1 # Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # ============================================================================== 15 """Contains the definition for inception v1 classification network.""" 16 17 from __future__ import absolute_import 18 from __future__ import division 19 from __future__ import print_function 20 21 from tensorflow.contrib import layers 22 from tensorflow.contrib.framework.python.ops import arg_scope 23 from tensorflow.contrib.layers.python.layers import initializers 24 from tensorflow.contrib.layers.python.layers import layers as layers_lib 25 from tensorflow.contrib.layers.python.layers import regularizers 26 from tensorflow.python.framework import ops 27 from tensorflow.python.ops import array_ops 28 from tensorflow.python.ops import init_ops 29 from tensorflow.python.ops import nn_ops 30 from tensorflow.python.ops import variable_scope 31 32 trunc_normal = lambda stddev: init_ops.truncated_normal_initializer(0.0, stddev) 33 34 35 def inception_v1_base(inputs, final_endpoint='Mixed_5c', scope='InceptionV1'): 36 """Defines the Inception V1 base architecture. 37 38 This architecture is defined in: 39 Going deeper with convolutions 40 Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, 41 Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich. 42 http://arxiv.org/pdf/1409.4842v1.pdf. 43 44 Args: 45 inputs: a tensor of size [batch_size, height, width, channels]. 46 final_endpoint: specifies the endpoint to construct the network up to. It 47 can be one of ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1', 48 'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c', 49 'MaxPool_4a_3x3', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d', 'Mixed_4e', 50 'Mixed_4f', 'MaxPool_5a_2x2', 'Mixed_5b', 'Mixed_5c'] 51 scope: Optional variable_scope. 52 53 Returns: 54 A dictionary from components of the network to the corresponding activation. 55 56 Raises: 57 ValueError: if final_endpoint is not set to one of the predefined values. 58 """ 59 end_points = {} 60 with variable_scope.variable_scope(scope, 'InceptionV1', [inputs]): 61 with arg_scope( 62 [layers.conv2d, layers_lib.fully_connected], 63 weights_initializer=trunc_normal(0.01)): 64 with arg_scope( 65 [layers.conv2d, layers_lib.max_pool2d], stride=1, padding='SAME'): 66 end_point = 'Conv2d_1a_7x7' 67 net = layers.conv2d(inputs, 64, [7, 7], stride=2, scope=end_point) 68 end_points[end_point] = net 69 if final_endpoint == end_point: 70 return net, end_points 71 end_point = 'MaxPool_2a_3x3' 72 net = layers_lib.max_pool2d(net, [3, 3], stride=2, scope=end_point) 73 end_points[end_point] = net 74 if final_endpoint == end_point: 75 return net, end_points 76 end_point = 'Conv2d_2b_1x1' 77 net = layers.conv2d(net, 64, [1, 1], scope=end_point) 78 end_points[end_point] = net 79 if final_endpoint == end_point: 80 return net, end_points 81 end_point = 'Conv2d_2c_3x3' 82 net = layers.conv2d(net, 192, [3, 3], scope=end_point) 83 end_points[end_point] = net 84 if final_endpoint == end_point: 85 return net, end_points 86 end_point = 'MaxPool_3a_3x3' 87 net = layers_lib.max_pool2d(net, [3, 3], stride=2, scope=end_point) 88 end_points[end_point] = net 89 if final_endpoint == end_point: 90 return net, end_points 91 92 end_point = 'Mixed_3b' 93 with variable_scope.variable_scope(end_point): 94 with variable_scope.variable_scope('Branch_0'): 95 branch_0 = layers.conv2d(net, 64, [1, 1], scope='Conv2d_0a_1x1') 96 with variable_scope.variable_scope('Branch_1'): 97 branch_1 = layers.conv2d(net, 96, [1, 1], scope='Conv2d_0a_1x1') 98 branch_1 = layers.conv2d( 99 branch_1, 128, [3, 3], scope='Conv2d_0b_3x3') 100 with variable_scope.variable_scope('Branch_2'): 101 branch_2 = layers.conv2d(net, 16, [1, 1], scope='Conv2d_0a_1x1') 102 branch_2 = layers.conv2d( 103 branch_2, 32, [3, 3], scope='Conv2d_0b_3x3') 104 with variable_scope.variable_scope('Branch_3'): 105 branch_3 = layers_lib.max_pool2d( 106 net, [3, 3], scope='MaxPool_0a_3x3') 107 branch_3 = layers.conv2d( 108 branch_3, 32, [1, 1], scope='Conv2d_0b_1x1') 109 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 110 end_points[end_point] = net 111 if final_endpoint == end_point: 112 return net, end_points 113 114 end_point = 'Mixed_3c' 115 with variable_scope.variable_scope(end_point): 116 with variable_scope.variable_scope('Branch_0'): 117 branch_0 = layers.