conciomith/RetinaFace_FaceDetector_Extractor
14
1import tensorflow as tf2import gdown3from pathlib import Path4import os5 6tf_version = int(tf.__version__.split(".")[0])7 8if tf_version == 1:9 from keras.models import Model10 from keras.layers import Input, BatchNormalization, ZeroPadding2D, Conv2D, ReLU, MaxPool2D, Add, UpSampling2D, concatenate, Softmax11 12else:13 from tensorflow.keras.models import Model14 from tensorflow.keras.layers import Input, BatchNormalization, ZeroPadding2D, Conv2D, ReLU, MaxPool2D, Add, UpSampling2D, concatenate, Softmax15 16def load_weights(model):17 18 exact_file = "retinaface.h5"19 20 model.load_weights(exact_file)21 22 return model23 24def build_model():25 26 data = Input(dtype=tf.float32, shape=(None, None, 3), name='data')27 28 bn_data = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn_data', trainable=False)(data)29 30 conv0_pad = ZeroPadding2D(padding=tuple([3, 3]))(bn_data)31 32 conv0 = Conv2D(filters = 64, kernel_size = (7, 7), name = 'conv0', strides = [2, 2], padding = 'VALID', use_bias = False)(conv0_pad)33 34 bn0 = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn0', trainable=False)(conv0)35 36 relu0 = ReLU(name='relu0')(bn0)37 38 pooling0_pad = ZeroPadding2D(padding=tuple([1, 1]))(relu0)39 40 pooling0 = MaxPool2D((3, 3), (2, 2), padding='VALID', name='pooling0')(pooling0_pad)41 42 stage1_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn1', trainable=False)(pooling0)43 44 stage1_unit1_relu1 = ReLU(name='stage1_unit1_relu1')(stage1_unit1_bn1)45 46 stage1_unit1_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu1)47 48 stage1_unit1_sc = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit1_sc', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu1)49 50 stage1_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn2', trainable=False)(stage1_unit1_conv1)51 52 stage1_unit1_relu2 = ReLU(name='stage1_unit1_relu2')(stage1_unit1_bn2)53 54 stage1_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit1_relu2)55 56 stage1_unit1_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit1_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_conv2_pad)57 58 stage1_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit1_bn3', trainable=False)(stage1_unit1_conv2)59 60 stage1_unit1_relu3 = ReLU(name='stage1_unit1_relu3')(stage1_unit1_bn3)61 62 stage1_unit1_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit1_relu3)63 64 plus0_v1 = Add()([stage1_unit1_conv3 , stage1_unit1_sc])65 66 stage1_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn1', trainable=False)(plus0_v1)67 68 stage1_unit2_relu1 = ReLU(name='stage1_unit2_relu1')(stage1_unit2_bn1)69 70 stage1_unit2_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_relu1)71 72 stage1_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn2', trainable=False)(stage1_unit2_conv1)73 74 stage1_unit2_relu2 = ReLU(name='stage1_unit2_relu2')(stage1_unit2_bn2)75 76 stage1_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit2_relu2)77 78 stage1_unit2_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_conv2_pad)79 80 stage1_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit2_bn3', trainable=False)(stage1_unit2_conv2)81 82 stage1_unit2_relu3 = ReLU(name='stage1_unit2_relu3')(stage1_unit2_bn3)83 84 stage1_unit2_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit2_relu3)85 86 plus1_v2 = Add()([stage1_unit2_conv3 , plus0_v1])87 88 stage1_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn1', trainable=False)(plus1_v2)89 90 stage1_unit3_relu1 = ReLU(name='stage1_unit3_relu1')(stage1_unit3_bn1)91 92 stage1_unit3_conv1 = Conv2D(filters = 64, kernel_size = (1, 1), name = 'stage1_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_relu1)93 94 stage1_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn2', trainable=False)(stage1_unit3_conv1)95 96 stage1_unit3_relu2 = ReLU(name='stage1_unit3_relu2')(stage1_unit3_bn2)97 98 stage1_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage1_unit3_relu2)99 100 stage1_unit3_conv2 = Conv2D(filters = 64, kernel_size = (3, 3), name = 'stage1_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_conv2_pad)101 