Tensorflow学习笔记（第三天）—卷积神经网络

     对CIFAR-10 数据集的分类是机器学习中一个公开的基准测试问题，其任务是对一组大小为32x32的RGB图像进行分类，这些图像涵盖了10个类别： 飞机， 汽车， 鸟， 猫， 鹿， 狗， 青蛙， 马， 船以及卡车。

一、cifar10.py代码：

# Copyright 2015 Google Inc. All Rights Reserved. 
# 
# Licensed under the Apache License, Version 2.0 (the "License"); 
# you may not use this file except in compliance with the License. 
# You may obtain a copy of the License at 
# 
#     http://www.apache.org/licenses/LICENSE-2.0 
# 
# Unless required by applicable law or agreed to in writing, software 
# distributed under the License is distributed on an "AS IS" BASIS, 
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
# See the License for the specific language governing permissions and 
# limitations under the License. 
# ============================================================================== 
"""Builds the CIFAR-10 network.
Summary of available functions:
 # Compute input images and labels for training. If you would like to run
 # evaluations, use input() instead.
 inputs, labels = distorted_inputs()
 # Compute inference on the model inputs to make a prediction.
 predictions = inference(inputs)
 # Compute the total loss of the prediction with respect to the labels.
 loss = loss(predictions, labels)
 # Create a graph to run one step of training with respect to the loss.
 train_op = train(loss, global_step)
""" 
# pylint: disable=missing-docstring 
from __future__ import absolute_import 
from __future__ import division 
from __future__ import print_function 
import gzip 
import os 
import re 
import sys 
import tarfile 
import tensorflow.python.platform 
from six.moves import urllib 
import tensorflow as tf 
import cifar10_input 
FLAGS = tf.app.flags.FLAGS 
# Basic model parameters. 
tf.app.flags.DEFINE_integer('batch_size', 128, 
                            """Number of images to process in a batch.""") 
tf.app.flags.DEFINE_string('data_dir', '/tmp/cifar10_data', 
                           """Path to the CIFAR-10 data directory.""") 
# Global constants describing the CIFAR-10 data set. 
IMAGE_SIZE = cifar10_input.IMAGE_SIZE 
NUM_CLASSES = cifar10_input.NUM_CLASSES 
NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN 
NUM_EXAMPLES_PER_EPOCH_FOR_EVAL = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_EVAL 
# Constants describing the training process. 
MOVING_AVERAGE_DECAY = 0.9999     # The decay to use for the moving average. 
NUM_EPOCHS_PER_DECAY = 350.0      # Epochs after which learning rate decays. 
LEARNING_RATE_DECAY_FACTOR = 0.1  # Learning rate decay factor. 
INITIAL_LEARNING_RATE = 0.1       # Initial learning rate. 
# If a model is trained with multiple GPU's prefix all Op names with tower_name 
# to differentiate the operations. Note that this prefix is removed from the 
# names of the summaries when visualizing a model. 
TOWER_NAME = 'tower' 
DATA_URL = 'http://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz' 
def _activation_summary(x): 
  """Helper to create summaries for activations.
  Creates a summary that provides a histogram of activations.
  Creates a summary that measure the sparsity of activations.
  Args:
    x: Tensor
  Returns:
    nothing
  """ 
  # Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training 
  # session. This helps the clarity of presentation on tensorboard. 
  tensor_name = re.sub('%s_[0-9]*/' % TOWER_NAME, '', x.op.name) 
  tf.summary.histogram(tensor_name + '/activations', x) 
  tf.summary.scalar(tensor_name + '/sparsity', tf.nn.zero_fraction(x)) 
def _variable_on_cpu(name, shape, initializer): 
  """Helper to create a Variable stored on CPU memory.
  Args:
    name: name of the variable
    shape: list of ints
    initializer: initializer for Variable
  Returns:
    Variable Tensor
  """ 
  with tf.device('/cpu:0'): 
    var = tf.get_variable(name, shape, initializer=initializer) 
  return var 
def _variable_with_weight_decay(name, shape, stddev, wd): 
  """Helper to create an initialized Variable with weight decay.
  Note that the Variable is initialized with a truncated normal distribution.
  A weight decay is added only if one is specified.
  Args:
    name: name of the variable
    shape: list of ints
    stddev: standard deviation of a truncated Gaussian
    wd: add L2Loss weight decay multiplied by this float. If None, weight
        decay is not added for this Variable.
  Returns:
    Variable Tensor
  """ 
  var = _variable_on_cpu(name, shape, 
                         tf.truncated_normal_initializer(stddev=stddev)) 
  if wd: 
    weight_decay = tf.multiply(tf.nn.l2_loss(var), wd, name='weight_loss') 
    tf.add_to_collection('losses', weight_decay) 
  return var 
def distorted_inputs(): 
