1 # Copyright 2015 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 # pylint: disable=invalid-name 16 # pylint: disable=unused-import 17 """ResNet50 model for Keras. 18 19 # Reference: 20 21 - [Deep Residual Learning for Image 22 Recognition](https://arxiv.org/abs/1512.03385) 23 24 Adapted from code contributed by BigMoyan. 25 """ 26 from __future__ import absolute_import 27 from __future__ import division 28 from __future__ import print_function 29 30 import os 31 32 from tensorflow.python.keras._impl.keras import backend as K 33 from tensorflow.python.keras._impl.keras import layers 34 from tensorflow.python.keras._impl.keras.applications.imagenet_utils import _obtain_input_shape 35 from tensorflow.python.keras._impl.keras.applications.imagenet_utils import decode_predictions 36 from tensorflow.python.keras._impl.keras.applications.imagenet_utils import preprocess_input 37 from tensorflow.python.keras._impl.keras.engine.topology import get_source_inputs 38 from tensorflow.python.keras._impl.keras.layers import Activation 39 from tensorflow.python.keras._impl.keras.layers import AveragePooling2D 40 from tensorflow.python.keras._impl.keras.layers import BatchNormalization 41 from tensorflow.python.keras._impl.keras.layers import Conv2D 42 from tensorflow.python.keras._impl.keras.layers import Dense 43 from tensorflow.python.keras._impl.keras.layers import Flatten 44 from tensorflow.python.keras._impl.keras.layers import GlobalAveragePooling2D 45 from tensorflow.python.keras._impl.keras.layers import GlobalMaxPooling2D 46 from tensorflow.python.keras._impl.keras.layers import Input 47 from tensorflow.python.keras._impl.keras.layers import MaxPooling2D 48 from tensorflow.python.keras._impl.keras.models import Model 49 from tensorflow.python.keras._impl.keras.utils import layer_utils 50 from tensorflow.python.keras._impl.keras.utils.data_utils import get_file 51 from tensorflow.python.platform import tf_logging as logging 52 from tensorflow.python.util.tf_export import tf_export 53 54 55 WEIGHTS_PATH = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels.h5' 56 WEIGHTS_PATH_NO_TOP = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5' 57 58 59 def identity_block(input_tensor, kernel_size, filters, stage, block): 60 """The identity block is the block that has no conv layer at shortcut. 61 62 Arguments: 63 input_tensor: input tensor 64 kernel_size: default 3, the kernel size of middle conv layer at main path 65 filters: list of integers, the filters of 3 conv layer at main path 66 stage: integer, current stage label, used for generating layer names 67 block: 'a','b'..., current block label, used for generating layer names 68 69 Returns: 70 Output tensor for the block. 71 """ 72 filters1, filters2, filters3 = filters 73 if K.image_data_format() == 'channels_last': 74 bn_axis = 3 75 else: 76 bn_axis = 1 77 conv_name_base = 'res' + str(stage) + block + '_branch' 78 bn_name_base = 'bn' + str(stage) + block + '_branch' 79 80 x = Conv2D(filters1, (1, 1), name=conv_name_base + '2a')(input_tensor) 81 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x) 82 x = Activation('relu')(x) 83 84 x = Conv2D( 85 filters2, kernel_size, padding='same', name=conv_name_base + '2b')( 86 x) 87 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x) 88 x = Activation('relu')(x) 89 90 x = Conv2D(filters3, (1, 1), name=conv_name_base + '2c')(x) 91 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x) 92 93 x = layers.add([x, input_tensor]) 94 x = Activation('relu')(x) 95 return x 96 97 98 def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 99 2)): 100 """A block that has a conv layer at shortcut. 101 102 Arguments: 103 input_tensor: input tensor 104 kernel_size: default 3, the kernel size of middle conv layer at main path 105 filters: list of integers, the filters of 3 conv layer at main path 106 stage: integer, current stage label, used for generating layer names 107 block: 'a','b'..., current block label, used for generating layer names 108 strides: Strides for the first conv layer in the block. 109 110 Returns: 111 Output tensor for the block. 112 113 Note that from stage 3, 114 the first conv layer at main path is with strides=(2, 2) 115 And the shortcut should have strides=(2, 2) as well 116 """ 117 filters1, filters2, filters3 = filters 118 if K.image_data_format() == 'channels_last': 119 bn_axis = 3 120 else: 121 bn_axis = 1 122 conv_name_base = 'res' + str(stage) + block + '_branch' 123 bn_name_base = 'bn' + str(stage) + block + '_branch' 124 125 x = Conv2D( 126 filters1, (1, 1), strides=strides, name=conv_name_base + '2a')( 127 input_tensor) 128 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x) 129 x = Activation('relu')(x) 130 131 x = Conv2D( 132 filters2, kernel_size, padding='same', name=conv_name_base + '2b')( 133 x) 134 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x) 135 x = Activation('relu')(x) 136 137 x = Conv2D(filters3, (1, 1), name=conv_name_base + '2c')(x) 138 x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x) 139 140 shortcut = Conv2D( 141 filters3, (1, 1), strides=strides, name=conv_name_base + '1')( 142 input_tensor) 143 shortcut = BatchNormalization(axis=bn_axis, name=bn_name_base + '1')(shortcut) 144 145 x = layers.add([x, shortcut]) 146 x = Activation('relu')(x) 147 return x 148 149 150 @tf_export('keras.applications.ResNet50', 151 'keras.applications.resnet50.ResNet50') 152 def ResNet50(include_top=True, 153 weights='imagenet', 154 input_tensor=None, 155 input_shape=None, 156 pooling=None, 157 classes=1000): 158 """Instantiates the ResNet50 architecture. 