102 stage1_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage1_unit3_bn3', trainable=False)(stage1_unit3_conv2)103 104 stage1_unit3_relu3 = ReLU(name='stage1_unit3_relu3')(stage1_unit3_bn3)105 106 stage1_unit3_conv3 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage1_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage1_unit3_relu3)107 108 plus2 = Add()([stage1_unit3_conv3 , plus1_v2])109 110 stage2_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn1', trainable=False)(plus2)111 112 stage2_unit1_relu1 = ReLU(name='stage2_unit1_relu1')(stage2_unit1_bn1)113 114 stage2_unit1_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit1_relu1)115 116 stage2_unit1_sc = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage2_unit1_relu1)117 118 stage2_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn2', trainable=False)(stage2_unit1_conv1)119 120 stage2_unit1_relu2 = ReLU(name='stage2_unit1_relu2')(stage2_unit1_bn2)121 122 stage2_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit1_relu2)123 124 stage2_unit1_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage2_unit1_conv2_pad)125 126 stage2_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit1_bn3', trainable=False)(stage2_unit1_conv2)127 128 stage2_unit1_relu3 = ReLU(name='stage2_unit1_relu3')(stage2_unit1_bn3)129 130 stage2_unit1_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit1_relu3)131 132 plus3 = Add()([stage2_unit1_conv3 , stage2_unit1_sc])133 134 stage2_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn1', trainable=False)(plus3)135 136 stage2_unit2_relu1 = ReLU(name='stage2_unit2_relu1')(stage2_unit2_bn1)137 138 stage2_unit2_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_relu1)139 140 stage2_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn2', trainable=False)(stage2_unit2_conv1)141 142 stage2_unit2_relu2 = ReLU(name='stage2_unit2_relu2')(stage2_unit2_bn2)143 144 stage2_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit2_relu2)145 146 stage2_unit2_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_conv2_pad)147 148 stage2_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit2_bn3', trainable=False)(stage2_unit2_conv2)149 150 stage2_unit2_relu3 = ReLU(name='stage2_unit2_relu3')(stage2_unit2_bn3)151 152 stage2_unit2_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit2_relu3)153 154 plus4 = Add()([stage2_unit2_conv3 , plus3])155 156 stage2_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn1', trainable=False)(plus4)157 158 stage2_unit3_relu1 = ReLU(name='stage2_unit3_relu1')(stage2_unit3_bn1)159 160 stage2_unit3_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_relu1)161 162 stage2_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn2', trainable=False)(stage2_unit3_conv1)163 164 stage2_unit3_relu2 = ReLU(name='stage2_unit3_relu2')(stage2_unit3_bn2)165 166 stage2_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit3_relu2)167 168 stage2_unit3_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_conv2_pad)169 170 stage2_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit3_bn3', trainable=False)(stage2_unit3_conv2)171 172 stage2_unit3_relu3 = ReLU(name='stage2_unit3_relu3')(stage2_unit3_bn3)173 174 stage2_unit3_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit3_relu3)175 176 plus5 = Add()([stage2_unit3_conv3 , plus4])177 178 stage2_unit4_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn1', trainable=False)(plus5)179 180 stage2_unit4_relu1 = ReLU(name='stage2_unit4_relu1')(stage2_unit4_bn1)181 182 stage2_unit4_conv1 = Conv2D(filters = 128, kernel_size = (1, 1), name = 'stage2_unit4_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_relu1)183 184 stage2_unit4_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn2', trainable=False)(stage2_unit4_conv1)185 186 stage2_unit4_relu2 = ReLU(name='stage2_unit4_relu2')(stage2_unit4_bn2)187 188 stage2_unit4_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage2_unit4_relu2)189 190 stage2_unit4_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'stage2_unit4_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_conv2_pad)191 