devbox
我家小花儿
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

基于深度学习的人脸识别系统：卷积神经网络实现(VIPLFaceNet、VGGNet、Xception、ResNet50、ResNet18)、AM-Softmax损失

作者：我家小花儿 | 2024-03-25 06:27:55

赞

踩

viplfacenet

日萌社

人工智能AI：Keras PyTorch MXNet TensorFlow PaddlePaddle 深度学习实战（不定时更新）

ResNet结构并分析其在计算机视觉方面取得成功的原因

Resnet通过引入残差单元来解决退化问题。
结构：
（1）通过增加 恒等快捷连接（identity shortcut connection）实现，直接跳过一个或多个层。优势是残差映射在实际中往往更容易优化。
（2）Resnet网络中短路连接shortcut的方式考虑到x的维度与F(X)维度可能不匹配情况，需进行维度匹配。通常采用两种方法解决这一问题:
zero_padding:对恒等层进行0填充的方式将维度补充完整。这种方法不会增加额外的参数
projection:在恒等层采用1x1的卷积核来增加维度。这种方法会增加额外的参数
（3）bottleneck 实现方式
使用1x1的卷积层来降维，最后一层使用1x1的卷积层来进行升维
（4）激活函数移动到残差部分可以提高模型的精度

  1. train.py
  2.  
  3. #!/usr/bin env python
  4. from keras.callbacks import EarlyStopping, ReduceLROnPlateau, CSVLogger, ModelCheckpoint
  5. # DO NOT REMOVE THIS:
  6. from model.cnn_models import *
  7. from utils.data_generator import DataGenerator
  8. from model.amsoftmax import wrap_cnn, amsoftmax_loss
  9.  
  10. input_shape = (64, 64, 1)
  11. batch_size = 64
  12. num_epochs = 1000
  13. #patience：生成被监视的验证数量，当training时无再改进时将停下来。验证数量不能为每个epoch
  14. patience = 100
  15. log_file_path = "./log.csv"
  16. cnn = "ResNet18"
  17. trained_models_path = "./trained_models/" + cnn
  18.  
  19. generator = DataGenerator(dataset="olivettifaces",
  20.                           path="./data/olivetti_faces/olivettifaces.jpg",
  21.                           batch_size=batch_size,
  22.                           input_size=input_shape,
  23.                           is_shuffle=True,
  24.                           data_augmentation=10,
  25.                           validation_split=.2)
  27. num_classes, num_images, training_set_size, validation_set_size = generator.get_number()
  28. #类型数40、images数8400、训练集大小6720、验证集大小1680
  29. print("get_number:",num_classes, num_images, training_set_size, validation_set_size)
  30.  
  31. #eval() 函数用来执行一个字符串表达式，并返回表达式的值。
  32. #此处eval("ResNet18") 实际执行的是 ./model/cnn_models文件中的def ResNet18()函数
  33. model = wrap_cnn(model=eval(cnn),
  34.                  feature_layer="feature",
  35.                  input_shape=input_shape,
  36.                  num_classes=num_classes)
  38. model.compile(optimizer='adam',
  39.               loss=amsoftmax_loss, #自定义amsoftmax_loss
  40.               metrics=['accuracy'])
  41. model.summary()
  43. #画出model网络模型的拓扑图
  44. from tensorflow.keras.utils import plot_model
  45. #如果是在默认的C:\Users\Administrator路径下打开的命令提示符窗口执行的代码的话，则保存图片到该当前默认路径下
  46. plot_model(model,to_file="model.png") #产生网络模型的拓扑图
  47. from IPython.display import Image #画生网络模型的拓扑图
  48. Image("model.png")
  50. # callbacks
  51. early_stop = EarlyStopping('loss', 0.1, patience=patience)
  52. reduce_lr = ReduceLROnPlateau('loss', factor=0.1, patience=int(patience / 2), verbose=1)
  53. csv_logger = CSVLogger(log_file_path, append=False)
  54. #epoch和accuracy这样的单词字眼必须对应fit训练时返回的字典中的key：accuracy、loss等，否则会报错Key Error找不到该key
  55. model_names = trained_models_path + '.{epoch:02d}-{accuracy:2f}.hdf5'
  56. model_checkpoint = ModelCheckpoint(model_names,
  57.                                    monitor='loss',
  58.                                    verbose=1,
  59.                                    save_best_only=True,
  60.                                    save_weights_only=False)
  61. callbacks = [model_checkpoint, csv_logger, early_stop, reduce_lr]
  63. # train model by generator
  64. model.fit_generator(generator=generator.flow('train'),
  65.                     steps_per_epoch=int(training_set_size / batch_size),
  66.                     epochs=num_epochs,
  67.                     verbose=1,
  68.                     callbacks=callbacks,
  69.                     validation_data=generator.flow('validate'),
  70.                     validation_steps=int(validation_set_size / batch_size)
  71.                     )
  73.  
  74.  