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1') 118 with variable_scope.variable_scope('Branch_1'): 119 branch_1 = layers.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1') 120 branch_1 = layers.conv2d( 121 branch_1, 192, [3, 3], scope='Conv2d_0b_3x3') 122 with variable_scope.variable_scope('Branch_2'): 123 branch_2 = layers.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1') 124 branch_2 = layers.conv2d( 125 branch_2, 96, [3, 3], scope='Conv2d_0b_3x3') 126 with variable_scope.variable_scope('Branch_3'): 127 branch_3 = layers_lib.max_pool2d( 128 net, [3, 3], scope='MaxPool_0a_3x3') 129 branch_3 = layers.conv2d( 130 branch_3, 64, [1, 1], scope='Conv2d_0b_1x1') 131 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 132 end_points[end_point] = net 133 if final_endpoint == end_point: 134 return net, end_points 135 136 end_point = 'MaxPool_4a_3x3' 137 net = layers_lib.max_pool2d(net, [3, 3], stride=2, scope=end_point) 138 end_points[end_point] = net 139 if final_endpoint == end_point: 140 return net, end_points 141 142 end_point = 'Mixed_4b' 143 with variable_scope.variable_scope(end_point): 144 with variable_scope.variable_scope('Branch_0'): 145 branch_0 = layers.conv2d(net, 192, [1, 1], scope='Conv2d_0a_1x1') 146 with variable_scope.variable_scope('Branch_1'): 147 branch_1 = layers.conv2d(net, 96, [1, 1], scope='Conv2d_0a_1x1') 148 branch_1 = layers.conv2d( 149 branch_1, 208, [3, 3], scope='Conv2d_0b_3x3') 150 with variable_scope.variable_scope('Branch_2'): 151 branch_2 = layers.conv2d(net, 16, [1, 1], scope='Conv2d_0a_1x1') 152 branch_2 = layers.conv2d( 153 branch_2, 48, [3, 3], scope='Conv2d_0b_3x3') 154 with variable_scope.variable_scope('Branch_3'): 155 branch_3 = layers_lib.max_pool2d( 156 net, [3, 3], scope='MaxPool_0a_3x3') 157 branch_3 = layers.conv2d( 158 branch_3, 64, [1, 1], scope='Conv2d_0b_1x1') 159 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 160 end_points[end_point] = net 161 if final_endpoint == end_point: 162 return net, end_points 163 164 end_point = 'Mixed_4c' 165 with variable_scope.variable_scope(end_point): 166 with variable_scope.variable_scope('Branch_0'): 167 branch_0 = layers.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1') 168 with variable_scope.variable_scope('Branch_1'): 169 branch_1 = layers.conv2d(net, 112, [1, 1], scope='Conv2d_0a_1x1') 170 branch_1 = layers.conv2d( 171 branch_1, 224, [3, 3], scope='Conv2d_0b_3x3') 172 with variable_scope.variable_scope('Branch_2'): 173 branch_2 = layers.conv2d(net, 24, [1, 1], scope='Conv2d_0a_1x1') 174 branch_2 = layers.conv2d( 175 branch_2, 64, [3, 3], scope='Conv2d_0b_3x3') 176 with variable_scope.variable_scope('Branch_3'): 177 branch_3 = layers_lib.max_pool2d( 178 net, [3, 3], scope='MaxPool_0a_3x3') 179 branch_3 = layers.conv2d( 180 branch_3, 64, [1, 1], scope='Conv2d_0b_1x1') 181 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 182 end_points[end_point] = net 183 if final_endpoint == end_point: 184 return net, end_points 185 186 end_point = 'Mixed_4d' 187 with variable_scope.variable_scope(end_point): 188 with variable_scope.variable_scope('Branch_0'): 189 branch_0 = layers.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1') 190 with variable_scope.variable_scope('Branch_1'): 191 branch_1 = layers.conv2d(net, 128, [1, 1], scope='Conv2d_0a_1x1') 192 branch_1 = layers.conv2d( 193 branch_1, 256, [3, 3], scope='Conv2d_0b_3x3') 194 with variable_scope.variable_scope('Branch_2'): 195 branch_2 = layers.conv2d(net, 24, [1, 1], scope='Conv2d_0a_1x1') 196 branch_2 = layers.conv2d( 197 branch_2, 64, [3, 3], scope='Conv2d_0b_3x3') 198 with variable_scope.variable_scope('Branch_3'): 199 branch_3 = layers_lib.max_pool2d( 200 net, [3, 3], scope='MaxPool_0a_3x3') 201 branch_3 = layers.conv2d( 202 branch_3, 64, [1, 1], scope='Conv2d_0b_1x1') 203 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 204 end_points[end_point] = net 205 if final_endpoint == end_point: 206 return net, end_points 207 208 end_point = 'Mixed_4e' 209 with variable_scope.variable_scope(end_point): 210 with variable_scope.variable_scope('Branch_0'): 211 branch_0 = layers.conv2d(net, 112, [1, 1], scope='Conv2d_0a_1x1') 212 with variable_scope.variable_scope('Branch_1'): 213 branch_1 = layers.conv2d(net, 144, [1, 1], scope='Conv2d_0a_1x1') 214 branch_1 = layers.conv2d( 215 branch_1, 288, [3, 3], scope='Conv2d_0b_3x3') 216 with variable_scope.variable_scope('Branch_2'): 217 branch_2 = layers.