해커가 알려주는 뉴럴 네트워크를 공부한 내용

  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

import sys
import random
from math import exp

class rgate:

	def mul(x,y):
		return x*y
	def add(x,y):
		return x+y

class unit:
	value = 0.0
	grad = 0.0
	def __init__(self, value=0.0,grad=0.0):
		self.value = value
		self.grad = grad

class u:
	value = 0.0
	grad = 0.0
	def __init__(self, value=0.0,grad=0.0):
		self.value = value
		self.grad = grad

class mg:
	u0 = None
	u1 = None
	ru = None
	def __init__(self, u0=None,u1=None):
		self.u0 = None
		self.u1 = None
		self.ru = None
	def f(self, u0, u1):
		self.u0 = u0
		self.u1 = u1
		self.ru = unit(u0.value*u1.value,0)
		return self.ru
	def b(self):
		self.u0.grad +=  self.ru.grad * self.u1.value
		self.u1.grad +=  self.ru.grad * self.u0.value

class ag:
	u0 = None
	u1 = None
	ru = None
	def f(self,u0, u1):
		self.u0 = u0
		self.u1 = u1
		self.ru = unit(u0.value+u1.value,0)
		return self.ru
	def b(self):
		self.u0.grad +=  self.ru.grad*1
		self.u1.grad +=  self.ru.grad*1

class sg:
	u0 = None
	ru = None
	def sig(self, x):
		return 1/(1+exp(-x))
	def f(self,u0):
		self.u0 = u0
		self.ru = unit(self.sig(self.u0.value), 0)
		return self.ru
	def b(self):
		s = self.sig(self.u0.value)
		self.u0.grad += (s*(1-s))*self.ru.grad

def fc(a,b,c,x,y):
	return 1/(1+exp(-(a*x+b*y+c)))

class sqg:
	u0 = None
	ru = None
	def f(self,u0):
		self.u0 = u0
		self.ru = unit(self.u0.value*self.u0.value,0.0)
		return self.ru
	def b(self):
		self.u0.grad += 2*self.u0.value*self.ru.grad

class dg:
	u0 = None
	u1 = None
	ru = None
	def f(self,u0=unit(1.0,0.0),u1=unit(1.0,0.0)):
		self.u0 = u0
		self.u1 = u1
		self.ru = unit(self.u0.value/self.u1.value,0.0)
		return self.ru
	def b(self):
		self.u0.grad += (-1/(self.u0.value*self.u0.value))

class mag:#max gate
	u0 = None
	u1 = None
	ru = None
	def f(self,u0,u1):
		self.u0 = u0
		self.u1 = u1
		self.ru = unit(self.u0.value > self.u1.value and self.u0.value or self.u1.value , 0.0)
		return self.ru
	def b(self):
		self.u0.grad += self.ru.value == self.u0.value and 1.0*self.ru.grad or 0
		self.u1.grad += self.ru.value == self.u1.value and 1.0*self.ru.grad or 0

class rg:#ReLU gate
	u0 = None
	ru = None
	def f(self,u0):
		self.u0 = u0
		self.ru = unit(self.u0.value > 0 and self.u0.value or 0 , 0.0)
		return self.ru
	def b(self):
		self.u0.grad += self.ru.value > 0 and 1.0*self.ru.grad or 0

def sig(a):
	return 1/(1+exp(-a))