159 160 Optionally loads weights pre-trained 161 on ImageNet. Note that when using TensorFlow, 162 for best performance you should set 163 `image_data_format='channels_last'` in your Keras config 164 at ~/.keras/keras.json. 165 166 The model and the weights are compatible with both 167 TensorFlow and Theano. The data format 168 convention used by the model is the one 169 specified in your Keras config file. 170 171 Arguments: 172 include_top: whether to include the fully-connected 173 layer at the top of the network. 174 weights: one of `None` (random initialization), 175 'imagenet' (pre-training on ImageNet), 176 or the path to the weights file to be loaded. 177 input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) 178 to use as image input for the model. 179 input_shape: optional shape tuple, only to be specified 180 if `include_top` is False (otherwise the input shape 181 has to be `(224, 224, 3)` (with `channels_last` data format) 182 or `(3, 224, 224)` (with `channels_first` data format). 183 It should have exactly 3 inputs channels, 184 and width and height should be no smaller than 197. 185 E.g. `(200, 200, 3)` would be one valid value. 186 pooling: Optional pooling mode for feature extraction 187 when `include_top` is `False`. 188 - `None` means that the output of the model will be 189 the 4D tensor output of the 190 last convolutional layer. 191 - `avg` means that global average pooling 192 will be applied to the output of the 193 last convolutional layer, and thus 194 the output of the model will be a 2D tensor. 195 - `max` means that global max pooling will 196 be applied. 197 classes: optional number of classes to classify images 198 into, only to be specified if `include_top` is True, and 199 if no `weights` argument is specified. 200 201 Returns: 202 A Keras model instance. 203 204 Raises: 205 ValueError: in case of invalid argument for `weights`, 206 or invalid input shape. 207 """ 208 if not (weights in {'imagenet', None} or os.path.exists(weights)): 209 raise ValueError('The `weights` argument should be either ' 210 '`None` (random initialization), `imagenet` ' 211 '(pre-training on ImageNet), ' 212 'or the path to the weights file to be loaded.') 213 214 if weights == 'imagenet' and include_top and classes != 1000: 215 raise ValueError('If using `weights` as imagenet with `include_top`' 216 ' as true, `classes` should be 1000') 217 218 # Determine proper input shape 219 input_shape = _obtain_input_shape( 220 input_shape, 221 default_size=224, 222 min_size=197, 223 data_format=K.image_data_format(), 224 require_flatten=include_top, 225 weights=weights) 226 227 if input_tensor is None: 228 img_input = Input(shape=input_shape) 229 else: 230 if not K.is_keras_tensor(input_tensor): 231 img_input = Input(tensor=input_tensor, shape=input_shape) 232 else: 233 img_input = input_tensor 234 if K.image_data_format() == 'channels_last': 235 bn_axis = 3 236 else: 237 bn_axis = 1 238 239 x = Conv2D( 240 64, (7, 7), strides=(2, 2), padding='same', name='conv1')( 241 img_input) 242 x = BatchNormalization(axis=bn_axis, name='bn_conv1')(x) 243 x = Activation('relu')(x) 244 x = MaxPooling2D((3, 3), strides=(2, 2))(x) 245 246 x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) 247 x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') 248 x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') 249 250 x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') 251 x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') 252 x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') 253 x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') 254 255 x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') 256 x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b') 257 x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c') 258 x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d') 259 x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e') 260 x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f') 261 262 x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') 263 x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') 264 x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') 265 266 x = AveragePooling2D((7, 7), name='avg_pool')(x) 267 268 if include_top: 269 x = Flatten()(x) 270 x = Dense(classes, activation='softmax', name='fc1000')(x) 271 else: 272 if pooling == 'avg': 273 x = GlobalAveragePooling2D()(x) 274 elif pooling == 'max': 275 x = GlobalMaxPooling2D()(x) 276 277 # Ensure that the model takes into account 278 # any potential predecessors of `input_tensor`. 279 if input_tensor is not None: 280 inputs = get_source_inputs(input_tensor) 281 else: 282 inputs = img_input 283 # Create model. 284 model = Model(inputs, x, name='resnet50') 285 286 # load weights 287 if weights == 'imagenet': 288 if include_top: 289 weights_path = get_file( 290 'resnet50_weights_tf_dim_ordering_tf_kernels.h5', 291 WEIGHTS_PATH, 292 cache_subdir='models', 293 md5_hash='a7b3fe01876f51b976af0dea6bc144eb') 294 else: 295 weights_path = get_file( 296 'resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5', 297 WEIGHTS_PATH_NO_TOP, 298 cache_subdir='models', 299 md5_hash='a268eb855778b3df3c7506639542a6af') 300 model.load_weights(weights_path) 301 elif weights is not None: 302 model.load_weights(weights) 303 304 return model 305
