-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsimulator_enose.py
1115 lines (1034 loc) · 57.6 KB
/
simulator_enose.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 numpy as np
import matplotlib.pyplot as plt
from pygenn import genn_model
from pygenn.genn_wrapper import NO_DELAY
from utils import random_shift, random_dilate, ID_jitter
import mnist
#import tonic
from models import *
import os
import urllib.request
import gzip, shutil
from tensorflow.keras.utils import get_file
import tables
import copy
from time import perf_counter
#from enose_data_loader import enose_data_load
#from src.data_loader import EnoseDataLoader
# ----------------------------------------------------------------------------
# Parameters
# ----------------------------------------------------------------------------
p= {}
p["NAME"]= "test"
p["DATASET"] = None
p["DEBUG"]= False
p["DEBUG_HIDDEN_N"]= False
p["OUT_DIR"]= "."
p["DT_MS"] = 0.1
p["BUILD"] = True
p["TIMING"] = True
p["TRAIN_DATA_SEED"]= 123
p["TEST_DATA_SEED"]= 456
p["MODEL_SEED"]= None
# Experiment parameters
p["TRIAL_MS"]= 20.0
p["N_MAX_SPIKE"]= 400 # make buffers for maximally 400 spikes (200 in a 30 ms trial) - should be safe
p["N_BATCH"]= 32
p["SUPER_BATCH"]= 1
p["N_TRAIN"]= 55000
p["N_VALIDATE"]= 5000
p["N_EPOCH"]= 10
p["SHUFFLE"]= True
p["N_TEST"]= 10000
# Network structure
p["NUM_HIDDEN"] = 350
p["RECURRENT"] = False
# Model parameters
p["TAU_SYN"] = 5.0
p["TAU_MEM"] = 20.0
p["V_THRESH"] = 1.0
p["V_RESET"] = 0.0
p["INPUT_HIDDEN_MEAN"]= 0.078
p["INPUT_HIDDEN_STD"]= 0.045
p["HIDDEN_OUTPUT_MEAN"]= 0.2
p["HIDDEN_OUTPUT_STD"]= 0.37
p["HIDDEN_HIDDEN_MEAN"]= 0.2 # only used when recurrent
p["HIDDEN_HIDDEN_STD"]= 0.37 # only used when recurrent
p["PDROP_INPUT"] = 0.2
p["PDROP_HIDDEN"] = 0.0
p["REG_TYPE"]= "none"
p["LBD_UPPER"]= 0.000005
p["LBD_LOWER"]= 0.001
p["NU_UPPER"]= 20*p["N_BATCH"]
p["NU_LOWER"]= 0.1*p["N_BATCH"]
p["RHO_UPPER"]= 5000.0
p["GLB_UPPER"]= 0.00001
# Learning parameters
p["ETA"]= 5e-3
p["ADAM_BETA1"]= 0.9
p["ADAM_BETA2"]= 0.999
p["ADAM_EPS"]= 1e-8
# applied every epoch
p["ETA_DECAY"]= 0.95
# try a step-down of learning rate after substantial training
p["ETA_FIDDELING"]= False
p["ETA_REDUCE"]= 0.1
p["ETA_REDUCE_PERIOD"]= 50
# recording
p["W_OUTPUT_EPOCH_TRIAL"] = []
p["SPK_REC_STEPS"]= int(p["TRIAL_MS"]/p["DT_MS"])
p["REC_SPIKES_EPOCH_TRIAL"] = []
p["REC_SPIKES"] = []
p["REC_NEURONS_EPOCH_TRIAL"] = []
p["REC_NEURONS"] = []
p["REC_SYNAPSES_EPOCH_TRIAL"] = []
p["REC_SYNAPSES"] = []
p["WRITE_TO_DISK"]= True
p["LOAD_LAST"]= False
# possible loss types: "max", "sum", "avg_xentropy"
p["LOSS_TYPE"]= "max"
p["EVALUATION"]= "random"
p["CUDA_VISIBLE_DEVICES"]= False
p["AVG_SNSUM"]= False
p["REDUCED_CLASSES"]= None
p["AUGMENTATION"]= {}
# ----------------------------------------------------------------------------
# Helper functions
# ----------------------------------------------------------------------------
rng= np.random.default_rng()
def update_adam(learning_rate, adam_step, optimiser_custom_updates):
first_moment_scale = 1.0 / (1.0 - (p["ADAM_BETA1"] ** adam_step))
second_moment_scale = 1.0 / (1.0 - (p["ADAM_BETA2"] ** adam_step))
# Loop through optimisers and set
for o in optimiser_custom_updates:
o.extra_global_params["alpha"].view[:] = learning_rate
o.extra_global_params["firstMomentScale"].view[:] = first_moment_scale
o.extra_global_params["secondMomentScale"].view[:] = second_moment_scale
class mnist_model:
def __init__(self, p):
if p["TRAIN_DATA_SEED"] is not None:
self.datarng= np.random.default_rng(p["TRAIN_DATA_SEED"])
else:
self.datarng= np.random.default_rng()
if p["TEST_DATA_SEED"] is not None:
self.tdatarng= np.random.default_rng(p["TEST_DATA_SEED"])
else:
self.tdatarng= np.random.default_rng()
if p["DATASET"] == "enose":
self.load_data_enose(p)
if p["REDUCED_CLASSES"] is not None:
self.X_train_orig, self.Y_train_orig, self.Z_train_orig= self.reduce_classes(self.X_train_orig, self.Y_train_orig, self.Z_train_orig, p["REDUCED_CLASSES"])
self.X_test_orig, self.Y_test_orig, self.Z_test_orig= self.reduce_classes(self.X_test_orig, self.Y_test_orig, self.Z_test_orig, p["REDUCED_CLASSES"])
def loss_func(self, Y, p):
expsum= self.output.vars["expsum"].view
exp_V= self.output.vars["exp_V"].view
exp_V_correct= np.array([ exp_V[i,y] for i, y in enumerate(Y) ])
if (np.sum(exp_V_correct == 0) > 0):
print("exp_V flushed to 0 exception!")
