-
Notifications
You must be signed in to change notification settings - Fork 5
/
model.py
155 lines (133 loc) · 5.99 KB
/
model.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
import tensorflow as tf
import tensorlayer as tl
from tensorlayer.layers import *
from tensorlayer.models import Model
from params import *
lrelu = lambda x: tl.act.lrelu(x, 0.2)
def residual(R, n_f, f_s):
w_init = tl.initializers.truncated_normal(stddev=0.01)
R_tmp = R
R = BatchNorm2d(act=tf.nn.relu)(Conv2d(n_f, (f_s, f_s), (1, 1), W_init=w_init)(R))
R = BatchNorm2d(act=None)(Conv2d(n_f, (f_s, f_s), (1, 1), W_init=w_init)(R))
R_tmp = Conv2d(n_f, (1, 1), (1, 1))(R_tmp)
return Elementwise(tf.add, act=tf.nn.relu)([R_tmp, R])
def Discriminator(input_shape, prefix = ""):
I = Input(input_shape)
D = Conv2d(
64, (4, 4), (2, 2), padding='SAME', act=lrelu, b_init=None, name=prefix+'D_conv_1')(I)
D = InstanceNorm2d(act=lrelu)(Conv2d(
128, (4, 4), (2, 2), padding='SAME', b_init=None, name=prefix+'D_conv_2')(D))
D = InstanceNorm2d(act=lrelu)(Conv2d(
256, (4, 4), (2, 2), padding='SAME', b_init=None, name=prefix+'D_conv_3')(D))
D = InstanceNorm2d(act=lrelu)(Conv2d(
512, (4, 4), (2, 2), padding='SAME', b_init=None, name=prefix+'D_conv_4')(D))
D = InstanceNorm2d(act=lrelu)(Conv2d(
512, (4, 4), (2, 2), padding='SAME', b_init=None, name=prefix+'D_conv_5')(D))
D = InstanceNorm2d(act=lrelu)(Conv2d(
512, (4, 4), (2, 2), padding='SAME', b_init=None, name=prefix+'D_conv_6')(D))
D = Conv2d(1, (4, 4), (1, 1), name=prefix+'D_conv_7')(D)
D = GlobalMeanPool2d()(D)
D_net = Model(inputs=I, outputs=D, name=prefix+'Discriminator')
return D_net
def Generator(input_shape, z_dim, prefix = ""):
w_init = tl.initializers.truncated_normal(stddev=0.01)
I = Input(input_shape)
Z = Input([input_shape[0], z_dim])
z = Reshape((input_shape[0], 1, 1, -1))(Z)
z = Tile([1, input_shape[1], input_shape[2], 1])(z)
conv_layers = []
G = Concat(concat_dim=-1)([I, z])
filters = [64, 128, 256, 512, 512, 512, 512]
if image_size == 256:
filters.append(512)
G = Conv2d(
filters[0], (4, 4), (2, 2), act=lrelu, W_init=w_init, b_init=None, name=prefix+'G_conv_1')(G)
conv_layers.append(G)
for i, n_f in enumerate(filters[1:]):
G = BatchNorm2d(act=lrelu)(Conv2d(
n_f, (4, 4), (2, 2), W_init=w_init, b_init=None, name=prefix+'G_conv_{}'.format(i + 2))(G))
conv_layers.append(G)
filters.pop()
filters.reverse()
conv_layers.pop()
for i, n_f in enumerate(filters):
G = BatchNorm2d(act=tf.nn.relu)(DeConv2d(
n_f, (4, 4), (2, 2), W_init=w_init, b_init=None, name=prefix+'G_deconv_{}'.format(len(filters)+1-i))(G))
G = Concat(concat_dim=-1)([G, conv_layers.pop()])
G = DeConv2d(3, (4, 4), (2, 2), act=tf.nn.tanh, W_init=w_init, b_init=None, name=prefix+'G_deconv_1')(G)
G_net = Model(inputs=[I, Z], outputs=G, name=prefix+'Generator')
return G_net
def Encoder(input_shape, z_dim, prefix=""):
I = Input(input_shape)
E = Conv2d(64, (4, 4), (2, 2), act=lrelu, name=prefix+'E_conv_1')(I)
E = MeanPool2d((2, 2), (2, 2), 'SAME')(residual(E, 128, 3))
E = MeanPool2d((2, 2), (2, 2), 'SAME')(residual(E, 256, 3))
