/
lvis.go
3449 lines (3103 loc) · 123 KB
/
lvis.go
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
// Copyright (c) 2021, The Emergent Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
/*
lvis explores how a hierarchy of areas in the ventral stream of visual
processing (up to inferotemporal (IT) cortex) can produce robust object
recognition that is invariant to changes in position, size, etc of retinal
input images.
*/
package main
import (
"flag"
"fmt"
"log"
"math/rand"
"os"
"strconv"
"strings"
"time"
"github.com/emer/axon/axon"
"github.com/emer/emergent/actrf"
"github.com/emer/emergent/confusion"
"github.com/emer/emergent/decoder"
"github.com/emer/emergent/emer"
"github.com/emer/emergent/env"
"github.com/emer/emergent/netview"
"github.com/emer/emergent/params"
"github.com/emer/emergent/prjn"
"github.com/emer/emergent/relpos"
"github.com/emer/empi/empi"
"github.com/emer/empi/mpi"
"github.com/emer/etable/agg"
"github.com/emer/etable/convolve"
"github.com/emer/etable/eplot"
"github.com/emer/etable/etable"
"github.com/emer/etable/etensor"
"github.com/emer/etable/etview" // include to get gui views
"github.com/emer/etable/metric"
"github.com/emer/etable/norm"
"github.com/emer/etable/pca"
"github.com/emer/etable/split"
"github.com/goki/gi/gi"
"github.com/goki/gi/gimain"
"github.com/goki/gi/giv"
"github.com/goki/ki/ki"
"github.com/goki/ki/kit"
"github.com/goki/mat32"
)
func main() {
TheSim.New()
if len(os.Args) > 1 {
TheSim.CmdArgs() // simple assumption is that any args = no gui -- could add explicit arg if you want
} else {
gimain.Main(func() { // this starts gui -- requires valid OpenGL display connection (e.g., X11)
guirun()
})
}
}
func guirun() {
TheSim.Config()
TheSim.Init()
win := TheSim.ConfigGui()
win.StartEventLoop()
}
// LogPrec is precision for saving float values in logs
const LogPrec = 4
// ParamSets is the default set of parameters -- Base is always applied, and others can be optionally
// selected to apply on top of that
var ParamSets = params.Sets{
{Name: "Base", Desc: "these are the best params", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: "Layer", Desc: "needs some special inhibition and learning params",
Params: params.Params{
"Layer.Inhib.Inhib.AvgTau": "30", // 30 > 20 >> 1 definitively
"Layer.Inhib.Inhib.GiSynThr": "0.0", // 0.01 shows effects
"Layer.Act.Dt.IntTau": "40", // 40 > 20
"Layer.Inhib.Layer.Gi": "1.1", // 1.1 > 1.0 > 1.2 -- all layers
"Layer.Inhib.Pool.Gi": "1.1", // 1.1 > 1.0 -- universal for all layers
"Layer.Inhib.Pool.FFEx0": "0.15", // .15 > .18; Ex .05 -- .2/.1, .2/.2, .3/.5 all blow up
"Layer.Inhib.Pool.FFEx": "0.05", // .05 best so far
"Layer.Inhib.Layer.FFEx0": "0.15",
"Layer.Inhib.Layer.FFEx": "0.05", // .05 best so far
"Layer.Act.Gbar.L": "0.2", // 0.2 orig > 0.1 new def
"Layer.Act.Decay.Act": "0.2", // 0.2 > 0 > 0.5 w/ glong.7 459
"Layer.Act.Decay.Glong": "0.6", // 0.6 > 0.7 > 0.8
"Layer.Act.KNa.Fast.Max": "0.1", // fm both .2 worse
"Layer.Act.KNa.Med.Max": "0.2", // 0.2 > 0.1 def
"Layer.Act.KNa.Slow.Max": "0.2", // 0.2 > higher
"Layer.Act.Noise.Dist": "Gaussian",
"Layer.Act.Noise.Mean": "0.0", // .05 max for blowup
"Layer.Act.Noise.Var": "0.01", // .01 a bit worse
"Layer.Act.Noise.Type": "NoNoise", // off for now
"Layer.Act.GTarg.GeMax": "1.2", // 1 > .8 -- rescaling not very useful.
"Layer.Act.Dt.LongAvgTau": "20", // 50 > 20 in terms of stability, but weird effect late
"Layer.Learn.ActAvg.MinLrn": "0.02", // sig improves "top5" hogging in pca strength
"Layer.Learn.ActAvg.SSTau": "40",
"Layer.Inhib.ActAvg.AdaptRate": "0.5", // 0.5 default for layers, except output
"Layer.Learn.TrgAvgAct.ErrLrate": "0.01", // 0.01 orig > 0.005
"Layer.Learn.TrgAvgAct.SynScaleRate": "0.005", // 0.005 orig > 0.01
"Layer.Learn.TrgAvgAct.TrgRange.Min": "0.5", // .5 > .2 overall
"Layer.Learn.TrgAvgAct.TrgRange.Max": "2.0", // objrec 2 > 1.8
"Layer.Learn.RLrate.On": "true", // true = essential -- prevents over rep of
"Layer.Learn.RLrate.ActThr": "0.1", // 0.1 > 0.15 > 0.05 > 0.2
"Layer.Learn.RLrate.ActDifThr": "0.02", // 0.02 > 0.05 in other models
"Layer.Learn.RLrate.Min": "0.001", // .001 best, adifthr.05
}},
{Sel: ".Input", Desc: "all V1 input layers",
Params: params.Params{
"Layer.Inhib.Pool.On": "true",
"Layer.Inhib.Layer.Gi": "1.1", // 0.9 > 1.1 def -- more activity
"Layer.Inhib.Pool.Gi": "1.1", // 0.9 > 1.1 def -- more activity
"Layer.Inhib.ActAvg.Init": "0.06", // .06 for !SepColor actuals: V1m8: .04, V1m16: .03
"Layer.Act.Clamp.Type": "GeClamp", // GeClamp better
"Layer.Act.Clamp.Ge": "1.0", // 1.0 > .6 -- more activity
"Layer.Act.Decay.Act": "1", // these make no diff
"Layer.Act.Decay.Glong": "1",
}},
{Sel: ".V2", Desc: "pool inhib, sparse activity",
Params: params.Params{
"Layer.Inhib.Topo.On": "false",
"Layer.Inhib.Topo.Width": "4",
"Layer.Inhib.Topo.Sigma": "1.0",
"Layer.Inhib.Topo.Gi": "0.002", // 0.002 best -- reduces Top5, keeps NStrong
"Layer.Inhib.Topo.FF0": "0.2", // 0.2 best -- test more
}},
{Sel: ".V2m", Desc: "pool inhib, sparse activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.FB": "1", // 0 possibly causes blowup at some point, no bene
"Layer.Inhib.ActAvg.Init": "0.02",
}},
{Sel: ".V2l", Desc: "pool inhib, sparse activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.FB": "1",
"Layer.Inhib.ActAvg.Init": "0.02",
"Layer.Inhib.Topo.Width": "2", // smaller
}},
{Sel: "#V2l16", Desc: "this layer is too active, drives V4f16 too strongly",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.02", // not clear if needed now..
