Skip to content

Commit 1aa95db

Browse files
DevilLord9967stokhos
DevilLord9967
authored and
stokhos
committed
Create GAN.py (TheAlgorithms#1445)
* Create GAN.py * gan update * Delete train-labels-idx1-ubyte.gz * Update GAN.py * Update GAN.py * Delete GAN.py * Create gan.py * Update gan.py * input_data import file
1 parent 90adc0c commit 1aa95db

File tree

2 files changed

+723
-0
lines changed

2 files changed

+723
-0
lines changed

‎neural_network/gan.py

+391
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,391 @@
1+
import matplotlib.gridspec as gridspec
2+
import matplotlib.pyplot as plt
3+
import numpy as np
4+
from sklearn.utils import shuffle
5+
import input_data
6+
7+
random_numer = 42
8+
9+
np.random.seed(random_numer)
10+
def ReLu(x):
11+
mask = (x>0) * 1.0
12+
return mask *x
13+
def d_ReLu(x):
14+
mask = (x>0) * 1.0
15+
return mask
16+
17+
def arctan(x):
18+
return np.arctan(x)
19+
def d_arctan(x):
20+
return 1 / (1 + x ** 2)
21+
22+
def log(x):
23+
return 1 / ( 1+ np.exp(-1*x))
24+
def d_log(x):
25+
return log(x) * (1 - log(x))
26+
27+
def tanh(x):
28+
return np.tanh(x)
29+
def d_tanh(x):
30+
return 1 - np.tanh(x) ** 2
31+
32+
def plot(samples):
33+
fig = plt.figure(figsize=(4, 4))
34+
gs = gridspec.GridSpec(4, 4)
35+
gs.update(wspace=0.05, hspace=0.05)
36+
37+
for i, sample in enumerate(samples):
38+
ax = plt.subplot(gs[i])
39+
plt.axis('off')
40+
ax.set_xticklabels([])
41+
ax.set_yticklabels([])
42+
ax.set_aspect('equal')
43+
plt.imshow(sample.reshape(28, 28), cmap='Greys_r')
44+
45+
return fig
46+
47+
48+
49+
# 1. Load Data and declare hyper
50+
print('--------- Load Data ----------')
51+
mnist = input_data.read_data_sets('MNIST_data', one_hot=False)
52+
temp = mnist.test
53+
images, labels = temp.images, temp.labels
54+
images, labels = shuffle(np.asarray(images),np.asarray(labels))
55+
num_epoch = 10
56+
learing_rate = 0.00009
57+
G_input = 100
58+
hidden_input,hidden_input2,hidden_input3 = 128,256,346
59+
hidden_input4,hidden_input5,hidden_input6 = 480,560,686
60+
61+
62+
63+
print('--------- Declare Hyper Parameters ----------')
64+
# 2. Declare Weights
65+
D_W1 = np.random.normal(size=(784,hidden_input),scale=(1. / np.sqrt(784 / 2.))) *0.002
66+
# D_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
67+
D_b1 = np.zeros(hidden_input)
68+
69+
D_W2 = np.random.normal(size=(hidden_input,1),scale=(1. / np.sqrt(hidden_input / 2.))) *0.002
70+
