inception v3迁移学习

1、

# -*- coding: utf-8 -*-  

"""
@author:
 
卷积神经网络 Inception-v3模型 迁移学习
"""  
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  
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'  
# 图像输入张量所对应的名称  
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 = './v3'  
# 下载的谷歌训练好的inception-v3模型文件名  
MODEL_FILE = 'tensorflow_inception_graph.pb'  
 
# 保存训练数据通过瓶颈层后提取的特征向量  
CACHE_DIR = './tmp/bottleneck'  
 
# 图片数据的文件夹  
INPUT_DATA = 'E:/flower_photos'  
 
# 验证的数据百分比  
VALIDATION_PERCENTAGE = 10  
# 测试的数据百分比  
TEST_PERCENTACE = 10  
 
# 定义神经网路的设置  
LEARNING_RATE = 0.01  
STEPS = 4000  
BATCH = 100  
 
 
# 这个函数把数据集分成训练，验证，测试三部分  
def create_image_lists(testing_percentage, validation_percentage):  
    """
    这个函数把数据集分成训练，验证，测试三部分
    :param testing_percentage:测试的数据百分比 10
    :param validation_percentage:验证的数据百分比 10
    :return:
    """  
    result = {}  
    # 获取目录下所有子目录  
    sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]  
    # ['/path/to/flower_data', '/path/to/flower_data\\daisy', '/path/to/flower_data\\dandelion',  
    # '/path/to/flower_data\\roses', '/path/to/flower_data\\sunflowers', '/path/to/flower_data\\tulips']  
 
    # 数组中的第一个目录是当前目录，这里设置标记，不予处理  
    is_root_dir = True  
 
    for sub_dir in sub_dirs:  # 遍历目录数组，每次处理一种  
        if is_root_dir:  
            is_root_dir = False  
            continue  
 
        # 获取当前目录下所有的有效图片文件  
        extensions = ['jpg', 'jepg', 'JPG', 'JPEG']  
        file_list = []  
        dir_name = os.path.basename(sub_dir)  # 返回路径名路径的基本名称，如：daisy|dandelion|roses|sunflowers|tulips  
        for extension in extensions:  
            file_glob = os.path.join(INPUT_DATA, dir_name, '*.' + extension)  # 将多个路径组合后返回  
            file_list.extend(glob.glob(file_glob))  # glob.glob返回所有匹配的文件路径列表，extend往列表中追加另一个列表  
        if not file_list: continue  
 
        # 通过目录名获取类别名称  
        label_name = dir_name.lower()  # 返回其小写  
        # 初始化当前类别的训练数据集、测试数据集、验证数据集  
        training_images = []  
        testing_images = []  
        validation_images = []  
 
        for file_name in file_list:  # 遍历此类图片的每张图片的路径  
            base_name = os.path.basename(file_name)  # 路径的基本名称也就是图片的名称，如：102841525_bd6628ae3c.jpg  
            # 随机讲数据分到训练数据集、测试集和验证集  
            chance = np.random.randint(100)  
            if chance < validation_percentage:  
                validation_images.append(base_name)  
            elif chance < (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  
 
 
# 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  
def get_image_path(image_lists, image_dir, label_name, index, category):  
    """
    :param image_lists:所有图片信息
    :param image_dir:根目录 （ 图片特征向量根目录 CACHE_DIR | 图片原始路径根目录 INPUT_DATA ）
    :param label_name:类别的名称（ daisy|dandelion|roses|sunflowers|tulips ）
    :param index:编号
    :param category:所属的数据集（ training|testing|validation ）
    :return: 一张图片的地址
    """  
    # 获取给定类别的图片集合  
    label_lists = image_lists[label_name]  
    # 获取这种类别的图片中，特定的数据集(base_name的一维数组)  
    category_list = label_lists[category]  
    mod_index = index % len(category_list)  # 图片的编号%此数据集中图片数量  
    # 获取图片文件名  
    base_name = category_list[mod_index]  
    sub_dir = label_lists['dir']  
    # 拼接地址  
    full_path = os.path.join(image_dir, sub_dir, base_name)  
    return full_path  
 
