1 # Copyright 2017 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 """Builds the CIFAR-10 network with additional variables to support pruning. 16 17 Summary of available functions: 18 19 # Compute input images and labels for training. If you would like to run 20 # evaluations, use inputs() instead. 21 inputs, labels = distorted_inputs() 22 23 # Compute inference on the model inputs to make a prediction. 24 predictions = inference(inputs) 25 26 # Compute the total loss of the prediction with respect to the labels. 27 loss = loss(predictions, labels) 28 29 # Create a graph to run one step of training with respect to the loss. 30 train_op = train(loss, global_step) 31 """ 32 # pylint: disable=missing-docstring 33 from __future__ import absolute_import 34 from __future__ import division 35 from __future__ import print_function 36 37 import os 38 import re 39 import sys 40 import tarfile 41 42 from six.moves import urllib 43 import tensorflow as tf 44 45 from tensorflow.contrib.model_pruning.examples.cifar10 import cifar10_input 46 from tensorflow.contrib.model_pruning.python import pruning 47 48 # Global constants describing the CIFAR-10 data set. 49 IMAGE_SIZE = cifar10_input.IMAGE_SIZE 50 NUM_CLASSES = cifar10_input.NUM_CLASSES 51 NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN # pylint: disable=line-too-long 52 NUM_EXAMPLES_PER_EPOCH_FOR_EVAL = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_EVAL 53 BATCH_SIZE = 128 54 DATA_DIR = '/tmp/cifar10_data' 55 56 # Constants describing the training process. 57 MOVING_AVERAGE_DECAY = 0.9999 # The decay to use for the moving average. 58 NUM_EPOCHS_PER_DECAY = 350.0 # Epochs after which learning rate decays. 59 LEARNING_RATE_DECAY_FACTOR = 0.1 # Learning rate decay factor. 60 INITIAL_LEARNING_RATE = 0.1 # Initial learning rate. 61 62 # If a model is trained with multiple GPUs, prefix all Op names with tower_name 63 # to differentiate the operations. Note that this prefix is removed from the 64 # names of the summaries when visualizing a model. 65 TOWER_NAME = 'tower' 66 67 DATA_URL = 'http://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz' 68 69 70 def _activation_summary(x): 71 """Helper to create summaries for activations. 72 73 Creates a summary that provides a histogram of activations. 74 Creates a summary that measures the sparsity of activations. 75 76 Args: 77 x: Tensor 78 Returns: 79 nothing 80 """ 81 # Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training 82 # session. This helps the clarity of presentation on tensorboard. 83 tensor_name = re.sub('%s_[0-9]*/' % TOWER_NAME, '', x.op.name) 84 tf.summary.histogram(tensor_name + '/activations', x) 85 tf.summary.scalar(tensor_name + '/sparsity', tf.nn.zero_fraction(x)) 86 87 88 def _variable_on_cpu(name, shape, initializer): 89 """Helper to create a Variable stored on CPU memory. 90 91 Args: 92 name: name of the variable 93 shape: list of ints 94 initializer: initializer for Variable 95 96 Returns: 97 Variable Tensor 98 """ 99 with tf.device('/cpu:0'): 100 dtype = tf.float32 101 var = tf.get_variable(name, shape, initializer=initializer, dtype=dtype) 102 return var 103 104 105 def _variable_with_weight_decay(name, shape, stddev, wd): 106 """Helper to create an initialized Variable with weight decay. 107 108 Note that the Variable is initialized with a truncated normal distribution. 109 A weight decay is added only if one is specified. 110 111 Args: 112 name: name of the variable 113 shape: list of ints 114 stddev: standard deviation of a truncated Gaussian 115 wd: add L2Loss weight decay multiplied by this float. If None, weight 116 decay is not added for this Variable. 117 118 Returns: 119 Variable Tensor 120 """ 121 dtype = tf.float32 122 var = _variable_on_cpu(name, shape, 123 tf.truncated_normal_initializer( 124 stddev=stddev, dtype=dtype)) 125 if wd is not None: 126 weight_decay = tf.multiply(tf.nn.l2_loss(var), wd, name='weight_loss') 127 tf.add_to_collection('losses', weight_decay) 128 return var 129 130 131 def distorted_inputs(): 132 """Construct distorted input for CIFAR training using the Reader ops. 133 134 Returns: 135 images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size. 136 labels: Labels. 