-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
200 lines (163 loc) · 9.28 KB
/
models.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
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
import tensorflow as tf
import tensorflow_addons as tfa
import numpy as np
from image_loss import PerceptualLoss
from unet import unet_model
from functools import partial
import os
from histogram_matching import match_histograms, make_cdfs
IMAGE_SIZE = (256, 256)
perceptualLoss = PerceptualLoss((*IMAGE_SIZE, 3))
def build_generator(image_size):
return unet_model()
# def build_discriminator(image_size):
# model = tf.keras.applications.ResNet50V2(include_top=False, weights='imagenet', input_shape=(*image_size, 3), pooling='max')
# model.trainable = True
# out = tf.keras.layers.Dense(1, activation='sigmoid')(model.output)
# return tf.keras.models.Model(model.input, out)
def build_discriminator(image_size):
layers = [
tf.keras.layers.Conv2D(64, kernel_size=4, strides=2),
tf.keras.layers.LeakyReLU(alpha=0.01),
tf.keras.layers.Conv2D(128, kernel_size=4, strides=2),
tf.keras.layers.LeakyReLU(alpha=0.01),
tf.keras.layers.Conv2D(256, kernel_size=4, strides=2),
tf.keras.layers.LeakyReLU(alpha=0.01),
tf.keras.layers.Conv2D(512, kernel_size=4, strides=1),
tf.keras.layers.LeakyReLU(alpha=0.01),
tf.keras.layers.Conv2D(1, kernel_size=4, strides=1, use_bias=False),
tf.keras.layers.Activation(tf.keras.activations.sigmoid)
]
layers = [tfa.layers.SpectralNormalization(l) if type(
l) == tf.keras.layers.Conv2D else l for l in layers]
model = tf.keras.models.Sequential(
[*layers, tf.keras.layers.GlobalAveragePooling2D()])
return model
class BeautyGAN:
def __init__(self, image_size) -> None:
self.image_size = image_size
self.generator = build_generator(image_size)
self.discriminatorA = build_discriminator(image_size)
self.discriminatorB = build_discriminator(image_size)
self.adversarial_loss_A = tf.function(
partial(BeautyGAN.adversarial_loss, discriminator=self.discriminatorA))
self.adversarial_loss_B = tf.function(
partial(BeautyGAN.adversarial_loss, discriminator=self.discriminatorB))
self.optim = tf.keras.optimizers.Adam(learning_rate=0.0001)
self.mae = tf.keras.losses.MeanAbsoluteError(
reduction=tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE)
@tf.function
def cycle_consistency_loss(self, src_reconstruction, src_ground_truth, ref_reconstruction, ref_ground_truth):
# dist1 = tf.sqrt(tf.reduce_sum(tf.square(src_reconstruction - src_ground_truth), [1, 2, 3]))
# dist2 = tf.sqrt(tf.reduce_sum(tf.square(ref_reconstruction - ref_ground_truth), [1, 2, 3]))
# dist1 = tf.reduce_mean(tf.square(src_reconstruction - src_ground_truth))
# dist2 = tf.reduce_mean(tf.square(ref_reconstruction - ref_ground_truth))
dist1 = tf.reduce_mean(self.mae(src_reconstruction, src_ground_truth))
dist2 = tf.reduce_mean(self.mae(ref_reconstruction, ref_ground_truth))
return dist1+dist2
@staticmethod
def adversarial_loss(true, fake, discriminator):
eps = 0.1
D_fake = discriminator(fake)
if true is None:
return tf.reduce_mean((D_fake-1.0)**2)
D_true = discriminator(true)
return 0.5 * (tf.reduce_mean((D_true-1.0)**2) + tf.reduce_mean(D_fake**2))
# return 0.5 * (tf.reduce_mean(tf.math.log(1-D_fake)) + tf.reduce_mean(tf.math.log(D_true)))
# return 0.5 * (tf.reduce_mean((D_fake-1.0-eps)**2-eps**2) + tf.reduce_mean((D_true-eps)**2-eps**2))
@tf.function
def makeup_loss(self, fake, face_matched, lips_matched, eyes_matched, masks):
masks = tf.cast(masks, tf.float32)
l_face = self.mae(fake, face_matched, sample_weight=masks[:, 0][..., tf.newaxis])
l_lips = self.mae(fake, lips_matched, sample_weight=masks[:, 1][..., tf.newaxis])
l_eyes = self.mae(fake, eyes_matched, sample_weight=masks[:, 2][..., tf.newaxis])
return 0.1 * tf.reduce_mean(l_face) + 1.0 * tf.reduce_mean(l_lips) + 1.0 * tf.reduce_mean(l_eyes)
@tf.function
def _train_on_batch(self, source_batch, ref_batch):
fake_A, fake_B = self.generator([source_batch, ref_batch])
with tf.GradientTape() as tape:
# discriminators maximize adversarial loss
D_A_loss = self.adversarial_loss_A(source_batch, fake_B)
D_B_loss = self.adversarial_loss_B(ref_batch, fake_A)
D_loss = 0.5*(D_A_loss + D_B_loss)