print(exp_V_correct)
print(exp_V[np.where(exp_V_correct == 0),:])
exp_V_correct[exp_V_correct == 0]+= 2e-45 # make sure all exp_V are > 0
loss= -np.sum(np.log(exp_V_correct)-np.log(expsum[:,0]))/p["N_BATCH"]
return loss
def loss_func_avg_xentropy(self, Y, p):
loss= self.output.vars["loss"].view
loss= np.sum(np.sum(loss))
return loss
def load_data_MNIST(self, p, shuffle= True):
X = mnist.train_images()
Y = mnist.train_labels()
self.data_full_length= 60000
self.N_class= 10
self.num_input= 28*28
self.num_output= 16 # first power of two greater than class number
idx= np.arange(self.data_full_length)
if (shuffle):
self.datarng.shuffle(idx)
X= X[idx]
self.X_val_orig= X[self.data_full_length-p["N_VALIDATE"]:,:,:]
self.X_train_orig= X[:p["N_TRAIN"],:,:]
Y= Y[idx]
self.Y_val_orig= Y[self.data_full_length-p["N_VALIDATE"]:]
self.Y_train_orig= Y[:p["N_TRAIN"]]
# also load some testing data
X = mnist.test_images()
Y = mnist.test_labels()
idx= np.arange(10000)
if (shuffle):
self.tdatarng.shuffle(idx)
X= X[idx]
self.X_test_orig= X[:p["N_TEST"],:,:]
Y= Y[idx]
self.Y_test_orig= Y[:p["N_TEST"]]
"""
For now I will disable this - uses tonic,which might be nicer but doesn't give access to speaker info
def load_data_SHD(self, p, shuffle= True):
if p["TRAIN_DATA_SEED"] is not None:
self.datarng= np.random.default_rng(p["TRAIN_DATA_SEED"])
else:
self.datarng= np.random.default_rng()
if p["TEST_DATA_SEED"] is not None:
self.tdatarng= np.random.default_rng(p["TEST_DATA_SEED"])
else:
self.tdatarng= np.random.default_rng()
dataset = tonic.datasets.SHD(save_to='./data', train=True)
sensor_size = dataset.sensor_size
self.data_full_length= len(dataset)
self.N_class= len(dataset.classes)
self.num_input= int(np.product(sensor_size))
self.num_output= 32 # first power of two greater than class number
idx= np.arange(self.data_full_length)
if (shuffle):
self.datarng.shuffle(idx)
train_idx= idx[np.arange(p["N_TRAIN"])]
eval_idx= idx[np.arange(p["N_VALIDATE"])+(self.data_full_length-p["N_VALIDATE"])]
self.Y_train_orig= np.empty(len(train_idx), dtype= int)
self.X_train_orig= []
for i, s in enumerate(train_idx):
events, label = dataset[s]
self.Y_train_orig[i]= label
self.X_train_orig.append(events)
self.Y_val_orig= np.empty(len(eval_idx), dtype= int)
self.X_val_orig= []
for i, s in enumerate(eval_idx):
events, label = dataset[s]
self.Y_val_orig[i]= label
self.X_val_orig.append(events)
dataset = tonic.datasets.SHD(save_to='./data', train=False)
self.data_full_length= max(self.data_full_length, len(dataset))
self.Y_test_orig= np.empty(len(dataset), dtype= int)
self.X_test_orig= []
for i in range(len(dataset)):
events, label = dataset[i]
self.Y_test_orig[i]= label
self.X_test_orig.append(events)
"""
def load_data_SHD_Zenke(self, p):
cache_dir=os.path.expanduser("~/data")
cache_subdir="SHD"
print("Using cache dir: %s"%cache_dir)
"""
#(uncomment this if you need to download the data and have internet access; comment when not connected to the public internet)
# The remote directory with the data files
base_url = "https://zenkelab.org/datasets"
# Retrieve MD5 hashes from remote
response = urllib.request.urlopen("%s/md5sums.txt"%base_url)
data = response.read()
lines = data.decode('utf-8').split("\n")
file_hashes = { line.split()[1]:line.split()[0] for line in lines if len(line.split())==2 }
def get_and_gunzip(origin, filename, md5hash=None):
gz_file_path = get_file(filename, origin, md5_hash=md5hash, cache_dir=cache_dir, cache_subdir=cache_subdir)
hdf5_file_path=gz_file_path[:-3]
if not os.path.isfile(hdf5_file_path) or os.path.getctime(gz_file_path) > os.path.getctime(hdf5_file_path):
print("Decompressing %s"%gz_file_path)
with gzip.open(gz_file_path, 'r') as f_in, open(hdf5_file_path, 'wb') as f_out:
shutil.copyfileobj(f_in, f_out)
return hdf5_file_path
# Download the Spiking Heidelberg Digits (SHD) dataset
files = [ "shd_train.h5.gz",
"shd_test.h5.gz",
]
hdf5_file_path= []
# (end of download code)
"""
self.num_input= 700
self.num_output= 32 # first power of two greater than class number
self.data_full_length= 0
# (use below when freshly downloading data)
#fn= files[0]
#origin= "%s/%s"%(base_url,fn)
#hdf5_file_path= get_and_gunzip(origin, fn, md5hash=file_hashes[fn])
hdf5_file_path= 'data/SHD/shd_train.h5'
fileh= tables.open_file(hdf5_file_path, mode='r')
units= fileh.root.spikes.units
times= fileh.root.spikes.times
self.Y_train_orig= fileh.root.labels
self.Z_train_orig= fileh.root.extra.speaker
self.data_full_length= max(self.data_full_length, len(units))
self.N_class= len(set(self.Y_train_orig))
self.X_train_orig= []
for i in range(len(units)):
self.X_train_orig.append({"x": units[i], "t": times[i]})
self.X_train_orig= np.array(self.X_train_orig)
# (use below when freshly downloading data)
#fn= files[1]
#origin= "%s/%s"%(base_url,fn)
#hdf5_file_path= get_and_gunzip(origin, fn, md5hash=file_hashes[fn])
hdf5_file_path= 'data/SHD/shd_test.h5'
fileh= tables.open_file(hdf5_file_path, mode='r')
units= fileh.root.spikes.units
times= fileh.root.spikes.times
self.Y_test_orig= fileh.root.labels
self.Z_test_orig= fileh.root.extra.speaker
self.data_full_length= max(self.data_full_length, len(units))
self.X_test_orig= []
for i in range(len(units)):
self.X_test_orig.append({"x": units[i], "t": times[i]})
self.X_test_orig= np.array(self.X_test_orig)
# def load_data_enose(self,p):
# self.num_input= 8
# self.num_output= 8 # first power of two greater than class number
# self.N_class= 5
# self.X_train_orig, self.Y_train_orig, self.X_test_orig, self.Y_test_orig= enose_data_load()
# self.data_full_length= len(self.Y_train_orig)
def reduce_classes(self, X, Y, Z, classes):
idx= [y in classes for y in Y]
newX= X[idx]
newY= Y[idx]
newZ= Z[idx]
return (newX, newY, newZ)
def split_SHD_random(self, X, Y, p, shuffle= True):
idx= np.arange(len(X),dtype= int)
if (shuffle):
self.datarng.shuffle(idx)
train_idx= idx[np.arange(p["N_TRAIN"])]
eval_idx= idx[np.arange(p["N_VALIDATE"])+(self.data_full_length-p["N_VALIDATE"])]
print(train_idx)
newX_t= X[train_idx]
newX_e= X[eval_idx]
newY_t= Y[train_idx]
newY_e= Y[eval_idx]
print(newX_t[0])
return (newX_t, newY_t, newX_e, newY_e)
# split off one speaker to form evaluation set
def split_SHD_speaker(self, X, Y, Z, speaker, p, shuffle= True):
speaker= np.array(speaker)
newX_t= X[Z != speaker]
newY_t= Y[Z != speaker]
train_idx= np.arange(len(newY_t))
if shuffle:
self.datarng.shuffle(train_idx)
train_idx= train_idx[:p["N_TRAIN"]]
newX_t= newX_t[train_idx]
newY_t= newY_t[train_idx]
newX_e= X[Z == speaker]
newY_e= Y[Z == speaker]
eval_idx= np.arange(len(newY_e))
if shuffle:
self.datarng.shuffle(eval_idx)
eval_idx= eval_idx[:p["N_VALIDATE"]]
newX_e= newX_e[eval_idx]
newY_e= newY_e[eval_idx]
return (newX_t, newY_t, newX_e, newY_e)
def spike_time_from_gray(self,t):
return (255.0-t)/255.0*(p["TRIAL_MS"]-4*p["DT_MS"])+2*p["DT_MS"] # make sure spikes are two timesteps within the presentation window
def spike_time_from_gray2(self,t):
t= t/255.0*10.0
return 10.0*np.log(t/(t-0.2))
"""
generate a spikeTimes array and startSpike and endSpike arrays to allow indexing into the
spikeTimes in a shuffled way
"""