}},
{Sel: ".V2h", Desc: "pool inhib, sparse activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.FB": "1", // 0 possibly causes blowup at some point, no bene
"Layer.Inhib.ActAvg.Init": "0.02",
}},
{Sel: ".V3h", Desc: "pool inhib, sparse activity -- only for h16",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.FB": "1", // 0 possibly causes blowup at some point, no bene
"Layer.Inhib.ActAvg.Init": "0.02", // .02 > .04
"Layer.Act.GTarg.GeMax": "1.2", // these need to get stronger?
}},
{Sel: ".V4", Desc: "pool inhib, sparse activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.FB": "1", // 1 >= 0 in lba
"Layer.Inhib.ActAvg.Init": "0.04", // .04 >= .03 > .05
"Layer.Inhib.Layer.Gi": "1.1", // was 1.1
"Layer.Inhib.Pool.Gi": "1.1", // was 1.1
"Layer.Inhib.Topo.On": "false",
"Layer.Inhib.Topo.Width": "4", // was 4
"Layer.Inhib.Topo.Sigma": "1.0",
"Layer.Inhib.Topo.Gi": "0.002", // 0.002 best -- reduces Top5, keeps NStrong
"Layer.Inhib.Topo.FF0": "0.2", // 0.2 best -- test more
}},
{Sel: ".TEO", Desc: "initial activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.On": "false", // no layer!
"Layer.Inhib.ActAvg.Init": "0.06", // .06 > .05 = .04
"Layer.Inhib.Pool.Gi": "1.1", // was 1.1
}},
{Sel: "#TE", Desc: "initial activity",
Params: params.Params{
"Layer.Inhib.Pool.On": "true", // needs pool-level
"Layer.Inhib.Layer.On": "false", // no layer!
"Layer.Inhib.ActAvg.Init": "0.06", // .03 actual with gi 1.2, was .06
"Layer.Inhib.Pool.Gi": "1.1", // was 1.1
}},
{Sel: "#Output", Desc: "general output, Localist default -- see RndOutPats, LocalOutPats",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.0", // old: 1.3 for 1/100
"Layer.Inhib.ActAvg.Init": "0.1", // .005 > .008 > .01 -- prevents loss of Ge over time..
"Layer.Inhib.ActAvg.Targ": "0.1", // .01 -- .005, .008 too low -- maybe not nec?
"Layer.Inhib.ActAvg.AdaptGi": "false", // true: it is essential -- too hard to balance manually
"Layer.Inhib.ActAvg.LoTol": "0.5",
"Layer.Inhib.ActAvg.AdaptRate": "0.02", // 0.01 >= 0.02 best in range 0.01..0.1
// "Layer.Act.Decay.Act": "0.5", // 0.5 makes no diff
// "Layer.Act.Decay.Glong": "1", // 1 makes no diff
"Layer.Act.Clamp.Type": "GeClamp",
"Layer.Act.Clamp.Ge": "0.6", // .6 = .7 > .5 (tiny diff) -- input has 1.0 now
"Layer.Act.Clamp.Burst": "false",
"Layer.Act.Clamp.BurstThr": "0.5", //
"Layer.Act.Clamp.BurstGe": "2", // 2, 20cyc with tr 2 or 3, ge .6 all about same
"Layer.Act.Clamp.BurstCyc": "20", // 20 > 15 > 10 -- maybe refractory?
"Layer.Act.Spike.Tr": "3", // 2 >= 3 > 1 > 0
"Layer.Act.GABAB.Gbar": "0.005", // .005 > .01 > .02 > .05 > .1 > .2
"Layer.Act.NMDA.Gbar": "0.03", // was .02
"Layer.Learn.RLrate.On": "true", // todo: try false
"Layer.Inhib.Pool.FFEx": "0.0", // no
"Layer.Inhib.Layer.FFEx": "0.0", //
}},
{Sel: "#Claustrum", Desc: "testing -- not working",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "0.8",
"Layer.Inhib.Pool.On": "false", // needs pool-level
"Layer.Inhib.Layer.On": "true",
"Layer.Inhib.ActAvg.Init": ".06",
}},
///////////////////////////////
// projections
{Sel: "Prjn", Desc: "exploring",
Params: params.Params{
"Prjn.PrjnScale.ScaleLrate": "2", // 2 = fast response, effective
"Prjn.PrjnScale.LoTol": "0.8", // good now...
"Prjn.PrjnScale.AvgTau": "500", // slower default
"Prjn.PrjnScale.Adapt": "false", // no adapt better?
"Prjn.SWt.Adapt.On": "true", // true > false, esp in cosdiff
"Prjn.SWt.Adapt.Lrate": "0.0002", // .0002, .001 > .01 > .1 after 250epc in NStrong
"Prjn.SWt.Adapt.SigGain": "6",
"Prjn.SWt.Adapt.DreamVar": "0.02", // 0.02 good overall, no ToOut
"Prjn.SWt.Init.SPct": "1", // 1 > lower
"Prjn.SWt.Init.Mean": "0.5", // .5 > .4 -- key, except v2?
"Prjn.SWt.Limit.Min": "0.2", // .2-.8 == .1-.9; .3-.7 not better -- 0-1 minor worse
"Prjn.SWt.Limit.Max": "0.8", //
"Prjn.Learn.Lrate.Base": "0.04", // 0.01 > 0.015 > 0.02 459 -- .04 for RLrate
"Prjn.Learn.XCal.SubMean": "1", // testing..
"Prjn.Learn.XCal.DWtThr": "0.0001", // 0.0001 > 0.001
"Prjn.Com.PFail": "0.0",
}},
{Sel: ".Back", Desc: "top-down back-projections MUST have lower relative weight scale, otherwise network hallucinates -- smaller as network gets bigger",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2",
// "Prjn.Learn.Lrate.Base": "0",
}},
{Sel: ".Forward", Desc: "use pfail only on forward cons?",
Params: params.Params{
// .2 max 1 = no diff, .5 max .8 = no diff
"Prjn.Com.PFail": "0.0", // 0 > .05 > .1 > .2
// "Prjn.SWt.Adapt.DreamVar": "0.02", // 0.01 big pca effects, no perf bene; 0.05 impairs perf
}},
{Sel: ".ToOut", Desc: "to output -- some things should be different..",
Params: params.Params{
"Prjn.Com.PFail": "0.0",
"Prjn.SWt.Adapt.DreamVar": "0.0", // nope
"Prjn.SWt.Adapt.On": "false", // off > on
"Prjn.SWt.Init.SPct": "0", // when off, 0
"Prjn.PrjnScale.LoTol": "0.5", // .5 > .8 -- needs extra kick at start!
