-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmultimodel.py
164 lines (156 loc) · 5.62 KB
/
multimodel.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
from scipy import misc
import matplotlib.pyplot as plt
import numpy as np
from numpy.random import randint
import math
import cv2
from keras.layers import Input, Dense, Dropout, Flatten, Activation, BatchNormalization, Reshape
from keras.layers import Conv2D, MaxPooling2D, UpSampling2D, Lambda
from keras.layers.noise import GaussianNoise
from keras.models import Model, Sequential, Input
from keras.optimizers import Adam , SGD
from keras.activations import elu ,selu
import keras
subtract_layer = Lambda(lambda inputs: inputs[0] - inputs[1],
output_shape=lambda shapes: shapes[0])
'''
This model uses an encoder decoder model with a cheat connection between the encoder and decoder, so it can erase and readd features
'''
s=96 #s must be a multiple of 8
savenum=0
def createpair(size): # generate toy data
data = np.zeros((size,size))
circcenter=(randint(0,size),randint(0,size))
r=randint(size//6,size//3)
squarecenter=(randint(0,size),randint(0,size))
squaresize=randint(size//8,size//4)
squareoffset=(squaresize,squaresize)
triverts=[(randint(0,size),randint(0,size)),(randint(0,size),randint(0,size)),(randint(0,size),randint(0,size))]
cv2.circle(data,circcenter,r,(255),thickness=-1)
cv2.rectangle(data,tuple(np.subtract(squarecenter,squareoffset)),tuple(np.add(squarecenter,squareoffset)),(255),thickness=-1)
notri=np.copy(data)
cv2.fillConvexPoly(data,np.array(triverts),(255))
data=np.expand_dims(data, axis=2)
notri=np.expand_dims(notri, axis=2)
return data,notri
def createpairs(size,num):
withtri=np.zeros((num,size,size,1),dtype=np.float32)
notri=np.zeros((num,size,size,1),dtype=np.float32)
for i in range(num):
c=createpair(size)
withtri[i]=c[0]/255
notri[i]=c[1]/255
return withtri,notri
encoderin=Input(shape=(s,s,1))
cheatencoderin=Input(shape=(s,s,1))
encoder=keras.layers.concatenate([encoderin,cheatencoderin])
encoder=BatchNormalization()(encoder)
encoder=Conv2D(32,(3,3),padding='same')(encoder)
for i in range(1,4): #for loops make it easer to change network structure, and are easier to read
encoder=MaxPooling2D()(encoder)
encoder=Conv2D(32*(2**i),(3,3),padding='same')(encoder)
encoder=BatchNormalization()(encoder)
encoder=Activation('elu')(encoder)
encoder=Conv2D(32*(2**i),(3,3),padding='same')(encoder)
encoder=BatchNormalization()(encoder)
encoder=Activation('elu')(encoder)
encodermodel=Model(inputs=[encoderin,cheatencoderin],outputs=encoder)
cheatin=Input(shape=(s//8,s//8,256))
cheat=Flatten()(cheatin)
cheat=Dense(100)(cheat)
cheat=BatchNormalization()(cheat)
cheat=Activation('elu')(cheat)
cheat=Dense(s*s)(cheat)
cheat=BatchNormalization()(cheat)
cheat=Activation('elu')(cheat)
cheat=Reshape((s,s,1))(cheat)
cheatskip=Model(inputs=cheatin,outputs=cheat)
zeros=Input(shape=(s,s,1))
maindecoderin=Input(shape=(s//8,s//8,256))
decoder=maindecoderin
for i in range(3,0,-1):
decoder=Conv2D(32*(2**i),(3,3),padding='same')(decoder)
decoder=BatchNormalization()(decoder)
decoder=Activation('elu')(decoder)
decoder=Conv2D(32*(2**i),(3,3),padding='same')(decoder)
decoder=BatchNormalization()(decoder)
decoder=Activation('elu')(decoder)
decoder=UpSampling2D()(decoder)
decoder=Conv2D(32,(3,3),padding='same')(decoder)
decoder=BatchNormalization()(decoder)
decoder=Activation('elu')(decoder)
decoder=Conv2D(1,(3,3),padding='same',activation='sigmoid')(decoder)
decodermodel=Model(inputs=[maindecoderin],outputs=[decoder])
eraserin=Input(shape=(s,s,1))
enc=encodermodel([eraserin,zeros])
skipped=cheatskip(enc)
notriangle=decodermodel(enc)
reenc=encodermodel([notriangle,skipped])
reconstructed=decodermodel(reenc)
eraser=Model(inputs=[eraserin,zeros],outputs=[notriangle,reconstructed]) #binary_crossentropy
eraser.compile(loss='binary_crossentropy', optimizer='adam')
discriminatorinput=Input(shape=[s,s,1]) # build the discrimiator
d=discriminatorinput
d=Conv2D(32,(3,3))(d)
#d=BatchNormalization()(d)
d=Activation('elu')(d)
for i in range(1,4):
d=MaxPooling2D()(d)
d=Conv2D(32*(2**i),(3,3))(d)
#d=BatchNormalization()(d)
d=Activation('elu')(d)
d=Conv2D(32*(2**i),(3,3))(d)
#d=BatchNormalization()(d)
d=Activation('elu')(d)
d=Flatten()(d)
for i in range(2):
d=Dense(512)(d)
#d=BatchNormalization()(d)
d=Activation('elu')(d)
d=Dense(1,activation='sigmoid')(d)
discriminator=Model(inputs=discriminatorinput,outputs=d)
discriminator.compile(loss='binary_crossentropy', optimizer='adam')
ganinput=Input(shape=[s,s,1]) # build the gan
img=eraser([ganinput,zeros])
ganout=discriminator(img[0])
gan=Model(inputs=[ganinput,zeros],outputs=[ganout,img[1]])
gan.compile(loss='binary_crossentropy', optimizer='adam')
def show(n): # show images
b=createpairs(s,n)
targets=np.squeeze(b[1])
inputs=np.squeeze(b[0])
out=eraser.predict([b[0],z])
outa=np.squeeze(out[0])
outb=np.squeeze(out[1])
for i in range(n):
plt.subplot(4,4,i+1)
plt.imshow(inputs[i])
plt.subplot(4,4,i+1+n)
plt.imshow(outa[i])
plt.subplot(4,4,i+1+n*2)
plt.imshow(outb[i])
#plt.show()
global savenum
plt.savefig(str(savenum)+".png")
savenum+=1
z=np.zeros(shape=(16,s,s,1))
for count in range(10000):
c=createpairs(s,16)
loss=eraser.train_on_batch([c[0],z],[c[1],c[0]])
print(loss)
if(count%100==0): # show every 50 iterations
print("")
show(4)
for count in range(10000):
real=np.array(createpairs(s,8)[1])
fake=np.array(createpairs(s,8)[0])
images=np.concatenate([real,fake])
labels=np.concatenate([np.ones(8),np.zeros(8)])
dloss=discriminator.train_on_batch(images,labels) # train the discriminator
ganlabels=np.ones(16)
fake=createpairs(s,16)[0]
ganloss=gan.train_on_batch(fake,ganlabels) # train the generator
print("discriminator loss",dloss,"generator loss",ganloss)
if(count%100==0): # show every 50 iterations
print("")
show(4)