192 stage2_unit4_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage2_unit4_bn3', trainable=False)(stage2_unit4_conv2)193 194 stage2_unit4_relu3 = ReLU(name='stage2_unit4_relu3')(stage2_unit4_bn3)195 196 stage2_unit4_conv3 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage2_unit4_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage2_unit4_relu3)197 198 plus6 = Add()([stage2_unit4_conv3 , plus5])199 200 stage3_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn1', trainable=False)(plus6)201 202 stage3_unit1_relu1 = ReLU(name='stage3_unit1_relu1')(stage3_unit1_bn1)203 204 stage3_unit1_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit1_relu1)205 206 stage3_unit1_sc = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage3_unit1_relu1)207 208 stage3_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn2', trainable=False)(stage3_unit1_conv1)209 210 stage3_unit1_relu2 = ReLU(name='stage3_unit1_relu2')(stage3_unit1_bn2)211 212 stage3_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit1_relu2)213 214 stage3_unit1_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage3_unit1_conv2_pad)215 216 ssh_m1_red_conv = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_m1_red_conv', strides = [1, 1], padding = 'VALID', use_bias = True)(stage3_unit1_relu2)217 218 stage3_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit1_bn3', trainable=False)(stage3_unit1_conv2)219 220 ssh_m1_red_conv_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_red_conv_bn', trainable=False)(ssh_m1_red_conv)221 222 stage3_unit1_relu3 = ReLU(name='stage3_unit1_relu3')(stage3_unit1_bn3)223 224 ssh_m1_red_conv_relu = ReLU(name='ssh_m1_red_conv_relu')(ssh_m1_red_conv_bn)225 226 stage3_unit1_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit1_relu3)227 228 plus7 = Add()([stage3_unit1_conv3 , stage3_unit1_sc])229 230 stage3_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn1', trainable=False)(plus7)231 232 stage3_unit2_relu1 = ReLU(name='stage3_unit2_relu1')(stage3_unit2_bn1)233 234 stage3_unit2_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_relu1)235 236 stage3_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn2', trainable=False)(stage3_unit2_conv1)237 238 stage3_unit2_relu2 = ReLU(name='stage3_unit2_relu2')(stage3_unit2_bn2)239 240 stage3_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit2_relu2)241 242 stage3_unit2_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_conv2_pad)243 244 stage3_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit2_bn3', trainable=False)(stage3_unit2_conv2)245 246 stage3_unit2_relu3 = ReLU(name='stage3_unit2_relu3')(stage3_unit2_bn3)247 248 stage3_unit2_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit2_relu3)249 250 plus8 = Add()([stage3_unit2_conv3 , plus7])251 252 stage3_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn1', trainable=False)(plus8)253 254 stage3_unit3_relu1 = ReLU(name='stage3_unit3_relu1')(stage3_unit3_bn1)255 256 stage3_unit3_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_relu1)257 258 stage3_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn2', trainable=False)(stage3_unit3_conv1)259 260 stage3_unit3_relu2 = ReLU(name='stage3_unit3_relu2')(stage3_unit3_bn2)261 262 stage3_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit3_relu2)263 264 stage3_unit3_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_conv2_pad)265 266 stage3_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit3_bn3', trainable=False)(stage3_unit3_conv2)267 268 stage3_unit3_relu3 = ReLU(name='stage3_unit3_relu3')(stage3_unit3_bn3)269 270 stage3_unit3_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit3_relu3)271 272 plus9 = Add()([stage3_unit3_conv3 , plus8])273 274 stage3_unit4_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn1', trainable=False)(plus9)275 276 stage3_unit4_relu1 = ReLU(name='stage3_unit4_relu1')(stage3_unit4_bn1)277 278 stage3_unit4_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit4_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_relu1)279 280 stage3_unit4_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn2', trainable=False)(stage3_unit4_conv1)281 