本文代码可在https://github.com/TimeIvyace/TensorFlow_Migration-learning_Inception-v3.git中下载,需要同时下载数据集和Inception-v3模型。
注:代码中文件夹放置位置需要自行修改。
迁移学习是将一个数据集上训练好的网络模型快速转移到另外一个数据集上,可以保留训练好的模型中倒数第一层之前的所有参数,替换最后一层即可,在最后层之前的网络层称之为瓶颈层。
下面代码是使用TensorFlow将ImageNet上训练好的Inception-v3模型转移到另外一个图像分类数据集上。
数据集,Inception-v3模型可在此点击下载。数据集文件夹包含5个子文件,每一个子文件夹的名称为一种花的名称,代表了不同的类别。平均每一种花有734张图片,每一张图片都是RGB色彩模式,大小也不相同,程序将直接处理没有整理过的图像数据。
注意:计算交叉熵损失函数时,sparse_softmax_cross_entropy_with_logits直接用标签就可以计算交叉熵,而softmax_cross_entropy_with_logits是需要标签的one hot向量来参与计算,并且需要argmax得到标签最大值位置,如此代码中第58行所示。
# -*- coding: utf-8 -*-
import glob # 返回一个包含有匹配文件/目录的数组
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile
# inception-v3瓶颈层的节点个数
BOTTLENECT_TENSOR_SIZE = 2048
# 在谷歌提供的inception-v3模型中,瓶颈层结果的张量名称为'pool_3/_reshape:0'
# 可以使用tensor.name来获取张量名称
BOTTLENECT_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inception-v3模型文件目录
MODEL_DIR = '/tensorflow_google/inception_model'
# 下载的训练好的模型文件名
MODEL_FILE = 'tensorflow_inception_graph.pb'
# 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,下面定义文件存放地址
CACHE_DIR = '/tensorflow_google/bottleneck'
# 图片数据文件夹,子文件为类别
INPUT_DATA = '/tensorflow_google/flower_photos'
# 验证的数据百分比
VALIDATION_PRECENTAGE = 10
# 测试的数据百分比
TEST_PRECENTAGE = 10
# 定义神经网络的参数
LEARNING_RATE = 0.01
STEPS = 4000
BATCH = 100
# 从数据文件夹中读取所有的图片列表并按训练、验证、测试数据分开
# testing_percentage和validation_percentage指定测试和验证数据集的大小
def create_image_lists(testing_percentage, validation_percentage):
# 得到的图片放到result字典中,key为类别名称,value为类别下的各个图片(也是字典)
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)]
# sub_dirs中第一个目录是当前目录,即flower_photos,不用考虑
is_root_dir = True
for sub_dir in sub_dirs:
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', 'jpeg', 'JPG', 'JPEG']
file_list = []
# 获取当前文件名
dir_name = os.path.basename(sub_dir)
for extension in extensions:
# 将分离的各部分组成一个路径名,如/flower_photos/roses/*.JPEG
file_glob = os.path.join(INPUT_DATA, dir_name, '*.'+extension)
# glob.glob()返回的是所有路径下的符合条件的文件名的列表
file_list.extend(glob.glob(file_glob))
if not file_list: continue
# 通过目录名获取类别的名称(全部小写)
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) #获取当前文件名
# 随机将数据分到训练数据集、测试数据集以及验证数据集
chance = np.random.randint(100) #随机返回一个整数
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)
# 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}
return result
# 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为所有图片信息,image_dir给出根目录,label_name为类别名称,index为图片编号,category指定图片是在哪个训练集
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的全部图片信息
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_path = os.path.join(image_dir, sub_dir, base_name)
return full_path
# 通过类别名称、所属数据集和图片编号经过inception-v3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'
# 使用加载的训练好的网络处理一张图片,得到这个图片的特征向量
def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):
# 将当前图片作为输入,计算瓶颈张量的值
# 这个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一维数组
bottleneck_values = np.squeeze(bottleneck_values) #从数组的形状中删除单维条目
return bottleneck_values
# 获取一张图片经过inception-v3模型处理之后的特征向量
# 先寻找已经计算并且保存的向量,若找不到则计算然后保存到文件
def get_or_create_bottleneck(sess, image_lists, label_name, index, category,
jpeg_data_tensor, bottleneck_tensor):
# 获取一张图片对应的特征向量文件路径
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) #若不存在则创建
bottleneck_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计算,并存入文件
if not os.path.exists(bottleneck_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,对图片的读取
image_data = gfile.FastGFile(image_path, 'rb').read()
# 通过inception-v3计算特征向量
bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)
# 将计算得到的特征向量存入文件,join()连接字符串
bottleneck_string = ','.join(str(x) for x in bottleneck_values)
with open(bottleneck_path, 'w') as bottleneck_file: #打开文件并写入
bottleneck_file.write(bottleneck_string)
else:
# 直接从文件中获取图片相应的特征向量
with open(bottleneck_path, 'r') as bottleneck_file:
bottleneck_string = bottleneck_file.read()
bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
# 返回特征向量
return bottleneck_values
# 随机选取一个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) # 返回指定递增基数集合中的一个随机数,基数缺省值为1,随机类别号
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,在最终测试的时候在所有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] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
return bottlenecks, ground_truths
def main(_):
# 读取所有图片
image_lists = create_image_lists(TEST_PRECENTAGE, VALIDATION_PRECENTAGE)