  """Construct distorted input for CIFAR training using the Reader ops.
  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.
  Raises:
    ValueError: If no data_dir
  """ 
  if not FLAGS.data_dir: 
    raise ValueError('Please supply a data_dir') 
  data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin') 
  return cifar10_input.distorted_inputs(data_dir=data_dir, 
                                        batch_size=FLAGS.batch_size) 
def inputs(eval_data): 
  """Construct input for CIFAR evaluation using the Reader ops.
  Args:
    eval_data: bool, indicating if one should use the train or eval data set.
  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.
  Raises:
    ValueError: If no data_dir
  """ 
  if not FLAGS.data_dir: 
    raise ValueError('Please supply a data_dir') 
  data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin') 
  return cifar10_input.inputs(eval_data=eval_data, data_dir=data_dir, 
                              batch_size=FLAGS.batch_size) 
def inference(images): 
  """Build the CIFAR-10 model.
  Args:
    images: Images returned from distorted_inputs() or inputs().
  Returns:
    Logits.
  """ 
  # We instantiate all variables using tf.get_variable() instead of 
  # tf.Variable() in order to share variables across multiple GPU training runs. 
  # If we only ran this model on a single GPU, we could simplify this function 
  # by replacing all instances of tf.get_variable() with tf.Variable(). 
  # 
  # conv1 
  with tf.variable_scope('conv1') as scope: 
    kernel = _variable_with_weight_decay('weights', shape=[5, 5, 3, 64], 
                                         stddev=1e-4, wd=0.0) 
    conv = tf.nn.conv2d(images, kernel, [1, 1, 1, 1], padding='SAME') 
    biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.0)) 
    bias = tf.nn.bias_add(conv, biases) 
    conv1 = tf.nn.relu(bias, name=scope.name) 
    _activation_summary(conv1) 
  # pool1 
  pool1 = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], 
                         padding='SAME', name='pool1') 
  # norm1 
  norm1 = tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, 
                    name='norm1') 
  # conv2 
  with tf.variable_scope('conv2') as scope: 
    kernel = _variable_with_weight_decay('weights', shape=[5, 5, 64, 64], 
                                         stddev=1e-4, wd=0.0) 
    conv = tf.nn.conv2d(norm1, kernel, [1, 1, 1, 1], padding='SAME') 
    biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.1)) 
    bias = tf.nn.bias_add(conv, biases) 
    conv2 = tf.nn.relu(bias, name=scope.name) 
    _activation_summary(conv2) 
  # norm2 
  norm2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, 
                    name='norm2') 
  # pool2 
  pool2 = tf.nn.max_pool(norm2, ksize=[1, 3, 3, 1], 
                         strides=[1, 2, 2, 1], padding='SAME', name='pool2') 
  # local3 
  with tf.variable_scope('local3') as scope: 
    # Move everything into depth so we can perform a single matrix multiply. 
    dim = 1 
    for d in pool2.get_shape()[1:].as_list(): 
      dim *= d 
    reshape = tf.reshape(pool2, [FLAGS.batch_size, dim]) 
    weights = _variable_with_weight_decay('weights', shape=[dim, 384], 
                                          stddev=0.04, wd=0.004) 
    biases = _variable_on_cpu('biases', [384], tf.constant_initializer(0.1)) 
    local3 = tf.nn.relu(tf.matmul(reshape, weights) + biases, name=scope.name) 
    _activation_summary(local3) 
  # local4 
  with tf.variable_scope('local4') as scope: 
    weights = _variable_with_weight_decay('weights', shape=[384, 192], 
                                          stddev=0.04, wd=0.004) 
    biases = _variable_on_cpu('biases', [192], tf.constant_initializer(0.1)) 
    local4 = tf.nn.relu(tf.matmul(local3, weights) + biases, name=scope.name) 
    _activation_summary(local4) 
  # softmax, i.e. softmax(WX + b) 
  with tf.variable_scope('softmax_linear') as scope: 
    weights = _variable_with_weight_decay('weights', [192, NUM_CLASSES], 
                                          stddev=1/192.0, wd=0.0) 
    biases = _variable_on_cpu('biases', [NUM_CLASSES], 
                              tf.constant_initializer(0.0)) 
    softmax_linear = tf.add(tf.matmul(local4, weights), biases, name=scope.name) 
    _activation_summary(softmax_linear) 
  return softmax_linear 
def loss(logits, labels): 
  """Add L2Loss to all the trainable variables.
  Add summary for for "Loss" and "Loss/avg".
  Args:
    logits: Logits from inference().
    labels: Labels from distorted_inputs or inputs(). 1-D tensor
            of shape [batch_size]
  Returns:
    Loss tensor of type float.
  """ 