282 stage3_unit4_relu2 = ReLU(name='stage3_unit4_relu2')(stage3_unit4_bn2)283 284 stage3_unit4_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit4_relu2)285 286 stage3_unit4_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit4_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_conv2_pad)287 288 stage3_unit4_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit4_bn3', trainable=False)(stage3_unit4_conv2)289 290 stage3_unit4_relu3 = ReLU(name='stage3_unit4_relu3')(stage3_unit4_bn3)291 292 stage3_unit4_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit4_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit4_relu3)293 294 plus10 = Add()([stage3_unit4_conv3 , plus9])295 296 stage3_unit5_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn1', trainable=False)(plus10)297 298 stage3_unit5_relu1 = ReLU(name='stage3_unit5_relu1')(stage3_unit5_bn1)299 300 stage3_unit5_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit5_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_relu1)301 302 stage3_unit5_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn2', trainable=False)(stage3_unit5_conv1)303 304 stage3_unit5_relu2 = ReLU(name='stage3_unit5_relu2')(stage3_unit5_bn2)305 306 stage3_unit5_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit5_relu2)307 308 stage3_unit5_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit5_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_conv2_pad)309 310 stage3_unit5_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit5_bn3', trainable=False)(stage3_unit5_conv2)311 312 stage3_unit5_relu3 = ReLU(name='stage3_unit5_relu3')(stage3_unit5_bn3)313 314 stage3_unit5_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit5_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit5_relu3)315 316 plus11 = Add()([stage3_unit5_conv3 , plus10])317 318 stage3_unit6_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn1', trainable=False)(plus11)319 320 stage3_unit6_relu1 = ReLU(name='stage3_unit6_relu1')(stage3_unit6_bn1)321 322 stage3_unit6_conv1 = Conv2D(filters = 256, kernel_size = (1, 1), name = 'stage3_unit6_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_relu1)323 324 stage3_unit6_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn2', trainable=False)(stage3_unit6_conv1)325 326 stage3_unit6_relu2 = ReLU(name='stage3_unit6_relu2')(stage3_unit6_bn2)327 328 stage3_unit6_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage3_unit6_relu2)329 330 stage3_unit6_conv2 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'stage3_unit6_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_conv2_pad)331 332 stage3_unit6_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage3_unit6_bn3', trainable=False)(stage3_unit6_conv2)333 334 stage3_unit6_relu3 = ReLU(name='stage3_unit6_relu3')(stage3_unit6_bn3)335 336 stage3_unit6_conv3 = Conv2D(filters = 1024, kernel_size = (1, 1), name = 'stage3_unit6_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage3_unit6_relu3)337 338 plus12 = Add()([stage3_unit6_conv3 , plus11])339 340 stage4_unit1_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn1', trainable=False)(plus12)341 342 stage4_unit1_relu1 = ReLU(name='stage4_unit1_relu1')(stage4_unit1_bn1)343 344 stage4_unit1_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit1_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit1_relu1)345 346 stage4_unit1_sc = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit1_sc', strides = [2, 2], padding = 'VALID', use_bias = False)(stage4_unit1_relu1)347 348 stage4_unit1_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn2', trainable=False)(stage4_unit1_conv1)349 350 stage4_unit1_relu2 = ReLU(name='stage4_unit1_relu2')(stage4_unit1_bn2)351 352 stage4_unit1_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit1_relu2)353 354 stage4_unit1_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit1_conv2', strides = [2, 2], padding = 'VALID', use_bias = False)(stage4_unit1_conv2_pad)355 356 ssh_c2_lateral = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_c2_lateral', strides = [1, 