-
Notifications
You must be signed in to change notification settings - Fork 0
/
activation.go
516 lines (436 loc) · 10.4 KB
/
activation.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
package deep
import math "github.com/chewxy/math32"
// Mode denotes inference mode
type Mode int
const (
// ModeDefault is unspecified mode
ModeDefault Mode = 0
// ModeMultiClass is for one-hot encoded classification, applies softmax output layer
ModeMultiClass Mode = 1
// ModeRegression is regression, applies linear output layer
ModeRegression Mode = 2
// ModeBinary is binary classification, applies sigmoid output layer
ModeBinary Mode = 3
// ModeMultiLabel is for multilabel classification, applies sigmoid output layer
ModeMultiLabel Mode = 4
)
// OutputActivation returns activation corresponding to prediction mode
func OutputActivation(c Mode) ActivationType {
switch c {
case ModeMultiClass:
return ActivationSoftmax
case ModeRegression:
return ActivationLinear
case ModeBinary, ModeMultiLabel:
return ActivationSigmoid
}
return ActivationNone
}
// GetActivation returns the concrete activation given an ActivationType
func GetActivation(act ActivationType) Differentiable {
switch act {
case ActivationSigmoid:
return Sigmoid{}
case ActivationTanh:
return Tanh{}
case ActivationReLU:
return ReLU{}
case ActivationELU:
return eLU{}
case ActivationSwish:
return Swish{}
case ActivationRootSwish:
return RootSwish{}
case ActivationMish:
return Mish{}
case ActivationCustom:
return Custom{}
case ActivationLinear:
return Linear{}
case ActivationSoftmax:
return Linear{}
case ActivationDoubleRoot:
return DoubleRoot{}
case ActivationRootX:
return RootX{}
case ActivationDivX:
return DivX{}
case ActivationDoubleDiv:
return DoubleDiv{}
case ActivationRootPow:
return RootPow{}
case ActivationDoublePow:
return RootPow{}
}
return Linear{}
}
// ActivationType is represents a neuron activation function
type ActivationType int
const (
// ActivationNone is no activation
ActivationNone ActivationType = 0
// ActivationSigmoid is a sigmoid activation
ActivationSigmoid ActivationType = 1
// ActivationTanh is hyperbolic activation
ActivationTanh ActivationType = 2
// ActivationReLU is rectified linear unit activation
ActivationReLU ActivationType = 3
// ActivationLinear is linear activation
ActivationLinear ActivationType = 4
// ActivationSoftmax is a softmax activation (per layer)
ActivationSoftmax ActivationType = 5
// ActivationELU is a Elu activation
ActivationELU ActivationType = 6
// ActivationSwish is a Swish activation
ActivationSwish ActivationType = 7
// ActivationMish is a Mish activation
ActivationMish ActivationType = 8
// ActivationCustom is a Custom activation
ActivationCustom ActivationType = 9
// ActivationCustom is a Custom activation
ActivationDoubleRoot ActivationType = 10
// ActivationCustom is a Custom activation
ActivationRootX ActivationType = 11
// ActivationMulDiv is a Custom activation
ActivationDivX ActivationType = 12
// ActivationMulDiv is a Custom activation
ActivationDoubleDiv ActivationType = 13
// ActivationMulDiv is a Custom activation
ActivationRootPow ActivationType = 14
// ActivationMulDiv is a Custom activation
ActivationDoublePow ActivationType = 15
// ActivationMulDiv is a Custom activation
ActivationRootSwish ActivationType = 16
)
// Differentiable is an activation function and its first order derivative,
// where the latter is expressed as a function of the former for efficiency
type Differentiable interface {