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1') 218 branch_2 = layers.conv2d( 219 branch_2, 64, [3, 3], scope='Conv2d_0b_3x3') 220 with variable_scope.variable_scope('Branch_3'): 221 branch_3 = layers_lib.max_pool2d( 222 net, [3, 3], scope='MaxPool_0a_3x3') 223 branch_3 = layers.conv2d( 224 branch_3, 64, [1, 1], scope='Conv2d_0b_1x1') 225 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 226 end_points[end_point] = net 227 if final_endpoint == end_point: 228 return net, end_points 229 230 end_point = 'Mixed_4f' 231 with variable_scope.variable_scope(end_point): 232 with variable_scope.variable_scope('Branch_0'): 233 branch_0 = layers.conv2d(net, 256, [1, 1], scope='Conv2d_0a_1x1') 234 with variable_scope.variable_scope('Branch_1'): 235 branch_1 = layers.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1') 236 branch_1 = layers.conv2d( 237 branch_1, 320, [3, 3], scope='Conv2d_0b_3x3') 238 with variable_scope.variable_scope('Branch_2'): 239 branch_2 = layers.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1') 240 branch_2 = layers.conv2d( 241 branch_2, 128, [3, 3], scope='Conv2d_0b_3x3') 242 with variable_scope.variable_scope('Branch_3'): 243 branch_3 = layers_lib.max_pool2d( 244 net, [3, 3], scope='MaxPool_0a_3x3') 245 branch_3 = layers.conv2d( 246 branch_3, 128, [1, 1], scope='Conv2d_0b_1x1') 247 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 248 end_points[end_point] = net 249 if final_endpoint == end_point: 250 return net, end_points 251 252 end_point = 'MaxPool_5a_2x2' 253 net = layers_lib.max_pool2d(net, [2, 2], stride=2, scope=end_point) 254 end_points[end_point] = net 255 if final_endpoint == end_point: 256 return net, end_points 257 258 end_point = 'Mixed_5b' 259 with variable_scope.variable_scope(end_point): 260 with variable_scope.variable_scope('Branch_0'): 261 branch_0 = layers.conv2d(net, 256, [1, 1], scope='Conv2d_0a_1x1') 262 with variable_scope.variable_scope('Branch_1'): 263 branch_1 = layers.conv2d(net, 160, [1, 1], scope='Conv2d_0a_1x1') 264 branch_1 = layers.conv2d( 265 branch_1, 320, [3, 3], scope='Conv2d_0b_3x3') 266 with variable_scope.variable_scope('Branch_2'): 267 branch_2 = layers.conv2d(net, 32, [1, 1], scope='Conv2d_0a_1x1') 268 branch_2 = layers.conv2d( 269 branch_2, 128, [3, 3], scope='Conv2d_0a_3x3') 270 with variable_scope.variable_scope('Branch_3'): 271 branch_3 = layers_lib.max_pool2d( 272 net, [3, 3], scope='MaxPool_0a_3x3') 273 branch_3 = layers.conv2d( 274 branch_3, 128, [1, 1], scope='Conv2d_0b_1x1') 275 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 276 end_points[end_point] = net 277 if final_endpoint == end_point: 278 return net, end_points 279 280 end_point = 'Mixed_5c' 281 with variable_scope.variable_scope(end_point): 282 with variable_scope.variable_scope('Branch_0'): 283 branch_0 = layers.conv2d(net, 384, [1, 1], scope='Conv2d_0a_1x1') 284 with variable_scope.variable_scope('Branch_1'): 285 branch_1 = layers.conv2d(net, 192, [1, 1], scope='Conv2d_0a_1x1') 286 branch_1 = layers.conv2d( 287 branch_1, 384, [3, 3], scope='Conv2d_0b_3x3') 288 with variable_scope.variable_scope('Branch_2'): 289 branch_2 = layers.conv2d(net, 48, [1, 1], scope='Conv2d_0a_1x1') 290 branch_2 = layers.conv2d( 291 branch_2, 128, [3, 3], scope='Conv2d_0b_3x3') 292 with variable_scope.variable_scope('Branch_3'): 293 branch_3 = layers_lib.max_pool2d( 294 net, [3, 3], scope='MaxPool_0a_3x3') 295 branch_3 = layers.conv2d( 296 branch_3, 128, [1, 1], scope='Conv2d_0b_1x1') 297 net = array_ops.concat([branch_0, branch_1, branch_2, branch_3], 3) 298 end_points[end_point] = net 299 if final_endpoint == end_point: 300 return net, end_points 301 raise ValueError('Unknown final endpoint %s' % final_endpoint) 302 303 304 def inception_v1(inputs, 305 num_classes=1000, 306 is_training=True, 307 dropout_keep_prob=0.8, 308 prediction_fn=layers_lib.softmax, 309 spatial_squeeze=True, 310 reuse=None, 311 scope='InceptionV1'): 312 """Defines the Inception V1 architecture. 