  75. """
  76. Model: "model_2"
  77. __________________________________________________________________________________________________
  78. Layer (type)                    Output Shape         Param #     Connected to                     
  79. ==================================================================================================
  80. input (InputLayer)              (None, 64, 64, 1)    0                                            
  81. __________________________________________________________________________________________________
  82. conv1_pad (ZeroPadding2D)       (None, 70, 70, 1)    0           input[0][0]                      
  83. __________________________________________________________________________________________________
  84. conv1 (Conv2D)                  (None, 32, 32, 64)   3200        conv1_pad[0][0]                  
  85. __________________________________________________________________________________________________
  86. batch_normalization_1 (BatchNor (None, 32, 32, 64)   256         conv1[0][0]                      
  87. __________________________________________________________________________________________________
  88. activation_1 (Activation)       (None, 32, 32, 64)   0           batch_normalization_1[0][0]      
  89. __________________________________________________________________________________________________
  90. pool1_pad (ZeroPadding2D)       (None, 34, 34, 64)   0           activation_1[0][0]               
  91. __________________________________________________________________________________________________
  92. max_pooling2d_1 (MaxPooling2D)  (None, 16, 16, 64)   0           pool1_pad[0][0]                  
  93. __________________________________________________________________________________________________
  94. zero_padding2d_1 (ZeroPadding2D (None, 18, 18, 64)   0           max_pooling2d_1[0][0]            
  95. __________________________________________________________________________________________________
  96. conv2d_1 (Conv2D)               (None, 8, 8, 64)     36928       zero_padding2d_1[0][0]           
  97. __________________________________________________________________________________________________
  98. batch_normalization_2 (BatchNor (None, 8, 8, 64)     256         conv2d_1[0][0]                   
  99. __________________________________________________________________________________________________
  100. activation_2 (Activation)       (None, 8, 8, 64)     0           batch_normalization_2[0][0]      
  101. __________________________________________________________________________________________________
  102. zero_padding2d_2 (ZeroPadding2D (None, 10, 10, 64)   0           activation_2[0][0]               
  103. __________________________________________________________________________________________________
  104. conv2d_2 (Conv2D)               (None, 8, 8, 64)     36928       zero_padding2d_2[0][0]           
  105. __________________________________________________________________________________________________
  106. conv2d_3 (Conv2D)               (None, 8, 8, 64)     4160        max_pooling2d_1[0][0]            
  107. __________________________________________________________________________________________________
  108. batch_normalization_3 (BatchNor (None, 8, 8, 64)     256         conv2d_2[0][0]                   
  109. __________________________________________________________________________________________________
  110. batch_normalization_4 (BatchNor (None, 8, 8, 64)     256         conv2d_3[0][0]                   
  111. __________________________________________________________________________________________________
  112. add_1 (Add)                     (None, 8, 8, 64)     0           batch_normalization_3[0][0]      
  113.                                                                  batch_normalization_4[0][0]      
  114. __________________________________________________________________________________________________
  115. activation_3 (Activation)       (None, 8, 8, 64)     0           add_1[0][0]                      
  116. __________________________________________________________________________________________________
  117. zero_padding2d_3 (ZeroPadding2D (None, 10, 10, 64)   0           activation_3[0][0]               
  118. __________________________________________________________________________________________________
  119. conv2d_4 (Conv2D)               (None, 8, 8, 64)     36928       zero_padding2d_3[0][0]           
  120. __________________________________________________________________________________________________
  121. batch_normalization_5 (BatchNor (None, 8, 8, 64)     256         conv2d_4[0][0]                   
  122. __________________________________________________________________________________________________
  123. activation_4 (Activation)       (None, 8, 8, 64)     0           batch_normalization_5[0][0]      
  124. __________________________________________________________________________________________________
  125. zero_padding2d_4 (ZeroPadding2D (None, 10, 10, 64)   0           activation_4[0][0]               
  126. __________________________________________________________________________________________________
  127. conv2d_5 (Conv2D)               (None, 8, 8, 64)     36928       zero_padding2d_4[0][0]           
  128. __________________________________________________________________________________________________
  129. batch_normalization_6 (BatchNor (None, 8, 8, 64)     256         conv2d_5[0][0]                   
  130. __________________________________________________________________________________________________
  131. add_2 (Add)                     (None, 8, 8, 64)     0           batch_normalization_6[0][0]      
  132.                                                                  activation_3[0][0]               
  133. __________________________________________________________________________________________________
  134. activation_5 (Activation)       (None, 8, 8, 64)     0           add_2[0][0]                      
  135. __________________________________________________________________________________________________
  136. zero_padding2d_5 (ZeroPadding2D (None, 10, 10, 64)   0           activation_5[0][0]               
  137. __________________________________________________________________________________________________
  138. conv2d_6 (Conv2D)               (None, 4, 4, 128)    73856       zero_padding2d_5[0][0]           
  139. __________________________________________________________________________________________________
  140. batch_normalization_7 (BatchNor (None, 4, 4, 128)    512         conv2d_6[0][0]                   
  141. __________________________________________________________________________________________________
  142. activation_6 (Activation)       (None, 4, 4, 128)    0           batch_normalization_7[0][0]      
  143. __________________________________________________________________________________________________
  144. zero_padding2d_6 (ZeroPadding2D (None, 6, 6, 128)    0           activation_6[0][0]               
  145. __________________________________________________________________________________________________
  146. conv2d_7 (Conv2D)               (None, 4, 4, 128)    147584      zero_padding2d_6[0][0]           
  147. __________________________________________________________________________________________________
  148. conv2d_8 (Conv2D)               (None, 4, 4, 128)    8320        activation_5[0][0]               
  149. __________________________________________________________________________________________________
  150. batch_normalization_8 (BatchNor (None, 4, 4, 128)    512         conv2d_7[0][0]                   
  151. __________________________________________________________________________________________________
  152. batch_normalization_9 (BatchNor (None, 4, 4, 128)    512         conv2d_8[0][0]                   
  153. __________________________________________________________________________________________________
  154. add_3 (Add)                     (None, 4, 4, 128)    0           batch_normalization_8[0][0]      
  155.                                                                  batch_normalization_9[0][0]      
  156. __________________________________________________________________________________________________
  157. activation_7 (Activation)       (None, 4, 4, 128)    0           add_3[0][0]                      
  158. __________________________________________________________________________________________________
  159. zero_padding2d_7 (ZeroPadding2D (None, 6, 6, 128)    0           activation_7[0][0]               
  160. __________________________________________________________________________________________________
  161. conv2d_9 (Conv2D)               (None, 4, 4, 128)    147584      zero_padding2d_7[0][0]           
  162. __________________________________________________________________________________________________
  163. batch_normalization_10 (BatchNo (None, 4, 4, 128)    512         conv2d_9[0][0]                   
  164. __________________________________________________________________________________________________
  165. activation_8 (Activation)       (None, 4, 4, 128)    0           batch_normalization_10[0][0]     
  166. __________________________________________________________________________________________________
  167. zero_padding2d_8 (ZeroPadding2D (None, 6, 6, 128)    0           activation_8[0][0]               
  168. __________________________________________________________________________________________________
  169. conv2d_10 (Conv2D)              (None, 4, 4, 128)    147584      zero_padding2d_8[0][0]           
  170. __________________________________________________________________________________________________
  171. batch_normalization_11 (BatchNo (None, 4, 4, 128)    512         conv2d_10[0][0]                  
  172. __________________________________________________________________________________________________
  173. add_4 (Add)                     (None, 4, 4, 128)    0           batch_normalization_11[0][0]     