# ISSUE: here we are not rounding to the multiples of batch size!
# When data loading, we are doing that for N_trial ...
# Needs tidying up!
def generate_input_spiketimes_shuffle_fast(self, p, Xtrain, Ytrain, Xeval, Yeval):
# N is the number of training/testing images: always use all images given
if Xtrain is None:
X= Xeval
Y= Yeval
else:
if Xeval is None:
X= Xtrain
Y= Ytrain
else:
X= np.append(Xtrain, Xeval, axis= 0)
Y= np.append(Ytrain, Yeval, axis= 0)
N= len(Y)
all_sts= []
all_input_end= []
all_input_start= []
stidx_offset= 0
self.max_stim_time= 0.0
for i in range(N):
if p["DATASET"] == "enose":
events= X[i]
spike_event_ids = events["x"]
i_end = np.cumsum(np.bincount(spike_event_ids.astype(int),
minlength=self.num_input))+stidx_offset
assert len(i_end) == self.num_input
tx = events["t"][np.lexsort((events["t"], spike_event_ids))].astype(float)
self.max_stim_time= max(self.max_stim_time, np.amax(tx))
all_sts.append(tx)
i_start= np.empty(i_end.shape)
i_start[0]= stidx_offset
i_start[1:]= i_end[:-1]
all_input_end.append(i_end)
all_input_start.append(i_start)
stidx_offset= i_end[-1]
X= np.hstack(all_sts)
input_end= np.hstack(all_input_end)
input_start= np.hstack(all_input_start)
return (X, Y, input_start, input_end)
def define_model(self, p, shuffle):
self.trial_steps= int(round(p["TRIAL_MS"]/p["DT_MS"]))
input_params= {"N_neurons": self.num_input,
"N_max_spike": p["N_MAX_SPIKE"]
}
self.input_init_vars= {"startSpike": 0.0, # to be set later
"endSpike": 0.0, # to be set later
"back_spike": 0,
"rp_ImV": p["N_MAX_SPIKE"]-1,
"wp_ImV": 0,
"fwd_start": p["N_MAX_SPIKE"]-1,
"new_fwd_start": p["N_MAX_SPIKE"]-1,
"rev_t": 0.0}
hidden_params= {"tau_m": p["TAU_MEM"],
"V_thresh": p["V_THRESH"],
"V_reset": p["V_RESET"],
"N_neurons": p["NUM_HIDDEN"],
"N_max_spike": p["N_MAX_SPIKE"],
"tau_syn": p["TAU_SYN"],
}
self.hidden_init_vars= {"V": p["V_RESET"],
"lambda_V": 0.0,
"lambda_I": 0.0,
"rev_t": 0.0,
"rp_ImV": p["N_MAX_SPIKE"]-1,
"wp_ImV": 0,
"fwd_start": p["N_MAX_SPIKE"]-1,
"new_fwd_start": p["N_MAX_SPIKE"]-1,
"back_spike": 0,
}
output_params= {"tau_m": p["TAU_MEM"],
"tau_syn": p["TAU_SYN"],
"N_batch": p["N_BATCH"],
"trial_t": p["TRIAL_MS"],
}
if p["LOSS_TYPE"] == "avg_xentropy":
output_params["N_neurons"]= self.num_output
output_params["trial_steps"]= self.trial_steps
output_params["N_class"]= self.N_class
self.output_init_vars= {"V": p["V_RESET"],
"lambda_V": 0.0,
"lambda_I": 0.0,
"trial": 0,
}
if p["LOSS_TYPE"] == "max":
self.output_init_vars["max_V"]= p["V_RESET"]
self.output_init_vars["new_max_V"]= p["V_RESET"]
self.output_init_vars["max_t"]= 0.0
self.output_init_vars["new_max_t"]= 0.0
self.output_init_vars["rev_t"]= 0.0
self.output_init_vars["expsum"]= 1.0
self.output_init_vars["exp_V"]= 1.0
if p["LOSS_TYPE"] == "sum":
self.output_init_vars["sum_V"]= 0.0
self.output_init_vars["new_sum_V"]= 0.0
self.output_init_vars["rev_t"]= 0.0
self.output_init_vars["expsum"]= 1.0
self.output_init_vars["exp_V"]= 1.0
if p["LOSS_TYPE"] == "avg_xentropy":
self.output_init_vars["sum_V"]= 0.0
self.output_init_vars["rp_V"]= 0
self.output_init_vars["wp_V"]= 0
self.output_init_vars["loss"]= 0
# ----------------------------------------------------------------------------
# Synapse initialisation
# ----------------------------------------------------------------------------
self.in_to_hid_init_vars= {"dw": 0}
self.in_to_hid_init_vars["w"]= genn_model.init_var("Normal", {"mean": p["INPUT_HIDDEN_MEAN"], "sd": p["INPUT_HIDDEN_STD"]})
self.hid_to_out_init_vars= {"dw": 0}
self.hid_to_out_init_vars["w"]= genn_model.init_var("Normal", {"mean": p["HIDDEN_OUTPUT_MEAN"], "sd": p["HIDDEN_OUTPUT_STD"]})
if p["RECURRENT"]:
self.hid_to_hid_init_vars= {"dw": 0}
self.hid_to_hid_init_vars["w"]= genn_model.init_var("Normal", {"mean": p["HIDDEN_HIDDEN_MEAN"], "sd": p["HIDDEN_HIDDEN_STD"]})
# ----------------------------------------------------------------------------
# Optimiser initialisation
# ----------------------------------------------------------------------------
adam_params = {"beta1": p["ADAM_BETA1"], "beta2": p["ADAM_BETA2"], "epsilon": p["ADAM_EPS"], "tau_syn": p["TAU_SYN"], "N_batch": p["N_BATCH"]}
self.adam_init_vars = {"m": 0.0, "v": 0.0}
# ----------------------------------------------------------------------------
# Model description
# ----------------------------------------------------------------------------
kwargs = {}
if p["CUDA_VISIBLE_DEVICES"]:
from pygenn.genn_wrapper.CUDABackend import DeviceSelect_MANUAL
kwargs["selectGPUByDeviceID"] = True
kwargs["deviceSelectMethod"] = DeviceSelect_MANUAL
self.model = genn_model.GeNNModel("float", p["NAME"], generateLineInfo=True, time_precision="double", **kwargs)
self.model.dT = p["DT_MS"]
self.model.timing_enabled = p["TIMING"]
self.model.batch_size = p["N_BATCH"]
if p["MODEL_SEED"] is not None:
self.model._model.set_seed(p["MODEL_SEED"])
# Add neuron populations
self.input = self.model.add_neuron_population("input", self.num_input, EVP_SSA_MNIST_SHUFFLE,
input_params, self.input_init_vars)
self.input.set_extra_global_param("t_k",-1e5*np.ones(p["N_BATCH"]*self.num_input*p["N_MAX_SPIKE"],dtype=np.float32))
self.input.set_extra_global_param("spikeTimes", np.zeros(200000000,dtype=np.float32)) # reserve enough space for any set of input spikes that is likely
if p["REG_TYPE"] == "simple":
hidden_params["N_batch"]= p["N_BATCH"]
hidden_params["lbd_upper"]= p["LBD_UPPER"]
hidden_params["lbd_lower"]= p["LBD_LOWER"]
hidden_params["nu_upper"]= p["NU_UPPER"]
self.hidden_init_vars["sNSum"]= 0.0
self.hidden_init_vars["new_sNSum"]= 0.0
self.hidden= self.model.add_neuron_population("hidden", p["NUM_HIDDEN"], EVP_LIF_reg, hidden_params, self.hidden_init_vars)
if p["REG_TYPE"] == "Thomas1":
hidden_params["N_batch"]= p["N_BATCH"]
hidden_params["lbd_lower"]= p["LBD_LOWER"]
hidden_params["nu_lower"]= p["NU_LOWER"]
hidden_params["lbd_upper"]= p["LBD_UPPER"]
hidden_params["nu_upper"]= p["NU_UPPER"]
hidden_params["rho_upper"]= p["RHO_UPPER"]
hidden_params["glb_upper"]= p["GLB_UPPER"]
hidden_params["N_batch"]= p["N_BATCH"]
self.hidden_init_vars["sNSum"]= 0.0
self.hidden_init_vars["new_sNSum"]= 0.0
self.hidden= self.model.add_neuron_population("hidden", p["NUM_HIDDEN"], EVP_LIF_reg_Thomas1, hidden_params, self.hidden_init_vars)
if p["REG_TYPE"] == "none":
self.hidden= self.model.add_neuron_population("hidden", p["NUM_HIDDEN"], EVP_LIF, hidden_params, self.hidden_init_vars)
self.hidden.set_extra_global_param("t_k",-1e5*np.ones(p["N_BATCH"]*p["NUM_HIDDEN"]*p["N_MAX_SPIKE"],dtype=np.float32))
self.hidden.set_extra_global_param("ImV",np.zeros(p["N_BATCH"]*p["NUM_HIDDEN"]*p["N_MAX_SPIKE"],dtype=np.float32))
if p["REG_TYPE"] == "Thomas1":
self.hidden.set_extra_global_param("sNSum_all", np.zeros(p["N_BATCH"]))
if p["LOSS_TYPE"] == "max":
self.output= self.model.add_neuron_population("output", self.num_output, EVP_LIF_output_MNIST, output_params, self.output_init_vars)
if p["LOSS_TYPE"] == "sum":
self.output= self.model.add_neuron_population("output", self.num_output, EVP_LIF_output_MNIST_sum, output_params, self.output_init_vars)
if p["LOSS_TYPE"] == "avg_xentropy":
self.output= self.model.add_neuron_population("output", self.num_output, EVP_LIF_output_avg_xentropy, output_params, self.output_init_vars)
self.output.set_extra_global_param("label", np.zeros(self.data_full_length,dtype=np.float32)) # reserve space for labels
if p["LOSS_TYPE"] == "avg_xentropy":
self.output.set_extra_global_param("Vbuf", np.zeros(p["N_BATCH"]*self.num_output*self.trial_steps*2,dtype=np.float32)) # reserve space for voltage buffer
input_var_refs= {"rp_ImV": genn_model.create_var_ref(self.input, "rp_ImV"),
"wp_ImV": genn_model.create_var_ref(self.input, "wp_ImV"),
"back_spike": genn_model.create_var_ref(self.input, "back_spike"),
"fwd_start": genn_model.create_var_ref(self.input, "fwd_start"),
"new_fwd_start": genn_model.create_var_ref(self.input, "new_fwd_start"),
"rev_t": genn_model.create_var_ref(self.input, "rev_t")
}
self.input_reset= self.model.add_custom_update("input_reset","neuronReset", EVP_input_reset_MNIST, {"N_max_spike": p["N_MAX_SPIKE"]}, {}, input_var_refs)
input_set_params= {"N_batch": p["N_BATCH"],
"num_input": self.num_input
}
input_var_refs= {"startSpike": genn_model.create_var_ref(self.input, "startSpike"),
"endSpike": genn_model.create_var_ref(self.input, "endSpike")
}
self.input_set= self.model.add_custom_update("input_set", "inputUpdate", EVP_input_set_MNIST_shuffle, input_set_params, {}, input_var_refs)
# reserving memory for the worst case of the full training set
self.input_set.set_extra_global_param("allStartSpike", np.zeros(self.data_full_length*self.num_input,dtype= int))
self.input_set.set_extra_global_param("allEndSpike", np.zeros(self.data_full_length*self.num_input,dtype= int))
self.input_set.set_extra_global_param("allInputID", np.zeros(self.data_full_length,dtype= int))
self.input_set.set_extra_global_param("trial", 0)
hidden_var_refs= {"rp_ImV": genn_model.create_var_ref(self.hidden, "rp_ImV"),
"wp_ImV": genn_model.create_var_ref(self.hidden, "wp_ImV"),
"V": genn_model.create_var_ref(self.hidden, "V"),
"lambda_V": genn_model.create_var_ref(self.hidden, "lambda_V"),
"lambda_I": genn_model.create_var_ref(self.hidden, "lambda_I"),
"rev_t": genn_model.create_var_ref(self.hidden, "rev_t"),
"fwd_start": genn_model.create_var_ref(self.hidden, "fwd_start"),
"new_fwd_start": genn_model.create_var_ref(self.hidden, "new_fwd_start"),
"back_spike": genn_model.create_var_ref(self.hidden, "back_spike")
}
if p["REG_TYPE"] == "simple" or p["REG_TYPE"] == "Thomas1":
hidden_var_refs["sNSum"]= genn_model.create_var_ref(self.hidden, "sNSum")
