理解深层神经网络中的迁移学习及TensorFlow实现

来源:互联网 发布:默沙东临床数据管理员 编辑:程序博客网 时间:2024/06/06 00:28

什么是迁移学习

在深度学习中,所谓的迁移学习是将一个问题A上训练好的模型通过简单的调整使其适应一个新的问题B。在实际使用中,往往是完成问题A的训练出的模型有更完善的数据,而问题B的数据量偏小。而调整的过程根据现实情况决定,可以选择保留前几层卷积层的权重,以保留低级特征的提取;也可以保留全部的模型,只根据新的任务改变其fc层。

迁移学习的作用

那么对于不同的任务,为什么不同的模型间可以做迁移呢?上面提到了,被迁移的模型往往是使用大量样本训练出来的,比如Google提供的Inception V3网络模型使用ImageNet数据集训练,而ImageNet中有120万标注图片,然后在实际应用中,很难收集到如此多的样本数据。而且收集的过程需要消耗大量的人力无力(其实深度学习解决实际问题时,最好费时间的往往不是训练的过程,而是数据标记的过程),所以一般情况下来说,问题B的数据量是较少的。
所以,同样一个模型在使用大样本很好的解决了问题A,那么有理由相信该模型中训练处的权重参数能够能够很好的完成特征提取任务(最起码前几层是这样),所以既然已经有了这样一个模型,那就拿过来用吧。
所以迁移学习具有如下优势:
更短的训练时间,更快的收敛速度,更精准的权重参数。
但是一般情况下如果任务B的数据量是足够的,那么迁移来的模型效果会不如训练的到,但是此时起码可以将底层的权重参数作为初始值来重新训练。

TensorFlow实现Inception V3迁移学习

下面的例子中使用Google提供的Inception V3模型完成花的分类任务,迁移的过程保留了Inception V3的全部卷积层,只修改了最后的全连接层以适应新的分类任务。

import globimport os.pathimport randomimport numpy as npimport tensorflow as tffrom tensorflow.python.platform import gfile#模型和样本路径的设置#inception-V3瓶颈层节点个数BOTTLENECK_TENSOR_SIZE = 2048#瓶颈层tenbsor nameBOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'#图像输入tensor nameJPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'# v3 pathMODEL_DIR = './datasets/inception_dec_2015'# v3 modefileMODEL_FILE= 'tensorflow_inception_graph.pb'#特征向量 save pathCACHE_DIR = './datasets/bottleneck'#数据pathINPUT_DATA = './datasets/flower_photos'#验证数据 percentageVALIDATION_PERCENTAGE = 10#测试数据 percentageTEST_PERCENTAGE = 10#神经网络参数的设置LEARNING_RATE = 0.01STEPS = 4000BATCH = 100#把样本中所有的图片列表并按训练、验证、测试数据分开def create_image_lists(testing_percentage, validation_percentage):    result = {}    sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]    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)        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        label_name = dir_name.lower()        # 初始化        training_images = []        testing_images = []        validation_images = []        for file_name in file_list:            base_name = os.path.basename(file_name)            # 随机划分数据            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#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):    label_lists = image_lists[label_name]    category_list = label_lists[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#函数获取Inception-v3模型处理之后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'#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_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):    bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})    bottleneck_values = np.squeeze(bottleneck_values)    return bottleneck_values#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, bottleneck_tensor):    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)    bottleneck_path = get_bottleneck_path(image_lists, label_name, index, category)    if not os.path.exists(bottleneck_path):        image_path = get_image_path(image_lists, INPUT_DATA, label_name, index, category)        image_data = gfile.FastGFile(image_path, 'rb').read()        bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)        bottleneck_string = ','.join(str(x) for x in bottleneck_values)        with open(bottleneck_path, 'w') as bottleneck_file:            bottleneck_file.write(bottleneck_string)    else:        with open(bottleneck_path, 'r') as bottleneck_file:            bottleneck_string = bottleneck_file.read()        bottleneck_values = [float(x) for x in bottleneck_string.split(',')]    return bottleneck_values#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):        label_index = random.randrange(n_classes)        label_name = list(image_lists.keys())[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] = 1.0        bottlenecks.append(bottleneck)        ground_truths.append(ground_truth)    return bottlenecks, ground_truths#获取全部的测试数据,并计算正确率def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):    bottlenecks = []    ground_truths = []    label_name_list = list(image_lists.keys())    for label_index, label_name in enumerate(label_name_list):        category = 'testing'        for index, unused_base_name in enumerate(image_lists[label_name][category]):            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_truthsdef main():    image_lists = create_image_lists(TEST_PERCENTAGE, VALIDATION_PERCENTAGE)    n_classes = len(image_lists.keys())    # 读取已经训练好的Inception-v3模型。    with gfile.FastGFile(os.path.join(MODEL_DIR, MODEL_FILE), 'rb') as f:        graph_def = tf.GraphDef()        graph_def.ParseFromString(f.read())    bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(        graph_def, return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME])    # 定义新的神经网络输入    bottleneck_input = tf.placeholder(tf.float32, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')    ground_truth_input = tf.placeholder(tf.float32, [None, n_classes], name='GroundTruthInput')    # 定义一层全链接层    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)    # 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):            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})            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))        # 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 = %.1f%%' % (test_accuracy * 100))if __name__ == '__main__':    main()

输出结果:
.
.
Step 1000: Validation accuracy on random sampled 100 examples = 92.0%
.
.
Step 2700: Validation accuracy on random sampled 100 examples = 94.0%
.
.
Step 3999: Validation accuracy on random sampled 100 examples = 94.0%
Final test accuracy = 92.7%

从结果可以看到,模型在很短的时间内即达到收敛并有着不错的准确率。最后点击这里下载整个工程,由于上传大小的限制,工程中的模型与数据集需要重新下载,路径下文件夹中已提供了下载方式。

原创粉丝点击