-
Notifications
You must be signed in to change notification settings - Fork 55
/
vargface.py
167 lines (138 loc) · 7.8 KB
/
vargface.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
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import backend as K
"""Building Block Functions"""
BATCH_NORM_DECAY = 0.9
BATCH_NORM_EPSILON = 1e-5
CONV_KERNEL_INITIALIZER = keras.initializers.VarianceScaling(scale=2.0, mode="fan_out", distribution="truncated_normal")
# CONV_KERNEL_INITIALIZER = 'glorot_uniform'
def hard_swish(inputs, name=None):
"""`out = xx * relu6(xx + 3) / 6`, arxiv: https://arxiv.org/abs/1905.02244"""
return keras.layers.Multiply(name=name)([inputs, tf.nn.relu6(inputs + 3) / 6])
def activation_by_name(inputs, activation="relu", name=None):
layer_name = name and activation and name + activation
if activation == "hard_swish":
return hard_swish(inputs, name=name)
elif activation.lower() == "prelu":
shared_axes = list(range(1, len(inputs.shape)))
shared_axes.pop(-1 if K.image_data_format() == "channels_last" else 0)
# print(f"{shared_axes = }")
return keras.layers.PReLU(shared_axes=shared_axes, alpha_initializer=tf.initializers.Constant(0.25), name=layer_name)(inputs)
elif activation:
return keras.layers.Activation(activation=activation, name=layer_name)(inputs)
else:
return inputs
def batchnorm_with_activation(inputs, activation="relu", zero_gamma=False, epsilon=BATCH_NORM_EPSILON, name=None):
"""Performs a batch normalization followed by an activation."""
bn_axis = -1 if K.image_data_format() == "channels_last" else 1
gamma_initializer = tf.zeros_initializer() if zero_gamma else tf.ones_initializer()
nn = keras.layers.BatchNormalization(
axis=bn_axis,
momentum=BATCH_NORM_DECAY,
epsilon=epsilon,
gamma_initializer=gamma_initializer,
name=name and name + "bn",
)(inputs)
if activation:
nn = activation_by_name(nn, activation=activation, name=name)
return nn
def conv2d_no_bias(inputs, filters, kernel_size, strides=1, padding="VALID", use_bias=False, groups=1, use_torch_padding=True, name=None, **kwargs):
pad = max(kernel_size) // 2 if isinstance(kernel_size, (list, tuple)) else kernel_size // 2
if use_torch_padding and padding.upper() == "SAME" and pad != 0:
inputs = keras.layers.ZeroPadding2D(padding=pad, name=name and name + "pad")(inputs)
padding = "VALID"
return keras.layers.Conv2D(
filters,
kernel_size,
strides=strides,
padding=padding,
use_bias=use_bias,
groups=groups,
kernel_initializer=CONV_KERNEL_INITIALIZER,
name=name and name + "conv",
**kwargs,
)(inputs)
def se_block(inputs, se_ratio=0.25, activation="relu", use_bias=True, name=None):
channel_axis = 1 if K.image_data_format() == "channels_first" else -1
h_axis, w_axis = [2, 3] if K.image_data_format() == "channels_first" else [1, 2]
filters = inputs.shape[channel_axis]
reduction = int(filters * se_ratio)
se = tf.reduce_mean(inputs, [h_axis, w_axis], keepdims=True)
se = keras.layers.Conv2D(reduction, kernel_size=1, use_bias=use_bias, kernel_initializer=CONV_KERNEL_INITIALIZER, name=name and name + "1_conv")(se)
se = activation_by_name(se, activation=activation, name=name)
se = keras.layers.Conv2D(filters, kernel_size=1, use_bias=use_bias, kernel_initializer=CONV_KERNEL_INITIALIZER, name=name and name + "2_conv")(se)
se = activation_by_name(se, activation="sigmoid", name=name)
return keras.layers.Multiply(name=name and name + "out")([inputs, se])
def separable_conv2d(inputs, hidden_ratio, out_channel, kernel_size=3, strides=1, group_size=8, activation="relu", name=None):
channel_axis = 1 if K.image_data_format() == "channels_first" else -1
input_filters = inputs.shape[channel_axis]
hidden_dim = int(input_filters * hidden_ratio)
nn = conv2d_no_bias(inputs, hidden_dim, kernel_size, strides, padding="SAME", groups=input_filters // group_size, name=name + "1_")