313 314 This architecture is defined in: 315 316 Going deeper with convolutions 317 Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, 318 Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich. 319 http://arxiv.org/pdf/1409.4842v1.pdf. 320 321 The default image size used to train this network is 224x224. 322 323 Args: 324 inputs: a tensor of size [batch_size, height, width, channels]. 325 num_classes: number of predicted classes. 326 is_training: whether is training or not. 327 dropout_keep_prob: the percentage of activation values that are retained. 328 prediction_fn: a function to get predictions out of logits. 329 spatial_squeeze: if True, logits is of shape is [B, C], if false logits is 330 of shape [B, 1, 1, C], where B is batch_size and C is number of classes. 331 reuse: whether or not the network and its variables should be reused. To be 332 able to reuse 'scope' must be given. 333 scope: Optional variable_scope. 334 335 Returns: 336 logits: the pre-softmax activations, a tensor of size 337 [batch_size, num_classes] 338 end_points: a dictionary from components of the network to the corresponding 339 activation. 340 """ 341 # Final pooling and prediction 342 with variable_scope.variable_scope( 343 scope, 'InceptionV1', [inputs, num_classes], reuse=reuse) as scope: 344 with arg_scope( 345 [layers_lib.batch_norm, layers_lib.dropout], is_training=is_training): 346 net, end_points = inception_v1_base(inputs, scope=scope) 347 with variable_scope.variable_scope('Logits'): 348 net = layers_lib.avg_pool2d( 349 net, [7, 7], stride=1, scope='MaxPool_0a_7x7') 350 net = layers_lib.dropout(net, dropout_keep_prob, scope='Dropout_0b') 351 logits = layers.conv2d( 352 net, 353 num_classes, [1, 1], 354 activation_fn=None, 355 normalizer_fn=None, 356 scope='Conv2d_0c_1x1') 357 if spatial_squeeze: 358 logits = array_ops.squeeze(logits, [1, 2], name='SpatialSqueeze') 359 360 end_points['Logits'] = logits 361 end_points['Predictions'] = prediction_fn(logits, scope='Predictions') 362 return logits, end_points 363 364 365 inception_v1.default_image_size = 224 366 367 368 def inception_v1_arg_scope(weight_decay=0.00004, 369 use_batch_norm=True, 370 batch_norm_var_collection='moving_vars'): 371 """Defines the default InceptionV1 arg scope. 372 373 Note: Althougth the original paper didn't use batch_norm we found it useful. 374 375 Args: 376 weight_decay: The weight decay to use for regularizing the model. 377 use_batch_norm: "If `True`, batch_norm is applied after each convolution. 378 batch_norm_var_collection: The name of the collection for the batch norm 379 variables. 380 381 Returns: 382 An `arg_scope` to use for the inception v3 model. 383 """ 384 batch_norm_params = { 385 # Decay for the moving averages. 386 'decay': 0.9997, 387 # epsilon to prevent 0s in variance. 388 'epsilon': 0.001, 389 # collection containing update_ops. 390 'updates_collections': ops.GraphKeys.UPDATE_OPS, 391 # collection containing the moving mean and moving variance. 392 'variables_collections': { 393 'beta': None, 394 'gamma': None, 395 'moving_mean': [batch_norm_var_collection], 396 'moving_variance': [batch_norm_var_collection], 397 } 398 } 399 if use_batch_norm: 400 normalizer_fn = layers_lib.batch_norm 401 normalizer_params = batch_norm_params 402 else: 403 normalizer_fn = None 404 normalizer_params = {} 405 # Set weight_decay for weights in Conv and FC layers. 406 with arg_scope( 407 [layers.conv2d, layers_lib.fully_connected], 408 weights_regularizer=regularizers.l2_regularizer(weight_decay)): 409 with arg_scope( 410 [layers.conv2d], 411 weights_initializer=initializers.variance_scaling_initializer(), 412 activation_fn=nn_ops.relu, 413 normalizer_fn=normalizer_fn, 414 normalizer_params=normalizer_params) as sc: 415 return sc 416