  174.                                                                  activation_7[0][0]               
  175. __________________________________________________________________________________________________
  176. activation_9 (Activation)       (None, 4, 4, 128)    0           add_4[0][0]                      
  177. __________________________________________________________________________________________________
  178. zero_padding2d_9 (ZeroPadding2D (None, 6, 6, 128)    0           activation_9[0][0]               
  179. __________________________________________________________________________________________________
  180. conv2d_11 (Conv2D)              (None, 2, 2, 256)    295168      zero_padding2d_9[0][0]           
  181. __________________________________________________________________________________________________
  182. batch_normalization_12 (BatchNo (None, 2, 2, 256)    1024        conv2d_11[0][0]                  
  183. __________________________________________________________________________________________________
  184. activation_10 (Activation)      (None, 2, 2, 256)    0           batch_normalization_12[0][0]     
  185. __________________________________________________________________________________________________
  186. zero_padding2d_10 (ZeroPadding2 (None, 4, 4, 256)    0           activation_10[0][0]              
  187. __________________________________________________________________________________________________
  188. conv2d_12 (Conv2D)              (None, 2, 2, 256)    590080      zero_padding2d_10[0][0]          
  189. __________________________________________________________________________________________________
  190. conv2d_13 (Conv2D)              (None, 2, 2, 256)    33024       activation_9[0][0]               
  191. __________________________________________________________________________________________________
  192. batch_normalization_13 (BatchNo (None, 2, 2, 256)    1024        conv2d_12[0][0]                  
  193. __________________________________________________________________________________________________
  194. batch_normalization_14 (BatchNo (None, 2, 2, 256)    1024        conv2d_13[0][0]                  
  195. __________________________________________________________________________________________________
  196. add_5 (Add)                     (None, 2, 2, 256)    0           batch_normalization_13[0][0]     
  197.                                                                  batch_normalization_14[0][0]     
  198. __________________________________________________________________________________________________
  199. activation_11 (Activation)      (None, 2, 2, 256)    0           add_5[0][0]                      
  200. __________________________________________________________________________________________________
  201. zero_padding2d_11 (ZeroPadding2 (None, 4, 4, 256)    0           activation_11[0][0]              
  202. __________________________________________________________________________________________________
  203. conv2d_14 (Conv2D)              (None, 2, 2, 256)    590080      zero_padding2d_11[0][0]          
  204. __________________________________________________________________________________________________
  205. batch_normalization_15 (BatchNo (None, 2, 2, 256)    1024        conv2d_14[0][0]                  
  206. __________________________________________________________________________________________________
  207. activation_12 (Activation)      (None, 2, 2, 256)    0           batch_normalization_15[0][0]     
  208. __________________________________________________________________________________________________
  209. zero_padding2d_12 (ZeroPadding2 (None, 4, 4, 256)    0           activation_12[0][0]              
  210. __________________________________________________________________________________________________
  211. conv2d_15 (Conv2D)              (None, 2, 2, 256)    590080      zero_padding2d_12[0][0]          
  212. __________________________________________________________________________________________________
  213. batch_normalization_16 (BatchNo (None, 2, 2, 256)    1024        conv2d_15[0][0]                  
  214. __________________________________________________________________________________________________
  215. add_6 (Add)                     (None, 2, 2, 256)    0           batch_normalization_16[0][0]     
  216.                                                                  activation_11[0][0]              
  217. __________________________________________________________________________________________________
  218. activation_13 (Activation)      (None, 2, 2, 256)    0           add_6[0][0]                      
  219. __________________________________________________________________________________________________
  220. zero_padding2d_13 (ZeroPadding2 (None, 4, 4, 256)    0           activation_13[0][0]              
  221. __________________________________________________________________________________________________
  222. conv2d_16 (Conv2D)              (None, 1, 1, 512)    1180160     zero_padding2d_13[0][0]          
  223. __________________________________________________________________________________________________
  224. batch_normalization_17 (BatchNo (None, 1, 1, 512)    2048        conv2d_16[0][0]                  
  225. __________________________________________________________________________________________________
  226. activation_14 (Activation)      (None, 1, 1, 512)    0           batch_normalization_17[0][0]     
  227. __________________________________________________________________________________________________
  228. zero_padding2d_14 (ZeroPadding2 (None, 3, 3, 512)    0           activation_14[0][0]              
  229. __________________________________________________________________________________________________
  230. conv2d_17 (Conv2D)              (None, 1, 1, 512)    2359808     zero_padding2d_14[0][0]          
  231. __________________________________________________________________________________________________
  232. conv2d_18 (Conv2D)              (None, 1, 1, 512)    131584      activation_13[0][0]              
  233. __________________________________________________________________________________________________
  234. batch_normalization_18 (BatchNo (None, 1, 1, 512)    2048        conv2d_17[0][0]                  
  235. __________________________________________________________________________________________________
  236. batch_normalization_19 (BatchNo (None, 1, 1, 512)    2048        conv2d_18[0][0]                  
  237. __________________________________________________________________________________________________
  238. add_7 (Add)                     (None, 1, 1, 512)    0           batch_normalization_18[0][0]     
  239.                                                                  batch_normalization_19[0][0]     
  240. __________________________________________________________________________________________________
  241. activation_15 (Activation)      (None, 1, 1, 512)    0           add_7[0][0]                      
  242. __________________________________________________________________________________________________
  243. zero_padding2d_15 (ZeroPadding2 (None, 3, 3, 512)    0           activation_15[0][0]              
  244. __________________________________________________________________________________________________
  245. conv2d_19 (Conv2D)              (None, 1, 1, 512)    2359808     zero_padding2d_15[0][0]          
  246. __________________________________________________________________________________________________
  247. batch_normalization_20 (BatchNo (None, 1, 1, 512)    2048        conv2d_19[0][0]                  
  248. __________________________________________________________________________________________________
  249. activation_16 (Activation)      (None, 1, 1, 512)    0           batch_normalization_20[0][0]     
  250. __________________________________________________________________________________________________
  251. zero_padding2d_16 (ZeroPadding2 (None, 3, 3, 512)    0           activation_16[0][0]              
  252. __________________________________________________________________________________________________
  253. conv2d_20 (Conv2D)              (None, 1, 1, 512)    2359808     zero_padding2d_16[0][0]          
  254. __________________________________________________________________________________________________
  255. batch_normalization_21 (BatchNo (None, 1, 1, 512)    2048        conv2d_20[0][0]                  
  256. __________________________________________________________________________________________________
  257. add_8 (Add)                     (None, 1, 1, 512)    0           batch_normalization_21[0][0]     
  258.                                                                  activation_15[0][0]              
  259. __________________________________________________________________________________________________
  260. activation_17 (Activation)      (None, 1, 1, 512)    0           add_8[0][0]                      
  261. __________________________________________________________________________________________________
  262. feature (GlobalAveragePooling2D (None, 512)          0           activation_17[0][0]              
  263. __________________________________________________________________________________________________
  264. dropout_1 (Dropout)             (None, 512)          0           feature[0][0]                    
  265. __________________________________________________________________________________________________
  266. predictions (AMSoftmax)         (None, 40)           20480       dropout_1[0][0]                  
  267. ==================================================================================================
  268. Total params: 11,209,536
  269. Trainable params: 11,199,808
  270. Non-trainable params: 9,728
  271. __________________________________________________________________________________________________
  272. """

  1. cnn_models.py
  2.  