"Prjn.PrjnScale.Adapt": "true", // was essential here
"Prjn.Learn.XCal.SubMean": "0",
}},
{Sel: ".FmOut", Desc: "from output -- some things should be different..",
Params: params.Params{
"Prjn.Learn.XCal.SubMean": "1",
}},
{Sel: ".Inhib", Desc: "inhibitory projection",
Params: params.Params{
"Prjn.Learn.Learn": "true", // learned decorrel is good
"Prjn.Learn.Lrate.Base": "0.0001", // .0001 > .001 -- slower better!
"Prjn.SWt.Init.Var": "0.0",
"Prjn.SWt.Init.Mean": "0.1",
"Prjn.SWt.Init.Sym": "false",
"Prjn.SWt.Adapt.On": "false",
"Prjn.PrjnScale.Abs": "0.2", // .2 > .1 for controlling PCA; .3 or.4 with GiSynThr .01
"Prjn.PrjnScale.Adapt": "false",
"Prjn.IncGain": "1", // .5 def
}},
{Sel: ".V1V2", Desc: "special SWt params",
Params: params.Params{
"Prjn.SWt.Init.Mean": "0.4", // .4 here is key!
"Prjn.SWt.Limit.Min": "0.1", // .1-.7
"Prjn.SWt.Limit.Max": "0.7", //
"Prjn.PrjnScale.Abs": "1.4", // 1.4 > 2.0 for color -- extra boost to get more v2 early on
}},
{Sel: ".V1V2fmSm", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2",
}},
{Sel: ".V2V4", Desc: "extra boost",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.0", // 1.0 prev, 1.2 not better
"Prjn.SWt.Init.Mean": "0.4", // .4 a tiny bit better overall
"Prjn.SWt.Limit.Min": "0.1", // .1-.7 def
"Prjn.SWt.Limit.Max": "0.7", //
}},
{Sel: ".V2V4sm", Desc: "extra boost",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.0", // 1.0 prev, 1.2 not better
}},
{Sel: "#V2m16ToV4f16", Desc: "weights into V416 getting too high",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.0", // was 0.8, but as of #680 1.0 better
}},
{Sel: "#V2l16ToV4f16", Desc: "weights into V416 getting too high",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.0", // see above
}},
{Sel: ".V4TEO", Desc: "stronger",
Params: params.Params{
// "Prjn.PrjnScale.Abs": "1.2", // trying bigger -- was low
}},
{Sel: ".V4TEOoth", Desc: "weaker rel",
Params: params.Params{
// "Prjn.PrjnScale.Abs": "1.2", // trying bigger -- was low
"Prjn.PrjnScale.Rel": "0.5",
}},
{Sel: ".V4Out", Desc: "NOT weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "1", // 1 > 0.5 > .2 -- v53 still
}},
{Sel: ".TEOTE", Desc: "too weak at start",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1", // 1.2 not better
}},
// back projections
{Sel: ".V4V2", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.05", // .05 > .02 > .1 v70
"Prjn.SWt.Init.Mean": "0.4", // .4 matches V2V4 -- not that big a diff on its own
"Prjn.SWt.Limit.Min": "0.1", // .1-.7 def
"Prjn.SWt.Limit.Max": "0.7", //
}},
{Sel: ".TEOV2", Desc: "weaker -- not used",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.05", // .05 > .02 > .1
}},
{Sel: ".TEOV4", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // .1 == .2
}},
{Sel: ".TETEO", Desc: "std",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // .1 orig
}},
{Sel: ".OutTEO", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.3", // .3 > .2 v53 in long run
}},
{Sel: ".OutV4", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // .1 > .2 v53
}},
{Sel: "#OutputToTE", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.3", // 0.3 > .2 v53 in long run
}},
{Sel: "#TEToOutput", Desc: "weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "1.0", // turn off for TE testing
}},
// shortcuts -- .5 > .2 (v32 still) -- all tested together
{Sel: "#V1l16ToClaustrum", Desc: "random fixed -- not useful",
Params: params.Params{
"Prjn.Learn.Learn": "false",
"Prjn.PrjnScale.Rel": "0.5", // .5 > .8 > 1 > .4 > .3 etc
"Prjn.SWt.Adapt.On": "false", // seems better
}},
{Sel: ".V1SC", Desc: "v1 shortcut",
Params: params.Params{
"Prjn.Learn.Lrate.Base": "0.001", //
// "Prjn.Learn.Learn": "false",
"Prjn.PrjnScale.Rel": "0.5", // .5 > .8 > 1 > .4 > .3 etc
"Prjn.SWt.Adapt.On": "false", // seems better
// "Prjn.SWt.Init.Var": "0.05",
}},
},
}},
{Name: "RndOutPats", Desc: "random output pattern", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: "#Output", Desc: "high inhib for one-hot output",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "0.9", // 0.9 > 1.0
"Layer.Inhib.ActAvg.Init": "0.1", // 0.1 seems good
}},
},
}},
{Name: "LocalOutPats", Desc: "localist output pattern", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: "#Output", Desc: "high inhib for one-hot output",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.5", // 1.5 = 1.6 > 1.4
"Layer.Inhib.ActAvg.Init": "0.01",
}},
},
}},
{Name: "ToOutTol", Desc: "delayed enforcement of low tolerance on .ToOut", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: ".ToOut", Desc: "to output -- some things should be different..",
Params: params.Params{
"Prjn.PrjnScale.LoTol": "0.5", // activation dropping off a cliff there at the end..
}},
},
}},
{Name: "OutAdapt", Desc: "delayed enforcement of output adaptation", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: "#Output", Desc: "general output, Localist default -- see RndOutPats, LocalOutPats",
Params: params.Params{
"Layer.Inhib.ActAvg.AdaptGi": "true", // true = definitely worse
}},
},
}},
}
// Sim encapsulates the entire simulation model, and we define all the
// functionality as methods on this struct. This structure keeps all relevant
// state information organized and available without having to pass everything around
// as arguments to methods, and provides the core GUI interface (note the view tags
// for the fields which provide hints to how things should be displayed).