E = MeanPool2d((2, 2), (2, 2), 'SAME')(residual(E, 512, 3))
E = MeanPool2d((2, 2), (2, 2), 'SAME')(residual(E, 512, 3))
E = MeanPool2d((2, 2), (2, 2), 'SAME')(residual(E, 512, 3))
E = Flatten()(MeanPool2d((8, 8), (8, 8), 'SAME')(E))
mu = Dense(z_dim)(E)
log_sigma = Dense(z_dim)(E)
z = Elementwise(tf.add)(
[mu, Lambda(lambda x:tf.random.normal(shape=[z_dim]) * tf.exp(x))(log_sigma)])
E_net = Model(inputs=I, outputs=[z, mu, log_sigma], name=prefix+'Encoder')
return E_net
class BicycleGAN(object):
count = 0
def __init__(self, LOAD = False, load_tag = model_tag):
BicycleGAN.count += 1
self.G = Generator(input_shape, z_dim, "model_{}/".format(self.count))
self.D = Discriminator(input_shape, "model_{}/".format(self.count))
self.E = Encoder(input_shape, z_dim, "model_{}/".format(self.count))
if LOAD:
self.load(load_tag)
self.G.train()
self.D.train()
self.E.train()
def load(self, model_tag):
tl.files.load_and_assign_npz(os.path.join(models_dir, "G_weights_{}.npz".format(model_tag)), self.G)
tl.files.load_and_assign_npz(os.path.join(models_dir, "D_weights_{}.npz".format(model_tag)), self.D)
tl.files.load_and_assign_npz(os.path.join(models_dir, "E_weights_{}.npz".format(model_tag)), self.E)
print("Model weights has been loaded.")
def save(self, model_tag):
tl.files.save_npz(self.G.all_weights, os.path.join(models_dir, "G_weights_{}.npz".format(model_tag)))
tl.files.save_npz(self.D.all_weights, os.path.join(models_dir, "D_weights_{}.npz".format(model_tag)))
tl.files.save_npz(self.E.all_weights, os.path.join(models_dir, "E_weights_{}.npz".format(model_tag)))
def calc_loss(self, image_A, image_B, z):
encoded_z, encoded_mu, encoded_log_sigma = self.E(image_B)
vae_img = self.G([image_A, encoded_z])
lr_img = self.G([image_A, z])
self.vae_img = vae_img
self.lr_img = lr_img
reconst_z, reconst_mu, reconst_log_sigma = self.E(lr_img)
P_real = self.D(image_B)
P_fake = self.D(lr_img)
P_fake_encoded = self.D(vae_img)
self.P_real = P_real
self.P_fake = P_fake
self.P_fake_encoded = P_fake_encoded
loss_vae_D = (tl.cost.mean_squared_error(P_real, 0.9, is_mean=True) +
tl.cost.mean_squared_error(P_fake_encoded, 0.0, is_mean=True))
loss_lr_D = (tl.cost.mean_squared_error(P_real, 0.9, is_mean=True) +
tl.cost.mean_squared_error(P_fake, 0.0, is_mean=True))
loss_vae_G = tl.cost.mean_squared_error(P_fake_encoded, 0.9, is_mean=True)
loss_lr_G = tl.cost.mean_squared_error(P_fake, 0.9, is_mean=True)
self.loss_G = loss_vae_G + loss_lr_G
self.loss_D = loss_vae_D + loss_lr_D
self.loss_vae_L1 = tl.cost.absolute_difference_error(
image_B, vae_img, is_mean=True, axis=[1, 2, 3])
self.loss_latent_L1 = tl.cost.absolute_difference_error(z, reconst_z, is_mean=True)
self.loss_kl_E = 0.5 * tf.reduce_mean(
-1 - 2 * encoded_log_sigma + encoded_mu**2 +
tf.exp(2 * tf.clip_by_value(encoded_log_sigma, -10, 10)))
loss = self.loss_D - reconst_C * self.loss_vae_L1 - latent_C * self.loss_latent_L1 - kl_C * self.loss_kl_E
return loss
if __name__ == '__main__':
x = BicycleGAN()
print(x.G)
print(x.D)
print(x.E)
y = BicycleGAN()
print(y.G)
print(y.D)
print(y.E)