    
# 图片的特征向量的文件地址  
def get_bottleneck_path(image_lists, label_name, index, category):  
    return get_image_path(image_lists, CACHE_DIR, label_name, index, category) + '.txt'  # CACHE_DIR 特征向量的根地址  
 
 
# 计算特征向量  
def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):  
    """
    :param sess:
    :param image_data:图片内容
    :param image_data_tensor:
    :param bottleneck_tensor:
    :return:
    """  
    bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})  
    bottleneck_values = np.squeeze(bottleneck_values)  
    return bottleneck_values  
 
 
# 获取一张图片对应的特征向量的路径  
def get_or_create_bottleneck(sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor):  
    """
    :param sess:
    :param image_lists:
    :param label_name:类别名
    :param index:图片编号
    :param category:
    :param jpeg_data_tensor:
    :param bottleneck_tensor:
    :return:
    """  
    label_lists = image_lists[label_name]  
    sub_dir = label_lists['dir']  
    sub_dir_path = os.path.join(CACHE_DIR, sub_dir)  # 到类别的文件夹  
    if not os.path.exists(sub_dir_path): os.makedirs(sub_dir_path)  
 
    bottleneck_path = get_bottleneck_path(image_lists, label_name, index, category)  # 获取图片特征向量的路径  
    if not os.path.exists(bottleneck_path):  # 如果不存在  
        # 获取图片原始路径  
        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()  
        # 计算图片特征向量  
        bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)  
        # 将特征向量存储到文件  
        bottleneck_string = ','.join(str(x) for x in bottleneck_values)  
        with open(bottleneck_path, 'w') as bottleneck_file:  
            bottleneck_file.write(bottleneck_string)  
    else:  
        # 读取保存的特征向量文件  
        with open(bottleneck_path, 'r') as bottleneck_file:  
            bottleneck_string = bottleneck_file.read()  
        # 字符串转float数组  
        bottleneck_values = [float(x) for x in bottleneck_string.split(',')]  
    return bottleneck_values  
 
 
# 随机获取一个batch的图片作为训练数据（特征向量，类别）  
def get_random_cached_bottlenecks(sess, n_classes, image_lists, how_many, category, jpeg_data_tensor,  
                                  bottleneck_tensor):  
    """
    :param sess:
    :param n_classes: 类别数量
    :param image_lists:
    :param how_many: 一个batch的数量
    :param category: 所属的数据集
    :param jpeg_data_tensor:
    :param bottleneck_tensor:
    :return: 特征向量列表，类别列表
    """  
    bottlenecks = []  
    ground_truths = []  
    for _ in range(how_many):  
        # 随机一个类别和图片编号加入当前的训练数据  
        label_index = random.randrange(n_classes)  
        label_name = list(image_lists.keys())[label_index]  # 随机图片的类别名  
        image_index = random.randrange(65536)  # 随机图片的编号  
        bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_index, category, jpeg_data_tensor,  
                                              bottleneck_tensor)  # 计算此图片的特征向量  
        ground_truth = np.zeros(n_classes, dtype=np.float32)  
        ground_truth[label_index] = 1.0  
        bottlenecks.append(bottleneck)  
        ground_truths.append(ground_truth)  
    return bottlenecks, ground_truths  
 
 
# 获取全部的测试数据  
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):  
    bottlenecks = []  
    ground_truths = []  
    label_name_list = list(image_lists.keys())  # ['dandelion', 'daisy', 'sunflowers', 'roses', 'tulips']  
    for label_index, label_name in enumerate(label_name_list):  # 枚举每个类别,如:0 sunflowers  
        category = 'testing'  
        for index, unused_base_name in enumerate(image_lists[label_name][category]):  # 枚举此类别中的测试数据集中的每张图片  
            '''''
            print(index, unused_base_name)
            0 10386503264_e05387e1f7_m.jpg
            1 1419608016_707b887337_n.jpg
            2 14244410747_22691ece4a_n.jpg
            ...
            105 9467543719_c4800becbb_m.jpg
            106 9595857626_979c45e5bf_n.jpg
            107 9922116524_ab4a2533fe_n.jpg
            '''  
            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] = 1.0  
            bottlenecks.append(bottleneck)  
            ground_truths.append(ground_truth)  
    return bottlenecks, ground_truths  
 