# D_b2 = np.random.normal(size=(1),scale=(1. / np.sqrt(1 / 2.))) *0.002
71+
D_b2 = np.zeros(1)
72+
73+
74+
G_W1 = np.random.normal(size=(G_input,hidden_input),scale=(1. / np.sqrt(G_input / 2.))) *0.002
75+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
76+
G_b1 = np.zeros(hidden_input)
77+
78+
G_W2 = np.random.normal(size=(hidden_input,hidden_input2),scale=(1. / np.sqrt(hidden_input / 2.))) *0.002
79+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
80+
G_b2 = np.zeros(hidden_input2)
81+
82+
G_W3 = np.random.normal(size=(hidden_input2,hidden_input3),scale=(1. / np.sqrt(hidden_input2 / 2.))) *0.002
83+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
84+
G_b3 = np.zeros(hidden_input3)
85+
86+
G_W4 = np.random.normal(size=(hidden_input3,hidden_input4),scale=(1. / np.sqrt(hidden_input3 / 2.))) *0.002
87+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
88+
G_b4 = np.zeros(hidden_input4)
89+
90+
G_W5 = np.random.normal(size=(hidden_input4,hidden_input5),scale=(1. / np.sqrt(hidden_input4 / 2.))) *0.002
91+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
92+
G_b5 = np.zeros(hidden_input5)
93+
94+
G_W6 = np.random.normal(size=(hidden_input5,hidden_input6),scale=(1. / np.sqrt(hidden_input5 / 2.))) *0.002
95+
# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.))) *0.002
96+
G_b6 = np.zeros(hidden_input6)
97+
98+
G_W7 = np.random.normal(size=(hidden_input6,784),scale=(1. / np.sqrt(hidden_input6 / 2.))) *0.002
99+
# G_b2 = np.random.normal(size=(784),scale=(1. / np.sqrt(784 / 2.))) *0.002
100+
G_b7 = np.zeros(784)
101+
102+
# 3. For Adam Optimzier
103+
v1,m1 = 0,0
104+
v2,m2 = 0,0
105+
v3,m3 = 0,0
106+
v4,m4 = 0,0
107+
108+
v5,m5 = 0,0
109+
v6,m6 = 0,0
110+
v7,m7 = 0,0
111+
v8,m8 = 0,0
112+
v9,m9 = 0,0
113+
v10,m10 = 0,0
114+
v11,m11 = 0,0
115+
v12,m12 = 0,0
116+
117+
v13,m13 = 0,0
118+
v14,m14 = 0,0
119+
120+
v15,m15 = 0,0
121+
v16,m16 = 0,0
122+
123+
v17,m17 = 0,0
124+
v18,m18 = 0,0
125+
126+
127+
beta_1,beta_2,eps = 0.9,0.999,0.00000001
128+
129+
print('--------- Started Training ----------')
130+
for iter in range(num_epoch):
131+
132+
random_int = np.random.randint(len(images) - 5)
133+
current_image = np.expand_dims(images[random_int],axis=0)
134+
135+
# Func: Generate The first Fake Data
136+
Z = np.random.uniform(-1., 1., size=[1, G_input])
137+
Gl1 = Z.dot(G_W1) + G_b1
138+
Gl1A = arctan(Gl1)
139+
Gl2 = Gl1A.dot(G_W2) + G_b2
140+
Gl2A = ReLu(Gl2)
141+
Gl3 = Gl2A.dot(G_W3) + G_b3
142+