if __name__=='__main__':
	"""
	#Random Local Search,출력을 임의로 변화시켜 더 나은 출력을 찾는다
	print(rgate.mul(1,2))
	x = -2
	y = 3
	best_x = x
	best_y = y
	x_try=0
	y_try=0
	tweak_amout = 0.01
	out = 0;
	best_out = -sys.maxsize-1
	for each in range(100):
		print(each)
		x_try = x + tweak_amout*(random.random()*2-1)
		y_try = y + tweak_amout*(random.random()*2-1)
		out = rgate.mul(x_try,y_try)
		if(out>best_out):
			best_out = out
			best_x = x_try
			best_y = y_try
	print(best_x,best_y,best_out)
	"""
	"""
	#Numerical Gradient, 간단한 계산을 통해 기울기를 찾고 더 나은 출력을 도출해낸다
	#기울기는 더 나은 출력을 위한 최선의 방향을 의미한다
	#단일 뉴런으로 볼때는 큰 스텝이 좋은 출력을 내지만, 복잡하게 꼬여있는 경우 스텝이 크면 예상을 벗어나는 값이 나올 수 있다.
	#스텝의 크기는 눈을 가리고 언덕을 오를때의 보폭의 크기로 비유할 수 있다. 작으면 느리지만 확실하게 언덕을 오를 수 있지만, 크면 빠를수 있지만 다칠수 있다.
	x =-2
	y = 3
	out = rgate.mul(x,y)
	h = 0.0001
	x_derivative = (rgate.mul(x+h,y)-out)/h
	y_derivative = (rgate.mul(x,y+h)-out)/h
	print(x_derivative)
	print(y_derivative)
	step_size = 0.01
	out = rgate.mul(x,y)
	print(out)
	x = x + step_size*x_derivative
	y = y + step_size*y_derivative
	new_out = rgate.mul(x,y)
	print(new_out)
	"""
	"""
	#Analytic Gradient
	#기울기를 입,출력의 변화로 부터 계산할 경우 입력을 개수에 따라 계산하는 비용이 선형적으로 증가한다.
	#수백만 수억개가 있을때는 큰 비용이 들게 된다.
	#이 방법을 입출력에 변화를 주어 계산할 필요 없이, 미분공식으로 기울기는 구한다
	x =-2
	y = 3
	out = rgate.mul(x,y)
	x_derivative = y #미분 결과에 의해
	y_derivative = x #미분 결과에 의해
	step_size = 0.01
	out = rgate.mul(x,y)
	print(out)
	x = x + step_size*x_derivative
	y = y + step_size*y_derivative
	new_out = rgate.mul(x,y)
	print(new_out)
	"""
	"""	
	뉴럴네트워크 라이브러리를 기울기를 구할때 #3 공식기울기를 사용하지만, 검증은 계산기울기를 통해서 한다.
	공식기울기는 효율적이지만 때로는 틀릴수도 있는 반면, 계산기울기는 비용은 크지만 확실한 값이다.
	"""
	"""
	#Backpropagation
	#연결된 게이트에서 #3공식기울기를 구할때는 체인룰을 적용한다. 체인룰은 곱셈으로 연결시키는 것이다.
	x = -2; y = 5; z =-4
	q = rgate.add(x,y)
	f = rgate.mul(q,z)
	#print(q)
	print(f)
	d_f_wrt_q = z
	d_f_wrt_z = q
	d_q_wrt_x = 1.0
	d_q_wrt_y = 1.0
	d_f_wrt_x = d_q_wrt_x*d_f_wrt_q
	d_f_wrt_y = d_q_wrt_y*d_f_wrt_q
	g = [d_f_wrt_x,d_f_wrt_y,d_f_wrt_z]
	step = 0.01
	x=x+step*g[0]
	y=y+step*g[1]
	z=z+step*g[2]
	q = rgate.add(x,y)
	f = rgate.mul(q,z)
	#print(q)
	print(f)
	"""
	
	a = unit(1.0, 0.0)
	b = unit(2.0, 0.0)
	c = unit(-3.0, 0.0)
	x = unit(-1.0, 0.0)
	y = unit(3.0 ,0.0)

	mg0 = mg()
	mg1 = mg()
	ag0 = ag()
	ag1 = ag()
	sg0 = sg()

	m1 = mg0.f(a, x)
	m2 = mg1.f(b, y)
	a1 = ag0.f(m1, m2)
	a2 = ag1.f(a1, c)
	s = sg0.f(a2)

	print(s.value)

	sg0.ru.grad = 1.0
	sg0.b()
	ag0.b()
	ag1.b()
	mg0.b()
	mg1.b()

	step = 0.01
	a.value += step*a.grad
	b.value += step*b.grad
	c.value += step*c.grad
	x.value += step*x.grad
	y.value += step*y.grad

	ax = mg0.f(a, x)
	by = mg1.f(b, y)
	ag0 = ag0.f(ax, by)
	ag1 = ag1.f(ag0, c)
	s = sg0.f(ag1)

	print(s.value)
	
	h = 0.001
	a = 1
	b = 2
	c = -3
	x = -1
	y = 3
	ga = (fc(a+h,b,c,x,y)-fc(a,b,c,x,y))/h
	gb = (fc(a,b+h,c,x,y)-fc(a,b,c,x,y))/h
	gc = (fc(a+h,b,c+h,x,y)-fc(a,b,c,x,y))/h
	gx = (fc(a+h,b,c,x+h,y)-fc(a,b,c,x,y))/h
	gy = (fc(a+h,b,c,x,y+h)-fc(a,b,c,x,y))/h
	
	print(ga)
	print(gb)
	print(gc)
	print(gx)
	print(gy)

	# * gate
	print()
	print("* gate")
	a = u(11.0,0.0)
	b = u(22.0,0.0)
	mg1 = mg()
	r1 = mg1.f(a,b)
	r1.grad = 1.0
	mg1.b()
	da = mg1.u0.grad; print(da)
	db = mg1.u1.grad; print(db)
	print('------')
	
	# + gate
	print("+ gate")
	a = u(11.0,0.0)
	b = u(22.0,0.0)
	ag1 = ag()
	r1 = ag1.f(a,b)
	r1.grad = 1.0
	ag1.b()
	da = ag1.u0.grad; print(da)
	db = ag1.u1.grad; print(db)
	print('------')

	# + gate, 3var
	print("+ gate, 3th var")
	#input def
	a = u(11.0,0.0)
	b = u(22.0,0.0)
	c = u(33.0,0.0)
	#gate def
	ag1 = ag()
	ag2 = ag()
	#cal f
	r1 = ag1.f(a,b)
	r2 = ag2.f(r1,c)
	#cal b
	r2.grad = 1.0
	ag2.b()
	#r1.grad = ag2.u0.grad
	ag1.b()
	#print derivative
	dc = ag2.u0.grad; print(dc)
	db = ag1.u1.grad; print(db)
	da = ag1.u0.grad; print(da)
	print('------')
	