  3. from keras import layers
  4. from keras.layers import Activation, Dropout, Conv2D, Dense
  5. from keras.layers import BatchNormalization
  6. from keras.layers import Flatten
  7. from keras.layers import GlobalAveragePooling2D
  8. from keras.layers import InputLayer, Input
  9. from keras.layers import MaxPooling2D
  10. from keras.layers import SeparableConv2D
  11. from keras.layers import ZeroPadding2D, Add
  12. from keras.models import Model
  13. from keras.models import Sequential
  14. from keras.regularizers import l2
  15.  
  16.  
  17. # 来自 VIPLFaceNet
  18. # 你可以在以下网址看到文献: https://arxiv.org/abs/1609.03892
  19. def VIPL_FaceNet(input_shape, num_classes):
  20.     model = Sequential()
  21.     model.add(InputLayer(input_shape=input_shape,
  22.                          name="input"))
  23.  
  24.     # Conv layer 1 output shape (55, 55, 48)
  25.     model.add(Conv2D(
  26.         kernel_size=(9, 9),
  27.         activation="relu",
  28.         filters=48,
  29.         strides=(4, 4)
  30.     ))
  31.     # pool1
  32.     model.add(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))
  33.  
  34.     # Conv layer 2 output shape (27, 27, 128)
  35.     model.add(Conv2D(
  36.         strides=(1, 1),
  37.         kernel_size=(3, 3),
  38.         activation="relu",
  39.         filters=128
  40.     ))
  41.  
  42.     # Conv layer 3 output shape (13, 13, 192)
  43.     model.add(Conv2D(
  44.         strides=(1, 1),
  45.         kernel_size=(3, 3),
  46.         activation="relu",
  47.         filters=128
  48.     ))
  49.  
  50.     # pool2
  51.     model.add(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))
  52.  
  53.     # conv4
  54.     model.add(Conv2D(
  55.         kernel_size=(3, 3),
  56.         activation="relu",
  57.         filters=256,
  58.         padding="same",
  59.         strides=(1, 1)
  60.     ))
  61.  
  62.     # conv5
  63.     model.add(Conv2D(
  64.         kernel_size=(3, 3),
  65.         activation="relu",
  66.         filters=192,
  67.         padding="same",
  68.         strides=(1, 1)
  69.     ))
  70.  
  71.     # conv6
  72.     model.add(Conv2D(
  73.         kernel_size=(3, 3),
  74.         activation="relu",
  75.         filters=192,
  76.         padding="same",
  77.         strides=(1, 1)
  78.     ))
  79.  
  80.     # conv7
  81.     model.add(Conv2D(
  82.         kernel_size=(3, 3),
  83.         activation="relu",
  84.         filters=128,
  85.         padding="same",
  86.         strides=(1, 1)
  87.     ))
  88.  
  89.     # pool3
  90.     model.add(MaxPooling2D((3, 3), strides=(2, 2), padding='same'))
  91.  
  92.     # fully connected layer 1
  93.     model.add(Flatten())
  94.     model.add(Dense(4096))
  95.     model.add(BatchNormalization())
  96.     model.add(Activation('relu'))
  97.     model.add(Dropout(0.5))
  98.  
  99.     # fully connected layer 2
  100.     model.add(Dense(2048))
  101.     model.add(BatchNormalization())
  102.     model.add(Activation('relu', name="feature"))
  103.     model.add(Dropout(0.5))
  104.  
  105.     # output
  106.     model.add(Dense(num_classes))
  107.     model.add(Activation('softmax', name='predictions'))
  108.  
  109.     # return
  110.     return model
  111.  
  112.  
  113. # 实现 VGGNet
  114. def VGGNet(input_shape, num_classes):
  115.     # 因为 VGGNet 更深，而且网络中有许多max pooling,
  116.     # 不建议input_shape太小，原input_shape是 (224, 224, 3)。
  117.     assert input_shape[0] >= 224 and input_shape[1] >= 224
  118.  
  119.     model = Sequential()
  120.     model.add(InputLayer(input_shape=input_shape,
  121.                          name="input"))
  122.  
  123.     # Conv1,2
  124.     model.add(Conv2D(
  125.         kernel_size=(3, 3),
  126.         activation="relu",
  127.         filters=64,
  128.         strides=(1, 1)))
  129.     model.add(Conv2D(
  130.         kernel_size=(3, 3),
  131.         activation="relu",
  132.         filters=64,
  133.         strides=(1, 1)))
  134.  
  135.     # pool1
  136.     model.add(MaxPooling2D((2, 2), strides=(2, 2),
  137.                            padding='same'))
  138.  
  139.     # Conv 3,4
  140.     model.add(Conv2D(
  141.         kernel_size=(3, 3),
  142.         activation="relu",
  143.         filters=128,
  144.         strides=(1, 1)))
  145.     model.add(Conv2D(
  146.         kernel_size=(3, 3),
  147.         activation="relu",
  148.         filters=128,
  149.         strides=(1, 1)))
  150.  
  151.     # pool2
  152.     model.add(MaxPooling2D((2, 2), strides=(2, 2),
  153.                            padding='same'))
  154.  
  155.     # Conv 5-7
  156.     model.add(Conv2D(
  157.         kernel_size=(3, 3),
  158.         activation="relu",
  159.         filters=256,
  160.         strides=(1, 1)))
  161.     model.add(Conv2D(
  162.         kernel_size=(3, 3),
  163.         activation="relu",
  164.         filters=256,
  165.         strides=(1, 1)))
  166.     model.add(Conv2D(
  167.         kernel_size=(3, 3),
  168.         activation="relu",
  169.         filters=256,
  170.         strides=(1, 1)))
  171.  