Gl3A = arctan(Gl3)
143+
144+
Gl4 = Gl3A.dot(G_W4) + G_b4
145+
Gl4A = ReLu(Gl4)
146+
Gl5 = Gl4A.dot(G_W5) + G_b5
147+
Gl5A = tanh(Gl5)
148+
Gl6 = Gl5A.dot(G_W6) + G_b6
149+
Gl6A = ReLu(Gl6)
150+
Gl7 = Gl6A.dot(G_W7) + G_b7
151+
152+
current_fake_data = log(Gl7)
153+
154+
# Func: Forward Feed for Real data
155+
Dl1_r = current_image.dot(D_W1) + D_b1
156+
Dl1_rA = ReLu(Dl1_r)
157+
Dl2_r = Dl1_rA.dot(D_W2) + D_b2
158+
Dl2_rA = log(Dl2_r)
159+
160+
# Func: Forward Feed for Fake Data
161+
Dl1_f = current_fake_data.dot(D_W1) + D_b1
162+
Dl1_fA = ReLu(Dl1_f)
163+
Dl2_f = Dl1_fA.dot(D_W2) + D_b2
164+
Dl2_fA = log(Dl2_f)
165+
166+
# Func: Cost D
167+
D_cost = -np.log(Dl2_rA) + np.log(1.0- Dl2_fA)
168+
169+
# Func: Gradient
170+
grad_f_w2_part_1 = 1/(1.0- Dl2_fA)
171+
grad_f_w2_part_2 = d_log(Dl2_f)
172+
grad_f_w2_part_3 = Dl1_fA
173+
grad_f_w2 = grad_f_w2_part_3.T.dot(grad_f_w2_part_1 * grad_f_w2_part_2)
174+
grad_f_b2 = grad_f_w2_part_1 * grad_f_w2_part_2
175+
176+
grad_f_w1_part_1 = (grad_f_w2_part_1 * grad_f_w2_part_2).dot(D_W2.T)
177+
grad_f_w1_part_2 = d_ReLu(Dl1_f)
178+
grad_f_w1_part_3 = current_fake_data
179+
grad_f_w1 = grad_f_w1_part_3.T.dot(grad_f_w1_part_1 * grad_f_w1_part_2)
180+
grad_f_b1 = grad_f_w1_part_1 * grad_f_w1_part_2
181+
182+
grad_r_w2_part_1 = - 1/Dl2_rA
183+
grad_r_w2_part_2 = d_log(Dl2_r)
184+
grad_r_w2_part_3 = Dl1_rA
185+
grad_r_w2 = grad_r_w2_part_3.T.dot(grad_r_w2_part_1 * grad_r_w2_part_2)
186+
grad_r_b2 = grad_r_w2_part_1 * grad_r_w2_part_2
187+
188+
grad_r_w1_part_1 = (grad_r_w2_part_1 * grad_r_w2_part_2).dot(D_W2.T)
189+
grad_r_w1_part_2 = d_ReLu(Dl1_r)
190+
grad_r_w1_part_3 = current_image
191+
grad_r_w1 = grad_r_w1_part_3.T.dot(grad_r_w1_part_1 * grad_r_w1_part_2)
192+
grad_r_b1 = grad_r_w1_part_1 * grad_r_w1_part_2
193+
194+
grad_w1 =grad_f_w1 + grad_r_w1
195+
grad_b1 =grad_f_b1 + grad_r_b1
196+
197+
grad_w2 =grad_f_w2 + grad_r_w2
198+
grad_b2 =grad_f_b2 + grad_r_b2
199+
200+
# ---- Update Gradient ----
201+
m1 = beta_1 * m1 + (1 - beta_1) * grad_w1
202+
v1 = beta_2 * v1 + (1 - beta_2) * grad_w1 ** 2
203+
204+
m2 = beta_1 * m2 + (1 - beta_1) * grad_b1
205+
v2 = beta_2 * v2 + (1 - beta_2) * grad_b1 ** 2
206+
207+
m3 = beta_1 * m3 + (1 - beta_1) * grad_w2
208+
v3 = beta_2 * v3 + (1 - beta_2) * grad_w2 ** 2
209+
210+
m4 = beta_1 * m4 + (1 - beta_1) * grad_b2
211+
v4 = beta_2 * v4 + (1 - beta_2) * grad_b2 ** 2
212+
213+
D_W1 = D_W1 - (learing_rate / (np.sqrt(v1 /(1-beta_2) ) + eps)) * (m1/(1-beta_1))
