=22,
verbose=True)
# Learn the features
cross.learn(local_response)
tmp = cross.features
# Predict the features
cross_response = cross.predict(local_response)
hists, norm_hist = cross.gen_histograms(cross_response)
sign, norm_sign= cross.gen_signatures(hists, norm_hist, classes, labels_test)
file_i = 0
local.response_plot(local_response, f_index=file_i, class_name=classes[labels_test[file_i]])
#%% Network cross layer as first
input_channels = 32 + 32*use_all_addr
channel_taus=1/np.array([2. , 2.09818492, 2.19739597, 2.30845776, 2.42444764,
2.54123647, 2.67160239, 2.80743423, 2.94395277, 3.09176148,
3.24720257, 3.41043852, 3.57786533, 3.75780207, 3.94362851,
4.14023903, 4.34601214, 4.56200522, 4.78905638, 5.02749048,
5.27770753, 5.53978449, 5.81554546, 6.10468077, 6.40818066,
6.72679013, 7.06129558, 7.41233318, 7.78086536, 8.16764435,
8.57370814, 9. ][::-1])
channel_taus = channel_taus/channel_taus[0]
tau_K = 10000
taus = tau_K*channel_taus
# Create the network
cross = Cross_Layer(n_features=20, cross_tv_width=3,
n_input_channels=input_channels, taus=taus,
n_input_features=1, n_batch_files=None,
dataset_runs=1, n_threads=22,
verbose=True)
# Learn the features
cross.learn(dataset_train)
tmp = cross.features
cross.predict(dataset_train)
# Predict the features
cross_response = cross.predict(dataset_train)
cross.learn(cross_response)
#%% Pool layer test
pool = Pool_Layer(n_input_channels=32, pool_factor=4)
tmp=pool.pool(dataset_train)
tmp=pool.pool(local_response)
#%% Network test
#First layer parameters
input_channels = 32 + 32*use_all_addr
df = pd.read_csv (r'whitenoise.csv')
mean_rate = np.asarray(df.columns[:], dtype=float)
channel_taus = 1/mean_rate
channel_taus = channel_taus/channel_taus[0]
tau_K = 500
taus = tau_K*channel_taus
n_features=20
local_tv_length=10
n_input_channels=input_channels
n_batch_files=None
dataset_runs=1
local_layer_parameters = [n_features, local_tv_length, n_input_channels, taus,\
n_batch_files, dataset_runs]
#Second layer parameters
n_input_features=n_features
n_input_channels=input_channels
n_features=64
cross_tv_width=6
taus=50e3
n_batch_files=None
dataset_runs=1
cross_layer_parameters = [n_features, cross_tv_width,
n_input_channels, taus,
n_input_features, n_batch_files,
dataset_runs]
Net = GORDONN(n_threads=24, verbose=True)
Net.add_layer("Local", local_layer_parameters)
Net.add_layer("Cross", cross_layer_parameters)
Net.learn(dataset_train,labels_train,classes)
Net.predict(dataset_test, labels_train, labels_test, classes)
print("Histogram accuracy: "+str(Net.layers[0].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[0].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[0].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[0].svm_norm_hist_accuracy))
print("Histogram accuracy: "+str(Net.layers[1].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[1].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[1].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[1].svm_norm_hist_accuracy))
#%% Parameter change and overload
#Second layer parameters
new_taus = 5e3
Net.layers[1].taus=new_taus
Net.learn(dataset_train,labels_train,classes, rerun_layer=1)
Net.predict(dataset_test, labels_train, labels_test, classes, rerun_layer=1)
print("Histogram accuracy: "+str(Net.layers[0].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[0].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[0].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[0].svm_norm_hist_accuracy))
