"""
@author: wangquaxiu
@time: 2018/8/27 19:13
"""
import glob
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile# Inception-v3模型瓶颈层的节点个数
BOTTLENECK_TENSOR_SIZE = 2048# Inception-v3模型中代表瓶颈层结果的张量名称。在谷歌提供的Inception-v3模型中，这
# 个张量名称就是'pool_3/_reshape:0'。在训练模型时，可以通过tensor.name来获取张量的名称。
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'# 图像输入张量多对应的名称。
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'# 下载的谷歌训练好的Inception-v3模型文件目录。
MODEL_DIR = 'path/to/model'# 下载的谷歌训练好的Inception-v3模型文件名
MODEL_FILE = 'tensorflow_inception_graph.pb'# 因为一个训练数据会被使用多次，所以可以将原始图像通过Inception-v3模型计算得到
# 的特征向量保存在文件中，免去重复计算。下面的变量定义了这些文件的存放地址。
CACHE_DIR = 'tmp/bottlebeck'# 图片数据文件夹。在这个文件夹中每一个子文件夹代表了一个需要区分的类别，每个子文件夹中
# 存放了对应类别的图片。
INPUT_DATA = 'path/to/flower_data'# 验证的数据百分比。
VALIDATION_PERCENTAGE = 10# 测试的数据百分比
TEST_PRECENTAGE = 10# 定义神经网络的设置。
LEARNING_RATE = 0.01
STEPS = 4000
BATCH = 100# 这个函数从数据文件夹中读取所有的图片列表并按照训练、验证、测试数据分开。
# testing_percentage和validation_percentage参数指定了测试数据集和验证数据集的
# 大小
def create_image_lists(testing_percentage, validation_perscentage):# 得到的所有图片都存在result这个字典（dictionary）里。这个字典的key为类别名称，# value也是一个字典，字典里储存；了所有图片的名称。result = {}# 获取当前目录下所有的子目录。sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]# 得到的第一个目录是当前目录，不需要考虑。is_root_dir = Truefor sub_dir in sub_dirs:if is_root_dir:is_root_dir = Falsecontinue# 获取当前目录下所有的有效图片文件extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']file_list = []dir_name = os.path.basename((sub_dir))for extension in extensions:file_glob = os.path.join(INPUT_DATA, dir_name, '*.' + extension)file_list.extend(glob.glob(file_glob))if not file_list:continue# 通过目录名称获取类别的名称label_name = dir_name.lower()# 初始化当前类别的训练数据集、测试数据集和验证数据集。training_images = []testing_images = []validation_images = []for file_name in file_list:base_name = os.path.basename(file_name)# 随即将数据分到训练数据集、测试数据集和验证数据集。chance = np.random.randint(100)if chance < validation_perscentage:validation_images.append(base_name)elif chance < (testing_percentage + validation_perscentage):testing_images.append(base_name)else:training_images.append(base_name)# 将当前类别的数据放入结果字典。result[label_name] = {'dir': dir_name,'training': training_images,'testing': testing_images,'validation': validation_images}# 返回整理好的所有数据。return result# 这个函数通过类别名称、所属数据集和图片编号获取一张图片的地址。
# image_lists参数给出了多有图片信息。
# image_dir参数给出了根目录。存放数据的根目录和存放图片特征向量的根目录地址不同。
# label_name参数给定了类别的名称。
# imdex参数给定了需要获取的图片的编号。
# category参数指定了需要获取的图片是在训练数据集、测试数据集还是验证数据集。
def get_image_path(image_lists, image_dir, label_name, index, category):# 获取给定类别中所有图片的信息。label_lists = image_lists[label_name]# 根据所属数据集的名称获取集合中全部图片信息。category_list = label_lists[category]mod_index = index % len(category_list)# 获取图片的文件名。base_name = category_list[mod_index]sub_dir = label_lists['dir']# 最终的地址为数据根目录的地址加上类别的文件夹加上图片的名称。full_apth = os.path.join(image_dir, sub_dir, base_name)return full_apth# 这个函数通过类别名称、所属数据集和图片编号获取经过Inception-v3模型处理之后的特征向量
# 文件地址
def get_bottleneck_path(image_lists, label_name, index, category):return get_image_path(image_lists, CACHE_DIR, label_name, index, category) + '.txt'# 这个函数使用加载的训练好的Inception-v3模型处理一张图片，得到这个图片的特征向量。
def run_bottlebeck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):# 这个过程实际上就是将当前图片作为输入计算瓶颈张量的值。这个瓶颈张量的值就是这张# 图片的新的特征向量bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})# 经过卷积神经网络处理的结果是一个四维数组，需要将这个结果压缩成一个特征向量bottleneck_values = np.squeeze(bottleneck_values)return bottleneck_values# 这个函数获取一张图片经过Inception-v3模型处理之后的特征向量。这个函数会先试图寻找