214+
D_b1 = D_b1 - (learing_rate / (np.sqrt(v2 /(1-beta_2) ) + eps)) * (m2/(1-beta_1))
215+
216+
D_W2 = D_W2 - (learing_rate / (np.sqrt(v3 /(1-beta_2) ) + eps)) * (m3/(1-beta_1))
217+
D_b2 = D_b2 - (learing_rate / (np.sqrt(v4 /(1-beta_2) ) + eps)) * (m4/(1-beta_1))
218+
219+
# Func: Forward Feed for G
220+
Z = np.random.uniform(-1., 1., size=[1, G_input])
221+
Gl1 = Z.dot(G_W1) + G_b1
222+
Gl1A = arctan(Gl1)
223+
Gl2 = Gl1A.dot(G_W2) + G_b2
224+
Gl2A = ReLu(Gl2)
225+
Gl3 = Gl2A.dot(G_W3) + G_b3
226+
Gl3A = arctan(Gl3)
227+
228+
Gl4 = Gl3A.dot(G_W4) + G_b4
229+
Gl4A = ReLu(Gl4)
230+
Gl5 = Gl4A.dot(G_W5) + G_b5
231+
Gl5A = tanh(Gl5)
232+
Gl6 = Gl5A.dot(G_W6) + G_b6
233+
Gl6A = ReLu(Gl6)
234+
Gl7 = Gl6A.dot(G_W7) + G_b7
235+
236+
current_fake_data = log(Gl7)
237+
238+
Dl1 = current_fake_data.dot(D_W1) + D_b1
239+
Dl1_A = ReLu(Dl1)
240+
Dl2 = Dl1_A.dot(D_W2) + D_b2
241+
Dl2_A = log(Dl2)
242+
243+
# Func: Cost G
244+
G_cost = -np.log(Dl2_A)
245+
246+
# Func: Gradient
247+
grad_G_w7_part_1 = ((-1/Dl2_A) * d_log(Dl2).dot(D_W2.T) * (d_ReLu(Dl1))).dot(D_W1.T)
248+
grad_G_w7_part_2 = d_log(Gl7)
249+
grad_G_w7_part_3 = Gl6A
250+
grad_G_w7 = grad_G_w7_part_3.T.dot(grad_G_w7_part_1 * grad_G_w7_part_1)
251+
grad_G_b7 = grad_G_w7_part_1 * grad_G_w7_part_2
252+
253+
grad_G_w6_part_1 = (grad_G_w7_part_1 * grad_G_w7_part_2).dot(G_W7.T)
254+
grad_G_w6_part_2 = d_ReLu(Gl6)
255+
grad_G_w6_part_3 = Gl5A
256+
grad_G_w6 = grad_G_w6_part_3.T.dot(grad_G_w6_part_1 * grad_G_w6_part_2)
257+
grad_G_b6 = (grad_G_w6_part_1 * grad_G_w6_part_2)
258+
259+
grad_G_w5_part_1 = (grad_G_w6_part_1 * grad_G_w6_part_2).dot(G_W6.T)
260+
grad_G_w5_part_2 = d_tanh(Gl5)
261+
grad_G_w5_part_3 = Gl4A
262+
grad_G_w5 = grad_G_w5_part_3.T.dot(grad_G_w5_part_1 * grad_G_w5_part_2)
263+
grad_G_b5 = (grad_G_w5_part_1 * grad_G_w5_part_2)
264+
265+
grad_G_w4_part_1 = (grad_G_w5_part_1 * grad_G_w5_part_2).dot(G_W5.T)
266+
grad_G_w4_part_2 = d_ReLu(Gl4)
267+
grad_G_w4_part_3 = Gl3A
268+
grad_G_w4 = grad_G_w4_part_3.T.dot(grad_G_w4_part_1 * grad_G_w4_part_2)
269+
grad_G_b4 = (grad_G_w4_part_1 * grad_G_w4_part_2)
270+
271+
grad_G_w3_part_1 = (grad_G_w4_part_1 * grad_G_w4_part_2).dot(G_W4.T)
272+
grad_G_w3_part_2 = d_arctan(Gl3)
273+
grad_G_w3_part_3 = Gl2A
274+
grad_G_w3 = grad_G_w3_part_3.T.dot(grad_G_w3_part_1 * grad_G_w3_part_2)
275+
grad_G_b3 = (grad_G_w3_part_1 * grad_G_w3_part_2)
276+
277+