type Sim struct {
Net *axon.Network `view:"no-inline" desc:"the network -- click to view / edit parameters for layers, prjns, etc"`
TrnTrlLog *etable.Table `view:"no-inline" desc:"training trial-level log data"`
TrnCycLog *etable.Table `view:"no-inline" desc:"training cycle-level log data"`
TrnTrlLogAll *etable.Table `view:"no-inline" desc:"all training trial-level log data (aggregated from MPI)"`
TrnTrlRepLog *etable.Table `view:"no-inline" desc:"training trial-level reps log data"`
TrnTrlRepLogAll *etable.Table `view:"no-inline" desc:"training trial-level reps log data"`
TrnEpcLog *etable.Table `view:"no-inline" desc:"training epoch-level log data"`
TstEpcLog *etable.Table `view:"no-inline" desc:"testing epoch-level log data"`
TstTrlLog *etable.Table `view:"no-inline" desc:"testing trial-level log data"`
TstTrlLogAll *etable.Table `view:"no-inline" desc:"all testing trial-level log data (aggregated from MPI)"`
TrnErrStats *etable.Table `view:"no-inline" desc:"training error stats"`
ActRFs actrf.RFs `view:"no-inline" desc:"activation-based receptive fields"`
RunLog *etable.Table `view:"no-inline" desc:"summary log of each run"`
RunStats *etable.Table `view:"no-inline" desc:"aggregate stats on all runs"`
Confusion confusion.Matrix `view:"no-inline" desc:"confusion matrix"`
ConfusionEpc int `desc:"epoch to start recording confusion matrix"`
MinusCycles int `desc:"number of minus-phase cycles"`
PlusCycles int `desc:"number of plus-phase cycles"`
NTests int `desc:"number of repeated tests of same input"`
SubPools bool `desc:"if true, organize layers and connectivity with 2x2 sub-pools within each topological pool"`
RndOutPats bool `desc:"if true, use random output patterns -- else localist"`
PostCycs int `desc:"number of cycles to run after main alphacyc cycles, between stimuli"`
PostDecay float32 `desc:"decay to apply at start of PostCycs"`
ErrLrMod axon.LrateMod `view:"inline" desc:"learning rate modulation as function of error"`
Params params.Sets `view:"no-inline" desc:"full collection of param sets"`
ParamSet string `desc:"which set of *additional* parameters to use -- always applies Base and optionaly this next if set -- can use multiple names separated by spaces (don't put spaces in ParamSet names!)"`
Tag string `desc:"extra tag string to add to any file names output from sim (e.g., weights files, log files, params for run)"`
Prjn4x4Skp2 *prjn.PoolTile `view:"-" desc:"Standard feedforward topographic projection, recv = 1/2 send size"`
Prjn4x4Skp2Recip *prjn.PoolTile `view:"-" desc:"Reciprocal"`
Prjn4x4Skp2Sub2 *prjn.PoolTileSub `view:"-" desc:"Standard feedforward topographic projection, recv = 1/2 send size"`
Prjn4x4Skp2Sub2Recip *prjn.PoolTileSub `view:"-" desc:"Reciprocal"`
Prjn4x4Skp2Sub2Send *prjn.PoolTileSub `view:"-" desc:"Standard feedforward topographic projection, recv = 1/2 send size"`
Prjn4x4Skp2Sub2SendRecip *prjn.PoolTileSub `view:"-" desc:"Standard feedforward topographic projection, recv = 1/2 send size"`
Prjn2x2Skp1 *prjn.PoolTile `view:"-" desc:"same-size prjn"`
Prjn2x2Skp1Recip *prjn.PoolTile `view:"-" desc:"same-size prjn reciprocal"`
Prjn2x2Skp1Sub2 *prjn.PoolTileSub `view:"-" desc:"same-size prjn"`
Prjn2x2Skp1Sub2Recip *prjn.PoolTileSub `view:"-" desc:"same-size prjn reciprocal"`
Prjn2x2Skp1Sub2Send *prjn.PoolTileSub `view:"-" desc:"same-size prjn"`
Prjn2x2Skp1Sub2SendRecip *prjn.PoolTileSub `view:"-" desc:"same-size prjn reciprocal"`
Prjn2x2Skp2 *prjn.PoolTileSub `view:"-" desc:"lateral inhib projection"`
Prjn4x4Skp0 *prjn.PoolTile `view:"-" desc:"for V4 <-> TEO"`
Prjn4x4Skp0Recip *prjn.PoolTile `view:"-" desc:"for V4 <-> TEO"`
Prjn4x4Skp0Sub2 *prjn.PoolTileSub `view:"-" desc:"for V4 <-> TEO"`
Prjn4x4Skp0Sub2Recip *prjn.PoolTileSub `view:"-" desc:"for V4 <-> TEO"`
Prjn1x1Skp0 *prjn.PoolTile `view:"-" desc:"for TE <-> TEO"`
Prjn1x1Skp0Recip *prjn.PoolTile `view:"-" desc:"for TE <-> TEO"`
Prjn6x6Skp2Lat *prjn.PoolTileSub `view:"-" desc:"lateral inhibitory connectivity for subpools"`
StartRun int `desc:"starting run number -- typically 0 but can be set in command args for parallel runs on a cluster"`
MaxRuns int `desc:"maximum number of model runs to perform"`
MaxEpcs int `desc:"maximum number of epochs to run per model run"`
MaxTrls int `desc:"maximum number of training trials per epoch"`
RepsInterval int `desc:"how often to analyze the representations"`
NZeroStop int `desc:"if a positive number, training will stop after this many epochs with zero Err"`
TrainEnv ImagesEnv `desc:"Training environment"`
TestEnv ImagesEnv `desc:"Testing environment"`
Decoder decoder.SoftMax `desc:"decoder for better output"`
Time axon.Time `desc:"axon timing parameters and state"`
TestInterval int `desc:"how often to run through the test patterns, in terms of training epochs -- can use 0 or -1 for no testing"`
ViewOn bool `desc:"whether to update the network view while running"`
TrainUpdt axon.TimeScales `desc:"at what time scale to update the display during training? Anything longer than Epoch updates at Epoch in this model"`
TestUpdt axon.TimeScales `desc:"at what time scale to update the display during testing? Anything longer than Epoch updates at Epoch in this model"`
InLays []string `view:"-" desc:"input layers -- for stats"`
OutLays []string `view:"-" desc:"output layers -- for stats"`
HidLays []string `view:"-" desc:"hidden layers -- for all main stats"`
ActRFNms []string `desc:"names of layers to compute activation rfields on"`
SpikeRastNms []string `view:"-" desc:"spike raster layers"`