# image_lists.keys()为dict_keys(['roses', 'sunflowers', 'daisy', 'dandelion', 'tulips'])
n_classes = len(image_lists.keys()) # 类别数
# 读取已经训练好的inception-v3模型,谷歌训练好的模型保存在了GraphDef Protocol Buffer中
# 里面保存了每一个节点取值的计算方法以及变量的取值
# 对模型的读取,二进制
with gfile.FastGFile(os.path.join(MODEL_DIR, MODEL_FILE), 'rb') as f:
# 新建GraphDef文件,用于临时载入模型中的图
graph_def = tf.GraphDef()
# 加载模型中的图
graph_def.ParseFromString(f.read())
# 加载读取的inception模型,并返回数据输出所对应的张量以及计算瓶颈层结果所对应的张量
# 从图上读取张量,同时把图设为默认图
# Tensor("import/pool_3/_reshape:0", shape=(1, 2048), dtype=float32)
# Tensor("import/DecodeJpeg/contents:0", shape=(), dtype=string)
bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, return_elements=[BOTTLENECT_TENSOR_NAME,
JPEG_DATA_TENSOR_NAME])
# 定义新的神经网络输入,这个输入就是新的图片经过inception模型前向传播达到瓶颈层的节点取值,None为了batch服务
bottleneck_input = tf.placeholder(tf.float32, [None, BOTTLENECT_TENSOR_SIZE],
name='BottleneckInputPlaceholder')
# 定义新的标准答案
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
with tf.name_scope('final_training_ops'):
weights = tf.Variable(tf.truncated_normal([BOTTLENECT_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
# tf.nn.softmax中dim默认为-1,即tf.nn.softmax会以最后一个维度作为一维向量计算softmax
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))
# 平均错误率,cast将bool值转成float
evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
with tf.Session() as sess:
init = tf.initialize_all_variables()
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 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__':
tf.app.run()
输出结果:
Step 0 :Validation accuracy on random sampled 100 examples = 45.0%
Step 100 :Validation accuracy on random sampled 100 examples = 78.0%
Step 200 :Validation accuracy on random sampled 100 examples = 85.0%
Step 300 :Validation accuracy on random sampled 100 examples = 82.0%
Step 400 :Validation accuracy on random sampled 100 examples = 84.0%
Step 500 :Validation accuracy on random sampled 100 examples = 84.0%
Step 600 :Validation accuracy on random sampled 100 examples = 84.0%
Step 700 :Validation accuracy on random sampled 100 examples = 87.0%
Step 800 :Validation accuracy on random sampled 100 examples = 90.0%
Step 900 :Validation accuracy on random sampled 100 examples = 85.0%
Step 1000 :Validation accuracy on random sampled 100 examples = 84.0%
Step 1100 :Validation accuracy on random sampled 100 examples = 88.0%
Step 1200 :Validation accuracy on random sampled 100 examples = 81.0%
Step 1300 :Validation accuracy on random sampled 100 examples = 85.0%
Step 1400 :Validation accuracy on random sampled 100 examples = 82.0%
Step 1500 :Validation accuracy on random sampled 100 examples = 82.0%
Step 1600 :Validation accuracy on random sampled 100 examples = 90.0%
Step 1700 :Validation accuracy on random sampled 100 examples = 90.0%
Step 1800 :Validation accuracy on random sampled 100 examples = 84.0%
Step 1900 :Validation accuracy on random sampled 100 examples = 88.0%
Step 2000 :Validation accuracy on random sampled 100 examples = 84.0%
Step 2100 :Validation accuracy on random sampled 100 examples = 88.0%
Step 2200 :Validation accuracy on random sampled 100 examples = 81.0%
Step 2300 :Validation accuracy on random sampled 100 examples = 92.0%
Step 2400 :Validation accuracy on random sampled 100 examples = 87.0%
Step 2500 :Validation accuracy on random sampled 100 examples = 80.0%
Step 2600 :Validation accuracy on random sampled 100 examples = 89.0%
Step 2700 :Validation accuracy on random sampled 100 examples = 87.0%
Step 2800 :Validation accuracy on random sampled 100 examples = 91.0%
Step 2900 :Validation accuracy on random sampled 100 examples = 90.0%
Step 3000 :Validation accuracy on random sampled 100 examples = 93.0%
Step 3100 :Validation accuracy on random sampled 100 examples = 88.0%
Step 3200 :Validation accuracy on random sampled 100 examples = 85.0%
Step 3300 :Validation accuracy on random sampled 100 examples = 91.0%
Step 3400 :Validation accuracy on random sampled 100 examples = 85.0%
Step 3500 :Validation accuracy on random sampled 100 examples = 87.0%
Step 3600 :Validation accuracy on random sampled 100 examples = 89.0%
Step 3700 :Validation accuracy on random sampled 100 examples = 88.0%
Step 3800 :Validation accuracy on random sampled 100 examples = 88.0%
Step 3900 :Validation accuracy on random sampled 100 examples = 85.0%
Step 3999 :Validation accuracy on random sampled 100 examples = 89.0%
Final test accuracy = 90.8%