1], padding = 'VALID', use_bias = True)(stage4_unit1_relu2)357 358 stage4_unit1_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit1_bn3', trainable=False)(stage4_unit1_conv2)359 360 ssh_c2_lateral_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c2_lateral_bn', trainable=False)(ssh_c2_lateral)361 362 stage4_unit1_relu3 = ReLU(name='stage4_unit1_relu3')(stage4_unit1_bn3)363 364 ssh_c2_lateral_relu = ReLU(name='ssh_c2_lateral_relu')(ssh_c2_lateral_bn)365 366 stage4_unit1_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit1_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit1_relu3)367 368 plus13 = Add()([stage4_unit1_conv3 , stage4_unit1_sc])369 370 stage4_unit2_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn1', trainable=False)(plus13)371 372 stage4_unit2_relu1 = ReLU(name='stage4_unit2_relu1')(stage4_unit2_bn1)373 374 stage4_unit2_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit2_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_relu1)375 376 stage4_unit2_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn2', trainable=False)(stage4_unit2_conv1)377 378 stage4_unit2_relu2 = ReLU(name='stage4_unit2_relu2')(stage4_unit2_bn2)379 380 stage4_unit2_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit2_relu2)381 382 stage4_unit2_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit2_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_conv2_pad)383 384 stage4_unit2_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit2_bn3', trainable=False)(stage4_unit2_conv2)385 386 stage4_unit2_relu3 = ReLU(name='stage4_unit2_relu3')(stage4_unit2_bn3)387 388 stage4_unit2_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit2_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit2_relu3)389 390 plus14 = Add()([stage4_unit2_conv3 , plus13])391 392 stage4_unit3_bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn1', trainable=False)(plus14)393 394 stage4_unit3_relu1 = ReLU(name='stage4_unit3_relu1')(stage4_unit3_bn1)395 396 stage4_unit3_conv1 = Conv2D(filters = 512, kernel_size = (1, 1), name = 'stage4_unit3_conv1', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_relu1)397 398 stage4_unit3_bn2 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn2', trainable=False)(stage4_unit3_conv1)399 400 stage4_unit3_relu2 = ReLU(name='stage4_unit3_relu2')(stage4_unit3_bn2)401 402 stage4_unit3_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(stage4_unit3_relu2)403 404 stage4_unit3_conv2 = Conv2D(filters = 512, kernel_size = (3, 3), name = 'stage4_unit3_conv2', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_conv2_pad)405 406 stage4_unit3_bn3 = BatchNormalization(epsilon=1.9999999494757503e-05, name='stage4_unit3_bn3', trainable=False)(stage4_unit3_conv2)407 408 stage4_unit3_relu3 = ReLU(name='stage4_unit3_relu3')(stage4_unit3_bn3)409 410 stage4_unit3_conv3 = Conv2D(filters = 2048, kernel_size = (1, 1), name = 'stage4_unit3_conv3', strides = [1, 1], padding = 'VALID', use_bias = False)(stage4_unit3_relu3)411 412 plus15 = Add()([stage4_unit3_conv3 , plus14])413 414 bn1 = BatchNormalization(epsilon=1.9999999494757503e-05, name='bn1', trainable=False)(plus15)415 416 relu1 = ReLU(name='relu1')(bn1)417 418 ssh_c3_lateral = Conv2D(filters = 256, kernel_size = (1, 1), name = 'ssh_c3_lateral', strides = [1, 1], padding = 'VALID', use_bias = True)(relu1)419 420 ssh_c3_lateral_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c3_lateral_bn', trainable=False)(ssh_c3_lateral)421 422 ssh_c3_lateral_relu = ReLU(name='ssh_c3_lateral_relu')(ssh_c3_lateral_bn)423 424 ssh_m3_det_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c3_lateral_relu)425 426 ssh_m3_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m3_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_conv1_pad)427 428 ssh_m3_det_context_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c3_lateral_relu)429 430 ssh_m3_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv1_pad)431 432 ssh_c3_up = UpSampling2D(size=(2, 2), interpolation="nearest", name="ssh_c3_up")(ssh_c3_lateral_relu)433 434 