CIFAR 10 TensorFlow 模型架构(一）

一.CIFAR 10数据集

二、CIFAR 10 TensorFlow模型架构

1.CIFAR-10_input.py ：读取 CIFAR-10 二进制文件格式

# -*- coding: utf-8 -*-
# Copyright 2015 The TensorFlow Authors. 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.
# ==============================================================================

"""Routine for decoding the CIFAR-10 binary file format."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os

from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

# Process images of this size. Note that this differs from the original CIFAR
# image size of 32 x 32. If one alters this number, then the entire model
# architecture will change and any model would need to be retrained.
IMAGE_SIZE = 24

# 关于 CIFAR-10 数据集的全局常量.
NUM_CLASSES = 10
NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN = 50000
NUM_EXAMPLES_PER_EPOCH_FOR_EVAL = 10000

def read_cifar10(filename_queue):
  """读取并解析从CIFAR10数据文件中的例子
  Recommendation: if you want N-way read parallelism, call this function
  N times.  This will give you N independent Readers reading different
  files & positions within those files, which will give better mixing of
  examples.
  @param  filename_queue  要读取的文件名队列
  @return 某个对象，具有以下字段:
        height: 结果中的行数 (32)
        width:  结果中的列数 (32)
        depth:  结果中颜色通道数(3)
        key:    一个描述当前抽样数据的文件名和记录数的标量字符串
        label:  一个 int32类型的标签，取值范围 0..9.
        uint8image: 一个[height, width, depth]维度的图像数据
  """

  class CIFAR10Record(object):
    #pass 不做任何事情，一般用做占位语句  
    pass
  result = CIFAR10Record()

  # Dimensions of the images in the CIFAR-10 dataset.
  # See http://www.cs.toronto.edu/~kriz/cifar.html for a description of the
  # input format.
  label_bytes = 1  # 2 for CIFAR-100
  result.height = 32 #图片的高度
  result.width = 32  #图片的宽度
  result.depth = 3   #图片的通道数
  image_bytes = result.height * result.width * result.depth
  # Every record consists of a label followed by the image, with a
  # fixed number of bytes for each.
  record_bytes = label_bytes + image_bytes

  # Read a record, getting filenames from the filename_queue.  No
  # header or footer in the CIFAR-10 format, so we leave header_bytes
  # and footer_bytes at their default of 0.
  # 读取固定长度字节数信息(针对bin文件使用FixedLengthRecordReader读取比较合适)
  reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
  result.key, value = reader.read(filename_queue)

  # 将字符串转为张量数据decode_raw
  record_bytes = tf.decode_raw(value, tf.uint8)

  # The first bytes represent the label, which we convert from uint8->int32.
  # tf.stride_slice(data, begin, end)：从张量中提取数据段，并用cast进行数据类型转换
  # 第一个字节表示标签，从uint8转换成int32
  result.label = tf.cast(
      tf.slice(record_bytes, [0],   [label_bytes]), tf.int32)

  # The remaining bytes after the label represent the image, which we reshape
  # from [depth * height * width] to [depth, height, width].
  # 除标签外的剩下的字节式图片，从[depth*height*widht]改变形状为[depth,height,width]
  depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),
                           [result.depth, result.height, result.width])
  # Convert from [depth, height, width] to [height, width, depth].
  # 从[depth,height,width]转变为[height,width,depth]
  result.uint8image = tf.transpose(depth_major, [1, 2, 0])

  return result

def _generate_image_and_label_batch(image, label, min_queue_examples,
                                    batch_size, shuffle):
  """构建一批已排队的图片和标签
  Args:
    image: 3-D张量 [height, width, 3] 类型为 type.float32.
    label: 1-D 张量，类型为 type.int32
    min_queue_examples: int32, 在提供成批例子的队列中，保留的最小样本数。
    batch_size: 每批的图片个数Number of images per batch.
    shuffle: boolean 指示是否使用混排队列。
  Returns:
    images: Images. 4D tensor of [batch_size, height, width, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.
  """
  # Create a queue that shuffles the examples, and then
  # read 'batch_size' images + labels from the example queue.
  num_preprocess_threads = 16
  if shuffle:
    #简单来说就是读取一个文件并且加载一个张量中的batch_size行  
    images, label_batch = tf.train.shuffle_batch(
        [image, label],
        batch_size=batch_size,
        num_threads=num_preprocess_threads,
        capacity=min_queue_examples + 3 * batch_size,
        min_after_dequeue=min_queue_examples)
  else:
    images, label_batch = tf.train.batch(
        [image, label],
        batch_size=batch_size,
        num_threads=num_preprocess_threads,
        capacity=min_queue_examples + 3 * batch_size)