  # Reshape the labels into a dense Tensor of 
  # shape [batch_size, NUM_CLASSES]. 
  sparse_labels = tf.reshape(labels, [FLAGS.batch_size, 1]) 
  indices = tf.reshape(tf.range(FLAGS.batch_size), [FLAGS.batch_size, 1]) 
  concated = tf.concat([indices, sparse_labels], 1) 
  dense_labels = tf.sparse_to_dense(concated, 
                                    [FLAGS.batch_size, NUM_CLASSES], 
                                    1.0, 0.0) 
  # Calculate the average cross entropy loss across the batch. 
  cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits = logits, labels = dense_labels, name='cross_entropy_per_example') 
  cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy') 
  tf.add_to_collection('losses', cross_entropy_mean) 
  # The total loss is defined as the cross entropy loss plus all of the weight 
  # decay terms (L2 loss). 
  return tf.add_n(tf.get_collection('losses'), name='total_loss') 
def _add_loss_summaries(total_loss): 
  """Add summaries for losses in CIFAR-10 model.
  Generates moving average for all losses and associated summaries for
  visualizing the performance of the network.
  Args:
    total_loss: Total loss from loss().
  Returns:
    loss_averages_op: op for generating moving averages of losses.
  """ 
  # Compute the moving average of all individual losses and the total loss. 
  loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg') 
  losses = tf.get_collection('losses') 
  loss_averages_op = loss_averages.apply(losses + [total_loss]) 
  # Attach a scalar summary to all individual losses and the total loss; do the 
  # same for the averaged version of the losses. 
  for l in losses + [total_loss]: 
    # Name each loss as '(raw)' and name the moving average version of the loss 
    # as the original loss name. 
    tf.summary.scalar(l.op.name +' (raw)', l) 
    tf.summary.scalar(l.op.name, loss_averages.average(l)) 
  return loss_averages_op 
def train(total_loss, global_step): 
  """Train CIFAR-10 model.
  Create an optimizer and apply to all trainable variables. Add moving
  average for all trainable variables.
  Args:
    total_loss: Total loss from loss().
    global_step: Integer Variable counting the number of training steps
      processed.
  Returns:
    train_op: op for training.
  """ 
  # Variables that affect learning rate. 
  num_batches_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / FLAGS.batch_size 
  decay_steps = int(num_batches_per_epoch * NUM_EPOCHS_PER_DECAY) 
  # Decay the learning rate exponentially based on the number of steps. 
  lr = tf.train.exponential_decay(INITIAL_LEARNING_RATE, 
                                  global_step, 
                                  decay_steps, 
                                  LEARNING_RATE_DECAY_FACTOR, 
                                  staircase=True) 
  tf.summary.scalar('learning_rate', lr) 
  # Generate moving averages of all losses and associated summaries. 
  loss_averages_op = _add_loss_summaries(total_loss) 
  # Compute gradients. 
  with tf.control_dependencies([loss_averages_op]): 
    opt = tf.train.GradientDescentOptimizer(lr) 
    grads = opt.compute_gradients(total_loss) 
  # Apply gradients. 
  apply_gradient_op = opt.apply_gradients(grads, global_step=global_step) 
  # Add histograms for trainable variables. 
  for var in tf.trainable_variables(): 
    tf.summary.histogram(var.op.name, var) 
  # Add histograms for gradients. 
  for grad, var in grads: 
    if grad is not None: 
      tf.summary.histogram(var.op.name + '/gradients', grad) 
  # Track the moving averages of all trainable variables. 
  variable_averages = tf.train.ExponentialMovingAverage( 
      MOVING_AVERAGE_DECAY, global_step) 
  variables_averages_op = variable_averages.apply(tf.trainable_variables()) 
  with tf.control_dependencies([apply_gradient_op, variables_averages_op]): 
    train_op = tf.no_op(name='train') 
  return train_op 
def maybe_download_and_extract(): 
  """Download and extract the tarball from Alex's website.""" 
  dest_directory = FLAGS.data_dir 
  if not os.path.exists(dest_directory): 
    os.makedirs(dest_directory) 
  filename = DATA_URL.split('/')[-1] 
  filepath = os.path.join(dest_directory, filename) 
  if not os.path.exists(filepath): 
    def _progress(count, block_size, total_size): 
      sys.stdout.write('\r>> Downloading %s %.1f%%' % (filename, 
          float(count * block_size) / float(total_size) * 100.0)) 
      sys.stdout.flush() 
    filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, 
                                             reporthook=_progress) 
    print() 
    statinfo = os.stat(filepath) 
    print('Successfully downloaded', filename, statinfo.st_size, 'bytes.') 
    tarfile.open(filepath, 'r:gz').extractall(dest_directory) 