grad_G_w2_part_1 = (grad_G_w3_part_1 * grad_G_w3_part_2).dot(G_W3.T)
278+
grad_G_w2_part_2 = d_ReLu(Gl2)
279+
grad_G_w2_part_3 = Gl1A
280+
grad_G_w2 = grad_G_w2_part_3.T.dot(grad_G_w2_part_1 * grad_G_w2_part_2)
281+
grad_G_b2 = (grad_G_w2_part_1 * grad_G_w2_part_2)
282+
283+
grad_G_w1_part_1 = (grad_G_w2_part_1 * grad_G_w2_part_2).dot(G_W2.T)
284+
grad_G_w1_part_2 = d_arctan(Gl1)
285+
grad_G_w1_part_3 = Z
286+
grad_G_w1 = grad_G_w1_part_3.T.dot(grad_G_w1_part_1 * grad_G_w1_part_2)
287+
grad_G_b1 = grad_G_w1_part_1 * grad_G_w1_part_2
288+
289+
# ---- Update Gradient ----
290+
m5 = beta_1 * m5 + (1 - beta_1) * grad_G_w1
291+
v5 = beta_2 * v5 + (1 - beta_2) * grad_G_w1 ** 2
292+
293+
m6 = beta_1 * m6 + (1 - beta_1) * grad_G_b1
294+
v6 = beta_2 * v6 + (1 - beta_2) * grad_G_b1 ** 2
295+
296+
m7 = beta_1 * m7 + (1 - beta_1) * grad_G_w2
297+
v7 = beta_2 * v7 + (1 - beta_2) * grad_G_w2 ** 2
298+
299+
m8 = beta_1 * m8 + (1 - beta_1) * grad_G_b2
300+
v8 = beta_2 * v8 + (1 - beta_2) * grad_G_b2 ** 2
301+
302+
m9 = beta_1 * m9 + (1 - beta_1) * grad_G_w3
303+
v9 = beta_2 * v9 + (1 - beta_2) * grad_G_w3 ** 2
304+
305+
m10 = beta_1 * m10 + (1 - beta_1) * grad_G_b3
306+
v10 = beta_2 * v10 + (1 - beta_2) * grad_G_b3 ** 2
307+
308+
m11 = beta_1 * m11 + (1 - beta_1) * grad_G_w4
309+
v11 = beta_2 * v11 + (1 - beta_2) * grad_G_w4 ** 2
310+
311+
m12 = beta_1 * m12 + (1 - beta_1) * grad_G_b4
312+
v12 = beta_2 * v12 + (1 - beta_2) * grad_G_b4 ** 2
313+
314+
m13 = beta_1 * m13 + (1 - beta_1) * grad_G_w5
315+
v13 = beta_2 * v13 + (1 - beta_2) * grad_G_w5 ** 2
316+
317+
m14 = beta_1 * m14 + (1 - beta_1) * grad_G_b5
318+
v14 = beta_2 * v14 + (1 - beta_2) * grad_G_b5 ** 2
319+
320+
m15 = beta_1 * m15 + (1 - beta_1) * grad_G_w6
321+
v15 = beta_2 * v15 + (1 - beta_2) * grad_G_w6 ** 2
322+
323+
m16 = beta_1 * m16 + (1 - beta_1) * grad_G_b6
324+
v16 = beta_2 * v16 + (1 - beta_2) * grad_G_b6 ** 2
325+
326+
m17 = beta_1 * m17 + (1 - beta_1) * grad_G_w7
327+
v17 = beta_2 * v17 + (1 - beta_2) * grad_G_w7 ** 2
328+
329+
m18 = beta_1 * m18 + (1 - beta_1) * grad_G_b7
330+
v18 = beta_2 * v18 + (1 - beta_2) * grad_G_b7 ** 2
331+
332+
G_W1 = G_W1 - (learing_rate / (np.sqrt(v5 /(1-beta_2) ) + eps)) * (m5/(1-beta_1))
333+
G_b1 = G_b1 - (learing_rate / (np.sqrt(v6 /(1-beta_2) ) + eps)) * (m6/(1-beta_1))
334+
335+
G_W2 = G_W2 - (learing_rate / (np.sqrt(v7 /(1-beta_2) ) + eps)) * (m7/(1-beta_1))
336+
G_b2 = G_b2 - (learing_rate / (np.sqrt(v8 /(1-beta_2) ) + eps)) * (m8/(1-beta_1))