# 已经计算且保存下来的特征向量，如果找不到则先计算这个特征向量，然后保存到文件。
def get_or_create_bottleneck(sess, image_lists, label_name, index,category, jpeg_data_tensor, bottlebeck_tensor):# 获取一张图片对应的特征向量文件的路径。label_lists = image_lists[label_name]sub_dir = label_lists['dir']sub_dir_path = os.path.join(CACHE_DIR, sub_dir)if not os.path.exists(sub_dir_path):os.makedirs(sub_dir_path)bottlebeck_path = get_bottleneck_path(image_lists, label_name, index, category)# 如果这个特征向量文件不存在，则通过Inception-v3模型来计算特征向量，并将计算的结果# 存入文件if not os.path.exists(bottlebeck_path):# 获取原始的图片路径。image_path = get_image_path(image_lists, INPUT_DATA, label_name, index, category)# 获取图片内容。image_data = gfile.FastGFile(image_path, 'rb').read()# 通过Inception-v3模型计算特征向量。bottlebeck_values = run_bottlebeck_on_image(sess, image_data, jpeg_data_tensor,bottlebeck_tensor)# 将计算得到的特征向量存入文件bottlebeck_string = ','.join(str(x) for x in bottlebeck_values)with open(bottlebeck_path, 'w') as bottlebeck_file:bottlebeck_file.write(bottlebeck_string)else:# 直接从文件中获取图片相应的特征向量。with open(bottlebeck_path, 'r') as bottlebeck_file:bottlebeck_string = bottlebeck_file.read()bottlebeck_values = [float(x) for x in bottlebeck_string.split(',')]# 返回得到的特征向量return bottlebeck_values# 这个函数随机获取一个batch的图片作为训练数据。
def get_random_cached_bottlenecks(sess, n_classes, image_lists, how_many, category,jpeg_data_tensor, bottleneck_tensor):bottlenecks = []ground_truths = []for _ in range(how_many):# 随机一个类别和图片的编号加入当前的训练数据。label_index = random.randrange(n_classes)label_name = list(image_lists)[label_index]image_index = random.randrange(65536)bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_index, category,jpeg_data_tensor, bottleneck_tensor)ground_truth = np.zeros(n_classes, dtype=np.float32)ground_truth[label_index] = 1.0bottlenecks.append(bottleneck)ground_truths.append(ground_truth)return bottlenecks, ground_truths# 这个函数获取全部的测试数据。在最终测试的时候需要在所有的测试数据上计算正确率。
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):bottlenecks = []ground_truths = []label_name_list = list(image_lists.keys())# 枚举所有的类别和每个类别中的测试图片。for label_index, label_name in enumerate(label_name_list):category = 'testing'for index, unused_base_name in enumerate(image_lists[label_name][category]):# 通过Inception-v3模型计算图片对应的特征向量，并将其加入最终数据的列表。bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, index, category,jpeg_data_tensor, bottleneck_tensor)ground_truth = np.zeros(n_classes, dtype=np.float32)ground_truth[label_index] = 1.0bottlenecks.append(bottleneck)ground_truths.append(ground_truth)return bottlenecks, ground_truthsdef main(_):# 读取所有图片image_lists = create_image_lists(TEST_PRECENTAGE, VALIDATION_PERCENTAGE)n_classes = len(image_lists.keys())# 读取已经训练好的Inception-v3模型。谷歌训练好的模型保存在了GraphDefProtocol# Buffer中，里面保存了每一个节点取值的计算方法以及变脸的取值。Tensorflow模型# 持久化的问题在第五章中有详细介绍with gfile.FastGFile(os.path.join(MODEL_DIR, MODEL_FILE), 'rb') as f:graph_def = tf.GraphDef()graph_def.ParseFromString(f.read())# 加载读取的Inception-v3模型，并返回数据输入多对应的张量以及计算瓶颈层结果# 所对应的张量。bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME])# 定义新的社宁网络输入，这个输入就是新的图片经过Inception-v3模型前向传播到达# 瓶颈层时的取值。可以将整个过程类似的理解为一种特征提取。bottleneck_input = tf.placeholder(tf.float32, [None, BOTTLENECK_TENSOR_SIZE],name='BottleneckInputPlaceholder')# 定义新的标准答案输入。ground_truth_input = tf.placeholder(tf.float32, [None, n_classes],name='GroundTruthInput')# 定义一层全连接层来解决新的图片分类问题。因为训练好的Inception-v3模型已经将原始的# 图片抽象为更加容易分类的特征向量了，所以不需要再训练那么复杂的神经网络来完成这个# 新的分类任务with tf.name_scope('final_training_ops'):weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, n_classes], stddev=0.001))biases = tf.Variable(tf.zeros([n_classes]))logits = tf.matmul(bottleneck_input, weights) + biasesfinal_tensor = tf.nn.softmax(logits)# 定义交叉熵损失函数。cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=ground_truth_input)cross_entropy_mean = tf.reduce_mean(cross_entropy)train_step = tf.train.GradientDescentOptimizer(LEARNING_RATE).minimize(cross_entropy_mean)# 计算正确率。with tf.name_scope('evaluation'):correct_prediction = tf.equal(tf.argmax(final_tensor, 1),tf.argmax(ground_truth_input, 1))evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))with tf.Session() as sess:init = tf.global_variables_initializer()sess.run(init)# 训练过程。for i in range(STEPS):# 每次获取一个batch的训练数据。