1D tensor of [batch_size] size. 137 138 Raises: 139 ValueError: If no data_dir 140 """ 141 if not DATA_DIR: 142 raise ValueError('Please supply a data_dir') 143 data_dir = os.path.join(DATA_DIR, 'cifar-10-batches-bin') 144 images, labels = cifar10_input.distorted_inputs( 145 data_dir=data_dir, batch_size=BATCH_SIZE) 146 return images, labels 147 148 149 def inputs(eval_data): 150 """Construct input for CIFAR evaluation using the Reader ops. 151 152 Args: 153 eval_data: bool, indicating if one should use the train or eval data set. 154 155 Returns: 156 images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size. 157 labels: Labels. 1D tensor of [batch_size] size. 158 159 Raises: 160 ValueError: If no data_dir 161 """ 162 if not DATA_DIR: 163 raise ValueError('Please supply a data_dir') 164 data_dir = os.path.join(DATA_DIR, 'cifar-10-batches-bin') 165 images, labels = cifar10_input.inputs( 166 eval_data=eval_data, data_dir=data_dir, batch_size=BATCH_SIZE) 167 return images, labels 168 169 170 def inference(images): 171 """Build the CIFAR-10 model. 172 173 Args: 174 images: Images returned from distorted_inputs() or inputs(). 175 176 Returns: 177 Logits. 178 """ 179 # We instantiate all variables using tf.get_variable() instead of 180 # tf.Variable() in order to share variables across multiple GPU training runs. 181 # If we only ran this model on a single GPU, we could simplify this function 182 # by replacing all instances of tf.get_variable() with tf.Variable(). 183 # 184 # While instantiating conv and local layers, we add mask and threshold 185 # variables to the layer by calling the pruning.apply_mask() function. 186 # Note that the masks are applied only to the weight tensors 187 # conv1 188 with tf.variable_scope('conv1') as scope: 189 kernel = _variable_with_weight_decay( 190 'weights', shape=[5, 5, 3, 64], stddev=5e-2, wd=0.0) 191 192 conv = tf.nn.conv2d( 193 images, pruning.apply_mask(kernel, scope), [1, 1, 1, 1], padding='SAME') 194 biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.0)) 195 pre_activation = tf.nn.bias_add(conv, biases) 196 conv1 = tf.nn.relu(pre_activation, name=scope.name) 197 _activation_summary(conv1) 198 199 # pool1 200 pool1 = tf.nn.max_pool( 201 conv1, 202 ksize=[1, 3, 3, 1], 203 strides=[1, 2, 2, 1], 204 padding='SAME', 205 name='pool1') 206 # norm1 207 norm1 = tf.nn.lrn( 208 pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm1') 209 210 # conv2 211 with tf.variable_scope('conv2') as scope: 212 kernel = _variable_with_weight_decay( 213 'weights', shape=[5, 5, 64, 64], stddev=5e-2, wd=0.0) 214 conv = tf.nn.conv2d( 215 norm1, pruning.apply_mask(kernel, scope), [1, 1, 1, 1], padding='SAME') 216 biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.1)) 217 pre_activation = tf.nn.bias_add(conv, biases) 218 conv2 = tf.nn.relu(pre_activation, name=scope.name) 219 _activation_summary(conv2) 220 221 # norm2 222 norm2 = tf.nn.lrn( 223 conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm2') 224 # pool2 225 pool2 = tf.nn.max_pool( 226 norm2, 227 ksize=[1, 3, 3, 1], 228 strides=[1, 2, 2, 1], 229 padding='SAME', 230 name='pool2') 231 232 # local3 233 with tf.variable_scope('local3') as scope: 234 # Move everything into depth so we can perform a single matrix multiply. 235 reshape = tf.reshape(pool2, [BATCH_SIZE, -1]) 236 dim = reshape.get_shape()[1].value 237 weights = _variable_with_weight_decay( 238 'weights', shape=[dim, 384], stddev=0.04, wd=0.004) 239 biases = _variable_on_cpu('biases', [384], tf.constant_initializer(0.1)) 240 local3 = tf.nn.relu( 241 tf.matmul(reshape, pruning.apply_mask(weights, scope)) + biases, 242 name=scope.name) 243 _activation_summary(local3) 244 245 # local4 246 with tf.variable_scope('local4') as scope: 247 weights = _variable_with_weight_decay( 248 'weights', shape=[384, 192], stddev=0.04, wd=0.004) 249 biases = _variable_on_cpu('biases', [192], tf.constant_initializer(0.1)) 250 local4 = tf.nn.relu( 251 tf.matmul(local3, pruning.apply_mask(weights, scope)) + biases, 252 name=scope.name) 253 _activation_summary(local4) 254 255 # linear layer(WX + b), 256 # We don't apply softmax here because 257 # tf.nn.sparse_softmax_cross_entropy_with_logits accepts the unscaled logits 258 # and performs the softmax internally for efficiency. 259 with tf.variable_scope('softmax_linear') as scope: 260 weights = _variable_with_weight_decay( 261 'weights', [192, NUM_CLASSES], stddev=1 / 192.0, wd=0.0) 262 biases = _variable_on_cpu('biases', [NUM_CLASSES], 263 tf.constant_initializer(0.0)) 264 softmax_linear = tf.add( 265 tf.matmul(local4, pruning.apply_mask(weights, scope)), 266 biases, 267 name=scope.name) 268 _activation_summary(softmax_linear) 269 270 return softmax_linear 271 272 273 def loss(logits, labels): 274 """Add L2Loss to all the trainable variables. 