/
ops.py
196 lines (178 loc) · 7.81 KB
/
ops.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
# -*- coding: utf-8 -*-
"""
Created on Tue Oct 30 16:07:54 2018
@author: Neoooli
"""
from __future__ import absolute_import, division, print_function, unicode_literals
import numpy as np
import tensorflow as tf
from collections import OrderedDict
from tensorflow.keras.layers import *
from tensorflow.keras.models import Sequential
from gdaldiy import *
def conv2d(input_,output_dim,kernel_size=3,stride=2,padding="SAME",biased=True):
return Conv2D(output_dim,kernel_size=[kernel_size,kernel_size],strides=[stride,stride],padding=padding,use_bias=biased)(input_)
def deconv2d(input_,output_dim,kernel_size=4,stride=2,padding="SAME",biased=True):
return Conv2DTranspose(output_dim,kernel_size=[kernel_size,kernel_size],strides=[stride,stride],padding=padding,use_bias=biased)(input_)
def DSC(input_, output_dim,kernel_size=3, stride=1, padding="SAME",scale=1, biased=True):
return SeparableConv2D(input_.shape[-1],kernel_size=[kernel_size,kernel_size],strides=[stride,stride],padding=padding,depth_multiplier=scale,use_bias=biased)(input_)
def MDSC(input_,output_dim,kernel_list=[3,5], stride=1, padding="SAME",scale=1,biased=True):
depthoutput_list=[]
for i in range(len(kernel_list)):
depth_output=DepthwiseConv2D(kernel_size=kernel_list[i],strides=stride,padding=padding,depth_multiplier=scale)
depthoutput_list.append(depth_output(input_))
output = concatenate(depthoutput_list,axis=-1)
output = conv2d(output,output_dim,kernel_size=1,stride=1,padding=padding,biased=biased)
return output
def SDC(input_,output_dim, kernel_size=3,stride=1,dilation=2,padding='SAME', biased=True):
"""
Smoothed dilated conv2d via the Separable and Shared Convolution (SSC) without BN or relu.
"""
input_dim = input_.shape[-1]
fix_w_size = dilation * 2 - 1
eo = tf.expand_dims(input_,-1)
o = Conv3D(1,kernel_size=[fix_w_size, fix_w_size,1],strides=[stride,stride,stride],padding=padding,use_bias=biased)(eo)
o = eo + o
o = tf.squeeze(o,-1)
o = Conv2D(output_dim,kernel_size=[kernel_size,kernel_size],strides=[stride,stride],padding=padding,dilation_rate=(dilation, dilation),use_bias=biased)(o)
return o
def SDRB(input_, kernel_size=3,stride=1,dilation=2,training=False, biased=True):
output_dim=input_.get_shape()[-1]
sconv1=SDC(input_,output_dim, kernel_size,stride,dilation,biased=biased)
sconv1=batch_norm(sconv1,training)
sconv1= relu(sconv1)
sconv2=SDC(sconv1,output_dim, kernel_size,stride,dilation,biased=biased)
sconv2=batch_norm(sconv2,training)
return relu(sconv2+input_)
def relu(input_):
return ReLU()(input_)
def lrelu(input_):
return LeakyReLU()(input_)
def avg_pooling(input_,kernel_size=2,stride=2,padding="same"):
return tf.keras.layers.AveragePooling2D((kernel_size,kernel_size),stride,padding)(input_)
def max_pooling(input_,kernel_size=2,stride=2,padding="same"):
return tf.keras.layers.MaxPool2D((kernel_size,kernel_size),stride,padding)(input_)
def dropout(input_,rate=0.2,training=True):
"""
rate是丢掉多少神经元.
