TensorFlow入门教程（十三）：利用inception-v3训练自己的模型

一、  训练自己模型的方法

inception-v3 最初是在ImageNet比赛训练的，其分类有1000多种，但对于自己的特定目标，分类效果并不好，需要自己再对自己的数据进行训练，有一下三种方法：

1. 赋予随机权值，从头开始训练。这种方法需要大量的数据集，并且运行时间较长。
2. 在 inception-v3 基础上，最后加一层进行训练，只修改最后一层的权重。
3. 在 inception-v3 基础上，对于训练好的权重，以很小的学习率做微调，主要训练自己添加的最后一层。

对于自己的模型，我们多采用第二种方法，相当于将之前训练好的模型去提取图像的特征，最后加一层，对这些特征做判断，分类，结构如图所示：

即在pool_3后再不作softmax处理，再送入一层中做分类识别。

二、  前期准备

1. 数据集：http://www.robots.ox.ac.uk/~vgg/data/
2. TensorFlow程序  https://github.com/tensorflow/tensorflow
3. inception-v3 模型

谷歌官方已经给出了迁移学习的代码，在retrain.py里面，谷歌好像推出新的retrain.py，导致老版的没有了，所以现在把老版的贴出来，供大家使用。

# 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.
# ==============================================================================
r"""Simple transfer learning with Inception v3 or Mobilenet models.

With support for TensorBoard.

This example shows how to take a Inception v3 or Mobilenet model trained on
ImageNet images, and train a new top layer that can recognize other classes of
images.

The top layer receives as input a 2048-dimensional vector (1001-dimensional for
Mobilenet) for each image. We train a softmax layer on top of this
representation. Assuming the softmax layer contains N labels, this corresponds
to learning N + 2048*N (or 1001*N)  model parameters corresponding to the
learned biases and weights.

Here's an example, which assumes you have a folder containing class-named
subfolders, each full of images for each label. The example folder flower_photos
should have a structure like this:

~/flower_photos/daisy/photo1.jpg
~/flower_photos/daisy/photo2.jpg
...
~/flower_photos/rose/anotherphoto77.jpg
...
~/flower_photos/sunflower/somepicture.jpg

The subfolder names are important, since they define what label is applied to
each image, but the filenames themselves don't matter. Once your images are
prepared, you can run the training with a command like this:


bash:
bazel build tensorflow/examples/image_retraining:retrain && \
bazel-bin/tensorflow/examples/image_retraining/retrain \
    --image_dir ~/flower_photos


Or, if you have a pip installation of tensorflow, `retrain.py` can be run
without bazel:

bash:
python tensorflow/examples/image_retraining/retrain.py \
    --image_dir ~/flower_photos


You can replace the image_dir argument with any folder containing subfolders of
images. The label for each image is taken from the name of the subfolder it's
in.

This produces a new model file that can be loaded and run by any TensorFlow
program, for example the label_image sample code.

By default this script will use the high accuracy, but comparatively large and
slow Inception v3 model architecture. It's recommended that you start with this
to validate that you have gathered good training data, but if you want to deploy
on resource-limited platforms, you can try the `--architecture` flag with a
Mobilenet model. For example:

bash:
python tensorflow/examples/image_retraining/retrain.py \
    --image_dir ~/flower_photos --architecture mobilenet_1.0_224


There are 32 different Mobilenet models to choose from, with a variety of file
size and latency options. The first number can be '1.0', '0.75', '0.50', or
'0.25' to control the size, and the second controls the input image size, either
'224', '192', '160', or '128', with smaller sizes running faster. See
https://research.googleblog.com/2017/06/mobilenets-open-source-models-for.html
for more information on Mobilenet.

To use with TensorBoard:

By default, this script will log summaries to /tmp/retrain_logs directory

Visualize the summaries with this command:

tensorboard --logdir /tmp/retrain_logs

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

import argparse
from datetime import datetime
import hashlib
import os.path
import random
import re
import sys
import tarfile

import numpy as np
from six.moves import urllib
import tensorflow as tf

from tensorflow.python.framework import graph_util
from tensorflow.python.framework import tensor_shape
from tensorflow.python.platform import gfile
from tensorflow.python.util import compat

FLAGS = None

# These are all parameters that are tied to the particular model architecture
# we're using for Inception v3. These include things like tensor names and their
# sizes. If you want to adapt this script to work with another model, you will
# need to update these to reflect the values in the network you're using.
MAX_NUM_IMAGES_PER_CLASS = 2 ** 27 - 1  # ~134M


def create_image_lists(image_dir, testing_percentage, validation_percentage):
  """Builds a list of training images from the file system.

  Analyzes the sub folders in the image directory, splits them into stable
  training, testing, and validation sets, and returns a data structure
  describing the lists of images for each label and their paths.

  Args:
    image_dir: String path to a folder containing subfolders of images.
    testing_percentage: Integer percentage of the images to reserve for tests.
    validation_percentage: Integer percentage of images reserved for validation.