SpikeRasters map[string]*etensor.Float32 `desc:"spike raster data for different layers"`
SpikeRastGrids map[string]*etview.TensorGrid `desc:"spike raster plots for different layers"`
// statistics: note use float64 as that is best for etable.Table
TrlCatIdx int `inactive:"+" desc:"current category index"`
TrlCat string `inactive:"+" desc:"current category name"`
TrlRespIdx int `inactive:"+" desc:"output response index of the network"`
TrlResp string `inactive:"+" desc:"output response name of the network"`
TrlErr float64 `inactive:"+" desc:"1 if trial was error, 0 if correct -- based on max out unit"`
TrlErr2 float64 `inactive:"+" desc:"1 if trial was error, 0 if correct -- based on second max out"`
TrlUnitErr float64 `inactive:"+" desc:"current trial's average sum squared error"`
TrlTrgAct float64 `inactive:"+" desc:"activity of target output unit on this trial"`
TrlCosDiff float64 `inactive:"+" desc:"current trial's cosine difference"`
TrlOutFirstCyc int `inactive:"+" desc:"first activation cycle of output"`
TrlFirstResp int `inactive:"+" desc:"response at OutFirstCyc"`
TrlFirstErr float64 `inactive:"+" desc:"error on first output response"`
TrlFirstErr2 float64 `inactive:"+" desc:"error2 on first output response"`
TrlDecRespIdx int `inactive:"+" desc:"output response index of the network"`
TrlDecErr float64 `inactive:"+" desc:"1 if trial was error, 0 if correct -- based on decoder max"`
TrlDecErr2 float64 `inactive:"+" desc:"1 if trial was error, 0 if correct -- based on decoder second max"`
EpcUnitErr float64 `inactive:"+" desc:"last epoch's average sum squared error (average over trials, and over units within layer)"`
EpcPctErr float64 `inactive:"+" desc:"last epoch's average TrlErr"`
EpcPctCor float64 `inactive:"+" desc:"1 - last epoch's average TrlErr"`
EpcPctErr2 float64 `inactive:"+" desc:"last epoch's average TrlErr2"`
EpcPctDecErr float64 `inactive:"+" desc:"last epoch's average TrlDecErr"`
EpcPctDecErr2 float64 `inactive:"+" desc:"last epoch's average TrlDecErr2"`
EpcCosDiff float64 `inactive:"+" desc:"last epoch's average cosine difference for output layer (a normalized error measure, maximum of 1 when the minus phase exactly matches the plus)"`
EpcErrTrgAct float64 `inactive:"+" desc:"avg activity of target output unit on err trials"`
EpcCorTrgAct float64 `inactive:"+" desc:"avg activity of target output unit on correct trials"`
EpcPerTrlMSec float64 `inactive:"+" desc:"how long did the epoch take per trial in wall-clock milliseconds"`
FirstZero int `inactive:"+" desc:"epoch at when Err first went to zero"`
NZero int `inactive:"+" desc:"number of epochs in a row with zero Err"`
PCA pca.PCA `view:"-" desc:"pca obj"`
GaussKernel []float64 `view:"-" desc:"gaussian smoothing kernel"`
SmoothData []float64 `view:"-" desc:"data for smoothing"`
// internal state - view:"-"
Win *gi.Window `view:"-" desc:"main GUI window"`
NetView *netview.NetView `view:"-" desc:"the network viewer"`
ToolBar *gi.ToolBar `view:"-" desc:"the master toolbar"`
CurImgGrid *etview.TensorGrid `view:"-" desc:"the current image grid view"`
ActRFGrids map[string]*etview.TensorGrid `view:"-" desc:"the act rf grid views"`
TrnTrlPlot *eplot.Plot2D `view:"-" desc:"the training trial plot"`
TrnCycPlot *eplot.Plot2D `view:"-" desc:"the training cycle plot"`
TrnEpcPlot *eplot.Plot2D `view:"-" desc:"the training epoch plot"`
TstEpcPlot *eplot.Plot2D `view:"-" desc:"the testing epoch plot"`
TstTrlPlot *eplot.Plot2D `view:"-" desc:"the test-trial plot"`
RunPlot *eplot.Plot2D `view:"-" desc:"the run plot"`
TrnEpcFile *os.File `view:"-" desc:"log file"`
TrnTrlFile *os.File `view:"-" desc:"log file"`
TstEpcFile *os.File `view:"-" desc:"log file"`
TstTrlFile *os.File `view:"-" desc:"log file"`
RunFile *os.File `view:"-" desc:"log file"`
ValsTsrs map[string]*etensor.Float32 `view:"-" desc:"for holding layer values"`
SaveWts bool `view:"-" desc:"for command-line run only, auto-save final weights after each run"`
NoGui bool `view:"-" desc:"if true, runing in no GUI mode"`
LogSetParams bool `view:"-" desc:"if true, print message for all params that are set"`
IsRunning bool `view:"-" desc:"true if sim is running"`
StopNow bool `view:"-" desc:"flag to stop running"`
NeedsNewRun bool `view:"-" desc:"flag to initialize NewRun if last one finished"`
RndSeeds []int64 `view:"-" desc:"the current random seeds to use for each run"`
LastEpcTime time.Time `view:"-" desc:"timer for last epoch"`
UseMPI bool `view:"-" desc:"if true, use MPI to distribute computation across nodes"`
SaveProcLog bool `view:"-" desc:"if true, save logs per processor"`
Comm *mpi.Comm `view:"-" desc:"mpi communicator"`
AllDWts []float32 `view:"-" desc:"buffer of all dwt weight changes -- for mpi sharing"`
SumDWts []float32 `view:"-" desc:"buffer of MPI summed dwt weight changes"`
}
// this registers this Sim Type and gives it properties that e.g.,
// prompt for filename for save methods.
var KiT_Sim = kit.Types.AddType(&Sim{}, SimProps)
// TheSim is the overall state for this simulation
var TheSim Sim
// New creates new blank elements and initializes defaults
func (ss *Sim) New() {
ss.Net = &axon.Network{}
ss.TrnTrlLog = &etable.Table{}
ss.TrnCycLog = &etable.Table{}
ss.TrnTrlLogAll = &etable.Table{}
ss.TrnTrlRepLog = &etable.Table{}
ss.TrnTrlRepLogAll = &etable.Table{}
ss.TrnEpcLog = &etable.Table{}
ss.TstEpcLog = &etable.Table{}
ss.TstTrlLog = &etable.Table{}
ss.TstTrlLogAll = &etable.Table{}
ss.RunLog = &etable.Table{}
ss.RunStats = &etable.Table{}
ss.Params = ParamSets
ss.TestInterval = 20 // maybe causing issues?