  172.     # pool3
  173.     model.add(MaxPooling2D((2, 2), strides=(2, 2),
  174.                            padding='same'))
  175.     # Conv 8-10
  176.     model.add(Conv2D(
  177.         kernel_size=(3, 3),
  178.         activation="relu",
  179.         filters=512,
  180.         strides=(1, 1)))
  181.  
  182.     model.add(Conv2D(
  183.         kernel_size=(3, 3),
  184.         activation="relu",
  185.         filters=512,
  186.         strides=(1, 1)))
  187.     model.add(Conv2D(
  188.         kernel_size=(3, 3),
  189.         activation="relu",
  190.         filters=512,
  191.         strides=(1, 1)))
  192.     # pool4
  193.     model.add(MaxPooling2D((2, 2), strides=(2, 2),
  194.                            padding='same'))
  195.     # Conv 11-13
  196.     model.add(Conv2D(
  197.         kernel_size=(3, 3),
  198.         activation="relu",
  199.         filters=512,
  200.         strides=(1, 1)))
  201.     model.add(Conv2D(
  202.         kernel_size=(3, 3),
  203.         activation="relu",
  204.         filters=512,
  205.         strides=(1, 1)))
  206.     model.add(Conv2D(
  207.         kernel_size=(3, 3),
  208.         activation="relu",
  209.         filters=512,
  210.         strides=(1, 1)))
  211.  
  212.     # pool5
  213.     model.add(MaxPooling2D((2, 2), strides=(2, 2), padding='same'))
  214.  
  215.     # fully connected layer 1
  216.     model.add(Flatten())
  217.     model.add(Dense(2048))
  218.     model.add(BatchNormalization())
  219.     model.add(Activation('relu'))
  220.     model.add(Dropout(0.5))
  221.  
  222.     # fully connected layer 2
  223.     model.add(Dense(2048))
  224.     model.add(BatchNormalization())
  225.     model.add(Activation('relu', name="feature"))
  226.     model.add(Dropout(0.5))
  227.  
  228.     model.add(Dense(num_classes))
  229.     model.add(Activation('softmax', name='predictions'))
  230.  
  231.     return model
  232.  
  233.  
  234. # 面部分类
  235. # 源代码: https://github.com/oarriaga/face_classification/blob/master/src/models/cnn.py
  236. def tiny_XCEPTION(input_shape, num_classes, l2_regularization=0.01):
  237.     regularization = l2(l2_regularization)
  238.  
  239.     # base
  240.     img_input = Input(input_shape, name="input")
  241.     x = Conv2D(5, (3, 3), strides=(1, 1), kernel_regularizer=regularization,
  242.                use_bias=False)(img_input)
  243.     x = BatchNormalization()(x)
  244.     x = Activation('relu')(x)
  245.     x = Conv2D(5, (3, 3), strides=(1, 1), kernel_regularizer=regularization,
  246.                use_bias=False)(x)
  247.     x = BatchNormalization()(x)
  248.     x = Activation('relu')(x)
  249.  
  250.     # module 1
  251.     residual = Conv2D(8, (1, 1), strides=(2, 2),
  252.                       padding='same', use_bias=False)(x)
  253.     residual = BatchNormalization()(residual)
  254.  
  255.     x = SeparableConv2D(8, (3, 3), padding='same',
  256.                         kernel_regularizer=regularization,
  257.                         use_bias=False)(x)
  258.     x = BatchNormalization()(x)
  259.     x = Activation('relu')(x)
  260.     x = SeparableConv2D(8, (3, 3), padding='same',
  261.                         kernel_regularizer=regularization,
  262.                         use_bias=False)(x)
  263.     x = BatchNormalization()(x)
  264.  
  265.     x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
  266.     x = layers.add([x, residual])
  267.  
  268.     # module 2
  269.     residual = Conv2D(16, (1, 1), strides=(2, 2),
  270.                       padding='same', use_bias=False)(x)
  271.     residual = BatchNormalization()(residual)
  272.  
  273.     x = SeparableConv2D(16, (3, 3), padding='same',
  274.                         kernel_regularizer=regularization,
  275.                         use_bias=False)(x)
  276.     x = BatchNormalization()(x)
  277.     x = Activation('relu')(x)
  278.     x = SeparableConv2D(16, (3, 3), padding='same',
  279.                         kernel_regularizer=regularization,
  280.                         use_bias=False)(x)
  281.     x = BatchNormalization()(x)
  282.  
  283.     x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
  284.     x = layers.add([x, residual])
  285.  
  286.     # module 3
  287.     residual = Conv2D(32, (1, 1), strides=(2, 2),
  288.                       padding='same', use_bias=False)(x)
  289.     residual = BatchNormalization()(residual)
  290.  
  291.     x = SeparableConv2D(32, (3, 3), padding='same',
  292.                         kernel_regularizer=regularization,
  293.                         use_bias=False)(x)
  294.     x = BatchNormalization()(x)
  295.     x = Activation('relu')(x)
  296.     x = SeparableConv2D(32, (3, 3), padding='same',
  297.                         kernel_regularizer=regularization,
  298.                         use_bias=False)(x)
  299.     x = BatchNormalization()(x)
  300.  
  301.     x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
  302.     x = layers.add([x, residual])
  303.  
  304.     # module 4
  305.     residual = Conv2D(64, (1, 1), strides=(2, 2),
  306.                       padding='same', use_bias=False)(x)
  307.     residual = BatchNormalization()(residual)
  308.  
  309.     x = SeparableConv2D(64, (3, 3), padding='same',
  310.                         kernel_regularizer=regularization,
  311.                         use_bias=False)(x)
  312.     x = BatchNormalization()(x)
  313.     x = Activation('relu')(x)
  314.     x = SeparableConv2D(64, (3, 3), padding='same',
  315.                         kernel_regularizer=regularization,
  316.                         use_bias=False)(x)
  317.     x = BatchNormalization()(x)
  318.  
  319.     x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
  320.     x = layers.add([x, residual])
  321.  
  322.     x = Conv2D(1024, (3, 3),
  323.                # kernel_regularizer=regularization,
  324.                padding='same')(x)
  325.     x = GlobalAveragePooling2D()(x)
  326.     x = Dense(1024, name="feature")(x)
  327.     x = Dropout(.5)(x)
  328.     x = Dense(num_classes)(x)
  329.     output = Activation('softmax', name='predictions')(x)
  330.     model = Model(img_input, output)
  331.  
  332.     return model
  333.  
  334.  
