代码如下：
1 from model import *
2 from data import *#导入这两个文件中的所有函数
3
4 #os.environ[“CUDA_VISIBLE_DEVICES”] = “0”
5
6
7 data_gen_args = dict(rotation_range=0.2,
8 width_shift_range=0.05,
9 height_shift_range=0.05,
10 shear_range=0.05,
11 zoom_range=0.05,
12 horizontal_flip=True,
13 fill_mode=‘nearest’)#数据增强时的变换方式的字典
14 myGene = trainGenerator(2,‘data/membrane/train’,‘image’,‘label’,data_gen_args,save_to_dir = None)
15 #得到一个生成器，以batch=2的速率无限生成增强后的数据
16
17 model = unet()
18 model_checkpoint = ModelCheckpoint(‘unet_membrane.hdf5’, monitor=‘loss’,verbose=1, save_best_only=True)
19 #回调函数，第一个是保存模型路径，第二个是检测的值，检测Loss是使它最小，第三个是只保存在验证集上性能最好的模型
20
21 model.fit_generator(myGene,steps_per_epoch=300,epochs=1,callbacks=[model_checkpoint])
22 #steps_per_epoch指的是每个epoch有多少个batch_size，也就是训练集总样本数除以batch_size的值
23 #上面一行是利用生成器进行batch_size数量的训练，样本和标签通过myGene传入
24 testGene = testGenerator(“data/membrane/test”)
25 results = model.predict_generator(testGene,30,verbose=1)
26 #30是step,steps: 在停止之前，来自 generator 的总步数 (样本批次)。 可选参数 Sequence：如果未指定，将使用len(generator) 作为步数。
27 #上面的返回值是：预测值的 Numpy 数组。
28 saveResult(“data/membrane/test”,results)#保存结果
29 1
30 data.py文件：
31
32 from __future__ import print_function
33 from keras.preprocessing.image import ImageDataGenerator
34 import numpy as np
35 import os
36 import glob
37 import skimage.io as io
38 import skimage.transform as trans
39
40 Sky = [128,128,128]
41 Building = [128,0,0]
42 Pole = [192,192,128]
43 Road = [128,64,128]
44 Pavement = [60,40,222]
45 Tree = [128,128,0]
46 SignSymbol = [192,128,128]
47 Fence = [64,64,128]
48 Car = [64,0,128]
49 Pedestrian = [64,64,0]
50 Bicyclist = [0,128,192]
51 Unlabelled = [0,0,0]
52
53 COLOR_DICT = np.array([Sky, Building, Pole, Road, Pavement,
54 Tree, SignSymbol, Fence, Car, Pedestrian, Bicyclist, Unlabelled])
55
56
57 def adjustData(img,mask,flag_multi_class,num_class):
58 if(flag_multi_class):#此程序中不是多类情况，所以不考虑这个
59 img = img / 255
60 mask = mask[:,:,:,0] if(len(mask.shape) == 4) else mask[:,:,0]
61 #if else的简洁写法，一行表达式，为真时放在前面，不明白mask.shape=4的情况是什么，由于有batch_size，所以mask就有3维[batch_size,wigth,heigh],估计mask[:,:,0]是写错了，应该写成[0,:,:],这样可以得到一片图片，
62 new_mask = np.zeros(mask.shape + (num_class,))
63 #np.zeros里面是shape元组，此目的是将数据厚度扩展到num_class层，以在层的方向实现one-hot结构
64
65 for i in range(num_class):
66 #for one pixel in the image, find the class in mask and convert it into one-hot vector
67 #index = np.where(mask == i)
68 #index_mask = (index[0],index[1],index[2],np.zeros(len(index[0]),dtype = np.int64) + i) if (len(mask.shape) == 4) else (index[0],index[1],np.zeros(len(index[0]),dtype = np.int64) + i)
69 #new_mask[index_mask] = 1
70 new_mask[mask == i,i] = 1#将平面的mask的每类，都单独变成一层，
71 new_mask = np.reshape(new_mask,(new_mask.shape[0],new_mask.shape[1]*new_mask.shape[2],new_mask.shape[3])) if flag_multi_class else np.reshape(new_mask,(new_mask.shape[0]*new_mask.shape[1],new_mask.shape[2]))
72 mask = new_mask
73 elif(np.max(img) > 1):
74 img = img / 255
75 mask = mask /255
76 mask[mask > 0.5] = 1
77 mask[mask <= 0.5] = 0
78 return (img,mask)
79 #上面这个函数主要是对训练集的数据和标签的像素值进行归一化
80
81
82 def trainGenerator(batch_size,train_path,image_folder,mask_folder,aug_dict,image_color_mode = "grayscale",
83 mask_color_mode = "grayscale",image_save_prefix = "image",mask_save_prefix = "mask",
84 flag_multi_class = False,num_class = 2,save_to_dir = None,target_size = (256,256),seed = 1):
85 '''
86 can generate image and mask at the same time
87 use the same seed for image_datagen and mask_datagen to ensure the transformation for image and mask is the same