ss.Time.Defaults()
ss.MinusCycles = 180
ss.PlusCycles = 50
ss.NTests = 10
ss.RepsInterval = 10
ss.SubPools = true // true
ss.RndOutPats = false // change here
if ss.RndOutPats {
ss.ParamSet = "RndOutPats"
}
ss.PostCycs = 0
ss.PostDecay = .5
ss.ErrLrMod.Defaults()
ss.ErrLrMod.Base = 0.05 // 0.05 > .02? > .1 in long-term output layer health
ss.ErrLrMod.Range.Set(0.2, 0.8)
ss.Prjn4x4Skp2 = prjn.NewPoolTile()
ss.Prjn4x4Skp2.Size.Set(4, 4)
ss.Prjn4x4Skp2.Skip.Set(2, 2)
ss.Prjn4x4Skp2.Start.Set(-1, -1)
ss.Prjn4x4Skp2.TopoRange.Min = 0.8
ss.Prjn4x4Skp2Recip = prjn.NewPoolTileRecip(ss.Prjn4x4Skp2)
ss.Prjn4x4Skp2Sub2 = prjn.NewPoolTileSub()
ss.Prjn4x4Skp2Sub2.Size.Set(4, 4)
ss.Prjn4x4Skp2Sub2.Skip.Set(2, 2)
ss.Prjn4x4Skp2Sub2.Start.Set(-1, -1)
ss.Prjn4x4Skp2Sub2.Subs.Set(2, 2)
ss.Prjn4x4Skp2Sub2.TopoRange.Min = 0.8
ss.Prjn4x4Skp2Sub2Recip = prjn.NewPoolTileSubRecip(ss.Prjn4x4Skp2Sub2)
ss.Prjn4x4Skp2Sub2Send = prjn.NewPoolTileSub()
*ss.Prjn4x4Skp2Sub2Send = *ss.Prjn4x4Skp2Sub2
ss.Prjn4x4Skp2Sub2Send.SendSubs = true
ss.Prjn4x4Skp2Sub2SendRecip = prjn.NewPoolTileSubRecip(ss.Prjn4x4Skp2Sub2Send)
ss.Prjn2x2Skp1 = prjn.NewPoolTile()
ss.Prjn2x2Skp1.Size.Set(2, 2)
ss.Prjn2x2Skp1.Skip.Set(1, 1)
ss.Prjn2x2Skp1.Start.Set(0, 0)
ss.Prjn2x2Skp1.TopoRange.Min = 0.8
ss.Prjn2x2Skp1Recip = prjn.NewPoolTileRecip(ss.Prjn2x2Skp1)
ss.Prjn2x2Skp1Sub2 = prjn.NewPoolTileSub()
ss.Prjn2x2Skp1Sub2.Size.Set(2, 2)
ss.Prjn2x2Skp1Sub2.Skip.Set(1, 1)
ss.Prjn2x2Skp1Sub2.Start.Set(0, 0)
ss.Prjn2x2Skp1Sub2.Subs.Set(2, 2)
ss.Prjn2x2Skp1Sub2.TopoRange.Min = 0.8
ss.Prjn2x2Skp1Sub2Recip = prjn.NewPoolTileSubRecip(ss.Prjn2x2Skp1Sub2)
ss.Prjn2x2Skp1Sub2Send = prjn.NewPoolTileSub()
ss.Prjn2x2Skp1Sub2Send.Size.Set(2, 2)
ss.Prjn2x2Skp1Sub2Send.Skip.Set(1, 1)
ss.Prjn2x2Skp1Sub2Send.Start.Set(0, 0)
ss.Prjn2x2Skp1Sub2Send.Subs.Set(2, 2)
ss.Prjn2x2Skp1Sub2Send.SendSubs = true
ss.Prjn2x2Skp1Sub2Send.TopoRange.Min = 0.8
ss.Prjn2x2Skp1Sub2SendRecip = prjn.NewPoolTileSub()
*ss.Prjn2x2Skp1Sub2SendRecip = *ss.Prjn2x2Skp1Sub2Send
ss.Prjn2x2Skp1Sub2SendRecip.Recip = true
ss.Prjn2x2Skp2 = prjn.NewPoolTileSub()
ss.Prjn2x2Skp2.Size.Set(2, 2)
ss.Prjn2x2Skp2.Skip.Set(2, 2)
ss.Prjn2x2Skp2.Start.Set(0, 0)
ss.Prjn2x2Skp2.Subs.Set(2, 2)
ss.Prjn4x4Skp0 = prjn.NewPoolTile()
ss.Prjn4x4Skp0.Size.Set(4, 4)
ss.Prjn4x4Skp0.Skip.Set(0, 0)
ss.Prjn4x4Skp0.Start.Set(0, 0)
ss.Prjn4x4Skp0.GaussFull.Sigma = 1.5
ss.Prjn4x4Skp0.GaussInPool.Sigma = 1.5
ss.Prjn4x4Skp0.TopoRange.Min = 0.8
ss.Prjn4x4Skp0Recip = prjn.NewPoolTileRecip(ss.Prjn4x4Skp0)
ss.Prjn4x4Skp0Sub2 = prjn.NewPoolTileSub()
ss.Prjn4x4Skp0Sub2.Size.Set(4, 4)
ss.Prjn4x4Skp0Sub2.Skip.Set(0, 0)
ss.Prjn4x4Skp0Sub2.Start.Set(0, 0)
ss.Prjn4x4Skp0Sub2.Subs.Set(2, 2)
ss.Prjn4x4Skp0Sub2.SendSubs = true
ss.Prjn4x4Skp0Sub2.GaussFull.Sigma = 1.5
ss.Prjn4x4Skp0Sub2.GaussInPool.Sigma = 1.5
ss.Prjn4x4Skp0Sub2.TopoRange.Min = 0.8
ss.Prjn4x4Skp0Sub2Recip = prjn.NewPoolTileSubRecip(ss.Prjn4x4Skp0Sub2)
ss.Prjn1x1Skp0 = prjn.NewPoolTile()
ss.Prjn1x1Skp0.Size.Set(1, 1)
ss.Prjn1x1Skp0.Skip.Set(0, 0)
ss.Prjn1x1Skp0.Start.Set(0, 0)
ss.Prjn1x1Skp0.GaussFull.Sigma = 1.5
ss.Prjn1x1Skp0.GaussInPool.Sigma = 1.5
ss.Prjn1x1Skp0.TopoRange.Min = 0.8
ss.Prjn1x1Skp0Recip = prjn.NewPoolTileRecip(ss.Prjn1x1Skp0)
ss.Prjn6x6Skp2Lat = prjn.NewPoolTileSub()
ss.Prjn6x6Skp2Lat.Size.Set(6, 6)
ss.Prjn6x6Skp2Lat.Skip.Set(2, 2)
ss.Prjn6x6Skp2Lat.Start.Set(-2, -2)
ss.Prjn6x6Skp2Lat.Subs.Set(2, 2)
ss.Prjn6x6Skp2Lat.TopoRange.Min = 0.8
ss.RndSeeds = make([]int64, 100) // make enough for plenty of runs
for i := 0; i < 100; i++ {
ss.RndSeeds[i] = int64(i) + 1 // exclude 0
}
ss.ViewOn = true
ss.TrainUpdt = axon.AlphaCycle
ss.TestUpdt = axon.GammaCycle
ss.ActRFNms = []string{"V4f16:Image", "V4f8:Output", "TEO8:Image", "TEO8:Output", "TEO16:Image", "TEO16:Output"}
ss.SpikeRastNms = []string{"V1l16", "V2l16", "V4f16", "TEOf16", "TE", "Output"}
ss.GaussKernel = convolve.GaussianKernel64(3, .5)
// fmt.Printf("%v\n", ss.GaussKernel)
}
////////////////////////////////////////////////////////////////////////////////////////////
// Configs
// Config configures all the elements using the standard functions
func (ss *Sim) Config() {
ss.ConfigEnv()
ss.ConfigNet(ss.Net)
ss.ConfigTrnTrlLog(ss.TrnTrlLog)
ss.ConfigTrnTrlLog(ss.TrnTrlLogAll)
ss.ConfigTrnTrlRepLog(ss.TrnTrlRepLog)
ss.ConfigTrnTrlRepLog(ss.TrnTrlRepLogAll)
ss.ConfigTrnCycLog(ss.TrnCycLog)
ss.ConfigTrnEpcLog(ss.TrnEpcLog)
ss.ConfigTstEpcLog(ss.TstEpcLog)
ss.ConfigTstTrlLog(ss.TstTrlLog)
ss.ConfigTstTrlLog(ss.TstTrlLogAll)
ss.ConfigSpikeRasts()
ss.ConfigRunLog(ss.RunLog)
}
func (ss *Sim) ConfigEnv() {
if ss.MaxRuns == 0 { // allow user override
ss.MaxRuns = 1
}
if ss.MaxEpcs == 0 { // allow user override
ss.MaxEpcs = 150 // was 1000
ss.NZeroStop = -1
}
if ss.MaxTrls == 0 { // allow user override
ss.MaxTrls = 512 / mpi.WorldSize()
}
path := "images/cifar-10-batches-bin"
ss.TrainEnv.Nm = "cifar10"
ss.TrainEnv.Dsc = "training params and state"
ss.TrainEnv.Defaults()
ss.TrainEnv.High16 = false // not useful -- may need more tuning?
