1.背景

        自1998年LeNet-5模型的提出一直到现在，卷积神经网络模型的层数和复杂度都发生了巨大的变化，下表中罗列了ILSVRC（Lareg Scale Visual Recognition Challenge）第一名模型的表现：

        可以看到随着层数的增加，模型的识别错误率也在降低。然而ImagNet图像分类数据集中有120万标注图片， 在真实的引用中，很难收集到海量的训练数据，要训练一个复杂的卷积神经网络需要几天甚至几周的时间。为了解决该问题，迁移学习的概念被提出来。所谓迁移学习，就是讲一个问题上训练好的模型通过调整使之适应于一个新的问题。

2.基于tensorflow实现

        该迁移学习用到的数据集合为http://download.tensorflow.org/example_images/flower_photos.tgz.

        数据预处理代码，将彩色的三通道的图片转化为矩阵形式并通过numpy保存到一个文件中，模型的下载地址为http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz，具体代码如下：

"""
将原始图像数据整理成模型需要的输入数据
"""
import glob
import osimport numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfileINPUT_DATA = r"E:\master_paper\DATA_SET\flower_photos"OUTPUT_DATA = r"E:\paper_exp\paper1\tansferlearning\flower_procssed_data.npy"VALIDATION_PRECENTAGE = 10
TEST_PERCENTAGE = 10def create_image_list(sess, test_percentage, validation_percentage):"""读取数据并将数据分割成训练数据，验证数据和测试数据:param sess::param test_percentage::param validation_percentage::return:"""sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]is_root_dir = True# 初始化各个数据集training_images = []training_labels = []testing_images = []testing_labels = []validation_images = []validation_labels = []current_label = 0# 读取子目录for sub_dir in sub_dirs:if is_root_dir:  # 过滤掉根目录is_root_dir = Falsecontinue# 获取一个子目录中的所有图片文件extensions = ['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))  # glob 获取满足正则表达式的文件名if not file_list:continue# 处理图片数据counter = 0total = len(file_list)for file_name in file_list:counter += 1print("total: [%s / %s] current process image is: %s" % (counter, total, file_name))image_raw_data = gfile.FastGFile(file_name, 'rb').read()image = tf.image.decode_jpeg(image_raw_data)if image.dtype != tf.float32:image = tf.image.convert_image_dtype(image, dtype=tf.float32)image = tf.image.resize_images(image, [299, 299])image_value = sess.run(image)# 随机划分数据集chance = np.random.randint(100)if chance < validation_percentage:validation_images.append(image_value)validation_labels.append(current_label)elif chance < (test_percentage + validation_percentage):testing_images.append(image_value)testing_labels.append(current_label)else:training_images.append(image_value)training_labels.append(current_label)current_label += 1# 将训练数据随机打乱以获得更好的训练效果state = np.random.get_state()np.random.shuffle(training_images)np.random.set_state(state)np.random.shuffle(training_labels)return np.asarray([training_images, training_labels,validation_images, validation_labels,testing_images, testing_labels])def main():with tf.Session() as sess:processed_data = create_image_list(sess, TEST_PERCENTAGE, VALIDATION_PRECENTAGE)np.save(OUTPUT_DATA, processed_data)if __name__ == '__main__':main()

        接着运行上述数据预处理代码，后会在本地文件夹下新增数据处理的结果，接下来，使用迁移学习将Inception-V3应用于图像分类任务中，具体实现代码如下：

"""
迁移学习
"""
import numpy as np
import tensorflow as tffrom tensorflow.contrib import slim# 处理结束之后的数据文件
import tensorflow.contrib.slim.python.slim.nets.inception_v3  as inception_v3INPUT_DATA = "./flower_procssed_data.npy"
# 保存训练好的模型路径
TRAIN_FILE = "./savemodel"
# 谷歌提供训练好的模型文件地址
CKPT_FILE = "./inception_v3.ckpt"# 定义训练中使用的参数
LEARNING_RAGE = 0.0001
STEPS = 300
BATCH = 32
N_CLASS = 5CHECKPOINT_EXCLUDE_SCOPES = "InceptionV3/Logits,InceptionV3/AuxLogits"
TRAINABLE_SCOPES = "InceptionV3/Logits,InceptionV3/AuxLogits"def get_tuned_variables():"""获取所有需要从谷歌训练好的模型中加载的参数:return:"""exclusions = [scope.strip() for scope in CHECKPOINT_EXCLUDE_SCOPES.split(',')]variables_to_restore = []for var in slim.get_model_variables():excluded = Falsefor exclusion in exclusions:if var.op.name.startswith(exclusion):excluded = Truebreakif not excluded:variables_to_restore.append(var)return variables_to_restoredef get_trainable_variables():"""获取所有需要训练的变量列表:return:"""scopes = [scope.strip() for scope in TRAINABLE_SCOPES.split(",")]variables_to_train = []for scope in scopes:variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)variables_to_train.extend(variables)return variables_to_traindef main(argv=None):processed_data = np.load(INPUT_DATA, allow_pickle=True)training_images = processed_data[0]training_labels = processed_data[1]n_training_example = len(training_labels)validation_images = processed_data[2]validation_labels = processed_data[3]test_images = processed_data[4]test_labels = processed_data[5]print("%d training examples, %d validation examples and %d testing examples."% (n_training_example, len(validation_labels), len(test_labels)))# 定义图片的输入images = tf.placeholder(tf.float32, [None, 299, 299, 3], name='input_images')labels = tf.placeholder(tf.int64, [None], name='labels')# 定义Inception-v3模型with slim.arg_scope(inception_v3.inception_v3_arg_scope()):logits, _ = inception_v3.inception_v3(images, num_classes=N_CLASS)# 获取要训练的变量trainable_variables = get_trainable_variables()tf.losses.softmax_cross_entropy(tf.one_hot(labels, N_CLASS), logits, weights=1.0)# 定义训练过程train_step = tf.train.RMSPropOptimizer(LEARNING_RAGE).minimize(tf.losses.get_total_loss())# 计算正确率with tf.name_scope('evaluation'):correct_prediction = tf.equal(tf.argmax(logits, 1), labels)evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))# 定义加载模型的函数load_fn = slim.assign_from_checkpoint_fn(CKPT_FILE, get_tuned_variables(), ignore_missing_vars=True)# 定义保存新训练好的模型的函数saver = tf.train.Saver()with tf.Session() as sess:init = tf.global_variables_initializer()sess.run(init)# 加载谷歌已经训练好的模型print("loading tunned variables from %s" % CKPT_FILE)load_fn(sess)start = 0end = BATCHfor i in range(STEPS):# 运行训练过程sess.run(train_step, feed_dict={images: training_images[start:end],labels: training_labels[start:end]})# 输出日志if i % 30 == 0 or i + 1 == STEPS:saver.save(sess, TRAIN_FILE, global_step=i)validation_accuracy = sess.run(evaluation_step, feed_dict={images: validation_images,labels: validation_labels})print("step %d: validation accuracy = %1.f%%" % (i, validation_accuracy * 100.0))start = endif start == n_training_example:start = 0end = start + BATCHif end > n_training_example:end = n_training_example# 在最后的测试集上测试正确率test_accuracy = sess.run(evaluation_step, feed_dict={images: test_images,labels: test_labels})print("final test accuracy = %.1f%%" % (test_accuracy * 100))if __name__ == '__main__':tf.app.run()