print("Histogram accuracy: "+str(Net.layers[1].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[1].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[1].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[1].svm_norm_hist_accuracy))
#%% Network test (only two layer like old times)
#First layer parameters
input_channels = 32 + 32*use_all_addr
df = pd.read_csv (r'whitenoise.csv')
mean_rate = np.asarray(df.columns[:], dtype=float)
channel_taus = 1/mean_rate
channel_taus = channel_taus/channel_taus[0]
tau_K = 500
taus = tau_K*channel_taus
n_features=20
local_tv_length=10
n_input_channels=input_channels
n_batch_files=None
dataset_runs=1
local_layer_parameters = [n_features, local_tv_length, n_input_channels, taus,\
n_batch_files, dataset_runs]
#Pool Layer
n_input_channels=32
pool_factor = 8
pool_layer_parameters = [n_input_channels, pool_factor]
#Second layer parameters
n_input_features=n_features
n_input_channels=4
n_features=64
cross_tv_width=None
taus=50e3
n_batch_files=None
dataset_runs=1
cross_layer_parameters = [n_features, cross_tv_width,
n_input_channels, taus,
n_input_features, n_batch_files,
dataset_runs]
Net = GORDONN(n_threads=24, verbose=True, low_memory_mode=False)
Net.add_layer("Local", local_layer_parameters)
Net.add_layer("Pool", pool_layer_parameters)
Net.add_layer("Cross", cross_layer_parameters)
Net.learn(dataset_train,labels_train,classes)
Net.predict(dataset_test, labels_train, labels_test, classes)
print("Histogram accuracy: "+str(Net.layers[0].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[0].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[0].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[0].svm_norm_hist_accuracy))
print("Histogram accuracy: "+str(Net.layers[2].hist_accuracy))
print("Norm Histogram accuracy: "+str(Net.layers[2].norm_hist_accuracy))
print("SVC Histogram accuracy: "+str(Net.layers[2].svm_hist_accuracy))
print("SVC norm Histogram accuracy: "+str(Net.layers[2].svm_norm_hist_accuracy))
#%% Last layer new classifier
#Actual decay
df = pd.read_csv (r'whitenoise.csv')
mean_rate =np.asarray(df.columns[:], dtype=float)
channel_taus = 1/mean_rate
channel_taus = channel_taus/channel_taus[0]
n_threads=24
#First Layer parameters
Tau_T=125
taus = (Tau_T*channel_taus)
input_channels = 32 + 32*use_all_addr
n_features=20
local_tv_length=10
n_input_channels=input_channels
n_batch_files=None
dataset_runs=1
local_layer_parameters = [n_features, local_tv_length, taus,\
n_batch_files, dataset_runs]
Net = GORDONN(n_threads=n_threads, verbose=True, server_mode=False, low_memory_mode=False)
Net.add_layer("Local", local_layer_parameters, n_input_channels)
#Second layer parameters
# n_features=64
n_features=32
cross_tv_width=3
taus=20e3
cross_layer_parameters = [n_features, cross_tv_width, taus, n_batch_files,
dataset_runs]
Net.add_layer("Cross", cross_layer_parameters)
# Net.learn(dataset_train,labels_train,classes)
# Net.predict(dataset_test, labels_train, labels_test, classes)
#%% Further addictions
#MIG Layers
MI_factor=99.99
Net.add_layer("MIG", [MI_factor])
#Pool Layer
pool_factor=2
Net.add_layer("Pool", [pool_factor])
#Third layer parameters
# n_features=128
n_features=64
cross_tv_width=3
taus=1e6
cross_layer_parameters = [n_features, cross_tv_width, taus, n_batch_files,
dataset_runs]
Net.add_layer("Cross", cross_layer_parameters)
#MIG Layer
MI_factor=95
Net.add_layer("MIG", [MI_factor])
Net.learn(dataset_train,labels_train,classes)
Net.predict(dataset_test, labels_train, labels_test, classes)
#%% Last layer new classifier (New)
#Actual decay
df = pd.read_csv (r'whitenoise.csv')
mean_rate =np.asarray(df.columns[:], dtype=float)