二、 执行并训练模型

cifar10_train.py代码：

# Copyright 2015 Google Inc. All Rights Reserved. 
# 
# Licensed under the Apache License, Version 2.0 (the "License"); 
# you may not use this file except in compliance with the License. 
# You may obtain a copy of the License at 
# 
#     http://www.apache.org/licenses/LICENSE-2.0 
# 
# Unless required by applicable law or agreed to in writing, software 
# distributed under the License is distributed on an "AS IS" BASIS, 
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
# See the License for the specific language governing permissions and 
# limitations under the License. 
# ============================================================================== 
"""A binary to train CIFAR-10 using a single GPU.
Accuracy:
cifar10_train.py achieves ~86% accuracy after 100K steps (256 epochs of
data) as judged by cifar10_eval.py.
Speed: With batch_size 128.
System        | Step Time (sec/batch)  |     Accuracy
------------------------------------------------------------------
1 Tesla K20m  | 0.35-0.60              | ~86% at 60K steps  (5 hours)
1 Tesla K40m  | 0.25-0.35              | ~86% at 100K steps (4 hours)
Usage:
Please see the tutorial and website for how to download the CIFAR-10
data set, compile the program and train the model.
http://tensorflow.org/tutorials/deep_cnn/
""" 
from __future__ import absolute_import 
from __future__ import division 
from __future__ import print_function 
from datetime import datetime 
import os.path 
import time 
import tensorflow.python.platform 
from tensorflow.python.platform import gfile 
import numpy as np 
from six.moves import xrange  # pylint: disable=redefined-builtin 
import tensorflow as tf 
import cifar10 
FLAGS = tf.app.flags.FLAGS 
tf.app.flags.DEFINE_string('train_dir', '/tmp/cifar10_train', 
                           """Directory where to write event logs """ 
                           """and checkpoint.""") 
tf.app.flags.DEFINE_integer('max_steps', 1000000, 
                            """Number of batches to run.""") 
tf.app.flags.DEFINE_boolean('log_device_placement', False, 
                            """Whether to log device placement.""") 
def train(): 
  """Train CIFAR-10 for a number of steps.""" 
  with tf.Graph().as_default(): 
    global_step = tf.Variable(0, trainable=False) 
    # Get images and labels for CIFAR-10. 
    images, labels = cifar10.distorted_inputs() 
    # Build a Graph that computes the logits predictions from the 
    # inference model. 
    logits = cifar10.inference(images) 
    # Calculate loss. 
    loss = cifar10.loss(logits, labels) 
    # Build a Graph that trains the model with one batch of examples and 
    # updates the model parameters. 
    train_op = cifar10.train(loss, global_step) 
    # Create a saver. 
    saver = tf.train.Saver(tf.global_variables()) 
    # Build the summary operation based on the TF collection of Summaries. 
    summary_op = tf.summary.merge_all() 
    # Build an initialization operation to run below. 
    init = tf.initialize_all_variables() 
    # Start running operations on the Graph. 
    sess = tf.Session(config=tf.ConfigProto( 
        log_device_placement=FLAGS.log_device_placement)) 
    sess.run(init) 
    # Start the queue runners. 
    tf.train.start_queue_runners(sess=sess) 
    summary_writer = tf.summary.FileWriter(FLAGS.train_dir, 
                                            graph_def=sess.graph_def) 
    for step in xrange(FLAGS.max_steps): 
      start_time = time.time() 
      _, loss_value = sess.run([train_op, loss]) 
      duration = time.time() - start_time 
      assert not np.isnan(loss_value), 'Model diverged with loss = NaN' 
      if step % 10 == 0: 
        num_examples_per_step = FLAGS.batch_size 
        examples_per_sec = num_examples_per_step / duration 
        sec_per_batch = float(duration) 
        format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f ' 
                      'sec/batch)') 
        print (format_str % (datetime.now(), step, loss_value, 
                             examples_per_sec, sec_per_batch)) 
      if step % 100 == 0: 
        summary_str = sess.run(summary_op) 
        summary_writer.add_summary(summary_str, step) 
      # Save the model checkpoint periodically. 
      if step % 1000 == 0 or (step + 1) == FLAGS.max_steps: 
        checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt') 
        saver.save(sess, checkpoint_path, global_step=step) 
def main(argv=None):  # pylint: disable=unused-argument 
  cifar10.maybe_download_and_extract() 
  if gfile.Exists(FLAGS.train_dir): 
    gfile.DeleteRecursively(FLAGS.train_dir) 
  gfile.MakeDirs(FLAGS.train_dir) 
  train() 
if __name__ == '__main__': 
  tf.app.run() 