F(float32, bool) float32
Df(float32) float32
}
// Sigmoid is a logistic activator in the special case of a = 1
type Sigmoid struct {
Mem map[float32]float32
}
// F is Sigmoid(x)
func (a Sigmoid) F(x float32, training bool) float32 { return Logistic(x, 1) }
// Df is Sigmoid'(y), where y = Sigmoid(x)
func (a Sigmoid) Df(y float32) float32 { return y * (1 - y) }
func Sqrt(N float32) float32 {
return math.Sqrt(N)
}
// DoubleRoot is a logistic activator in the special case of a = 1
type DoubleRoot struct {
Mem map[float32]float32
}
// F is DoubleRoot(x)
func (a DoubleRoot) F(x float32, training bool) float32 {
if x == 0 {
return 0
} else if x > 0 {
return Sqrt(x)
} else {
return -Sqrt(-x)
}
}
// Df is DoubleRoot'(y), where y = DoubleRoot(x)
func (a DoubleRoot) Df(x float32) float32 {
if x == 0 {
return 0
} else if x > 0 {
return 1 / (2 * Sqrt(x))
} else {
return 1 / (2 * Sqrt(-x))
}
}
// RootX is a logistic activator in the special case of a = 1
type DoublePow struct {
Mem map[float32]float32
}
// F is RootX(x)
func (a DoublePow) F(x float32, training bool) float32 {
if x == 0 {
return 0
} else if x > 0 {
return x * x
} else {
return x * (-x)
}
}
// Df is DoubleRoot'(y), where y = DoubleRoot(x)
func (a DoublePow) Df(x float32) float32 {
if x == 0 {
return 0
} else if x > 0 {
return (2 * x)
} else {
return -(2 * x)
}
}
// RootX is a logistic activator in the special case of a = 1
type RootPow struct {
Mem map[float32]float32
}
// F is RootX(x)
func (a RootPow) F(x float32, training bool) float32 {
if x == 0 {
return 0
} else if x > 0 {
return ((x + 0.5) * (x + 0.5)) - 0.25
} else {
return 0.5 - Sqrt(0.25-x)
}
}
// Df is DoubleRoot'(y), where y = DoubleRoot(x)
func (a RootPow) Df(x float32) float32 {
if x == 0 {
return 0
} else if x > 0 {
return (2 * x) + 1
} else {
return 1 / (2 * Sqrt(0.25-x))
}
}
// RootX is a logistic activator in the special case of a = 1
type RootX struct {
Mem map[float32]float32
}
// F is RootX(x)
func (a RootX) F(x float32, training bool) float32 {
if x == 0 {
return 0
} else if x > 0 {
return x
} else {
return 0.5 - Sqrt(0.25-x)
}
}
// Df is DoubleRoot'(y), where y = DoubleRoot(x)
func (a RootX) Df(x float32) float32 {
if x == 0 {
return 0
} else if x > 0 {
return 1
} else {
return 1 / (2 * Sqrt(0.25-x))
}
}
// MulDiv is a logistic activator in the special case of a = 1
type DivX struct {
Mem map[float32]float32
}
// F is MulDiv(x)
func (a DivX) F(x float32, training bool) float32 {
if x >= 0 {
return x
} else {
return (1/(x-1) + 1) * -1
}
}
// Df is MulDiv'(y), where y = MulDiv(x)
func (a DivX) Df(x float32) float32 {
if x >= 0 {
return 1
} else {
return (1 / ((x - 1) * (x - 1)))
}
}
// MulDiv is a logistic activator in the special case of a = 1
type DoubleDiv struct {
Mem map[float32]float32
}
// F is MulDiv(x)
func (a DoubleDiv) F(x float32, training bool) float32 {
if x == 0 {
return 0
} else if x > 0 {
return (1/(x+1) - 1) * -1
} else {
return (1/(x-1) + 1) * -1
}
}
// Df is MulDiv'(y), where y = MulDiv(x)
func (a DoubleDiv) Df(x float32) float32 {
if x == 0 {
return 0
} else if x > 0 {
return (1 / ((x + 1) * (x + 1)))
} else {
return (1 / ((x - 1) * (x - 1)))
}
}
// Logistic is the logistic function
func Logistic(x, a float32) float32 {
return 1 / (1 + math.Exp(-a*x))
}
// Tanh is a hyperbolic activator
type Tanh struct {
Mem map[float32]float32
}
// F is Tanh(x)
func (a Tanh) F(x float32, training bool) float32 { return (1 - math.Exp(-2*x)) / (1 + math.Exp(-2*x)) }