 
def main(_):  
    image_lists = create_image_lists(TEST_PERCENTACE, VALIDATION_PERCENTAGE)  
    n_classes = len(image_lists.keys())  
    # 读取模型  
    with gfile.FastGFile(os.path.join(MODEL_DIR, MODEL_FILE), 'rb') as f:  
        graph_def = tf.GraphDef()  
        graph_def.ParseFromString(f.read())  
    # 加载模型，返回对应名称的张量  
    bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, return_elements=[BOTTLENECK_TENSOR_NAME,  
                                                                                          JPEG_DATA_TENSOR_NAME])  
    # 输入  
    bottleneck_input = tf.placeholder(tf.float32, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')  
    ground_truth_input = tf.placeholder(tf.float32, [None, n_classes], name='GroundTruthInput')  
 
    # 全连接层  
    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) + biases  
        final_tensor = tf.nn.softmax(logits)  
 
    # 损失  
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=ground_truth_input)  
    cross_entropy_mean = tf.reduce_mean(cross_entropy)  
    # 优化  
    train_step = tf.train.GradientDescentOptimizer(LEARNING_RATE).minimize(cross_entropy_mean)  
 
    # 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:  
        # 初始化参数  
        init = tf.global_variables_initializer()
        saver = tf.train.Saver()
        sess.run(init)  
 
        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)  
            # 训练  
            sess.run(train_step,  
                     feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth})  
            saver.save(sess, "./model/model.ckpt")
            # 验证  
            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 examples = %.1f%%' % (  
                    i, BATCH, validation_accuracy * 100))  
 
        # 测试  
        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 = %.1f%%' % (test_accuracy * 100))  
 
 
if __name__ == '__main__':  

    tf.app.run()

2.预测

# -*- coding: utf-8 -*-
import tensorflow as tf
import numpy as np

# 模型目录
CHECKPOINT_DIR = './model/'  #训练后生成的检查点文件夹，在当前工程下。
INCEPTION_MODEL_FILE = './v3/tensorflow_inception_graph.pb'

# inception-v3模型参数
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'  # inception-v3模型中代表瓶颈层结果的张量名称
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'  # 图像输入张量对应的名称

# 测试数据
path = 'C:/Users/28761/tensorflow/path/imags/xiang.jpg' #这里选择一张图片用于测试，该图片属于sunflower类别的花。
#类别字典
flower_dict={0:'daisy',1:'dandelion',2:'roses',3:'sunflowers',4:'tulips'}

# 读取数据
image_data = tf.gfile.FastGFile(path, 'rb').read()

# 评估
checkpoint_file = tf.train.latest_checkpoint(CHECKPOINT_DIR)
with tf.Graph().as_default() as graph:
    with tf.Session().as_default() as sess:

        # 读取训练好的inception-v3模型
        with tf.gfile.FastGFile(INCEPTION_MODEL_FILE, 'rb') as f:
            graph_def = tf.GraphDef()
            graph_def.ParseFromString(f.read())

        # 加载inception-v3模型，并返回数据输入张量和瓶颈层输出张量
        bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(
            graph_def,
            return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME])

        # 使用inception-v3处理图片获取特征向量
        bottleneck_values = sess.run(bottleneck_tensor,{jpeg_data_tensor: image_data})
        # 将四维数组压缩成一维数组，由于全连接层输入时有batch的维度，所以用列表作为输入
        bottleneck_values = [np.squeeze(bottleneck_values)]

        # 加载图和变量（这里我选择的是step=900的图，使用的是绝对路径。）
        saver = tf.train.import_meta_graph('./model/model.ckpt.meta')
        saver.restore(sess, tf.train.latest_checkpoint('./model/'))

        # 通过名字从图中获取输入占位符
        input_x = graph.get_operation_by_name(
            'BottleneckInputPlaceholder').outputs[0]

        # 我们想要评估的tensors
        predictions = graph.get_operation_by_name('evaluation/ArgMax').outputs[0]

        # 收集预测值
        all_predictions = []
        all_predictions = sess.run(predictions, {input_x: bottleneck_values})


        # 打印出预测结果
        index=str(all_predictions)[1]
        index=int(index)
        print(path+' '+'预测为：'+flower_dict[index])