hidden_var_refs["new_sNSum"]= genn_model.create_var_ref(self.hidden, "new_sNSum")
if p["REG_TYPE"] == "simple":
self.hidden_reset= self.model.add_custom_update("hidden_reset","neuronReset", EVP_neuron_reset_reg, {"V_reset": p["V_RESET"], "N_max_spike": p["N_MAX_SPIKE"], "N_neurons": p["NUM_HIDDEN"]}, {}, hidden_var_refs)
if p["REG_TYPE"] == "Thomas1":
self.hidden_reset= self.model.add_custom_update("hidden_reset","neuronReset", EVP_neuron_reset_reg_global, {"V_reset": p["V_RESET"], "N_max_spike": p["N_MAX_SPIKE"], "N_neurons": p["NUM_HIDDEN"]}, {}, hidden_var_refs)
self.hidden_reset.set_extra_global_param("sNSum_all", np.zeros(p["N_BATCH"]))
if (p["REG_TYPE"] == "simple" or p["REG_TYPE"] == "Thomas1") and p["AVG_SNSUM"]:
var_refs= {"sNSum": genn_model.create_var_ref(self.hidden, "sNSum")}
self.hidden_reg_reduce= self.model.add_custom_update("hidden_reg_reduce","sNSumReduce", EVP_reg_reduce, {}, {"reduced_sNSum": 0.0}, var_refs)
var_refs= {
"reduced_sNSum": genn_model.create_var_ref(self.hidden_reg_reduce, "reduced_sNSum"),
"sNSum": genn_model.create_var_ref(self.hidden, "sNSum")
}
self.hidden_redSNSum_apply= self.model.add_custom_update("hidden_redSNSum_apply","sNSumApply", EVP_sNSum_apply, {"N_batch": p["N_BATCH"]}, {}, var_refs)
if p["REG_TYPE"] == "none":
self.hidden_reset= self.model.add_custom_update("hidden_reset","neuronReset", EVP_neuron_reset, {"V_reset": p["V_RESET"], "N_max_spike": p["N_MAX_SPIKE"]}, {}, hidden_var_refs)
output_reset_params= {"V_reset": p["V_RESET"],
"N_class": self.N_class
}
if p["LOSS_TYPE"] == "avg_xentropy":
output_reset_params["trial_steps"]= self.trial_steps
output_var_refs= {"V": genn_model.create_var_ref(self.output, "V"),
"lambda_V": genn_model.create_var_ref(self.output, "lambda_V"),
"lambda_I": genn_model.create_var_ref(self.output, "lambda_I"),
"trial": genn_model.create_var_ref(self.output, "trial")
}
if p["LOSS_TYPE"] == "max":
output_var_refs["max_V"]= genn_model.create_var_ref(self.output, "max_V")
output_var_refs["new_max_V"]= genn_model.create_var_ref(self.output, "new_max_V")
output_var_refs["max_t"]= genn_model.create_var_ref(self.output, "max_t")
output_var_refs["new_max_t"]= genn_model.create_var_ref(self.output, "new_max_t")
output_var_refs["rev_t"]= genn_model.create_var_ref(self.output, "rev_t")
output_var_refs["expsum"]= genn_model.create_var_ref(self.output, "expsum")
output_var_refs["exp_V"]= genn_model.create_var_ref(self.output, "exp_V")
if p["LOSS_TYPE"] == "sum":
output_var_refs["sum_V"]= genn_model.create_var_ref(self.output, "sum_V")
output_var_refs["new_sum_V"]= genn_model.create_var_ref(self.output, "new_sum_V")
output_var_refs["rev_t"]= genn_model.create_var_ref(self.output, "rev_t")
output_var_refs["expsum"]= genn_model.create_var_ref(self.output, "expsum")
output_var_refs["exp_V"]= genn_model.create_var_ref(self.output, "exp_V")
if p["LOSS_TYPE"] == "avg_xentropy":
output_var_refs["sum_V"]= genn_model.create_var_ref(self.output, "sum_V")
output_var_refs["rp_V"]= genn_model.create_var_ref(self.output, "rp_V")
output_var_refs["wp_V"]= genn_model.create_var_ref(self.output, "wp_V")
output_var_refs["loss"]= genn_model.create_var_ref(self.output, "loss")
if p["DATASET"] == "enose":
if p["LOSS_TYPE"] == "max":
self.output_reset= self.model.add_custom_update("output_reset","neuronReset", EVP_neuron_reset_output_SHD, output_reset_params, {}, output_var_refs)
if p["LOSS_TYPE"] == "sum":
self.output_reset= self.model.add_custom_update("output_reset","neuronReset", EVP_neuron_reset_output_SHD_sum, output_reset_params, {}, output_var_refs)
if p["LOSS_TYPE"] == "avg_xentropy":
self.output_reset= self.model.add_custom_update("output_reset","neuronReset", EVP_neuron_reset_output_avg_xentropy, output_reset_params, {}, output_var_refs)
# synapse populations
self.in_to_hid= self.model.add_synapse_population("in_to_hid", "DENSE_INDIVIDUALG", NO_DELAY, self.input, self.hidden, EVP_input_synapse,
{}, self.in_to_hid_init_vars, {}, {}, my_Exp_Curr, {"tau": p["TAU_SYN"]}, {})
self.hid_to_out= self.model.add_synapse_population("hid_to_out", "DENSE_INDIVIDUALG", NO_DELAY, self.hidden, self.output, EVP_synapse,
{}, self.hid_to_out_init_vars, {}, {}, my_Exp_Curr, {"tau": p["TAU_SYN"]}, {})
if p["RECURRENT"]:
self.hid_to_hid= self.model.add_synapse_population("hid_to_hid", "DENSE_INDIVIDUALG", NO_DELAY, self.hidden, self.hidden, EVP_synapse,
{}, self.hid_to_hid_init_vars, {}, {}, my_Exp_Curr, {"tau": p["TAU_SYN"]}, {})
self.optimisers= []
var_refs = {"dw": genn_model.create_wu_var_ref(self.in_to_hid, "dw")}
self.in_to_hid_reduce= self.model.add_custom_update("in_to_hid_reduce","EVPReduce", EVP_grad_reduce, {}, {"reduced_dw": 0.0}, var_refs)
var_refs = {"gradient": genn_model.create_wu_var_ref(self.in_to_hid_reduce, "reduced_dw"),
"variable": genn_model.create_wu_var_ref(self.in_to_hid, "w")}