  335. def ResNet50(input_shape, num_classes):
  336.     # wrap ResNet50 from keras, because ResNet50 is so deep.
  337.     from keras.applications.resnet50 import ResNet50
  338.     input_tensor = Input(shape=input_shape, name="input")
  339.     x = ResNet50(include_top=False,
  340.                  weights=None,
  341.                  input_tensor=input_tensor,
  342.                  input_shape=None,
  343.                  pooling="avg",
  344.                  classes=num_classes)
  345.     x = Dense(units=2048, name="feature")(x.output)
  346.     return Model(inputs=input_tensor, outputs=x)
  347.  
  348.  
  349. # 实现于 ResNet's block.
  350. # 实现了两个 classes block: 一个是基本块basic block，另一个是瓶颈块bottleneck block
  351. def basic_block(filters, kernel_size=3, is_first_block=True):
  352.     stride = 1
  353.     if is_first_block:
  354.         stride = 2
  356.     def f(x):
  357.         # f(x) named y
  358.         # 1st Conv
  359.         y = ZeroPadding2D(padding=1)(x)
  360.         y = Conv2D(filters, kernel_size, strides=stride, kernel_initializer='he_normal')(y)
  361.         y = BatchNormalization()(y)
  362.         y = Activation("relu")(y)
  363.         # 2nd Conv
  364.         y = ZeroPadding2D(padding=1)(y)
  365.         y = Conv2D(filters, kernel_size, kernel_initializer='he_normal')(y)
  366.         y = BatchNormalization()(y)
  368.         # f(x) + x
  369.         if is_first_block:
  370.             shortcut = Conv2D(filters, kernel_size=1, strides=stride, kernel_initializer='he_normal')(x)
  371.             shortcut = BatchNormalization()(shortcut)
  372.         else:
  373.             shortcut = x
  375.         y = Add()([y, shortcut])
  376.         y = Activation("relu")(y)
  378.         return y
  380.     return f
  383. # ResNet v1, we can see the paper at:
  384. # https://arxiv.org/abs/1512.03385
  385. def ResNet18(input_shape, num_classes):
  386.     input_layer = Input(shape=input_shape, name="input")
  388.     # Conv1
  389.     x = layers.ZeroPadding2D(padding=(3, 3), name='conv1_pad')(input_layer)
  390.     x = layers.Conv2D(64, (7, 7),
  391.                       strides=(2, 2),
  392.                       padding='valid',
  393.                       kernel_initializer='he_normal',
  394.                       name='conv1')(x)
  395.     x = layers.BatchNormalization()(x)
  396.     x = layers.Activation('relu')(x)
  397.     x = layers.ZeroPadding2D(padding=(1, 1), name='pool1_pad')(x)
  398.     x = layers.MaxPooling2D((3, 3), strides=(2, 2))(x)
  400.     # Conv2
  401.     x = basic_block(filters=64)(x)
  402.     x = basic_block(filters=64, is_first_block=False)(x)
  404.     # Conv3
  405.     x = basic_block(filters=128)(x)
  406.     x = basic_block(filters=128, is_first_block=False)(x)
  408.     # Conv4
  409.     x = basic_block(filters=256)(x)
  410.     x = basic_block(filters=256, is_first_block=False)(x)
  412.     # Conv5
  413.     x = basic_block(filters=512)(x)
  414.     x = basic_block(filters=512, is_first_block=False)(x)
  416.     x = GlobalAveragePooling2D(name="feature")(x)
  417.     output_layer = Dense(num_classes, activation='softmax')(x)
  419.     model = Model(input_layer, output_layer)
  420.     return model
  1. amsoftmax.py
  2.  
  3. import tensorflow as tf
  4. from keras import backend as K
  5. from keras.layers import Dropout
  6. from keras.engine.topology import Layer
  7. from keras.models import Model
  8.  
  9.  
  10. class AMSoftmax(Layer):
  11.     def __init__(self, units, **kwargs):
  12.         self.units = units
  13.         self.kernel = None
  14.         super(AMSoftmax, self).__init__(**kwargs)
  15.  
  16.     # 类(初始化值)(call函数传入值)：即创建完实例对象后就执行call函数，
  17.     # 那么类中的执行顺序为先执行 _init__函数，然后执行 build函数，最后执行 call函数
  18.     def build(self, input_shape):
  19.         assert len(input_shape) >= 2
  20.         self.kernel = self.add_weight(name='kernel',
  21.                                       shape=(input_shape[1], self.units),
  22.                                       initializer='uniform',
  23.                                       trainable=True)
  24.         super(AMSoftmax, self).build(input_shape)
  25.  
  26.     #作为模型输出层的输出
  27.     def call(self, inputs, **kwargs):
  28.         # 得到余弦相似性
  29.         # cosine余弦 = x * w / (||x|| * ||w||)
  30.         inputs = K.l2_normalize(inputs, axis=1)
  31.         kernel = K.l2_normalize(self.kernel, axis=0)
  32.         cosine = K.dot(inputs, kernel)
  33.         return cosine
  34.  
  35.     def compute_output_shape(self, input_shape):
  36.         return input_shape[0], self.units
  37.  
  38.     def get_config(self):
  39.         config = {
  40.             'units': self.units}
  41.         base_config = super(AMSoftmax, self).get_config()
  42.  
  43.         return dict(list(base_config.items())
  44.                     + list(config.items()))
  45.  
  46.  
  47. # 参考自: https://github.com/hao-qiang/AM-Softmax/blob/master/AM-Softmax.ipynb
  48. def amsoftmax_loss(y_true, y_pred, scale=30.0, margin=0.35):
  49.     # 定义两个常量张量
  50.     m = K.constant(margin, name='m')
  51.     s = K.constant(scale, name='s')
  52.     # 重塑标签label
  53.     label = K.reshape(K.argmax(y_true, axis=-1), shape=(-1, 1))
  54.     label = K.cast(label, dtype=tf.int32)
  55.     #pred_batch批预处理
  56.     pred_batch = K.reshape(tf.range(K.shape(y_pred)[0]), shape=(-1, 1))
  57.     #concat两个列向量，一个是pred_batch批预处理，另一个是标签label。
  58.     ground_truth_indices = tf.concat([pred_batch,
  59.                                       K.reshape(label, shape=(-1, 1))], axis=1)
  60.     """
  61.     参考自:https://tensorflow.google.cn/api_docs/python/tf/gather_nd?hl=en
  62.     tf.gather_nd(params, indices, batch_dims=0, name=None)
  63.     根据indices索引对params矩阵/向量进行元素操作
  64.         indices = [[0, 0], [1, 1]]
  65.         params = [['a', 'b'], ['c', 'd']]
  66.         output = ['a', 'd']
  67.         