train_bottlenecks, train_ground_truth = get_random_cached_bottlenecks(sess, n_classes, image_lists, BATCH, 'training',jpeg_data_tensor, bottleneck_tensor)sess.run(train_step, feed_dict={bottleneck_input: train_bottlenecks,ground_truth_input: train_ground_truth})# 在验证数据上测试正确率。if i % 100 == 0 or i + 1 == STEPS:validation_bottlenecks, validation_ground_truth = get_random_cached_bottlenecks(sess, n_classes, image_lists, BATCH,'validation', jpeg_data_tensor,bottleneck_tensor)validation_accuracy = sess.run(evaluation_step, feed_dict={bottleneck_input: validation_bottlenecks,ground_truth_input: validation_ground_truth})print('Step % d: Validation accuracy on random sampled %d example = %.lf%%' %(i, BATCH, validation_accuracy * 100))# 在最后的测试数据上测试正确率。test_bottlenecks, test_ground_truth = get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor)test_accuracy = sess.run(evaluation_step, feed_dict = {bottleneck_input: test_bottlenecks,ground_truth_input: test_ground_truth})print('Final test accuracy = %.lf%%' % (test_accuracy*100))if __name__ == '__main__':tf.app.run()
结果：
Step 0: Validation accuracy on random sampled 100 example = 30%
Step 100: Validation accuracy on random sampled 100 example = 77%
Step 200: Validation accuracy on random sampled 100 example = 90%
Step 300: Validation accuracy on random sampled 100 example = 89%
Step 400: Validation accuracy on random sampled 100 example = 92%
Step 500: Validation accuracy on random sampled 100 example = 94%
Step 600: Validation accuracy on random sampled 100 example = 88%
Step 700: Validation accuracy on random sampled 100 example = 93%
Step 800: Validation accuracy on random sampled 100 example = 92%
Step 900: Validation accuracy on random sampled 100 example = 92%
Step 1000: Validation accuracy on random sampled 100 example = 93%
Step 1100: Validation accuracy on random sampled 100 example = 88%
Step 1200: Validation accuracy on random sampled 100 example = 89%
Step 1300: Validation accuracy on random sampled 100 example = 88%
Step 1400: Validation accuracy on random sampled 100 example = 90%
Step 1500: Validation accuracy on random sampled 100 example = 94%
Step 1600: Validation accuracy on random sampled 100 example = 93%
Step 1700: Validation accuracy on random sampled 100 example = 94%
Step 1800: Validation accuracy on random sampled 100 example = 91%
Step 1900: Validation accuracy on random sampled 100 example = 93%
Step 2000: Validation accuracy on random sampled 100 example = 98%
Step 2100: Validation accuracy on random sampled 100 example = 94%
Step 2200: Validation accuracy on random sampled 100 example = 94%
Step 2300: Validation accuracy on random sampled 100 example = 94%
Step 2400: Validation accuracy on random sampled 100 example = 91%
Step 2500: Validation accuracy on random sampled 100 example = 92%
Step 2600: Validation accuracy on random sampled 100 example = 90%
Step 2700: Validation accuracy on random sampled 100 example = 94%
Step 2800: Validation accuracy on random sampled 100 example = 95%
Step 2900: Validation accuracy on random sampled 100 example = 95%
Step 3000: Validation accuracy on random sampled 100 example = 95%
Step 3100: Validation accuracy on random sampled 100 example = 94%
Step 3200: Validation accuracy on random sampled 100 example = 97%
Step 3300: Validation accuracy on random sampled 100 example = 94%
Step 3400: Validation accuracy on random sampled 100 example = 86%
Step 3500: Validation accuracy on random sampled 100 example = 94%
Step 3600: Validation accuracy on random sampled 100 example = 93%
Step 3700: Validation accuracy on random sampled 100 example = 91%
Step 3800: Validation accuracy on random sampled 100 example = 92%
Step 3900: Validation accuracy on random sampled 100 example = 95%
Step 3999: Validation accuracy on random sampled 100 example = 92%
Final test accuracy = 94%