  # 在visualizer中显示正在训练的图片
  tf.summary.image('images',images)
  #tf.image('images',images)
  #tf.image_summary('images', images)

  return images, tf.reshape(label_batch, [batch_size])

def distorted_inputs(data_dir, batch_size):
  """CIFAR训练使用Reader操作,构建变形的输入
  Args:
    data_dir: CIFAR-10 数据的目录.
    batch_size: 每批次的图片数量.
  Returns:
    images: Images. 4D 张量 of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D 张量 of [batch_size] size.
  """
  filenames = [os.path.join(data_dir, 'data_batch_%d.bin' % i)
               for i in xrange(1, 6)]
  for f in filenames:
    if not tf.gfile.Exists(f):
      raise ValueError('Failed to find file: ' + f)

  # 创建生成文件名的队列.
  filename_queue = tf.train.string_input_producer(filenames)

  # Read examples from files in the filename queue.
  read_input = read_cifar10(filename_queue)
  reshaped_image = tf.cast(read_input.uint8image, tf.float32)

  height = IMAGE_SIZE
  width = IMAGE_SIZE

  # Image processing for training the network. Note the many random
  # distortions applied to the image.
  # 为了训练网络图片处理，多种随机变形应用到图片中
  
  # 随机裁剪成crop a [height, width] section of the image.
  distorted_image = tf.random_crop(reshaped_image, [height, width, 3])

  # 随机水平翻转
  distorted_image = tf.image.random_flip_left_right(distorted_image)

  # Because these operations are not commutative, consider randomizing
  # the order their operation.
  # 因为这些操作不是交替的，考虑操作的随机顺序
  distorted_image = tf.image.random_brightness(distorted_image,
                                               max_delta=63)
  distorted_image = tf.image.random_contrast(distorted_image,
                                             lower=0.2, upper=1.8)

  # Subtract off the mean and divide by the variance of the pixels.
  float_image = tf.image.per_image_standardization(distorted_image)
  #float_image = tf.image.ptrain_direr_image_whitening(distorted_image)

  # Ensure that the random shuffling has good mixing properties.
  min_fraction_of_examples_in_queue = 0.4
  min_queue_examples = int(NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN *
                           min_fraction_of_examples_in_queue)
  print ('Filling queue with %d CIFAR images before starting to train. '
         'This will take a few minutes.' % min_queue_examples)

  # Generate a batch of images and labels by building up a queue of examples.
  return _generate_image_and_label_batch(float_image, read_input.label,
                                         min_queue_examples, batch_size,
                                         shuffle=True)

def inputs(eval_data, data_dir, batch_size):
  """Construct input for CIFAR evaluation using the Reader ops.
  Args:
    eval_data: bool, indicating if one should use the train or eval data set.
    data_dir: Path to the CIFAR-10 data directory.
    batch_size: Number of images per batch.
  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.
  """
  if not eval_data:
    filenames = [os.path.join(data_dir, 'data_batch_%d' % i)
                 for i in xrange(1, 6)]
    num_examples_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN
  else:
    filenames = [os.path.join(data_dir, 'test_batch')]
    num_examples_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_EVAL

  for f in filenames:
    if not tf.gfile.Exists(f):
      raise ValueError('找不到文件: ' + f)

  # Create a queue that produces the filenames to read.
  filename_queue = tf.train.string_input_producer(filenames)

  # Read examples from files in the filename queue.
  read_input = read_cifar10(filename_queue)
  reshaped_image = tf.cast(read_input.uint8image, tf.float32)

  height = IMAGE_SIZE
  width = IMAGE_SIZE

  # Image processing for evaluation.
  # Crop the central [height, width] of the image.
  resized_image = tf.image.resize_image_with_crop_or_pad(reshaped_image,
                                                         width, height)

