所谓迁移学习,就是将一个问题上训练好的模型通过简单的调整使其适用于一个新的问题。针对inceptionV3模型,我们可以保留训练好的inceptionV3模型中所有卷积层的参数,只是替换最后一个全连接层。在最后这一层全连接层之前的网络层称之为瓶颈层。
将新的图像通过训练好的卷积神经网络直到瓶颈层的过程,可以看成是对图像进行特征提取的过程。在训练好的inceptionV3模型中,因为将瓶颈层的输出再通过一个单层的全连接层神经网络可以很好地区分1000种类别的图像,所以有理由认为瓶颈层输出的节点向量可以被作为任何图像的一个更加精简且表达能力更强的特征向量。
一般来说,在数据量足够的情况下,迁移学习的效果不如完全重新训练。
案例来源于 《TensorFlow实战Google深度学习框架》。在源代码基础上,做了一些修改后,能够保存完整的pb文件。
一、资料下载
1、谷歌提供的训练好的Inception-v3模型 https://storage.googleapis.com/download.tensorflow.org/models/inception_dec_2015.zip
2、案例使用的数据集 http://download.tensorflow.org/example_images/flower_photos.tgz
二、代码
1、训练代码
# -*- coding: utf-8 -*- """ 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 from tensorflow.python.framework import graph_util # 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 = './inceptionV3' # 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 = './inceptionV3/tmp/bottleneck' # 图片数据的文件夹 INPUT_DATA = './inceptionV3/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 = 1000 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 create_inception_graph(): with tf.Graph().as_default() as graph: model_filename = os.path.join( MODEL_DIR, MODEL_FILE) with gfile.FastGFile(model_filename, 'rb') as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, name='', return_elements=[ BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME]) return graph, bottleneck_tensor, jpeg_data_tensor def add_final_training_ops(class_count, bottleneck_tensor): # 输入 bottleneck_input = tf.placeholder_with_default(bottleneck_tensor, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder') ground_truth_input = tf.placeholder(tf.float32, [None, class_count], name='GroundTruthInput') # 全连接层 with tf.name_scope('output'): weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, class_count], stddev=0.001)) biases = tf.Variable(tf.zeros([class_count])) logits = tf.matmul(bottleneck_input, weights) + biases final_tensor = tf.nn.softmax(logits, name='prob') # 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) 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)) return (train_step,evaluation_step, cross_entropy_mean, bottleneck_input, ground_truth_input) def train(): image_lists = create_image_lists(TEST_PERCENTACE, VALIDATION_PERCENTAGE) n_classes = len(image_lists.keys()) print('n_classes:',n_classes) graph, bottleneck_tensor, jpeg_data_tensor=create_inception_graph() print(bottleneck_tensor.graph is tf.get_default_graph()) with tf.Session(graph=graph) as sess: train_step,evaluation_step,cross_entropy_mean,bottleneck_input,ground_truth_input=add_final_training_ops(n_classes,bottleneck_tensor) # 初始化参数 init = tf.global_variables_initializer() 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}) # 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)) constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["output/prob"]) with tf.gfile.FastGFile("./pbtxt/nn.pb", mode='wb') as f: f.write(constant_graph.SerializeToString())
2、预测代码
import cv2 def predict(): strings = ['daisy', 'dandelion', 'roses', 'sunflowers', 'tulips'] def id_to_string(node_id): return strings[node_id] with tf.gfile.FastGFile('./pbtxt/nn.pb', 'rb') as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) tf.import_graph_def(graph_def, name='') with tf.Session() as sess: softmax_tensor = sess.graph.get_tensor_by_name('output/prob:0') # 遍历目录 for root, dirs, files in os.walk('./inceptionV3/predict_images/'): for file in files: # 载入图片 image_data = tf.gfile.FastGFile(os.path.join(root, file), 'rb').read() predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': image_data}) # 图片格式是jpg格式 predictions = np.squeeze(predictions) # 把结果转为1维数据 # 打印图片路径及名称 image_path = os.path.join(root, file) print(image_path) # 排序 top_k = predictions.argsort()[::-1] print(top_k) for node_id in top_k: # 获取分类名称 human_string = id_to_string(node_id) # 获取该分类的置信度 score = predictions[node_id] print('%s (score = %.5f)' % (human_string, score)) print() img = cv2.imread(image_path) cv2.imshow('image', img) cv2.waitKey(0) cv2.destroyAllWindows()
三、参考博文
1、http://blog.csdn.net/tz_zs/article/details/77728391?ABstrategy=codes_snippets_optimize_v3
2、https://blog.csdn.net/l18930738887/article/details/72812689