ssh_m3_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_conv1_bn', trainable=False)(ssh_m3_det_conv1)435 436 ssh_m3_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv1_bn', trainable=False)(ssh_m3_det_context_conv1)437 438 x1_shape = tf.shape(ssh_c3_up)439 x2_shape = tf.shape(ssh_c2_lateral_relu)440 offsets = [0, (x1_shape[1] - x2_shape[1]) // 2, (x1_shape[2] - x2_shape[2]) // 2, 0]441 size = [-1, x2_shape[1], x2_shape[2], -1]442 crop0 = tf.slice(ssh_c3_up, offsets, size, "crop0")443 444 ssh_m3_det_context_conv1_relu = ReLU(name='ssh_m3_det_context_conv1_relu')(ssh_m3_det_context_conv1_bn)445 446 plus0_v2 = Add()([ssh_c2_lateral_relu , crop0])447 448 ssh_m3_det_context_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv1_relu)449 450 ssh_m3_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv2_pad)451 452 ssh_m3_det_context_conv3_1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv1_relu)453 454 ssh_m3_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv3_1_pad)455 456 ssh_c2_aggr_pad = ZeroPadding2D(padding=tuple([1, 1]))(plus0_v2)457 458 ssh_c2_aggr = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_c2_aggr', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_c2_aggr_pad)459 460 ssh_m3_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv2_bn', trainable=False)(ssh_m3_det_context_conv2)461 462 ssh_m3_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv3_1_bn', trainable=False)(ssh_m3_det_context_conv3_1)463 464 ssh_c2_aggr_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c2_aggr_bn', trainable=False)(ssh_c2_aggr)465 466 ssh_m3_det_context_conv3_1_relu = ReLU(name='ssh_m3_det_context_conv3_1_relu')(ssh_m3_det_context_conv3_1_bn)467 468 ssh_c2_aggr_relu = ReLU(name='ssh_c2_aggr_relu')(ssh_c2_aggr_bn)469 470 ssh_m3_det_context_conv3_2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m3_det_context_conv3_1_relu)471 472 ssh_m3_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m3_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_context_conv3_2_pad)473 474 ssh_m2_det_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c2_aggr_relu)475 476 ssh_m2_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m2_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_conv1_pad)477 478 ssh_m2_det_context_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c2_aggr_relu)479 480 ssh_m2_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv1_pad)481 482 ssh_m2_red_up = UpSampling2D(size=(2, 2), interpolation="nearest", name="ssh_m2_red_up")(ssh_c2_aggr_relu)483 484 ssh_m3_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m3_det_context_conv3_2_bn', trainable=False)(ssh_m3_det_context_conv3_2)485 486 ssh_m2_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_conv1_bn', trainable=False)(ssh_m2_det_conv1)487 488 ssh_m2_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv1_bn', trainable=False)(ssh_m2_det_context_conv1)489 490 x1_shape = tf.shape(ssh_m2_red_up)491 x2_shape = tf.shape(ssh_m1_red_conv_relu)492 offsets = [0, (x1_shape[1] - x2_shape[1]) // 2, (x1_shape[2] - x2_shape[2]) // 2, 0]493 size = [-1, x2_shape[1], x2_shape[2], -1]494 crop1 = tf.slice(ssh_m2_red_up, offsets, size, "crop1")495 496 ssh_m3_det_concat = concatenate([ssh_m3_det_conv1_bn, ssh_m3_det_context_conv2_bn, ssh_m3_det_context_conv3_2_bn], 3, name='ssh_m3_det_concat')497 498 ssh_m2_det_context_conv1_relu = ReLU(name='ssh_m2_det_context_conv1_relu')(ssh_m2_det_context_conv1_bn)499 500 plus1_v1 = Add()([ssh_m1_red_conv_relu , crop1])501 502 ssh_m3_det_concat_relu = ReLU(name='ssh_m3_det_concat_relu')(ssh_m3_det_concat)503 504 ssh_m2_det_context_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv1_relu)505 506 ssh_m2_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv2_pad)507 508 ssh_m2_det_context_conv3_1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv1_relu)509 510 ssh_m2_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv3_1_pad)511 512 ssh_c1_aggr_pad = ZeroPadding2D(padding=tuple([1, 1]))(plus1_v1)513 514 ssh_c1_aggr = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_c1_aggr', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_c1_aggr_pad)515 516 