  # Subtract off the mean and divide by the variance of the pixels.
  float_image = tf.image.per_image_whitening(resized_image)

  # Ensure that the random shuffling has good mixing properties.
  min_fraction_of_examples_in_queue = 0.4
  min_queue_examples = int(num_examples_per_epoch *
                           min_fraction_of_examples_in_queue)

  # Generate a batch of images and labels by building up a queue of examples.
  return _generate_image_and_label_batch(float_image, read_input.label,
                                         min_queue_examples, batch_size,
                                         shuffle=False)

2.cifar10.py：构建 CIFAR-10模型

# -*- coding: utf-8 -*-
# Copyright 2015 The TensorFlow Authors. 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 inputs() 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 os
import re
import sys
import tarfile

from six.moves import urllib
import tensorflow as tf

#from tensorflow.models.image.cifar10 import cifar10_input
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.""")
tf.app.flags.DEFINE_boolean('use_fp16', False,
                            """Train the model using fp16.""")

# 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 GPUs, 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: 张量
  Returns:
    无
  """
  # 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.histogram_summary(tensor_name + '/activations', x)
  tf.summary.histogram(tensor_name +'/activations',x)
  tf.summary.scalar(tensor_name + '/sparsity', tf.nn.zero_fraction(x))
  #tf.scalar_summary(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: 变量的名字
    shape: 整数列表
    initializer: initializer for Variable
  Returns:
    Variable Tensor
  """
  with tf.device('/cpu:0'):
    dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
    var = tf.get_variable(name, shape, initializer=initializer, dtype=dtype)
  return var

def _variable_with_weight_decay(name, shape, stddev, wd):
  """Helper to create an initialized Variable with weight decay权值递减.
  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
  """
  dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
  var = _variable_on_cpu(
      name,
      shape,
      tf.truncated_normal_initializer(stddev=stddev, dtype=dtype))
  if wd is not None:
    #weight_decay = tf.mul(tf.nn.l2_loss(var), wd, name='weight_loss')
    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')
  images, labels = cifar10_input.distorted_inputs(data_dir=data_dir,
                                                  batch_size=FLAGS.batch_size)
  if FLAGS.use_fp16:
    images = tf.cast(images, tf.float16)
    labels = tf.cast(labels, tf.float16)
  return images, labels

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')
  images, labels = cifar10_input.inputs(eval_data=eval_data,
                                        data_dir=data_dir,
                                        batch_size=FLAGS.batch_size)
  if FLAGS.use_fp16:
    images = tf.cast(images, tf.float16)
    labels = tf.cast(labels, tf.float16)
  return images, labels

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().
  #
  # 卷积1
  with tf.variable_scope('conv1') as scope:
    kernel = _variable_with_weight_decay('weights',
                                         shape=[5, 5, 3, 64],
                                         stddev=5e-2,
                                         wd=0.0)
    #images:输入；kernel：核；[1,1,1,1]：步长;padding:填充方式
    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)

  # 池化1
  pool1 = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
                         padding='SAME', name='pool1')
  # 规范化1
  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=5e-2,
                                         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.
    reshape = tf.reshape(pool2, [FLAGS.batch_size, -1])
    dim = reshape.get_shape()[1].value
    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 "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.
  """
  # Calculate the average cross entropy loss across the batch.
  labels = tf.cast(labels, tf.int64)
  cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
      labels=labels,logits=logits,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.scalar_summary(l.op.name +' (raw)', l)
    #tf.scalar_summary(l.op.name, loss_averages.average(l))
    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.scalar_summary('learning_rate', lr)
  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.histogram_summary(var.op.name, var)
    tf.summary.histogram(var.op.name,var)

  # Add histograms for gradients.
  for grad, var in grads:
    if grad is not None:
      #tf.histogram_summary(var.op.name + '/gradients', grad)
      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, _progress)
    print()
    statinfo = os.stat(filepath)
    print('Successfully downloaded', filename, statinfo.st_size, 'bytes.')
    tarfile.open(filepath, 'r:gz').extractall(dest_directory)