  Returns:
    A dictionary containing an entry for each label subfolder, with images split
    into training, testing, and validation sets within each label.
  """
  if not gfile.Exists(image_dir):
    tf.logging.error("Image directory '" + image_dir + "' not found.")
    return None
  result = {}
  sub_dirs = [x[0] for x in gfile.Walk(image_dir)]
  # The root directory comes first, so skip it.
  is_root_dir = True
  for sub_dir in sub_dirs:
    if is_root_dir:
      is_root_dir = False
      continue
    extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
    file_list = []
    dir_name = os.path.basename(sub_dir)
    if dir_name == image_dir:
      continue
    tf.logging.info("Looking for images in '" + dir_name + "'")
    for extension in extensions:
      file_glob = os.path.join(image_dir, dir_name, '*.' + extension)
      file_list.extend(gfile.Glob(file_glob))
    if not file_list:
      tf.logging.warning('No files found')
      continue
    if len(file_list) < 20:
      tf.logging.warning(
          'WARNING: Folder has less than 20 images, which may cause issues.')
    elif len(file_list) > MAX_NUM_IMAGES_PER_CLASS:
      tf.logging.warning(
          'WARNING: Folder {} has more than {} images. Some images will '
          'never be selected.'.format(dir_name, MAX_NUM_IMAGES_PER_CLASS))
    label_name = re.sub(r'[^a-z0-9]+', ' ', dir_name.lower())
    training_images = []
    testing_images = []
    validation_images = []
    for file_name in file_list:
      base_name = os.path.basename(file_name)
      # We want to ignore anything after '_nohash_' in the file name when
      # deciding which set to put an image in, the data set creator has a way of
      # grouping photos that are close variations of each other. For example
      # this is used in the plant disease data set to group multiple pictures of
      # the same leaf.
      hash_name = re.sub(r'_nohash_.*$', '', file_name)
      # This looks a bit magical, but we need to decide whether this file should
      # go into the training, testing, or validation sets, and we want to keep
      # existing files in the same set even if more files are subsequently
      # added.
      # To do that, we need a stable way of deciding based on just the file name
      # itself, so we do a hash of that and then use that to generate a
      # probability value that we use to assign it.
      hash_name_hashed = hashlib.sha1(compat.as_bytes(hash_name)).hexdigest()
      percentage_hash = ((int(hash_name_hashed, 16) %
                          (MAX_NUM_IMAGES_PER_CLASS + 1)) *
                         (100.0 / MAX_NUM_IMAGES_PER_CLASS))
      if percentage_hash < validation_percentage:
        validation_images.append(base_name)
      elif percentage_hash < (testing_percentage + validation_percentage):
        testing_images.append(base_name)
      else:
        training_images.append(base_name)
    result[label_name] = {
        'dir': dir_name,
        'training': training_images,
        'testing': testing_images,
        'validation': validation_images,
    }
  return result


def get_image_path(image_lists, label_name, index, image_dir, category):
  """"Returns a path to an image for a label at the given index.

  Args:
    image_lists: Dictionary of training images for each label.
    label_name: Label string we want to get an image for.
    index: Int offset of the image we want. This will be moduloed by the
    available number of images for the label, so it can be arbitrarily large.
    image_dir: Root folder string of the subfolders containing the training
    images.
    category: Name string of set to pull images from - training, testing, or
    validation.

  Returns:
    File system path string to an image that meets the requested parameters.

  """
  if label_name not in image_lists:
    tf.logging.fatal('Label does not exist %s.', label_name)
  label_lists = image_lists[label_name]
  if category not in label_lists:
    tf.logging.fatal('Category does not exist %s.', category)
  category_list = label_lists[category]
  if not category_list:
    tf.logging.fatal('Label %s has no images in the category %s.',
                     label_name, category)
  mod_index = index % len(category_list)
  base_name = category_list[mod_index]
  sub_dir = label_lists['dir']
  full_path = os.path.join(image_dir, sub_dir, base_name)
  return full_path


def get_bottleneck_path(image_lists, label_name, index, bottleneck_dir,
                        category, architecture):
  """"Returns a path to a bottleneck file for a label at the given index.

  Args:
    image_lists: Dictionary of training images for each label.
    label_name: Label string we want to get an image for.
    index: Integer offset of the image we want. This will be moduloed by the
    available number of images for the label, so it can be arbitrarily large.
    bottleneck_dir: Folder string holding cached files of bottleneck values.
    category: Name string of set to pull images from - training, testing, or
    validation.
    architecture: The name of the model architecture.

  Returns:
    File system path string to an image that meets the requested parameters.
  """
  return get_image_path(image_lists, label_name, index, bottleneck_dir,
                        category) + '_' + architecture + '.txt'


def create_model_graph(model_info):
  """"Creates a graph from saved GraphDef file and returns a Graph object.

  Args:
    model_info: Dictionary containing information about the model architecture.

  Returns:
    Graph holding the trained Inception network, and various tensors we'll be
    manipulating.
  """
  with tf.Graph().as_default() as graph:
    model_path = os.path.join(FLAGS.model_dir, mode