88 if you want to visualize the results of generator, set save_to_dir = "your path"
89 '''
90 image_datagen = ImageDataGenerator(**aug_dict)
91 mask_datagen = ImageDataGenerator(**aug_dict)
92 image_generator = image_datagen.flow_from_directory(#https://blog.csdn.net/nima1994/article/details/80626239
93 train_path,#训练数据文件夹路径
94 classes = [image_folder],#类别文件夹,对哪一个类进行增强
95 class_mode = None,#不返回标签
96 color_mode = image_color_mode,#灰度，单通道模式
97 target_size = target_size,#转换后的目标图片大小
98 batch_size = batch_size,#每次产生的（进行转换的）图片张数
99 save_to_dir = save_to_dir,#保存的图片路径
100 save_prefix = image_save_prefix,#生成图片的前缀，仅当提供save_to_dir时有效
101 seed = seed)
102 mask_generator = mask_datagen.flow_from_directory(
103 train_path,
104 classes = [mask_folder],
105 class_mode = None,
106 color_mode = mask_color_mode,
107 target_size = target_size,
108 batch_size = batch_size,
109 save_to_dir = save_to_dir,
110 save_prefix = mask_save_prefix,
111 seed = seed)
112 train_generator = zip(image_generator, mask_generator)#组合成一个生成器
113 for (img,mask) in train_generator:
114 #由于batch是2，所以一次返回两张，即img是一个2张灰度图片的数组，[2,256,256]
115 img,mask = adjustData(img,mask,flag_multi_class,num_class)#返回的img依旧是[2,256,256]
116 yield (img,mask)
117 #每次分别产出两张图片和标签，不懂yield的请看https://blog.csdn.net/mieleizhi0522/article/details/82142856
118 # yield相当于生成器
119 #上面这个函数主要是产生一个数据增强的图片生成器，方便后面使用这个生成器不断生成图片
120
121
122 def testGenerator(test_path,num_image = 30,target_size = (256,256),flag_multi_class = False,as_gray = True):
123 for i in range(num_image):
124 img = io.imread(os.path.join(test_path,"%d.png"%i),as_gray = as_gray)
125 img = img / 255
126 img = trans.resize(img,target_size)
127 img = np.reshape(img,img.shape+(1,)) if (not flag_multi_class) else img
128 img = np.reshape(img,(1,)+img.shape)
129 #将测试图片扩展一个维度，与训练时的输入[2,256,256]保持一致
130 yield img
131
132 #上面这个函数主要是对测试图片进行规范，使其尺寸和维度上和训练图片保持一致
133
134 def geneTrainNpy(image_path,mask_path,flag_multi_class = False,num_class = 2,image_prefix = "image",mask_prefix = "mask",image_as_gray = True,mask_as_gray = True):
135 image_name_arr = glob.glob(os.path.join(image_path,"%s*.png"%image_prefix))
136 #相当于文件搜索，搜索某路径下与字符匹配的文件https://blog.csdn.net/u010472607/article/details/76857493/
137 image_arr = []
138 mask_arr = []
139 for index,item in enumerate(image_name_arr):#enumerate是枚举，输出[(0,item0),(1,item1),(2,item2)]
140 img = io.imread(item,as_gray = image_as_gray)
141 img = np.reshape(img,img.shape + (1,)) if image_as_gray else img
142 mask = io.imread(item.replace(image_path,mask_path).replace(image_prefix,mask_prefix),as_gray = mask_as_gray)
143 #重新在mask_path文件夹下搜索带有mask字符的图片（标签图片）
144 mask = np.reshape(mask,mask.shape + (1,)) if mask_as_gray else mask
145 img,mask = adjustData(img,mask,flag_multi_class,num_class)
146 image_arr.append(img)
147 mask_arr.append(mask)
148 image_arr = np.array(image_arr)
149 mask_arr = np.array(mask_arr)#转换成array
150 return image_arr,mask_arr
151 #该函数主要是分别在训练集文件夹下和标签文件夹下搜索图片，然后扩展一个维度后以array的形式返回，是为了在没用数据增强时的读取文件夹内自带的数据
152
153
154 def labelVisualize(num_class,color_dict,img):
155 img = img[:,:,0] if len(img.shape) == 3 else img
156 img_out = np.zeros(img.shape + (3,))
157 #变成RGB空间，因为其他颜色只能再RGB空间才会显示
158 for i in range(num_class):
159 img_out[img == i,:] = color_dict[i]
160 #为不同类别涂上不同的颜色，color_dict[i]是与类别数有关的颜色，img_out[img == i,:]是img_out在img中等于i类的位置上的点
161 return img_out / 255
162
163 #上面函数是给出测试后的输出之后，为输出涂上不同的颜色，多类情况下才起作用，两类的话无用
164
165 def saveResult(save_path,npyfile,flag_multi_class = False,num_class = 2):