self.in_to_hid_learn= self.model.add_custom_update("in_to_hid_learn","EVPLearn", adam_optimizer_model, adam_params, self.adam_init_vars, var_refs)
self.optimisers.append(self.in_to_hid_learn)
var_refs = {"dw": genn_model.create_wu_var_ref(self.hid_to_out, "dw")}
self.hid_to_out_reduce= self.model.add_custom_update("hid_to_out_reduce","EVPReduce", EVP_grad_reduce, {}, {"reduced_dw": 0.0}, var_refs)
var_refs = {"gradient": genn_model.create_wu_var_ref(self.hid_to_out_reduce, "reduced_dw"),
"variable": genn_model.create_wu_var_ref(self.hid_to_out, "w")}
self.hid_to_out_learn= self.model.add_custom_update("hid_to_out_learn","EVPLearn", adam_optimizer_model, adam_params, self.adam_init_vars, var_refs)
self.hid_to_out.pre_target_var= "revIsyn"
self.optimisers.append(self.hid_to_out_learn)
if p["RECURRENT"]:
var_refs = {"dw": genn_model.create_wu_var_ref(self.hid_to_hid, "dw")}
self.hid_to_hid_reduce= self.model.add_custom_update("hid_to_hid_reduce","EVPReduce", EVP_grad_reduce, {}, {"reduced_dw": 0.0}, var_refs)
var_refs = {"gradient": genn_model.create_wu_var_ref(self.hid_to_hid_reduce, "reduced_dw"),
"variable": genn_model.create_wu_var_ref(self.hid_to_hid, "w")}
self.hid_to_hid_learn= self.model.add_custom_update("hid_to_hid_learn","EVPLearn", adam_optimizer_model, adam_params, self.adam_init_vars, var_refs)
self.hid_to_hid.pre_target_var= "revIsyn"
self.optimisers.append(self.hid_to_hid_learn)
# DEBUG hidden layer spike numbers
if p["DEBUG_HIDDEN_N"]:
if p["REG_TYPE"] != "Thomas1":
self.model.neuron_populations["hidden"].spike_recording_enabled= True
# enable buffered spike recording where desired
for pop in p["REC_SPIKES"]:
self.model.neuron_populations[pop].spike_recording_enabled= True
"""
----------------------------------------------------------------------------
Run the model
----------------------------------------------------------------------------
"""
def run_model(self, number_epochs, p, shuffle, X_t_orig= None, labels= None, X_t_eval= None, labels_eval= None, resfile= None):
if p["LOAD_LAST"]:
self.in_to_hid.vars["w"].view[:]= np.load(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_input_hidden_last.npy"))
self.in_to_hid.push_var_to_device("w")
self.hid_to_out.vars["w"].view[:]= np.load(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_hidden_output_last.npy"))
self.hid_to_out.push_var_to_device("w")
if p["RECURRENT"]:
self.hid_to_hid.vars["w"].view[:]= np.load(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_hidden_hidden_last.npy"))
self.hid_to_hid.push_var_to_device("w")
else:
# zero the weights of synapses to "padding output neurons" - this should remove them from influencing the backward pass
mask= np.zeros((p["NUM_HIDDEN"],self.num_output))
mask[:,self.N_class:]= 1
mask= np.array(mask, dtype= bool).flatten()
self.hid_to_out.pull_var_from_device("w")
self.hid_to_out.vars["w"].view[mask]= 0
self.hid_to_out.push_var_to_device("w")
print("connections zeroed")
# set up run
N_trial= 0
if X_t_orig is not None:
assert(labels is not None)
N_train= len(X_t_orig) // p["N_BATCH"]
N_trial+= N_train
else:
N_train= 0
if X_t_eval is not None:
assert(labels_eval is not None)
N_eval= len(X_t_eval) // p["N_BATCH"]
N_trial+= N_eval
else:
N_eval= 0
adam_step= 1
learning_rate= p["ETA"]
# set up recording if required
spike_t= {}
spike_ID= {}
for pop in p["REC_SPIKES"]:
spike_t[pop]= []
spike_ID[pop]= []
rec_spk_lbl= []
rec_spk_pred= []
rec_vars_n= {}
for pop, var in p["REC_NEURONS"]:
rec_vars_n[var+pop]= []
#rec_exp_V= []
#rec_expsum= []
rec_n_t= []
rec_n_lbl= []
rec_n_pred= []
rec_vars_s= {}
for pop, var in p["REC_SYNAPSES"]:
rec_vars_s[var+pop]= []
rec_s_t= []
rec_s_lbl= []
rec_s_pred= []
"""
for x,y in zip(lX[:10],X_t_orig[:10]):
fig,ax = plt.subplots(1,2,sharex= True, sharey= True)
ax[0].scatter(x["t"],x["x"],s=0.2)
ax[1].scatter(y["t"],y["x"],s=0.2)
plt.show()
exit(1)
"""
# build and assign the input spike train and corresponding labels
X, Y, input_start, input_end= self.generate_input_spiketimes_shuffle_fast(p, X_t_orig, labels, X_t_eval, labels_eval)
self.input.extra_global_params["spikeTimes"].view[:len(X)]= X
self.input.push_extra_global_param_to_device("spikeTimes")
self.input_set.extra_global_params["allStartSpike"].view[:len(input_start)]= input_start
self.input_set.push_extra_global_param_to_device("allStartSpike")
self.input_set.extra_global_params["allEndSpike"].view[:len(input_end)]= input_end
self.input_set.push_extra_global_param_to_device("allEndSpike")
if labels is not None:
input_id= np.arange(labels.shape[0])
else:
input_id= []
all_input_id= np.arange(Y.shape[0])
self.input_set.extra_global_params["allInputID"].view[:len(all_input_id)]= all_input_id
self.input_set.push_extra_global_param_to_device("allInputID")
for epoch in range(number_epochs):