vars = tape.watched_variables()
grads = tape.gradient(D_loss, vars)
self.optim.apply_gradients(zip(grads, vars))
# grads = tape.gradient(D_A_loss, self.discriminatorA.trainable_variables)
# self.optim.apply_gradients(zip(grads, self.discriminatorA.trainable_variables))
# grads = tape.gradient(D_B_loss, self.discriminatorB.trainable_variables)
# self.optim.apply_gradients(zip(grads, self.discriminatorB.trainable_variables))
face_matched_A, lips_matched_A, eyes_matched_A, src_masks, face_matched_B, lips_matched_B, eyes_matched_B, ref_masks = tf.py_function(
self.histogram_matching, inp=[fake_A, fake_B], Tout=[tf.float32, tf.float32, tf.float32, tf.bool, tf.float32, tf.float32, tf.float32, tf.bool])
with tf.GradientTape() as tape:
fake_A, fake_B = self.generator([source_batch, ref_batch])
D_A_loss2 = self.adversarial_loss_A(None, fake_B)
D_B_loss2 = self.adversarial_loss_B(None, fake_A)
adversarial_loss = 1*0.5*(D_A_loss2 + D_B_loss2)
perceptual_loss = 0.005 * \
(perceptualLoss(source_batch, fake_A) +
perceptualLoss(ref_batch, fake_B))
makeup_loss_A = self.makeup_loss(
fake_A, face_matched_A, lips_matched_A, eyes_matched_A, src_masks)
makeup_loss_B = self.makeup_loss(
fake_B, face_matched_B, lips_matched_B, eyes_matched_B, ref_masks)
makeup_loss = 50*0.5*(makeup_loss_A+makeup_loss_B)
rec_B, rec_A = self.generator([fake_B, fake_A])
cycle_loss = 10 * \
self.cycle_consistency_loss(
rec_A, source_batch, rec_B, ref_batch)
total_loss = cycle_loss + perceptual_loss + adversarial_loss + makeup_loss
grads = tape.gradient(total_loss, self.generator.trainable_variables)
self.optim.apply_gradients(
zip(grads, self.generator.trainable_variables))
return D_A_loss, D_B_loss, adversarial_loss, perceptual_loss, cycle_loss, makeup_loss, total_loss
def histogram_matching(self, fake_A, fake_B):
fake_A = fake_A.numpy()
fake_B = fake_B.numpy()
source_batch, ref_batch, src_masks, ref_masks, source_cdfs, ref_cdfs = self.current_args
face_matched_A = np.empty_like(fake_A)
lips_matched_A = np.empty_like(fake_A)
eyes_matched_A = np.empty_like(fake_A)
face_matched_B = np.empty_like(fake_B)
lips_matched_B = np.empty_like(fake_B)
eyes_matched_B = np.empty_like(fake_B)
fake_cdfs_A = make_cdfs(fake_A, src_masks)
fake_cdfs_B = make_cdfs(fake_B, ref_masks)
for i in range(fake_A.shape[0]):
face_matched_A[i] = match_histograms(
fake_A[i], fake_cdfs_A[i, 0], ref_cdfs[i, 0], src_masks[i, 0])
lips_matched_A[i] = match_histograms(
fake_A[i], fake_cdfs_A[i, 1], ref_cdfs[i, 1], src_masks[i, 1])
eyes_matched_A[i] = match_histograms(
fake_A[i], fake_cdfs_A[i, 2], ref_cdfs[i, 2], src_masks[i, 2])
face_matched_B[i] = match_histograms(
fake_B[i], fake_cdfs_B[i, 0], source_cdfs[i, 0], ref_masks[i, 0])
lips_matched_B[i] = match_histograms(
fake_B[i], fake_cdfs_B[i, 1], source_cdfs[i, 1], ref_masks[i, 1])
eyes_matched_B[i] = match_histograms(
fake_B[i], fake_cdfs_B[i, 2], source_cdfs[i, 2], ref_masks[i, 2])
return face_matched_A, lips_matched_A, eyes_matched_A, src_masks, face_matched_B, lips_matched_B, eyes_matched_B, ref_masks
def train_on_batch(self, source_batch, ref_batch, src_masks, ref_masks, source_cdfs, ref_cdfs):
self.current_args = (source_batch, ref_batch,
src_masks, ref_masks, source_cdfs, ref_cdfs)
return [x.numpy() for x in self._train_on_batch(source_batch, ref_batch)]
def predict(self, source_batch, ref_batch):
fake_A, fake_B = self.generator([source_batch, ref_batch])
rec_B, rec_A = self.generator([fake_B, fake_A])
return fake_A.numpy(), fake_B.numpy(), rec_A.numpy(), rec_B.numpy()
def save(self, filename):
self.discriminatorA.save_weights(
os.path.join(filename, 'discriminatorA' + '.h5'))
self.discriminatorB.save_weights(
os.path.join(filename, 'discriminatorB' + '.h5'))
self.generator.save_weights(
os.path.join(filename, 'generator' + '.h5'))
# with open(os.path.join(filename, 'optimizer.json'), 'w') as f:
# json.dump(self.optim.get_config(), f)
def load(self, filename):
self.discriminatorA.load_weights(
os.path.join(filename, 'discriminatorA' + '.h5'))
self.discriminatorB.load_weights(
os.path.join(filename, 'discriminatorB' + '.h5'))
self.generator.load_weights(
os.path.join(filename, 'generator' + '.h5'))