channel_taus = 1/mean_rate
channel_taus = channel_taus/channel_taus[0]
n_threads=24
#First Layer parameters
Tau_T=125
taus = (Tau_T*channel_taus)
input_channels = 32 + 32*use_all_addr
n_features=20
local_tv_length=10
n_input_channels=input_channels
n_batch_files=None
dataset_runs=1
local_layer_parameters = [n_features, local_tv_length, taus,\
n_batch_files, dataset_runs]
Net = GORDONN(n_threads=n_threads, verbose=True, server_mode=False, low_memory_mode=False)
Net.add_layer("Local", local_layer_parameters, n_input_channels)
#Second layer parameters
n_features=16
cross_tv_width=3
taus=20e3
cross_layer_parameters = [n_features, cross_tv_width, taus, n_batch_files,
dataset_runs]
Net.add_layer("Cross", cross_layer_parameters)
#MIG Layers
MI_factor=76
Net.add_layer("MIG", [MI_factor])
#Pool Layer
pool_factor=2
Net.add_layer("Pool", [pool_factor])
#Third layer parameters
n_features=32
cross_tv_width=5
taus=16e4
cross_layer_parameters = [n_features, cross_tv_width, taus, n_batch_files,
dataset_runs]
Net.add_layer("Cross", cross_layer_parameters)
#MIG Layers
# MI_factor=50
# Net.add_layer("MIG", [MI_factor])
Net.learn(dataset_train,labels_train,classes)
Net.predict(dataset_test, labels_train, labels_test, classes)
#%%
Net.layers[5].MI_factor=95
Net.layers[4].cross_tv_width=5
Net.layers[4].n_features=64
Net.learn(dataset_train,labels_train,classes,rerun_layer=4)
Net.predict(dataset_test, labels_train, labels_test, classes, rerun_layer=4)
#%%
Net.layers[4].n_features = 128
Net.layers[4].n_output_features = 128
Net.layers[5].MI_factor = 75
Net.learn(dataset_train,labels_train,classes,rerun_layer=4)
Net.predict(dataset_test, labels_train, labels_test, classes, rerun_layer=4)
#%% Add layers
#Pool Layer
pool_factor=2
Net.add_layer("Pool", [pool_factor])
#Third layer parameters
n_features=128
cross_tv_width=3
taus=1e6
cross_layer_parameters = [n_features, cross_tv_width, taus, n_batch_files,
dataset_runs]
Net.add_layer("Cross", cross_layer_parameters)
Net.learn(dataset_train,labels_train,classes,rerun_layer=5)
Net.predict(dataset_test, labels_train, labels_test, classes, rerun_layer=5)
#%%
rec = 300
n_layers = len(Net.architecture)
fig, axs = plt.subplots(1, n_layers)
for layer in range(n_layers):
word = classes[labels_train[rec]]
fig_title = "Word: '"+word+"' Layer: "+str(layer)+" "+Net.architecture[layer]
timestamps = Net.net_response_train[layer][rec][0]
channels = Net.net_response_train[layer][rec][1]
features = Net.net_response_train[layer][rec][2]
axs[layer].scatter(timestamps,channels, c=features)
axs[layer].set_title(fig_title)
axs[layer].set_ylabel("Channel Index")
axs[layer].set_xlabel("Microseconds")
#%%
#Third layer parameters
n_input_features=n_features
# n_input_channels=16
n_input_channels=11
n_hidden_units=128
cross_tv_width=3
taus=120e3
n_labels=number_of_labels
learning_rate=1e-3
mlp_epochs=50
mlp_ts_batch_size=128
from Libs.GORDONN.Layers.Cross_class_layer import Cross_class_layer
class_layer = Cross_class_layer(n_hidden_units, cross_tv_width, taus, n_labels,
learning_rate, mlp_ts_batch_size, mlp_epochs,
n_input_channels, n_input_features,
n_batch_files, dataset_runs, n_threads, True)
class_layer.learn(Net.net_response_train[-1],labels_train)
mlp = class_layer.mlp
#%%
train_tv, train_labels_one_hot = class_layer.tv_generation(Net.net_response_train[-1],labels_train)
test_tv, test_labels_one_hot = class_layer.tv_generation(Net.net_response_test[-1],labels_test)
#%%% Save
with open('train_tv.npy', 'wb') as f:
np.save(f, train_tv)
with open('train_labels_one_hot.npy', 'wb') as f:
np.save(f, train_labels_one_hot)
with open('test_tv.npy', 'wb') as f:
np.save(f, test_tv)
with open('test_labels_one_hot.npy', 'wb') as f:
np.save(f, test_labels_one_hot)
with open('labels_train.npy', 'wb') as f:
np.save(f, labels_train)
with open('labels_test.npy', 'wb') as f:
np.save(f, labels_test)
# with open('dataset_train.npy', 'wb') as f:
# np.save(f, dataset_train)
# with open('dataset_test.npy', 'wb') as f:
# np.save(f, dataset_test)
#%%% Load
with open('train_tv.npy', 'rb') as f:
train_tv=np.load(f)
with open('train_labels_one_hot.npy', 'rb') as f:
train_labels_one_hot=np.load(f)
with open('test_tv.npy', 'rb') as f:
test_tv=np.load(f)
with open('test_labels_one_hot.npy', 'rb') as f:
test_labels_one_hot=np.load(f)
with open('labels_train.npy', 'rb') as f:
labels_train=np.load(f)
with open('labels_test.npy', 'rb') as f:
labels_test=np.load(f)
# with open('dataset_train.npy', 'rb') as f:
# dataset_train=np.load(f)
# with open('dataset_test.npy', 'rb') as f:
# dataset_test=np.load(f)
#%%
#Third layer parameters
n_input_features=32
# n_input_channels=16
n_input_channels=11
n_hidden_units=128
cross_tv_width=3
taus=120e36
n_labels=number_of_labels
learning_rate=1e-3
mlp_epochs=50
mlp_ts_batch_size=128
mlp = create_mlp(n_input_features*cross_tv_width, n_hidden_units, n_labels, learning_rate)
mlp.fit(train_tv, train_labels_one_hot, epochs=mlp_epochs, batch_size=mlp_ts_batch_size,
shuffle=True)
#%% Test the code
pred_test=mlp.predict(test_tv)
ev_count = 0
n_test_recs = len(dataset_test)
responses = np.zeros([n_test_recs, len(classes)])
for test_rec_i in range(n_test_recs):
n_ev = len(dataset_test[test_rec_i][0])
responses[test_rec_i] = sum(pred_test[ev_count:ev_count+n_ev,:])
predicted_labels = np.argmax(responses,1)
percent_correct=100*sum((labels_test-predicted_labels)==0)/n_test_recs
#%%
# =============================================================================
def create_mlp(input_size, hidden_size, output_size, learning_rate):
"""
Function used to create a small mlp used for classification purposes
Arguments :
input_size (int) : size of the input layer
hidden_size (int) : size of the hidden layer
output_size (int) : size of the output layer
learning_rate (int) : the learning rate for the optimization alg.
Returns :
mlp (keras model) : the freshly baked network
"""
def relu_advanced(x):
return K.activations.relu(x, alpha=0.3)
inputs = Input(shape=(input_size,), name='encoder_input')
# x = BatchNormalization()(inputs)
x = Dense(hidden_size, activation='sigmoid')(inputs)
# x = Dropout(0.3)(x)#0.3
x = Dense(hidden_size, activation='sigmoid')(x)
# x = Dropout(0.7)(x)#0.7
# x = Dense(hidden_size, activation='sigmoid')(x)
# x = Dropout(0.9)(x)
# x = Dense(hidden_size, activation='sigmoid')(x)
# x = Dropout(0.5)(x)
# x = Dense(hidden_size, activation='sigmoid')(x)
# x = Dropout(0.5)(x)
outputs = Dense(output_size, activation='sigmoid')(x)
adam=optimizers.Adam(lr=learning_rate, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
mlp = Model(inputs, outputs, name='mlp')
# mlp.compile(optimizer=adam,
# loss='categorical_crossentropy', metrics=['accuracy'])
mlp.compile(optimizer=adam,
loss='MSE', metrics=['accuracy'])
return mlp
#%% Net Mutual info layer 2
i_layer = 1
MI_factor = 99
n_recordings = len(Net.net_response_train[i_layer])
n_clusters = len(Net.layers[i_layer].features)
labels = labels_train
activity = np.zeros([n_recordings, n_clusters])
for recording in range(n_recordings):
clusters = np.unique(Net.net_response_train[i_layer][recording][-1])
activity[recording, clusters]=1;