face_rpn_cls_score_stride32 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)517 518 inter_1 = concatenate([face_rpn_cls_score_stride32[:, :, :, 0], face_rpn_cls_score_stride32[:, :, :, 1]], axis=1)519 inter_2 = concatenate([face_rpn_cls_score_stride32[:, :, :, 2], face_rpn_cls_score_stride32[:, :, :, 3]], axis=1)520 final = tf.stack([inter_1, inter_2])521 face_rpn_cls_score_reshape_stride32 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride32")522 523 face_rpn_bbox_pred_stride32 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)524 525 face_rpn_landmark_pred_stride32 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride32', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m3_det_concat_relu)526 527 ssh_m2_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv2_bn', trainable=False)(ssh_m2_det_context_conv2)528 529 ssh_m2_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv3_1_bn', trainable=False)(ssh_m2_det_context_conv3_1)530 531 ssh_c1_aggr_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_c1_aggr_bn', trainable=False)(ssh_c1_aggr)532 533 ssh_m2_det_context_conv3_1_relu = ReLU(name='ssh_m2_det_context_conv3_1_relu')(ssh_m2_det_context_conv3_1_bn)534 535 ssh_c1_aggr_relu = ReLU(name='ssh_c1_aggr_relu')(ssh_c1_aggr_bn)536 537 face_rpn_cls_prob_stride32 = Softmax(name = 'face_rpn_cls_prob_stride32')(face_rpn_cls_score_reshape_stride32)538 539 input_shape = [tf.shape(face_rpn_cls_prob_stride32)[k] for k in range(4)]540 sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)541 inter_1 = face_rpn_cls_prob_stride32[:, 0:sz, :, 0]542 inter_2 = face_rpn_cls_prob_stride32[:, 0:sz, :, 1]543 inter_3 = face_rpn_cls_prob_stride32[:, sz:, :, 0]544 inter_4 = face_rpn_cls_prob_stride32[:, sz:, :, 1]545 final = tf.stack([inter_1, inter_3, inter_2, inter_4])546 face_rpn_cls_prob_reshape_stride32 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride32")547 548 ssh_m2_det_context_conv3_2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m2_det_context_conv3_1_relu)549 550 ssh_m2_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m2_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_context_conv3_2_pad)551 552 ssh_m1_det_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c1_aggr_relu)553 554 ssh_m1_det_conv1 = Conv2D(filters = 256, kernel_size = (3, 3), name = 'ssh_m1_det_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_conv1_pad)555 556 ssh_m1_det_context_conv1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_c1_aggr_relu)557 558 ssh_m1_det_context_conv1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv1_pad)559 560 ssh_m2_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m2_det_context_conv3_2_bn', trainable=False)(ssh_m2_det_context_conv3_2)561 562 ssh_m1_det_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_conv1_bn', trainable=False)(ssh_m1_det_conv1)563 564 ssh_m1_det_context_conv1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv1_bn', trainable=False)(ssh_m1_det_context_conv1)565 566 ssh_m2_det_concat = concatenate([ssh_m2_det_conv1_bn, ssh_m2_det_context_conv2_bn, ssh_m2_det_context_conv3_2_bn], 3, name='ssh_m2_det_concat')567 568 ssh_m1_det_context_conv1_relu = ReLU(name='ssh_m1_det_context_conv1_relu')(ssh_m1_det_context_conv1_bn)569 570 ssh_m2_det_concat_relu = ReLU(name='ssh_m2_det_concat_relu')(ssh_m2_det_concat)571 572 ssh_m1_det_context_conv2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv1_relu)573 574 ssh_m1_det_context_conv2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv2_pad)575 576 ssh_m1_det_context_conv3_1_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv1_relu)577 578 ssh_m1_det_context_conv3_1 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv3_1', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv3_1_pad)579 580 face_rpn_cls_score_stride16 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)581 582 inter_1 = concatenate([face_rpn_cls_score_stride16[:, :, :, 0], face_rpn_cls_score_stride16[:, :, :, 1]], axis=1)583 inter_2 = concatenate([face_rpn_cls_score_stride16[:, :, :, 2], face_rpn_cls_score_stride16[:, :, :, 3]], axis=1)584 final = tf.stack([inter_1, inter_2])585 face_rpn_cls_score_reshape_stride16 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride16")586 587 face_rpn_bbox_pred_stride16 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)588 589 face_rpn_landmark_pred_stride16 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride16', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m2_det_concat_relu)590 591 ssh_m1_det_context_conv2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv2_bn', trainable=False)(ssh_m1_det_context_conv2)592 593 ssh_m1_det_context_conv3_1_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv3_1_bn', trainable=False)(ssh_m1_det_context_conv3_1)594 595 ssh_m1_det_context_conv3_1_relu = ReLU(name='ssh_m1_det_context_conv3_1_relu')(ssh_m1_det_context_conv3_1_bn)596 597 face_rpn_cls_prob_stride16 = Softmax(name = 'face_rpn_cls_prob_stride16')(face_rpn_cls_score_reshape_stride16)598 599 input_shape = [tf.shape(face_rpn_cls_prob_stride16)[k] for k in range(4)]600 sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)601 inter_1 = face_rpn_cls_prob_stride16[:, 0:sz, :, 0]602 inter_2 = face_rpn_cls_prob_stride16[:, 0:sz, :, 1]603 inter_3 = face_rpn_cls_prob_stride16[:, sz:, :, 0]604 inter_4 = face_rpn_cls_prob_stride16[:, sz:, :, 1]605 final = tf.stack([inter_1, inter_3, inter_2, inter_4])606 face_rpn_cls_prob_reshape_stride16 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride16")607 608 ssh_m1_det_context_conv3_2_pad = ZeroPadding2D(padding=tuple([1, 1]))(ssh_m1_det_context_conv3_1_relu)609 610 ssh_m1_det_context_conv3_2 = Conv2D(filters = 128, kernel_size = (3, 3), name = 'ssh_m1_det_context_conv3_2', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_context_conv3_2_pad)611 612 ssh_m1_det_context_conv3_2_bn = BatchNormalization(epsilon=1.9999999494757503e-05, name='ssh_m1_det_context_conv3_2_bn', trainable=False)(ssh_m1_det_context_conv3_2)613 614 ssh_m1_det_concat = concatenate([ssh_m1_det_conv1_bn, ssh_m1_det_context_conv2_bn, ssh_m1_det_context_conv3_2_bn], 3, name='ssh_m1_det_concat')615 616 ssh_m1_det_concat_relu = ReLU(name='ssh_m1_det_concat_relu')(ssh_m1_det_concat)617 face_rpn_cls_score_stride8 = Conv2D(filters = 4, kernel_size = (1, 1), name = 'face_rpn_cls_score_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)618 619 inter_1 = concatenate([face_rpn_cls_score_stride8[:, :, :, 0], face_rpn_cls_score_stride8[:, :, :, 1]], axis=1)620 inter_2 = concatenate([face_rpn_cls_score_stride8[:, :, :, 2], face_rpn_cls_score_stride8[:, :, :, 3]], axis=1)621 final = tf.stack([inter_1, inter_2])622 face_rpn_cls_score_reshape_stride8 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_score_reshape_stride8")623 624 face_rpn_bbox_pred_stride8 = Conv2D(filters = 8, kernel_size = (1, 1), name = 'face_rpn_bbox_pred_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)625 626 face_rpn_landmark_pred_stride8 = Conv2D(filters = 20, kernel_size = (1, 1), name = 'face_rpn_landmark_pred_stride8', strides = [1, 1], padding = 'VALID', use_bias = True)(ssh_m1_det_concat_relu)627 628 face_rpn_cls_prob_stride8 = Softmax(name = 'face_rpn_cls_prob_stride8')(face_rpn_cls_score_reshape_stride8)629 630 input_shape = [tf.shape(face_rpn_cls_prob_stride8)[k] for k in range(4)]631 sz = tf.dtypes.cast(input_shape[1] / 2, dtype=tf.int32)632 inter_1 = face_rpn_cls_prob_stride8[:, 0:sz, :, 0]633 inter_2 = face_rpn_cls_prob_stride8[:, 0:sz, :, 1]634 inter_3 = face_rpn_cls_prob_stride8[:, sz:, :, 0]635 inter_4 = face_rpn_cls_prob_stride8[:, sz:, :, 1]636 final = tf.stack([inter_1, inter_3, inter_2, inter_4])637 face_rpn_cls_prob_reshape_stride8 = tf.transpose(final, (1, 2, 3, 0), name="face_rpn_cls_prob_reshape_stride8")638 639 model = Model(inputs=data,640 outputs=[face_rpn_cls_prob_reshape_stride32,641 face_rpn_bbox_pred_stride32,642 face_rpn_landmark_pred_stride32,643 face_rpn_cls_prob_reshape_stride16,644 face_rpn_bbox_pred_stride16,645 face_rpn_landmark_pred_stride16,646 face_rpn_cls_prob_reshape_stride8,647 face_rpn_bbox_pred_stride8,648 face_rpn_landmark_pred_stride8649 ])650 model = load_weights(model)651 652 return model653