  68.         indices = [[1], [0]]
  69.         params = [['a', 'b'], ['c', 'd']]
  70.         output = [['c', 'd'], ['a', 'b']]
  71.     """
  72.     # get ground truth scores by indices
  73.     ground_truth_scores = tf.gather_nd(y_pred, ground_truth_indices)
  74.  
  75.     # 如果 ground_truth_scores > m, 则 ground_truth_scores = group_truth_score - m
  76.     #K.greater(x, y)用法：如果x>y，则返回 布尔张量
  77.     added_margin = K.cast(K.greater(ground_truth_scores, m), dtype=tf.float32) * m
  78.     added_margin = K.reshape(added_margin, shape=(-1, 1))
  79.     #tf.subtract 减法
  80.     added_embedding_feature = tf.subtract(y_pred, y_true * added_margin) * s
  81.     #计算logits和labels之间的softmax cross entropy交叉熵
  82.     cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true, logits=added_embedding_feature)
  83.     loss = tf.reduce_mean(cross_entropy)
  84.     return loss
  85.  
  86.  
  87. def wrap_cnn(model, feature_layer, input_shape, num_classes):
  88.     cnn = model(input_shape, num_classes)
  89.     assert isinstance(cnn, Model)
  90.     #获取feature (GlobalAveragePooling2D)的输出
  91.     x = cnn.get_layer(name=feature_layer).output
  92.     x = Dropout(.5)(x)
  93.     #1.类(初始化值)(call函数传入值)：即创建完实例对象后就执行call函数，
  94.     # 那么类中的执行顺序为先执行 _init__函数，然后执行 build函数，最后执行 call函数
  95.     #2.predictions (AMSoftmax) 作为模型输出，自定义AM-Softmax损失函数和AM-Softmax网络层
  96.     output_layer = AMSoftmax(num_classes, name="predictions")(x)
  97.     return Model(inputs=cnn.input, outputs=output_layer)
  98.  
  99.  
  100. def load_model(filepath):
  101.     import keras.models
  102.     model = keras.models.load_model(filepath,
  103.                                     {"AMSoftmax": AMSoftmax,
  104.                                      "amsoftmax_loss": amsoftmax_loss})
  105.     return model
  106.  
  1. evaluate.py
  2.  
  3. # /usr/bin env python
  4. import os
  5.  
  6. import cv2
  7. import matplotlib.pyplot as plt
  8. import numpy as np
  9. from sklearn.metrics import roc_curve, auc
  10.  
  11. from data.olivetti_faces.split_img import split_to_dataset
  12. from model.amsoftmax import load_model
  13. from utils.feature import get_feature_function
  14. from utils.measure import kappa, cosine_similarity
  15.  
  16. model_path = "./trained_models/tiny_XCEPTION.hdf5"
  17. img_path = "./data/olivetti_faces/olivettifaces.jpg"
  18. test_data_path = "./olive"
  19. input_shape = (64, 64, 1)
  20.  
  21. def classifier():
  22.     model = load_model(filepath=model_path)
  23.     files = list(os.walk(test_data_path))[0][2]
  24.     x_list = []
  25.     total = 0
  26.     correct = 0
  27.     matrix = np.zeros(shape=(20, 20))
  28.     for file in files:
  29.         label = file.split("_")[0].replace("olive", "")
  30.         img = cv2.imread(test_data_path + file)
  31.         img = cv2.resize(img, (64, 64))
  32.         img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
  33.         img = img / 256
  34.         img = np.expand_dims(img, -1)
  35.         img = np.expand_dims(img, 0)
  36.         x_list.append(img)
  37.         y = model.predict(x=img)
  38.         y = int(np.argmax(y) / 2)
  39.  
  40.         y_correct = int(label)
  41.         total += 1
  42.         if y == y_correct:
  43.             correct += 1
  44.         matrix[y_correct][y] += 1
  45.  
  46.     k = kappa(matrix=matrix)
  47.     print("total is {0}, precise is {1}, kappa is {2}."
  48.           .format(total, correct / total, k))
  49.  
  50.  
  51. def recognition():
  52.     # This threshold is used to determine if two face images belong to the same person.
  53.     threshold = 0.80
  54.  
  55.     model = load_model(filepath=model_path)
  56.     f = get_feature_function(model)
  57.     base_feature = f(cv2.imread("./olive/0_0.jpg"))
  58.  
  59.     y_true = []
  60.     for i in range(200):
  61.         if i < 10:
  62.             y_true.append(1)  # True
  63.         else:
  64.             y_true.append(0)  # False
  65.     y_score = []
  66.     for label in range(20):
  67.         for photo in range(10):
  68.             file = "./olive/" + str(label) + "_" + str(photo) + ".jpg"
  69.             img_feature = f(cv2.imread(file))
  70.             sim = cosine_similarity(base_feature, img_feature)
  71.             print("label:{0} - {1} ,sim : {2}".format(label, photo, sim))
  72.             if sim > threshold:
  73.                 y_score.append(1)  # True
  74.             else:
  75.                 y_score.append(0)  # False
  76.     correct = 0
  77.     for i in range(200):
  78.         if y_true[i] == y_score[i]:
  79.             correct += 1
  80.  
  81.     print("acc is " + str(correct / 200))
  82.     fpr, tpr, t = roc_curve(y_true, y_score)
  83.     roc_auc = auc(fpr, tpr)
  84.  
  85.     plt.figure()
  86.     lw = 2
  87.     plt.figure(figsize=(10, 10))
  88.     plt.plot(fpr, tpr, color='darkorange',
  89.              lw=lw, label='ROC curve (area = %0.2f)' % roc_auc)
  90.     plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='-.')
  91.     plt.xlim([0.0, 1.0])
  92.     plt.ylim([0.0, 1.05])
  93.     plt.xlabel('False Positive Rate')
  94.     plt.ylabel('True Positive Rate')
  95.     plt.title('ROC curve')
  96.     plt.legend(loc="lower right")
  97.     plt.show()
  98.  
  99.  
  100. if __name__ == '__main__':
  101.     if not os.path.exists(test_data_path):
  102.         os.mkdir(test_data_path)
  103.     # generate_more_faces(, test_data_path)
  104.     split_to_dataset(img_path, test_data_path)
  105.     recognition()
  106.  