三 、使用tensorboard

1.键入命令行，启动TensorBoard

输入tensorboard --logdir=E:\tmp\cifar10_train

这个后面的地址是这个函数的地址

tf.app.flags.DEFINE_string('train_dir', '/tmp/cifar10_train', 
                           """Directory where to write event logs """ 
                           """and checkpoint.""") 

之后得到这个地址在浏览器中打开。最好是使用谷歌浏览器

四、评估模型

cifar10_eval.py代码：# Copyright 2015 Google Inc. All Rights Reserved. 
# 
# Licensed under the Apache License, Version 2.0 (the "License"); 
# you may not use this file except in compliance with the License. 
# You may obtain a copy of the License at 
# 
#     http://www.apache.org/licenses/LICENSE-2.0 
# 
# Unless required by applicable law or agreed to in writing, software 
# distributed under the License is distributed on an "AS IS" BASIS, 
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
# See the License for the specific language governing permissions and 
# limitations under the License. 
# ============================================================================== 
"""Evaluation for CIFAR-10.
Accuracy:
cifar10_train.py achieves 83.0% accuracy after 100K steps (256 epochs
of data) as judged by cifar10_eval.py.
Speed:
On a single Tesla K40, cifar10_train.py processes a single batch of 128 images
in 0.25-0.35 sec (i.e. 350 - 600 images /sec). The model reaches ~86%
accuracy after 100K steps in 8 hours of training time.
Usage:
Please see the tutorial and website for how to download the CIFAR-10
data set, compile the program and train the model.
http://tensorflow.org/tutorials/deep_cnn/
""" 
from __future__ import absolute_import 
from __future__ import division 
from __future__ import print_function 
from datetime import datetime 
import math 
import time 
import tensorflow.python.platform 
from tensorflow.python.platform import gfile 
import numpy as np 
import tensorflow as tf 
import cifar10 
FLAGS = tf.app.flags.FLAGS 
tf.app.flags.DEFINE_string('eval_dir', '/tmp/cifar10_eval', 
                           """Directory where to write event logs.""") 
tf.app.flags.DEFINE_string('eval_data', 'test', 
                           """Either 'test' or 'train_eval'.""") 
tf.app.flags.DEFINE_string('checkpoint_dir', '/tmp/cifar10_train', 
                           """Directory where to read model checkpoints.""") 
tf.app.flags.DEFINE_integer('eval_interval_secs', 60 * 5, 
                            """How often to run the eval.""") 
tf.app.flags.DEFINE_integer('num_examples', 10000, 
                            """Number of examples to run.""") 
tf.app.flags.DEFINE_boolean('run_once', False, 
                         """Whether to run eval only once.""") 
def eval_once(saver, summary_writer, top_k_op, summary_op): 
  """Run Eval once.
  Args:
    saver: Saver.
    summary_writer: Summary writer.
    top_k_op: Top K op.
    summary_op: Summary op.
  """ 
  with tf.Session() as sess: 
    ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir) 
    if ckpt and ckpt.model_checkpoint_path: 
      # Restores from checkpoint 
      saver.restore(sess, ckpt.model_checkpoint_path) 