275 276 Add summary for "Loss" and "Loss/avg". 277 Args: 278 logits: Logits from inference(). 279 labels: Labels from distorted_inputs or inputs(). 1-D tensor 280 of shape [batch_size] 281 282 Returns: 283 Loss tensor of type float. 284 """ 285 # Calculate the average cross entropy loss across the batch. 286 labels = tf.cast(labels, tf.int64) 287 cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits( 288 labels=labels, logits=logits, name='cross_entropy_per_example') 289 cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy') 290 tf.add_to_collection('losses', cross_entropy_mean) 291 292 # The total loss is defined as the cross entropy loss plus all of the weight 293 # decay terms (L2 loss). 294 return tf.add_n(tf.get_collection('losses'), name='total_loss') 295 296 297 def _add_loss_summaries(total_loss): 298 """Add summaries for losses in CIFAR-10 model. 299 300 Generates moving average for all losses and associated summaries for 301 visualizing the performance of the network. 302 303 Args: 304 total_loss: Total loss from loss(). 305 Returns: 306 loss_averages_op: op for generating moving averages of losses. 307 """ 308 # Compute the moving average of all individual losses and the total loss. 309 loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg') 310 losses = tf.get_collection('losses') 311 loss_averages_op = loss_averages.apply(losses + [total_loss]) 312 313 # Attach a scalar summary to all individual losses and the total loss; do the 314 # same for the averaged version of the losses. 315 for l in losses + [total_loss]: 316 # Name each loss as '(raw)' and name the moving average version of the loss 317 # as the original loss name. 318 tf.summary.scalar(l.op.name + ' (raw)', l) 319 tf.summary.scalar(l.op.name, loss_averages.average(l)) 320 321 return loss_averages_op 322 323 324 def train(total_loss, global_step): 325 """Train CIFAR-10 model. 326 327 Create an optimizer and apply to all trainable variables. Add moving 328 average for all trainable variables. 329 330 Args: 331 total_loss: Total loss from loss(). 332 global_step: Integer Variable counting the number of training steps 333 processed. 334 Returns: 335 train_op: op for training. 336 """ 337 # Variables that affect learning rate. 338 num_batches_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / BATCH_SIZE 339 decay_steps = int(num_batches_per_epoch * NUM_EPOCHS_PER_DECAY) 340 341 # Decay the learning rate exponentially based on the number of steps. 342 lr = tf.train.exponential_decay( 343 INITIAL_LEARNING_RATE, 344 global_step, 345 decay_steps, 346 LEARNING_RATE_DECAY_FACTOR, 347 staircase=True) 348 tf.summary.scalar('learning_rate', lr) 349 350 # Generate moving averages of all losses and associated summaries. 351 loss_averages_op = _add_loss_summaries(total_loss) 352 353 # Compute gradients. 354 with tf.control_dependencies([loss_averages_op]): 355 opt = tf.train.GradientDescentOptimizer(lr) 356 grads = opt.compute_gradients(total_loss) 357 358 # Apply gradients. 359 apply_gradient_op = opt.apply_gradients(grads, global_step=global_step) 360 361 # Add histograms for trainable variables. 362 for var in tf.trainable_variables(): 363 tf.summary.histogram(var.op.name, var) 364 365 # Add histograms for gradients. 366 for grad, var in grads: 367 if grad is not None: 368 tf.summary.histogram(var.op.name + '/gradients', grad) 369 370 # Track the moving averages of all trainable variables. 371 variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, 372 global_step) 373 variables_averages_op = variable_averages.apply(tf.trainable_variables()) 374 375 with tf.control_dependencies([apply_gradient_op, variables_averages_op]): 376 train_op = tf.no_op(name='train') 377 378 return train_op 379 380 381 def maybe_download_and_extract(): 382 """Download and extract the tarball from Alex's website.""" 383 dest_directory = DATA_DIR 384 if not os.path.exists(dest_directory): 385 os.makedirs(dest_directory) 386 filename = DATA_URL.split('/')[-1] 387 filepath = os.path.join(dest_directory, filename) 388 if not os.path.exists(filepath): 389 390 def _progress(count, block_size, total_size): 391 sys.stdout.write('\r>> Downloading %s %.1f%%' % 392 (filename, 393 float(count * block_size) / float(total_size) * 100.0)) 394 sys.stdout.flush() 395 396 filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, _progress) 397 print() 398 statinfo = os.stat(filepath) 399 print('Successfully downloaded', filename, statinfo.st_size, 'bytes.') 400 401 tarfile.open(filepath, 'r:gz').extractall(dest_directory) 402