# if we are doing augmentation, the entore spike time array needs to be set up anew.
if N_train > 0 and len(p["AUGMENTATION"]) > 0:
lX= copy.deepcopy(X_t_orig)
for aug in p["AUGMENTATION"]:
if aug == "random_shift":
lX= random_shift(lX,self.datarng, p["AUGMENTATION"][aug])
if aug == "random_dilate":
lX= random_dilate(lX,self.datarng, p["AUGMENTATION"][aug][0], p["AUGMENTATION"][aug][1])
if aug == "ID_jitter":
lX= ID_jitter(lX,self.datarng, p["AUGMENTATION"][aug])
X, Y, input_start, input_end= self.generate_input_spiketimes_shuffle_fast(p, lX, labels, X_t_eval, labels_eval)
self.input.extra_global_params["spikeTimes"].view[:len(X)]= X
self.input.push_extra_global_param_to_device("spikeTimes")
self.input_set.extra_global_params["allStartSpike"].view[:len(input_start)]= input_start
self.input_set.push_extra_global_param_to_device("allStartSpike")
self.input_set.extra_global_params["allEndSpike"].view[:len(input_end)]= input_end
self.input_set.push_extra_global_param_to_device("allEndSpike")
if N_train > 0 and shuffle:
self.datarng.shuffle(input_id)
all_input_id[:len(input_id)]= input_id
Y[:len(input_id)]= labels[input_id]
self.output.extra_global_params["label"].view[:len(Y)]= Y
self.output.push_extra_global_param_to_device("label")
self.input_set.extra_global_params["allInputID"].view[:len(all_input_id)]= all_input_id
self.input_set.push_extra_global_param_to_device("allInputID")
predict= {
"train": [],
"eval": []
}
the_loss= {
"train": [],
"eval": []
}
good= {
"train": 0.0,
"eval": 0.0
}
self.model.t= 0.0
self.model.timestep= 0
for var, val in self.input_init_vars.items():
self.input.vars[var].view[:]= val
self.input.push_state_to_device()
self.input.extra_global_params["pDrop"].view[:]= p["PDROP_INPUT"]
for var, val in self.hidden_init_vars.items():
self.hidden.vars[var].view[:]= val
self.hidden.push_state_to_device()
for var, val in self.output_init_vars.items():
self.output.vars[var].view[:]= val
self.output.push_state_to_device()
self.model.custom_update("EVPReduce") # this zeros dw (so as to ignore eval gradients from last epoch!
if p["DEBUG_HIDDEN_N"]:
all_hidden_n= []
all_sNSum= []
for trial in range(N_trial):
trial_end= (trial+1)*p["TRIAL_MS"]
# assign the input spike train and corresponding labels
if trial < N_train:
phase= "train"
else:
phase= "eval"
self.input.extra_global_params["pDrop"].view[:]= 0.0
self.input_set.extra_global_params["trial"].view[:]= trial
self.model.custom_update("inputUpdate")
self.input.extra_global_params["t_offset"].view[:]= self.model.t
int_t= 0
if p["DEBUG_HIDDEN_N"]:
if p["REG_TYPE"] != "Thomas1":
spike_N_hidden= np.zeros(p["N_BATCH"])
while (self.model.t < trial_end-1e-1*p["DT_MS"]):
self.model.step_time()
int_t += 1
# DEBUG of middle layer activity
if p["DEBUG_HIDDEN_N"]:
if int_t%p["SPK_REC_STEPS"] == 0:
if p["REG_TYPE"] != "Thomas1":
self.model.pull_recording_buffers_from_device()
x= self.model.neuron_populations["hidden"].spike_recording_data
for btch in range(p["N_BATCH"]):
spike_N_hidden[btch]+= len(x[btch][0])
if ((epoch,trial) in p["REC_SPIKES_EPOCH_TRIAL"]) and (len(p["REC_SPIKES"]) > 0):
if int_t%p["SPK_REC_STEPS"] == 0:
self.model.pull_recording_buffers_from_device()
for pop in p["REC_SPIKES"]:
the_pop= self.model.neuron_populations[pop]
x= the_pop.spike_recording_data
if p["N_BATCH"] > 1:
for i in range(p["N_BATCH"]):
spike_t[pop].append(x[i][0]+(epoch*N_trial*p["N_BATCH"]+trial*p["N_BATCH"]+i-trial)*p["TRIAL_MS"]) # subtracting trial to compensate the progression of model.t by p["TRIAL_MS"] each trial
spike_ID[pop].append(x[i][1])
else:
spike_t[pop].append(x[0]+epoch*N_trial*p["TRIAL_MS"])
spike_ID[pop].append(x[1])
if ((epoch,trial) in p["REC_NEURONS_EPOCH_TRIAL"]):
for pop, var in p["REC_NEURONS"]:
the_pop= self.model.neuron_populations[pop]
the_pop.pull_var_from_device(var)
rec_vars_n[var+pop].append(the_pop.vars[var].view.copy())
rec_n_t.append(self.model.t)
if ((epoch,trial) in p["REC_SYNAPSES_EPOCH_TRIAL"]):
for pop, var in p["REC_SYNAPSES"]:
the_pop= self.model.synapse_populations[pop]
if var == "in_syn":
the_pop.pull_in_syn_from_device()
rec_vars_s[var+pop].append(the_pop.in_syn.copy())
else:
the_pop.pull_var_from_device(var)
rec_vars_s[var+pop].append(the_pop.vars[var].view.copy())
rec_s_t.append(self.model.t)
# clamp in_syn to 0 one timestep before trial end to avoid bleeding spikes into the next trial
if np.abs(self.model.t + p["DT_MS"] - trial_end) < 1e-1*p["DT_MS"]:
self.in_to_hid.in_syn[:]= 0.0
self.in_to_hid.push_in_syn_to_device()
self.hid_to_out.in_syn[:]= 0.0
self.hid_to_out.push_in_syn_to_device()
# do not learn after the 0th trial where lambdas are meaningless
if (phase == "train") and trial > 0 and ((trial+1)%p["SUPER_BATCH"]) == 0:
update_adam(learning_rate, adam_step, self.optimisers)
adam_step += 1
self.model.custom_update("EVPReduce")
#if trial%2 == 1:
self.model.custom_update("EVPLearn")
self.in_to_hid.in_syn[:]= 0.0
self.in_to_hid.push_in_syn_to_device()
self.hid_to_out.in_syn[:]= 0.0
self.hid_to_out.push_in_syn_to_device()
if p["REG_TYPE"] == "Thomas1":
# for hidden regularistation prepare "sNSum_all"
self.hidden_reset.extra_global_params["sNSum_all"].view[:]= np.zeros(p["N_BATCH"])
self.hidden_reset.push_extra_global_param_to_device("sNSum_all")
if p["LOSS_TYPE"] == "avg_xentropy": # need to copy sum_V and loss from device before neuronReset!