	# mixed gate
	print("mixed gate")

	#input def
	a = u(1.0,0.0)
	b = u(2.0,0.0)
	c = u(3.0,0.0)

	#gate def
	mg1 = mg()
	ag1 = ag()

	#cal f
	r1 = mg1.f(a,b)
	r2 = ag1.f(r1,c)

	#cal b
	r2.grad = 1.0
	ag1.b()
	#r1.grad = ag1.u0.grad
	mg1.b()

	#print derivative
	dc = ag1.u0.grad; print(dc)
	db = mg1.u1.grad; print(db)
	da = mg1.u0.grad; print(da)
	print('------')

	# square gate
	print("square gate")

	#input def
	a = u(11.0,0.0)

	#gate def
	s1 = sqg()

	#cal f
	r1 = s1.f(a)

	#cal b
	r1.grad = 1.0
	s1.b()

	#print derivative
	da = s1.u0.grad; print(da)
	print('------')

	# single neuron,ax+by+c
	print("single neuron,ax+by+c")

	#input def
	a = u(1.0,0.0)
	b = u(2.0,0.0)
	c = u(3.0,0.0)
	x = u(4.0,0.0)
	y = u(5.0,0.0)

	#gate def
	mg1 = mg()
	mg2 = mg()
	ag1 = ag()
	ag2 = ag()
	sg1 = sg()
	

	#cal f
	r1 = mg1.f(a,x)
	r2 = mg2.f(b,y)
	r3 = ag1.f(r1,r2)
	r4 = ag2.f(r3,c)
	r5 = sg1.f(r4)

	#cal b
	r5.grad = 1.0
	sg1.b()
	#r4.grad = sg1.u0.grad
	ag2.b()
	#r3.grad = ag2.u0.grad
	ag1.b()
	#r2.grad = ag1.u1.grad
	mg2.b()
	#r1.grad = ag1.u0.grad
	mg1.b()

	#print derivative
	da = mg1.u0.grad; print(da)
	db = mg2.u0.grad; print(db)
	dc = ag2.u1.grad; print(dc)
	dx = mg1.u1.grad; print(dx)
	dy = mg2.u1.grad; print(dy)
	print('------')

	# Math.pow(((a*b+c),2); 
	print("Math.pow(((a*b+c),2) gate")

	#input def
	a = u(3.0,0.0)
	b = u(2.0,0.0)
	c = u(1.0,0.0)

	#gate def
	mg1 = mg()
	ag1 = ag()
	sq1 = sqg()

	#cal f
	r1 = mg1.f(a,b)
	r2 = ag1.f(r1,c)
	r3 = sq1.f(r2)

	#cal b
	r3.grad = 1.0
	sq1.b()
	ag1.b()
	mg1.b()

	#print derivative
	da = mg1.u0.grad; print(da)
	db = mg1.u1.grad; print(db)
	dc = ag1.u1.grad; print(dc)
	print('------')

	# division gate
	print("division gate")

	#input def
	a = u(3.0,0.0)

	#gate def
	dg1 = dg()

	#cal f
	r1 = dg1.f(a)

	#cal b
	r1.grad = 1.0
	dg1.b()

	#print derivative
	da = dg1.u0.grad; print(da)
	print('------')
	# max gate
	print("max gate")

	#input def
	a = u(1.0,0.0)
	b = u(2.0,0.0)

	#gate def
	mag1 = mag()

	#cal f
	r1 = mag1.f(a,b)

	#cal b
	r1.grad = 1.0
	mag1.b()

	#print derivative
	da = mag1.u0.grad; print(da)
	db = mag1.u1.grad; print(db)
	print('------')

	# ReLU gate
	print("ReLU gate")

	#input def
	a = u(1.0,0.0)

	#gate def
	rg1 = rg()

	#cal f
	r1 = rg1.f(a)

	#cal b
	r1.grad = 1.0
	rg1.b()

	#print derivative
	da = rg1.u0.grad; print(da)
	print('------')

	# mixed gate, (a+b)/(c+d)
	print("mixed gate, (a+b)/(c+d)")

	#input def
	a = u(1.0,0.0)
	b = u(2.0,0.0)
	c = u(3.0,0.0)
	d = u(4.0,0.0)

	#gate def
	ag1 = ag()
	ag2 = ag()
	dg1 = dg()
	mg1 = mg()

	#cal f
	r1 = ag1.f(a,b)
	r2 = ag2.f(c,d)
	r3 = dg1.f(r2)
	r4 = mg1.f(r1,r3)

	#cal b
	r4.grad = 1.0
	mg1.b()
	dg1.b()
	ag1.b()
	ag2.b()

	#print derivative
	da = ag1.u0.grad; print(da)
	db = ag1.u1.grad; print(db)
	dc = ag2.u0.grad; print(dc)
	dd = ag2.u1.grad; print(dd)
	print('------')