  1. predict.py
  2.  
  3. import os
  4. import cv2
  5. from utils.feature import get_feature_function
  6. from utils.measure import *
  7.  
  8. model_path = "./trained_models/tiny_XCEPTION.hdf5"
  9.  
  10. def main():
  11.     get_feature = get_feature_function(model=model_path)
  12.     features = []
  13.     base_feature = None
  14.     dir_list = list(list(os.walk("./data/manual_testing"))[0])[2]
  15.     dir_list.sort()
  16.     for file in dir_list:
  17.         path = "./data/manual_testing/" + file
  18.         img = cv2.imread(path)
  19.         feature = get_feature(img)
  20.         features.append((file, feature))
  21.         if file == "base.jpg":
  22.             base_feature = feature
  23.  
  24.     for file, feature in features:
  25.         print(file, '\t',
  26.               cosine_similarity(feature, base_feature), '\t',
  27.               euclidean_metric(feature, base_feature), '\t',
  28.               pearson_correlation(feature, base_feature))
  29.  
  30.  
  31. if __name__ == '__main__':
  32.     main()
  33.  
  34. """
  35. base.jpg 	 1.0000000596046448 	 1.0 	 0.99999994
  36. base_full.jpg 	 0.8650757670402527 	 0.0014323909546049918 	 0.73107094
  37. lena.jpg 	 0.7991840243339539 	 0.0010432298559344395 	 0.59929365
  38. man1.jpg 	 0.6370709836483002 	 0.0031209503507164146 	 0.2738905
  39. man1_2.jpg 	 0.5672858357429504 	 0.002656866875466825 	 0.13452913
  40. man2.jpg 	 0.48662818130105734 	 0.0027977924693378836 	 -0.026793957
  41. woman1_crop.jpg 	 0.8841563165187836 	 0.00508468014331387 	 0.76889646
  42. woman1_full.jpg 	 0.8546876013278961 	 0.0016342116788881774 	 0.70945275
  43. woman2_crop.jpg 	 0.7393457293510437 	 0.002660944596353025 	 0.47877395
  44. woman2_full.jpg 	 0.8256216049194336 	 0.0009339273743781038 	 0.65218705
  45. woman3_crop.jpg 	 0.8040041327476501 	 0.003071616047994001 	 0.6079379
  46. woman3_full.jpg 	 0.8212800323963165 	 0.0013429052489530274 	 0.6433183
  47. woman4_crop.jpg 	 0.696440264582634 	 0.0032429208812613745 	 0.3926806
  48. woman4_full.jpg 	 0.8566377758979797 	 0.0015620006726975552 	 0.71384853
  49. woman5.jpg 	 0.9146099090576172 	 0.001529003613932187 	 0.8297442
  50. woman5_crop.jpg 	 0.9101028144359589 	 0.005750450657876265 	 0.8201517
  51. """
  1. web.py
  2.  
  3. import cv2
  4. from tempfile import SpooledTemporaryFile
  5. import numpy as np
  6. from flask import Flask
  7. from flask import request
  8. from utils.feature import get_feature_function
  9. from utils.measure import cosine_similarity
  10.  
  11. model_path = "./trained_models/tiny_XCEPTION.hdf5"
  12. get_feature = get_feature_function(model=model_path)
  13.  
  14. app = Flask(__name__, static_folder="web_static")
  15. # if we save file to disk, we must use the following configuration.
  16. # upload_folder = './web_static/uploads/'
  17. # app.config['UPLOAD_FOLDER'] = upload_folder
  18. app.config['ALLOWED_EXTENSIONS'] = {'png', 'jpg', 'jpeg', 'gif', 'bmp'}
  19.  
  20.  
  21. @app.route("/")
  22. def hello():
  23.     return "Hello, SmooFaceEngine!"
  24.  
  25.  
  26. @app.route("/test")
  27. def test():
  28.     html = '''
  29.         <!DOCTYPE html>
  30.         <html lang="en">
  31.         <head>
  32.             <meta charset="UTF-8">
  33.             <title>Uploading</title>
  34.         </head>
  35.         <body>
  36.             <form action="/data" method="post" enctype="multipart/form-data">
  37.                 <input type="file" name="pic1" value="Pic1" /><br>
  38.                 <input type="file" name="pic2" value="Pic2" /><br>
  39.                 <input type="submit" value="upload">
  40.             </form>
  41.         </body>
  42.         </html>
  43.     '''
  44.     return html
  45.  
  46.  
  47. def get_feature_from_client(request_filename):
  48.     # If we want to save this file to disk, we can use the
  49.     # following code. But if we should save the binary file from client to disk,
  50.     # the program would run slowly.
  51.     """
  52.     import random
  53.     def get_random_string(length):
  54.         string = ""
  55.         for i in range(0, length):
  56.             code = random.randint(97, 122)
  57.             string += chr(code)
  58.         return string
  60.     pic = request.files[request_filename]
  61.     img_type = pic.filename.split('.')[1]
  62.     filename = get_random_string(10) + "." + img_type
  63.     filepath = os.path.join(app.root_path,
  64.                             app.config['UPLOAD_FOLDER'],
  65.                             filename)
  66.     pic.save(filepath)
  67.     vector = get_feature(filepath)
  68.     os.unlink(filepath)
  70.     return vector
  71.     """
  72.  
  73.     # the following codes:
  74.     # We will save the file from client to memory, then
  75.     # the program run much faster than saving it to disk.
  76.     file = request.files[request_filename]
  77.     stream = file.stream
  78.     # for old version flask:
  79.     """
  80.      if isinstance(stream, SpooledTemporaryFile):
  81.          stream = stream.file
  82.     """
  83.     value = bytearray(stream.read())
  84.     value = np.asarray(value, dtype='uint8')
  85.     img = cv2.imdecode(value, 1)
  86.     vector = get_feature(img)
  87.     return vector
  88.  
  89.  
  90. @app.route("/data", methods=["POST"])
  91. def predict():
  92.     vector1 = get_feature_from_client('pic1')
  93.     vector2 = get_feature_from_client('pic2')
  94.     similarity = cosine_similarity(vector1, vector2)
  95.     return str(similarity)
  96.  
  97.  
  98. if __name__ == "__main__":
  99.     app.run(host='0.0.0.0', port=8080, debug=True)
  100.

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/我家小花儿/article/detail/307597
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号