utils.py
5.84 KB
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
# -*- coding: utf-8 -*-
# @Author : lk
# @Email : 9428.al@gmail.com
# @Create Date : 2022-04-19 17:24:28
# @Last Modified : 2022-04-23 15:07:48
# @Description :
import os
import cv2
import time
import numpy as np
import tensorflow as tf
def MobileUNet():
def dw_conv_bn_relu(inputs, strides):
kernel_init = tf.initializers.RandomNormal(0.0, 0.01)
x = tf.keras.layers.DepthwiseConv2D(kernel_size=3,
strides=strides,
padding="same",
kernel_initializer=kernel_init)(inputs)
x = tf.keras.layers.BatchNormalization(axis=-1, momentum=0.99, epsilon=0.001)(x)
x = tf.keras.layers.ReLU(6.)(x)
return x
def deconv_bn_relu(inputs, filters, strides):
kernel_init = tf.initializers.RandomNormal(0.0, 0.01)
x = tf.keras.layers.Conv2DTranspose(filters,
kernel_size=3,
strides=strides,
padding="same",
kernel_initializer=kernel_init)(inputs)
x = tf.keras.layers.BatchNormalization(axis=-1, momentum=0.99, epsilon=0.001)(x)
x = tf.keras.layers.ReLU(6.)(x)
return x
def build_head(inputs, output_filters, bias_init):
kernel_init = tf.initializers.RandomNormal(0.0, 0.01)
x = dw_conv_bn_relu(inputs, strides=1)
x = dw_conv_bn_relu(x, strides=1)
x = dw_conv_bn_relu(x, strides=1)
x = dw_conv_bn_relu(x, strides=1)
x = tf.keras.layers.Conv2D(output_filters,
kernel_size=1,
strides=1,
padding="same",
kernel_initializer=kernel_init,
bias_initializer=bias_init)(x)
return x
mobilenetv2 = tf.keras.applications.MobileNetV2(
include_top=False, input_shape=[1024, 1024, 3], alpha=1.0
)
c3_output, c4_output, c5_output = [
mobilenetv2.get_layer(layer_name).output
for layer_name in ["block_5_add", "block_12_add", "block_15_add"]
]
c6_output = dw_conv_bn_relu(c5_output, strides=2)
c7_output = dw_conv_bn_relu(c6_output, strides=2)
# FPN
c7_output_up = deconv_bn_relu(c7_output, filters=c6_output.shape[-1], strides=2)
c6_output = tf.keras.layers.Add()([c6_output, c7_output_up])
c6_output_up = deconv_bn_relu(c6_output, filters=c5_output.shape[-1], strides=2)
c5_output = tf.keras.layers.Add()([c5_output, c6_output_up])
c5_output_up = deconv_bn_relu(c5_output, filters=c4_output.shape[-1], strides=2)
c4_output = tf.keras.layers.Add()([c4_output, c5_output_up])
c4_output_up = deconv_bn_relu(c4_output, filters=c3_output.shape[-1], strides=2)
c3_output = tf.keras.layers.Add()([c3_output, c4_output_up])
quarter_maps = deconv_bn_relu(c3_output, filters=256, strides=2)
a_half_maps = deconv_bn_relu(quarter_maps, filters=128, strides=2)
one_one_maps = deconv_bn_relu(a_half_maps, filters=128, strides=2)
prob_maps = tf.keras.layers.Conv2D(filters=1, kernel_size=1, strides=1, activation='sigmoid', name='probability_maps_layer')(one_one_maps)
thre_maps = tf.keras.layers.Conv2D(filters=1, kernel_size=1, strides=1, activation='sigmoid', name='threshold_maps_layer')(one_one_maps)
out = tf.keras.layers.Concatenate(axis=-1)([prob_maps, thre_maps])
# binary_maps = 1 / (1 + tf.math.exp(-50 * (prob_maps - thre_maps)))
model = tf.keras.Model(inputs=mobilenetv2.inputs, outputs=out)
return model
class DBNetLoss(tf.losses.Loss):
"""Wrapper to combine both the losses"""
def __init__(self):
super(DBNetLoss, self).__init__(reduction="auto", name="DBNetLoss")
self._bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=False, reduction=tf.keras.losses.Reduction.NONE)
def call(self, y_true, y_pred):
prob_map_label, thre_map_label, binary_map_label = tf.split(y_true, num_or_size_splits=3, axis=3)
prob_map_pred, thre_map_pred = tf.split(y_pred, num_or_size_splits=2, axis=3)
binary_map_pred = 1 / (1 + tf.math.exp(-50 * (prob_map_pred - thre_map_pred)))
prob_loss = self._bce_loss(prob_map_label, prob_map_pred)
binary_loss = self._bce_loss(binary_map_label, binary_map_pred)
loss = 10 * prob_loss + binary_loss
return loss
def resize_with_padding(src, limit_max=1024):
'''限制长边不大于 limit_max 短边等比例缩放，以 0 填充'''
img = src.copy()
h, w, _ = img.shape
max_side = max(h, w)
ratio = limit_max / max_side if max_side > limit_max else 1
h, w = int(h * ratio), int(w * ratio)
proc = cv2.resize(img, (w, h))
canvas = np.zeros((limit_max, limit_max, 3), dtype=np.float32)
canvas[0:h, 0:w, :] = proc
return canvas, ratio
if __name__ == '__main__':
model = MobileUNet()
model.summary()
model.save('model.h5')
file_dir = '/home/lk/MyProject/BMW/数据集/文件分类/其他/其他'
mean_resize_time = list()
mean_infer_time = list()
for fn in os.listdir(file_dir):
file_path = os.path.join(file_dir, fn)
image = cv2.imread(file_path)
t0 = time.time()
image, ratio = resize_with_padding(image)
inputs = np.expand_dims(np.float32(image/255.), axis=0)
t1 = time.time()
mean_resize_time.append(t1-t0)
# print(int(np.mean(mean_resize_time)*1000))
# save_path = os.path.join('./tmp', fn)
# cv2.imwrite(save_path, image)
t2 = time.time()
predictions = model.predict(inputs)
t3 = time.time()
mean_infer_time.append(t3-t2)
print(int((t3-t2)*1000))
# print(int(np.mean(mean_infer_time)*1000))