"""
return tf.keras.layers.Dropout(rate)(input_,training)
def GAP(input_):
return GlobalAveragePooling2D()(input_)
def batch_norm(input_,training=True):
return BatchNormalization()(input_,training)
class InstanceNormalization(tf.keras.layers.Layer):
def __init__(self, epsilon=1e-5):
super(InstanceNormalization, self).__init__()
self.epsilon = epsilon
def build(self, input_shape):
self.scale = self.add_weight(
name='scale',
shape=input_shape[-1:],
initializer=tf.random_normal_initializer(1., 0.02),
trainable=True)
self.offset = self.add_weight(
name='offset',
shape=input_shape[-1:],
initializer='zeros',
trainable=True)
def call(self, x):
mean, variance = tf.nn.moments(x, axes=[1, 2], keepdims=True)
inv = tf.math.rsqrt(variance + self.epsilon)
normalized = (x - mean) * inv
return self.scale * normalized + self.offset
def instance_norm(input_):
return InstanceNormalization()(input_)
def norm(input_,norm='batch_norm',training=True):
if norm==None:
return input_
elif norm=='batch_norm':
return BatchNormalization()(input_,training)
elif norm=='instance_norm':
return InstanceNormalization()(input_)
def act(input_,activation='relu'):
if activation==None:
return input_
elif activation=='relu':
return ReLU()(input_)
elif activation=='lrelu':
return LeakyReLU(alpha=0.2)(input_)
def diydecay(steps,baselr,cycle_step=100000,decay_steps=100,decay_rate=0.96):
n=steps//cycle_step
clr=baselr*(0.8**n)
steps=steps-n*cycle_step
k=steps//decay_steps
i=(-1)**k
step=((i+1)/2)*steps-i*((k+1)//2)*decay_steps
dlr = clr*decay_rate**(int(step))
return dlr
def decay(global_steps,baselr,start_decay_step=100000,cycle_step=100000,decay_steps=100,decay_rate=0.96):
lr=np.where(np.greater_equal(global_steps,start_decay_step),
diydecay(global_steps-start_decay_step,baselr,cycle_step,decay_steps,decay_rate),
baselr)
return lr
def l2_loss(x):
return tf.sqrt(tf.reduce_sum(x**2))
def randomflip(input_,n):
#生成-3到2的随机整数，-1顺时针90度，-2顺时针180，-3顺时针270,0垂直翻转，1水平翻转，2不变
if n<0:
return np.rot90(input_,n)
elif -1<n<2:
return np.flip(input_,n)
else:
return input_
def read_img(datapath,scale=255):
img=imgread(datapath)/scale
return img
def read_imgs(datapath,scale=255,k=2):
img_list=[]
l=len(datapath)
for i in range(l):
img=read_img(datapath[i],scale)
img = randomflip(img,k)
img=img[np.newaxis,:]
img_list.append(img)
imgs=np.concatenate(img_list,axis=0)
return tf.convert_to_tensor(imgs,tf.float32)
def read_labels(datapath,k=2):
img_list=[]
l=len(datapath)
for i in range(l):
img=imgread(datapath[i])
img=randomflip(img,k)
img=img[np.newaxis,:]
img_list.append(img)
imgs=np.concatenate(img_list,axis=0)
imgs = rgb_to_gray(imgs)
return tf.convert_to_tensor(imgs,tf.float32)
rgb_colors=OrderedDict([
("cloud-free",np.array([0],dtype=np.uint8)),
("cloud",np.array([255],dtype=np.uint8))])
def rgb_to_gray(rgb_mask):
label = (np.zeros(rgb_mask.shape[:3]+tuple([1]))).astype(np.uint8)
if len(rgb_mask.shape)==4:
for gray, (class_name,rgb_values) in enumerate(rgb_colors.items()):
match_pixs = np.where((rgb_mask == np.asarray(rgb_values)).astype(np.uint8).sum(-1) == 3)
label[match_pixs] = gray
else:
for gray, (class_name,rgb_values) in enumerate(rgb_colors.items()):
match_pixs = np.where((rgb_mask == np.asarray(rgb_values)).astype(np.uint8) == 1)
label[match_pixs] = gray
return label.astype(np.uint8)
#输入shape=(w,h,c)/(batch_size,w,h,c)
def label_to_rgb(labels):
max_index=np.argmax(labels,axis=-1)#第三维上最大值的索引，返回其他维度，并在并对位置填上最大值之索引
n=len(labels.shape)-1
if labels.shape[-1]<3:
rgb = (np.zeros(labels.shape[:n])).astype(np.uint8)
else:
rgb = (np.zeros(labels.shape[:n]+tuple([3]))).astype(np.uint8)
for gray, (class_name,rgb_values) in enumerate(rgb_colors.items()):
match_pixs = np.where(max_index == gray)
rgb[match_pixs] = rgb_values
return rgb.astype(np.uint8)
def down_sample(input_,kernel_size,classes):
onehot=tf.one_hot(tf.cast(input_,dtype=tf.int32),classes)
onehot=avg_pooling(onehot,kernel_size,kernel_size)
onehot=tf.argmax(onehot,axis=-1)
return onehot