self.output.pull_var_from_device("sum_V")
self.output.pull_var_from_device("loss")
self.model.custom_update("neuronReset")
if (p["REG_TYPE"] == "simple" or p["REG_TYPE"] == "Thomas1") and p["AVG_SNSUM"]:
self.model.custom_update("sNSumReduce")
self.model.custom_update("sNSumApply")
if p["REG_TYPE"] == "Thomas1":
self.hidden_reset.pull_extra_global_param_from_device("sNSum_all")
#self.hidden.extra_global_params["sNSum_all"].view[:]= np.mean(self.hidden_reset.extra_global_params["sNSum_all"].view)
self.hidden.extra_global_params["sNSum_all"].view[:]= self.hidden_reset.extra_global_params["sNSum_all"].view[:]
self.hidden.push_extra_global_param_to_device("sNSum_all")
if p["DEBUG_HIDDEN_N"]:
spike_N_hidden= self.hidden_reset.extra_global_params["sNSum_all"].view[:].copy()
# collect data for rewiring rule for silent neurons
# record training loss and error
# NOTE: the neuronReset does the calculation of expsum and updates exp_V for loss types sum and max
if p["LOSS_TYPE"] == "max" or p["LOSS_TYPE"] == "sum":
self.output.pull_var_from_device("exp_V")
#print(self.output.vars["exp_V"].view)
pred= np.argmax(self.output.vars["exp_V"].view[:,:self.N_class], axis=-1)
if p["LOSS_TYPE"] == "avg_xentropy":
pred= np.argmax(self.output.vars["sum_V"].view[:,:self.N_class], axis=-1)
lbl= Y[trial*p["N_BATCH"]:(trial+1)*p["N_BATCH"]]
if ((epoch, trial) in p["REC_SPIKES_EPOCH_TRIAL"]):
rec_spk_lbl.append(lbl.copy())
rec_spk_pred.append(pred.copy())
if ((epoch, trial) in p["REC_NEURONS_EPOCH_TRIAL"]):
rec_n_lbl.append(lbl.copy())
rec_n_pred.append(pred.copy())
if ((epoch, trial) in p["REC_SYNAPSES_EPOCH_TRIAL"]):
rec_s_lbl.append(lbl.copy())
rec_s_pred.append(pred.copy())
if p["DEBUG"]:
print(pred)
print(lbl)
print("---------------------------------------")
if p["LOSS_TYPE"] == "max" or p["LOSS_TYPE"] == "sum":
self.output.pull_var_from_device("expsum")
losses= self.loss_func(lbl,p) # uses self.output.vars["exp_V"].view and self.output.vars["expsum"].view
if p["LOSS_TYPE"] == "avg_xentropy":
losses= self.loss_func_avg_xentropy(lbl,p) # uses self.output.vars["loss"].view
if ((epoch, trial) in p["REC_NEURONS_EPOCH_TRIAL"]):
print(pred)
print(lbl)
print("---------------------------------------")
#rec_exp_V.append(self.output.vars["exp_V"].view.copy())
#rec_expsum.append(self.output.vars["expsum"].view.copy())
#print(f'{np.min(self.output.vars["expsum"].view)} {np.mean(self.output.vars["expsum"].view)} {np.max(self.output.vars["expsum"].view)}')
good[phase] += np.sum(pred == lbl)
predict[phase].append(pred)
the_loss[phase].append(losses)
if p["DEBUG_HIDDEN_N"]:
all_hidden_n.append(spike_N_hidden)
self.hidden.pull_var_from_device("sNSum")
all_sNSum.append(np.mean(self.hidden.vars['sNSum'].view.copy(),axis= 0))
if ((epoch, trial) in p["W_OUTPUT_EPOCH_TRIAL"]):
self.in_to_hid.pull_var_from_device("w")
np.save(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_input_hidden_e{}_t{}.npy".format(epoch,trial)), self.in_to_hid.vars["w"].view.copy())
self.hid_to_out.pull_var_from_device("w")
np.save(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_hidden_output_e{}_t{}.npy".format(epoch,trial)), self.hid_to_out.vars["w"].view.copy())
if p["RECURRENT"]:
self.hid_to_hid.pull_var_from_device("w")
np.save(os.path.join(p["OUT_DIR"], p["NAME"]+"_w_hidden_hidden_e{}_t{}.npy".format(epoch,trial)), self.hid_to_hid.vars["w"].view.copy())
if N_train > 0:
correct= good["train"]/(N_train*p["N_BATCH"])
else:
correct= 0
if N_eval > 0:
correct_eval= good["eval"]/(N_eval*p["N_BATCH"])
else:
correct_eval= 0
if p["DEBUG_HIDDEN_N"]:
all_hidden_n= np.hstack(all_hidden_n)
all_sNSum= np.hstack(all_sNSum)
print("Hidden spikes in model per trial: {} +/- {}, min {}, max {}".format(np.mean(all_hidden_n),np.std(all_hidden_n),np.amin(all_hidden_n),np.amax(all_hidden_n)))
print("Hidden spikes per trial per neuron across batches: {} +/- {}, min {}, max {}".format(np.mean(all_sNSum),np.std(all_sNSum),np.amin(all_sNSum),np.amax(all_sNSum)))
print("{} Training Correct: {}, Training Loss: {}, Evaluation Correct: {}, Evaluation Loss: {}".format(epoch, correct, np.mean(the_loss["train"]), correct_eval, np.mean(the_loss["eval"])))
if resfile is not None:
resfile.write("{} {} {} {} {}".format(epoch, correct, np.mean(the_loss["train"]), correct_eval, np.mean(the_loss["eval"])))
if p["DEBUG_HIDDEN_N"]:
resfile.write(" {} {} {} {}".format(np.mean(all_hidden_n),np.std(all_hidden_n),np.amin(all_hidden_n),np.amax(all_hidden_n)))
resfile.write(" {} {} {} {}".format(np.mean(all_sNSum),np.std(all_sNSum),np.amin(all_sNSum),np.amax(all_sNSum)))
resfile.write("\n")
resfile.flush()
predict[phase]= np.hstack(predict[phase])
learning_rate *= p["ETA_DECAY"]
if p["ETA_FIDDELING"]:
if (epoch+1) % p["ETA_REDUCE_PERIOD"] == 0:
learning_rate *= p["ETA_REDUCE"]
adam_step= 1
for pop in p["REC_SPIKES"]:
spike_t[pop]= np.hstack(spike_t[pop])
spike_ID[pop]= np.hstack(spike_ID[pop])
for rec_t, rec_var, rec_list in [ (rec_n_t, rec_vars_n, p["REC_NEURONS"]), (rec_s_t, rec_vars_s, p["REC_SYNAPSES"])]:
for pop, var in rec_list:
rec_var[var+pop]= np.array(rec_var[var+pop])
print(rec_var[var+pop].shape)
rec_t= np.array(rec_t)
#rec_exp_V= np.array(rec_exp_V)
#rec_expsum= np.array(rec_expsum)
rec_spk_lbl= np.array(rec_spk_lbl)