ss.TrainEnv.ColorDoG = true
ss.TrainEnv.Images.ImgSize = 32
ss.TrainEnv.Images.TestBatch = 5
ss.TrainEnv.OutRandom = ss.RndOutPats
ss.TrainEnv.OutSize.Set(10, 1)
ss.TrainEnv.Images.OpenPath(path)
ss.TrainEnv.TransMax.Set(0.1, 0.1) // 0.2, 0.2 for CU3D100
ss.TrainEnv.ScaleRange.Set(0.9, 1.0) // 0.8, 1.1 for CU3D100
ss.TrainEnv.RotateMax = 0 // 8 for CU3D100
ss.TrainEnv.Validate()
ss.TrainEnv.Run.Max = ss.MaxRuns // note: we are not setting epoch max -- do that manually
ss.TrainEnv.Trial.Max = ss.MaxTrls
ss.TestEnv.Nm = ss.TrainEnv.Nm
ss.TestEnv.Dsc = "testing params and state"
ss.TestEnv.Defaults()
ss.TestEnv.High16 = ss.TrainEnv.High16
ss.TestEnv.ColorDoG = ss.TrainEnv.ColorDoG
ss.TestEnv.Images.ImgSize = 32
ss.TestEnv.Images.TestBatch = 5
ss.TestEnv.OutRandom = ss.RndOutPats
ss.TestEnv.OutSize.Set(10, 1)
ss.TestEnv.Test = true
ss.TestEnv.Images.OpenPath(path)
ss.TestEnv.Trial.Max = ss.MaxTrls
ss.TestEnv.TransMax.Set(0.0, 0.0) // 0.2, 0.2 for CU3D100
ss.TestEnv.ScaleRange.Set(1.0, 1.0) // 0.8, 1.1 for CU3D100
ss.TestEnv.RotateMax = 0 // 8 for CU3D100
ss.TestEnv.Validate()
if ss.UseMPI {
ss.TrainEnv.MPIAlloc()
ss.TestEnv.MPIAlloc()
}
ss.TrainEnv.Init(0)
ss.TestEnv.Init(0)
}
func (ss *Sim) ConfigNet(net *axon.Network) {
net.InitName(net, "Lvis")
v1nrows := 5
v2mNp := 8
v2lNp := 4
v2Nu := 7
v4Np := 4
v4Nu := 8
if ss.SubPools {
v2mNp *= 2
v2lNp *= 2
v2Nu = 6
v4Np = 8
v4Nu = 7
}
deg8 := true
var v1m8, v1l8, v2m8, v2l8, v4f8, teo8 emer.Layer
v1m16 := net.AddLayer4D("V1m16", 16, 16, v1nrows, 4, emer.Input)
v1l16 := net.AddLayer4D("V1l16", 8, 8, v1nrows, 4, emer.Input)
v1m16.SetClass("V1m")
v1l16.SetClass("V1l")
// not useful so far..
// clst := net.AddLayer2D("Claustrum", 5, 5, emer.Hidden)
v2m16 := net.AddLayer4D("V2m16", v2mNp, v2mNp, v2Nu, v2Nu, emer.Hidden)
v2l16 := net.AddLayer4D("V2l16", v2lNp, v2lNp, v2Nu, v2Nu, emer.Hidden)
v2m16.SetClass("V2m V2")
v2l16.SetClass("V2l V2")
v4f16 := net.AddLayer4D("V4f16", v4Np, v4Np, v4Nu, v4Nu, emer.Hidden)
v4f16.SetClass("V4")
teNu := 10
teo16 := net.AddLayer4D("TEOf16", 2, 2, teNu, teNu, emer.Hidden)
teo16.SetClass("TEO")
if deg8 {
v1m8 = net.AddLayer4D("V1m8", 16, 16, v1nrows, 4, emer.Input)
v1l8 = net.AddLayer4D("V1l8", 8, 8, v1nrows, 4, emer.Input)
v1l8.SetClass("V1l")
v1m8.SetClass("V1m")
v2m8 = net.AddLayer4D("V2m8", v2mNp, v2mNp, v2Nu, v2Nu, emer.Hidden)
v2l8 = net.AddLayer4D("V2l8", v2lNp, v2lNp, v2Nu, v2Nu, emer.Hidden)
v2m8.SetClass("V2m V2")
v2l8.SetClass("V2l V2")
v4f8 = net.AddLayer4D("V4f8", v4Np, v4Np, v4Nu, v4Nu, emer.Hidden)
v4f8.SetClass("V4")
teo8 = net.AddLayer4D("TEOf8", 2, 2, teNu, teNu, emer.Hidden)
teo8.SetClass("TEO")
}
te := net.AddLayer4D("TE", 2, 2, teNu, teNu, emer.Hidden)
var out emer.Layer
if ss.RndOutPats {
out = net.AddLayer2D("Output", ss.TrainEnv.OutSize.Y, ss.TrainEnv.OutSize.X, emer.Target)
} else {
out = net.AddLayer4D("Output", ss.TrainEnv.OutSize.Y, ss.TrainEnv.OutSize.X, ss.TrainEnv.NOutPer, 1, emer.Target)
}
full := prjn.NewFull()
_ = full
rndcut := prjn.NewUnifRnd()
rndcut.PCon = 0.1 // 0.2 == .1 459
// rndprjn := prjn.NewUnifRnd()
// rndprjn.PCon = 0.5 // 0.2 > .1
pool1to1 := prjn.NewPoolOneToOne()
_ = pool1to1
var p4x4s2, p2x2s1, p4x4s2send, p2x2s1send, p4x4s2recip, p2x2s1recip, v4toteo, teotov4 prjn.Pattern
p4x4s2 = ss.Prjn4x4Skp2
p2x2s1 = ss.Prjn2x2Skp1
p4x4s2send = ss.Prjn4x4Skp2
p2x2s1send = ss.Prjn2x2Skp1
p4x4s2recip = ss.Prjn4x4Skp2Recip
p2x2s1recip = ss.Prjn2x2Skp1Recip
v4toteo = full
teotov4 = full
if ss.SubPools {
p4x4s2 = ss.Prjn4x4Skp2Sub2
p2x2s1 = ss.Prjn2x2Skp1Sub2
p4x4s2send = ss.Prjn4x4Skp2Sub2Send
p2x2s1send = ss.Prjn2x2Skp1Sub2Send
p4x4s2recip = ss.Prjn4x4Skp2Sub2SendRecip
p2x2s1recip = ss.Prjn2x2Skp1Sub2SendRecip
v4toteo = ss.Prjn4x4Skp0Sub2
teotov4 = ss.Prjn4x4Skp0Sub2Recip
}
net.ConnectLayers(v1m16, v2m16, p4x4s2, emer.Forward).SetClass("V1V2")
net.ConnectLayers(v1l16, v2m16, p2x2s1, emer.Forward).SetClass("V1V2fmSm V1V2")
net.ConnectLayers(v1l16, v2l16, p4x4s2, emer.Forward).SetClass("V1V2")
v2v4, v4v2 := net.BidirConnectLayers(v2m16, v4f16, p4x4s2send)
v2v4.SetClass("V2V4")
v4v2.SetClass("V4V2").SetPattern(p4x4s2recip)
v2v4, v4v2 = net.BidirConnectLayers(v2l16, v4f16, p2x2s1send)
v2v4.SetClass("V2V4sm")
v4v2.SetClass("V4V2").SetPattern(p2x2s1recip)
v4teo, teov4 := net.BidirConnectLayers(v4f16, teo16, v4toteo)
v4teo.SetClass("V4TEO")
teov4.SetClass("TEOV4").SetPattern(teotov4)
teote, teteo := net.BidirConnectLayers(teo16, te, full)
teote.SetClass("TEOTE")
teteo.SetClass("TETEO")
// full connections to output are key
teoout, outteo := net.BidirConnectLayers(teo16, out, full)
teoout.SetClass("TEOOut ToOut")
outteo.SetClass("OutTEO FmOut")
teout, _ := net.BidirConnectLayers(te, out, full)
teout.SetClass("ToOut FmOut")
// v59 459 -- only useful later -- TEO maybe not doing as well later?