// Df is Tanh'(y), where y = Tanh(x)
func (a Tanh) Df(y float32) float32 { return 1 - math.Pow(y, 2) }
// ReLU is a rectified linear unit activator
type ReLU struct {
Mem map[float32]float32
}
// F is ReLU(x)
func (a ReLU) F(x float32, training bool) float32 {
return math.Max(x, 0)
}
// Df is ReLU'(y), where y = ReLU(x)
func (a ReLU) Df(y float32) float32 {
if y > 0 {
return 1
}
return 0
}
type eLU struct {
Mem map[float32]float32
}
// F is ELU(x)
func (a eLU) F(x float32, training bool) float32 {
if x >= 0 {
// elu formula
return x + 0.0000001
} else {
return 1.0*math.Pow(math.E, x)*-1 + float32(math.SmallestNonzeroFloat32)
}
}
// Df is ReLU'(y), where y = ReLU(x)
func (a eLU) Df(y float32) float32 {
if y > 0 {
return 1 - 0.0000001
} else {
return 1.0*math.Exp(y) - float32(math.SmallestNonzeroFloat32)
}
}
type Swish struct {
Mem map[float32]float32
}
// F is Swish(x)
func (a Swish) F(x float32, training bool) float32 {
// if a.Mem == nil {
// a.Mem = map[float32]float32{}
// }
// ans := x * Logistic(x, 1)
// if training {
// a.Mem[ans] = x
// }
// return ans
return x / (math.Exp(-x) + 1)
}
// Df is swish'(y), where y = Swish(x)
func (a Swish) Df(y float32) float32 {
// x := a.Mem[y]
// delete(a.Mem, y)
// sigX := Logistic(x, 1)
// return y * (sigX * (1 + x*(1-sigX)))
ey := math.Exp(y)
ey1 := ey + 1
return (ey * (ey1 + y)) / (ey1 * ey1)
}
type RootSwish struct {
Mem map[float32]float32
}
// F is Swish(x)
func (a RootSwish) F(x float32, training bool) float32 {
// if a.Mem == nil {
// a.Mem = map[float32]float32{}
// }
// ans := x * Logistic(x, 1)
// if training {
// a.Mem[ans] = x
// }
// return ans
if x > 0 {
return x / (math.Exp(-x) + 1)
} else {
return 0.5 - math.Sqrt(0.25-(0.5*x))
}
}
// Df is swish'(y), where y = Swish(x)
func (a RootSwish) Df(y float32) float32 {
// x := a.Mem[y]
// delete(a.Mem, y)
// sigX := Logistic(x, 1)
// return y * (sigX * (1 + x*(1-sigX)))
if y > 0 {
ey := math.Exp(y)
ey1 := ey + 1
return (ey * (ey1 + y)) / (ey1 * ey1)
} else {
return 1 / (2 * math.Sqrt(1-(2*y)))
}
}
type Mish struct {
Mem map[float32]float32
}
// F is Mish(x)
func (a Mish) F(x float32, training bool) float32 {
if a.Mem == nil {
a.Mem = map[float32]float32{}
}
ans := x * math.Tanh(math.Log(1+math.Exp(x)))
if training {
a.Mem[ans] = x
}
return ans
}
// Df is Mish'(y), where y = Mish(x)
func (a Mish) Df(y float32) float32 {
x := a.Mem[y]
delete(a.Mem, y)
sigX := Logistic(x, 1)
xTanhSp := math.Tanh(math.Log(1 + math.Exp(x)))
return y * (xTanhSp + x*sigX*(1-xTanhSp*xTanhSp))
}
type Custom struct {
Mem map[float32]float32
}
var customF func(float32) float32
var customDf func(float32, float32) float32
func SetCustomF(F func(float32) float32) {
customF = F
}
func SetCustomDf(Df func(float32, float32) float32) {
customDf = Df
}
// F is Custom(x)
func (a Custom) F(x float32, training bool) float32 {
if a.Mem == nil {
a.Mem = map[float32]float32{}
}
if customF != nil {
ans := customF(x)
if training {
a.Mem[ans] = x
}
return ans
} else {
ans := x
if training {
a.Mem[ans] = x
}
return x
}
}
// Df is Custom'(y), where y = Custom(x)
func (a Custom) Df(y float32) float32 {
x := a.Mem[y]
delete(a.Mem, y)
if customDf != nil {
return customDf(y, x)
} else {
return x
}
}
// Linear is a linear activator
type Linear struct {
Mem map[float32]float32
}
// F is the identity function
func (a Linear) F(x float32, training bool) float32 { return x }
// Df is constant
func (a Linear) Df(x float32) float32 { return 1 }