337+
338+
G_W3 = G_W3 - (learing_rate / (np.sqrt(v9 /(1-beta_2) ) + eps)) * (m9/(1-beta_1))
339+
G_b3 = G_b3 - (learing_rate / (np.sqrt(v10 /(1-beta_2) ) + eps)) * (m10/(1-beta_1))
340+
341+
G_W4 = G_W4 - (learing_rate / (np.sqrt(v11 /(1-beta_2) ) + eps)) * (m11/(1-beta_1))
342+
G_b4 = G_b4 - (learing_rate / (np.sqrt(v12 /(1-beta_2) ) + eps)) * (m12/(1-beta_1))
343+
344+
G_W5 = G_W5 - (learing_rate / (np.sqrt(v13 /(1-beta_2) ) + eps)) * (m13/(1-beta_1))
345+
G_b5 = G_b5 - (learing_rate / (np.sqrt(v14 /(1-beta_2) ) + eps)) * (m14/(1-beta_1))
346+
347+
G_W6 = G_W6 - (learing_rate / (np.sqrt(v15 /(1-beta_2) ) + eps)) * (m15/(1-beta_1))
348+
G_b6 = G_b6 - (learing_rate / (np.sqrt(v16 /(1-beta_2) ) + eps)) * (m16/(1-beta_1))
349+
350+
G_W7 = G_W7 - (learing_rate / (np.sqrt(v17 /(1-beta_2) ) + eps)) * (m17/(1-beta_1))
351+
G_b7 = G_b7 - (learing_rate / (np.sqrt(v18 /(1-beta_2) ) + eps)) * (m18/(1-beta_1))
352+
353+
# --- Print Error ----
354+
#print("Current Iter: ",iter, " Current D cost:",D_cost, " Current G cost: ", G_cost,end='\r')
355+
356+
if iter == 0:
357+
learing_rate = learing_rate * 0.01
358+
if iter == 40:
359+
learing_rate = learing_rate * 0.01
360+
361+
# ---- Print to Out put ----
362+
if iter%10 == 0:
363+
364+
print("Current Iter: ",iter, " Current D cost:",D_cost, " Current G cost: ", G_cost,end='\r')
365+
print('--------- Show Example Result See Tab Above ----------')
366+
print('--------- Wait for the image to load ---------')
367+
Z = np.random.uniform(-1., 1., size=[16, G_input])
368+
369+
Gl1 = Z.dot(G_W1) + G_b1
370+
Gl1A = arctan(Gl1)
371+
Gl2 = Gl1A.dot(G_W2) + G_b2
372+
Gl2A = ReLu(Gl2)
373+
Gl3 = Gl2A.dot(G_W3) + G_b3
374+
Gl3A = arctan(Gl3)
375+
376+
Gl4 = Gl3A.dot(G_W4) + G_b4
377+
Gl4A = ReLu(Gl4)
378+
Gl5 = Gl4A.dot(G_W5) + G_b5
379+
Gl5A = tanh(Gl5)
380+
Gl6 = Gl5A.dot(G_W6) + G_b6
381+
Gl6A = ReLu(Gl6)
382+
Gl7 = Gl6A.dot(G_W7) + G_b7
383+
384+
current_fake_data = log(Gl7)
385+
386+
fig = plot(current_fake_data)
387+
fig.savefig('Click_Me_{}.png'.format(str(iter).zfill(3)+"_Ginput_"+str(G_input)+ \
388+
"_hiddenone"+str(hidden_input) + "_hiddentwo"+str(hidden_input2) + "_LR_" + str(learing_rate)
389+
), bbox_inches='tight')
390+
#for complete explanation visit https://towardsdatascience.com/only-numpy-implementing-gan-general-adversarial-networks-and-adam-optimizer-using-numpy-with-2a7e4e032021
391+
# -- end code --

0 commit comments

Comments
 (0)