v4out, outv4 := net.BidirConnectLayers(v4f16, out, full)
v4out.SetClass("V4Out ToOut")
outv4.SetClass("OutV4 FmOut")
var v2inhib, v4inhib prjn.Pattern
v2inhib = pool1to1
v4inhib = pool1to1
if ss.SubPools {
v2inhib = ss.Prjn2x2Skp2 // ss.Prjn6x6Skp2Lat
v4inhib = ss.Prjn2x2Skp2
}
// this extra inhibition drives decorrelation, produces significant learning benefits
net.LateralConnectLayerPrjn(v2m16, v2inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(v2l16, v2inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(v4f16, v4inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(teo16, pool1to1, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(te, pool1to1, &axon.HebbPrjn{}).SetType(emer.Inhib)
///////////////////////
// Shortcuts:
// clst not useful
// net.ConnectLayers(v1l16, clst, full, emer.Forward)
// V1 shortcuts best for syncing all layers -- like the pulvinar basically
net.ConnectLayers(v1l16, v4f16, rndcut, emer.Forward).SetClass("V1SC")
net.ConnectLayers(v1l16, teo16, rndcut, emer.Forward).SetClass("V1SC")
net.ConnectLayers(v1l16, teo16, rndcut, emer.Forward).SetClass("V1SC")
net.ConnectLayers(v1l16, te, rndcut, emer.Forward).SetClass("V1SC")
if deg8 {
net.ConnectLayers(v1m8, v2m8, p4x4s2, emer.Forward).SetClass("V1V2")
net.ConnectLayers(v1l8, v2m8, p2x2s1, emer.Forward).SetClass("V1V2fmSm V1V2")
net.ConnectLayers(v1l8, v2l8, p4x4s2, emer.Forward).SetClass("V1V2")
v2v4, v4v2 := net.BidirConnectLayers(v2m8, v4f8, p4x4s2send)
v2v4.SetClass("V2V4")
v4v2.SetClass("V4V2").SetPattern(p4x4s2recip)
v2v4, v4v2 = net.BidirConnectLayers(v2l8, v4f8, p2x2s1send)
v2v4.SetClass("V2V4sm")
v4v2.SetClass("V4V2").SetPattern(p2x2s1recip)
net.ConnectLayers(v4f8, teo16, v4toteo, emer.Forward).SetClass("V4TEOoth")
v4teo, teov4 := net.BidirConnectLayers(v4f8, teo8, v4toteo)
v4teo.SetClass("V4TEO")
teov4.SetClass("TEOV4").SetPattern(teotov4)
net.ConnectLayers(v4f16, teo8, v4toteo, emer.Forward).SetClass("V4TEOoth")
teote, teteo := net.BidirConnectLayers(teo8, te, full)
teote.SetClass("TEOTE")
teteo.SetClass("TETEO")
teoout, outteo := net.BidirConnectLayers(teo8, out, full)
teoout.SetClass("TEOOut ToOut")
outteo.SetClass("OutTEO FmOut")
v4out, outv4 := net.BidirConnectLayers(v4f8, out, full)
v4out.SetClass("V4Out ToOut")
outv4.SetClass("OutV4 FmOut")
// this extra inhibition drives decorrelation, produces significant learning benefits
net.LateralConnectLayerPrjn(v2m8, v2inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(v2l8, v2inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(v4f8, v4inhib, &axon.HebbPrjn{}).SetType(emer.Inhib)
net.LateralConnectLayerPrjn(teo8, pool1to1, &axon.HebbPrjn{}).SetType(emer.Inhib)
// V1 shortcuts best for syncing all layers -- like the pulvinar basically
net.ConnectLayers(v1l8, v4f8, rndcut, emer.Forward).SetClass("V1SC")
net.ConnectLayers(v1l8, teo8, rndcut, emer.Forward).SetClass("V1SC")
net.ConnectLayers(v1l8, teo8, rndcut, emer.Forward).SetClass("V1SC")
}
//////////////////////
// Positioning
v1l16.SetRelPos(relpos.Rel{Rel: relpos.Behind, Other: v1m16.Name(), XAlign: relpos.Left, Space: 4})
v2m16.SetRelPos(relpos.Rel{Rel: relpos.Above, Other: v1m16.Name(), XAlign: relpos.Left, YAlign: relpos.Front})
v2l16.SetRelPos(relpos.Rel{Rel: relpos.Behind, Other: v2m16.Name(), XAlign: relpos.Left, Space: 4})
v4f16.SetRelPos(relpos.Rel{Rel: relpos.Above, Other: v2m16.Name(), XAlign: relpos.Left, YAlign: relpos.Front})
teo16.SetRelPos(relpos.Rel{Rel: relpos.RightOf, Other: v4f16.Name(), YAlign: relpos.Front, Space: 4})
out.SetRelPos(relpos.Rel{Rel: relpos.Behind, Other: te.Name(), XAlign: relpos.Left, Space: 15})
if deg8 {
v1m8.SetRelPos(relpos.Rel{Rel: relpos.RightOf, Other: v1m16.Name(), YAlign: relpos.Front, Space: 4})
v1l8.SetRelPos(relpos.Rel{Rel: relpos.Behind, Other: v1m8.Name(), XAlign: relpos.Left, Space: 4})