      # Assuming model_checkpoint_path looks something like: 
      #   /my-favorite-path/cifar10_train/model.ckpt-0, 
      # extract global_step from it. 
      global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1] 
    else: 
      print('No checkpoint file found') 
      return 
    # Start the queue runners. 
    coord = tf.train.Coordinator() 
    try: 
      threads = [] 
      for qr in tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS): 
        threads.extend(qr.create_threads(sess, coord=coord, daemon=True, 
                                         start=True)) 
      num_iter = int(math.ceil(FLAGS.num_examples / FLAGS.batch_size)) 
      true_count = 0  # Counts the number of correct predictions. 
      total_sample_count = num_iter * FLAGS.batch_size 
      step = 0 
      while step < num_iter and not coord.should_stop(): 
        predictions = sess.run([top_k_op]) 
        true_count += np.sum(predictions) 
        step += 1 
      # Compute precision @ 1. 
      precision = true_count / total_sample_count 
      print('%s: precision @ 1 = %.3f' % (datetime.now(), precision)) 
      summary = tf.Summary() 
      summary.ParseFromString(sess.run(summary_op)) 
      summary.value.add(tag='Precision @ 1', simple_value=precision) 
      summary_writer.add_summary(summary, global_step) 
    except Exception as e:  # pylint: disable=broad-except 
      coord.request_stop(e) 
    coord.request_stop() 
    coord.join(threads, stop_grace_period_secs=10) 
def evaluate(): 
  """Eval CIFAR-10 for a number of steps.""" 
  with tf.Graph().as_default(): 
    # Get images and labels for CIFAR-10. 
    eval_data = FLAGS.eval_data == 'test' 
    images, labels = cifar10.inputs(eval_data=eval_data) 
    # Build a Graph that computes the logits predictions from the 
    # inference model. 
    logits = cifar10.inference(images) 
    # Calculate predictions. 
    top_k_op = tf.nn.in_top_k(logits, labels, 1) 
    # Restore the moving average version of the learned variables for eval. 
    variable_averages = tf.train.ExponentialMovingAverage( 
        cifar10.MOVING_AVERAGE_DECAY) 
    variables_to_restore = variable_averages.variables_to_restore() 
    saver = tf.train.Saver(variables_to_restore) 
    # Build the summary operation based on the TF collection of Summaries. 
    summary_op = tf.summary.merge_all() 
    graph_def = tf.get_default_graph().as_graph_def() 
    summary_writer = tf.summary.FileWriter(FLAGS.eval_dir, 
                                            graph_def=graph_def) 
    while True: 
      eval_once(saver, summary_writer, top_k_op, summary_op) 
      if FLAGS.run_once: 
        break 
      time.sleep(FLAGS.eval_interval_secs) 
def main(argv=None):  # pylint: disable=unused-argument 
  cifar10.maybe_download_and_extract() 
  if gfile.Exists(FLAGS.eval_dir): 
    gfile.DeleteRecursively(FLAGS.eval_dir) 
  gfile.MakeDirs(FLAGS.eval_dir) 
  evaluate() 
if __name__ == '__main__': 
  tf.app.run() 

要注意这个注意事项，最好挂起训练程序先

注意：不要在同一块GPU上同时运行训练程序和评估程序，因为可能会导致内存耗尽。尽可能的在其它单独的GPU上运行评估程序，或者在同一块GPU上运行评估程序时先挂起训练程序。