p_r_s = np.zeros([number_of_labels, n_clusters])
p_s = 1/number_of_labels
p_r = np.sum(activity,axis=0)/(n_recordings)
p_not_r = 1-p_r
for label in range(number_of_labels):
indx = labels==label
p_r_s[label]+=np.sum(activity[indx,:],axis=0)/sum(labels==label)
p_not_r_s = 1-p_r_s
MI_cluster_on = p_r_s*np.log2(p_r_s/((p_r)+np.logical_not(p_r)) + np.logical_not(p_r_s))
MI_cluster_off = p_not_r_s*np.log2(p_not_r_s/((p_not_r)+np.logical_not(p_not_r)) + np.logical_not(p_not_r_s))
MI = p_s*(MI_cluster_on+MI_cluster_off)
featMI = np.sum(MI, axis=0)
sortfeats = np.argsort(featMI)[::-1]
featMI = featMI[sortfeats]
cumMI = (np.cumsum(featMI)/sum(featMI))*100
topfeats = sortfeats[cumMI<=MI_factor]
plt.figure()
plt.plot(np.cumsum(featMI)/sum(featMI))
plt.ylabel("Bits of information (over 3)")
plt.xlabel("#Features")
plt.title("Sorted MI per cluster")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
# topclusters = np.argsort(clusterMI)[-150:]
rel_clusterMI = featMI/(1+0.01*np.arange(n_clusters))
plt.figure()
plt.plot(rel_clusterMI/n_clusters)
plt.ylabel("Bits of information (over 3)/#Features")
plt.xlabel("#Features")
plt.title("Relative sorted MI")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
#%% Net Mutual info layer 2
i_layer = 1
MI_factor = 99
n_recordings = len(Net.net_response_train[i_layer])
n_clusters = len(Net.layers[i_layer].features)
labels = labels_train
activity = np.zeros([n_recordings, n_clusters])
n_labels=number_of_labels
n_features=n_clusters
for recording in range(n_recordings):
clusters = np.unique(Net.net_response_train[i_layer][recording][-1])
activity[recording, clusters]=1;
p_r = np.sum(activity,axis=0)/(n_recordings)
r_s_table = np.zeros([n_labels, n_features])
for label in range(number_of_labels):
indx = labels==label
p_r_s[label]+=np.sum(activity[indx,:],axis=0)/sum(labels==label)
for label in range(n_labels):
indx = labels==label
r_s_table[label] += np.sum(activity[indx,:],axis=0)
# Entropy of label
H_s = np.log2(n_labels)
featMI = np.zeros(n_features)
for i_feat in range(n_features):
#probability of the cluster being on
p_on = p_r[i_feat]
p_off = 1-p_on
#probability of all labels conditioned to the on cluster
p_s_on = r_s_table[:,i_feat]/sum(r_s_table[:,i_feat])
r_s_table_off = 800-r_s_table#TODO find this value
p_s_off = r_s_table_off[:,i_feat]/sum(r_s_table_off[:,i_feat]+1e-9)
# p_s_off = 1-p_s_on
H_s_on = -p_s_on*np.log2(p_s_on+1e-9)#to avoid zero log
H_s_off = -p_s_off*np.log2(p_s_off+1e-9)#to avoid zero log
H_s_r = p_on*sum(H_s_on) + p_off*sum(H_s_off)
featMI[i_feat] = H_s - H_s_r
sortfeats = np.argsort(featMI)[::-1]
featMI = featMI[sortfeats]
cumMI = (np.cumsum(featMI)/sum(featMI))*100
topfeats = sortfeats[cumMI<=MI_factor]
plt.figure()
plt.plot(np.cumsum(featMI)/sum(featMI))
plt.ylabel("Bits of information (over 3)")
plt.xlabel("#Features")
plt.title("Sorted MI per cluster")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
# topclusters = np.argsort(clusterMI)[-150:]
rel_clusterMI = featMI/(1+0.01*np.arange(n_clusters))
plt.figure()
plt.plot(rel_clusterMI/n_clusters)
plt.ylabel("Bits of information (over 3)/#Features")
plt.xlabel("#Features")
plt.title("Relative sorted MI")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
#%% Net Mutual info layer 2 (rate)
i_layer = 1
MI_factor = 75
n_recordings = len(Net.net_response_train[i_layer])
n_clusters = len(Net.layers[i_layer].features)
labels = labels_train
activity = np.zeros([n_recordings, n_clusters])
n_labels=number_of_labels
n_features=n_clusters
norm_hists = np.mean(Net.layers[i_layer].norm_hists,1)