166 for i,item in enumerate(npyfile):
167 img = labelVisualize(num_class,COLOR_DICT,item) if flag_multi_class else item[:,:,0]
168 #多类的话就图成彩色，非多类（两类）的话就是黑白色
169 io.imsave(os.path.join(save_path,"%d_predict.png"%i),img)19
194 #下面是model.py：
195
196
197
198 import numpy as np
199 import os
200 import skimage.io as io
201 import skimage.transform as trans
202 import numpy as np
203 from keras.models import *
204 from keras.layers import *
205 from keras.optimizers import *
206 from keras.callbacks import ModelCheckpoint, LearningRateScheduler
207 from keras import backend as keras
208
209
210 def unet(pretrained_weights = None,input_size = (256,256,1)):
211 inputs = Input(input_size)
212 conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(inputs)
213 conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv1)
214 pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
215 conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool1)
216 conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv2)
217 pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
218 conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool2)
219 conv3 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv3)
220 pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
221 conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool3)
222 conv4 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv4)
223 drop4 = Dropout(0.5)(conv4)
224 pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)
225
226 conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(pool4)
227 conv5 = Conv2D(1024, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv5)
228 drop5 = Dropout(0.5)(conv5)
229
230 up6 = Conv2D(512, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(drop5))#上采样之后再进行卷积，相当于转置卷积操作！
231 merge6 = concatenate([drop4,up6],axis=3)
232 conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge6)
233 conv6 = Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv6)
234
235 up7 = Conv2D(256, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv6))
236 merge7 = concatenate([conv3,up7],axis = 3)
237 conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge7)
238 conv7 = Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv7)
239
240 up8 = Conv2D(128, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv7))
241 merge8 = concatenate([conv2,up8],axis = 3)
242 conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge8)
243 conv8 = Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv8)
244
245 up9 = Conv2D(64, 2, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(UpSampling2D(size = (2,2))(conv8))
246 merge9 = concatenate([conv1,up9],axis = 3)
247 conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(merge9)
248 conv9 = Conv2D(64, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)
249 conv9 = Conv2D(2, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal')(conv9)
250 conv10 = Conv2D(1, 1, activation = 'sigmoid')(conv9)#我怀疑这个sigmoid激活函数是多余的，因为在后面的loss中用到的就是二进制交叉熵，包含了sigmoid
251
252 model = Model(input = inputs, output = conv10)
253
254 model.compile(optimizer = Adam(lr = 1e-4), loss = 'binary_crossentropy', metrics = ['accuracy'])#模型执行之前必须要编译https://keras-cn.readthedocs.io/en/latest/getting_started/sequential_model/
255 #利用二进制交叉熵，也就是sigmoid交叉熵，metrics一般选用准确率，它会使准确率往高处发展
256 #model.summary()
257
258 if(pretrained_weights):
259 model.load_weights(pretrained_weights)
260
261 return model
262
263
来源：CSDN
作者：树莓派派酒
链接：https://blog.csdn.net/qq_35054151/article/details/104742897