nn = batchnorm_with_activation(nn, activation=activation, name=name + "1_")
nn = conv2d_no_bias(nn, out_channel, 1, 1, name=name + "2_")
nn = batchnorm_with_activation(nn, activation=None, name=name + "2_")
return nn
def vargface_block(inputs, out_channel, conv_shortcut=False, is_merge=False, strides=1, hidden_ratio=2, se_ratio=0, activation="relu", name=None):
kernel_size = 3
if conv_shortcut:
shortcut = separable_conv2d(inputs, hidden_ratio, out_channel, kernel_size, strides, activation=activation, name=name + "shortcut_")
else:
shortcut = inputs
if is_merge:
sep_1 = separable_conv2d(inputs, hidden_ratio, out_channel, kernel_size, strides, activation=activation, name=name + "sep1_1_")
sep_2 = separable_conv2d(inputs, hidden_ratio, out_channel, kernel_size, strides, activation=activation, name=name + "sep1_2_")
nn = keras.layers.Add(name=name + "merge_")([sep_1, sep_2])
else:
nn = separable_conv2d(inputs, hidden_ratio, out_channel, kernel_size, strides, activation=activation, name=name + "sep1_")
nn = activation_by_name(nn, activation, name=name + "sep1_")
nn = separable_conv2d(nn, hidden_ratio, out_channel, kernel_size, 1, activation=activation, name=name + "sep2_")
if se_ratio > 0:
nn = se_block(nn, se_ratio, activation=activation, name=name + "se_")
# print(shortcut.shape, nn.shape)
out = keras.layers.Add()([shortcut, nn])
out = activation_by_name(out, activation, name=name + "output_")
return out
def vargface_stack(inputs, num_block, out_channel, strides=2, hidden_ratio=2, se_ratio=0, activation="relu", name=None):
nn = inputs
for id in range(num_block):
block_name = name + "block{}_".format(id + 1)
is_merge = True if id == 0 else False
conv_shortcut = True if id == 0 else False
cur_strides = strides if id == 0 else 1
cur_se_ratio = 0 if id == 0 else se_ratio
nn = vargface_block(nn, out_channel, conv_shortcut, is_merge, cur_strides, hidden_ratio, cur_se_ratio, activation, block_name)
return nn
def vargface_stem(inputs, stem_width=32, strides=1, se_ratio=0, activation="relu", name=None):
nn = conv2d_no_bias(inputs, stem_width, kernel_size=3, strides=strides, padding="SAME", name=name + "1_")
nn = batchnorm_with_activation(nn, activation=activation, name=name + "1_")
nn = vargface_block(nn, stem_width, conv_shortcut=True, strides=2, hidden_ratio=1, se_ratio=se_ratio, activation=activation, name=name)
return nn
def vargface_output(inputs, output_num_features, group_size=8, activation="relu", name=None):
nn = conv2d_no_bias(inputs, output_num_features, 1, 1, name=name + "1_")
nn = batchnorm_with_activation(nn, activation=activation, name=name + "1_")
nn = conv2d_no_bias(nn, output_num_features, kernel_size=int(nn.shape[1]), groups=output_num_features // group_size, name=name + "2_")
nn = batchnorm_with_activation(nn, activation=None, name=name + "2_")
return nn
def VargFace(
num_blocks=[3, 7, 4],
out_channels=[64, 128, 256],
channel_expand=1,
hidden_ratio=2,
stem_width=32,
stem_strides=1,
se_ratio=0,
input_shape=(112, 112, 3),
activation="relu",
output_num_features=1024,
model_name="vargface",
):
inputs = keras.layers.Input(input_shape)
stem_width = int(stem_width * channel_expand)
out_channels = [int(ii * channel_expand) for ii in out_channels]
nn = vargface_stem(inputs, stem_width, stem_strides, se_ratio=se_ratio, activation=activation, name="stem_")
for id, (num_block, out_channel) in enumerate(zip(num_blocks, out_channels)):
name = "stack{}_".format(id + 1)
nn = vargface_stack(nn, num_block, out_channel, hidden_ratio=hidden_ratio, se_ratio=se_ratio, activation=activation, name=name)
nn = vargface_output(nn, output_num_features, activation=activation, name="output_")
model = keras.models.Model(inputs, nn, name=model_name)
return model