n_bins = int(1e3)
# n_bins = int(1e5)
#Check histogram too ge the bins
norm_hists_concat = np.concatenate(norm_hists)
plt.figure()
plt.hist(norm_hists_concat, bins=n_bins)
max_response = np.max(norm_hists)
binsize = max_response/n_bins
norm_hists_binned = np.asarray(norm_hists//binsize, dtype=int)# Maybe try fractional bins
# norm_hists_binned+=1
# norm_hists_binned[norm_hists==0]=0
max_rate = np.max(norm_hists_binned)
r_s_table = np.zeros([max_rate, n_labels, n_features])
for rate in range(max_rate):
for label in range(n_labels):
label_indx = labels==label
rate_pos=norm_hists_binned[label_indx]==rate
r_s_table[rate,label] = np.sum(rate_pos,axis=0)
p_r = np.sum(r_s_table,axis=1)/n_recordings
# Entropy of label
H_s = np.log2(n_labels)
featMI = np.zeros(n_features)+H_s
for rate in range(max_rate):
for i_feat in range(n_features):
p_x = p_r[rate, i_feat]
#probability of all labels conditioned to the on cluster
p_s_x = r_s_table[rate,:,i_feat]/sum(r_s_table[rate,:,i_feat]+1e-9)
# p_s_off = 1-p_s_on
H_s_x = -p_s_x*np.log2(p_s_x+1e-9)#to avoid zero log
H_s_r = p_x*sum(H_s_x)
featMI[i_feat] -= H_s_r
sortfeats = np.argsort(featMI)[::-1]
featMI = featMI[sortfeats]
cumMI = (np.cumsum(featMI)/sum(featMI))*100
topfeats = sortfeats[cumMI<=MI_factor]
plt.figure()
plt.plot(np.cumsum(featMI)/sum(featMI))
plt.ylabel("Bits of information (over 3)")
plt.xlabel("#Features")
plt.title("Sorted MI per cluster")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
# topclusters = np.argsort(clusterMI)[-150:]
rel_clusterMI = featMI/(1+0.01*np.arange(n_clusters))
plt.figure()
plt.plot(rel_clusterMI/n_clusters)
plt.ylabel("Bits of information (over 3)/#Features")
plt.xlabel("#Features")
plt.title("Relative sorted MI")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])
#%% Net Mutual info layer 2 (rate) HISTS abs
i_layer = 1
MI_factor = 75
n_recordings = len(Net.net_response_train[i_layer])
n_clusters = len(Net.layers[i_layer].features)
labels = labels_train
activity = np.zeros([n_recordings, n_clusters])
n_labels=number_of_labels
n_features=n_clusters
hists = np.sum(Net.layers[i_layer].hists,1)
norm_hists = hists/np.max(hists)
#Check histogram too ge the bins
norm_hists_concat = np.concatenate(norm_hists)
plt.figure()
plt.hist(norm_hists_concat, bins=100)
binsize = 0.01
#set the zero apart
norm_hists_binned = np.asarray(norm_hists//binsize, dtype=int)# Maybe try fractional bins
norm_hists_binned+=1
norm_hists_binned[norm_hists==0]=0
max_rate = np.max(norm_hists_binned)
r_s_table = np.zeros([max_rate, n_labels, n_features])
for rate in range(max_rate):
for label in range(n_labels):
label_indx = labels==label
rate_pos=norm_hists_binned[label_indx]==rate
r_s_table[rate,label] = np.sum(rate_pos,axis=0)
p_r = np.sum(r_s_table,axis=1)/n_recordings
# Entropy of label
H_s = np.log2(n_labels)
featMI = np.zeros(n_features)+H_s
for rate in range(max_rate):
for i_feat in range(n_features):
p_x = p_r[rate, i_feat]
#probability of all labels conditioned to the on cluster
p_s_x = r_s_table[rate,:,i_feat]/sum(r_s_table[rate,:,i_feat]+1e-9)
# p_s_off = 1-p_s_on
H_s_x = -p_s_x*np.log2(p_s_x+1e-9)#to avoid zero log
H_s_r = p_x*sum(H_s_x)
featMI[i_feat] -= H_s_r
sortfeats = np.argsort(featMI)[::-1]
featMI = featMI[sortfeats]
cumMI = (np.cumsum(featMI)/sum(featMI))*100
topfeats = sortfeats[cumMI<=MI_factor]
plt.figure()
plt.plot(np.cumsum(featMI)/sum(featMI))
plt.ylabel("Bits of information (over 3)")
plt.xlabel("#Features")
plt.title("Sorted MI per cluster")
plt.vlines(len(topfeats),plt.ylim()[0],plt.ylim()[1])