简介
- Inceptionv3论文地址：https://arxiv.org/abs/1512.00567
- Inceptionv3 tensorflow 实现代码:https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v3.py
- 谷歌训练好的InceptionV3模型：https://storage.googleapis.com/download.tensorflow.org/models/inception_dec_2015.zip
- 所谓迁移训练就是将已经训练好的某个模型做简单的调整使其能够适应一个全新的问题（没有机器的痛）。我们保留InceptionV3当中的图像特征提取的部分，对最后一层的fc进行再训练，也就是对分类器做再训练。我们将fc的前一层称作 瓶颈层。
- 什么时候使用迁移训练？
- 数据量小
- 没有机器
CODING
- 参数定义
# 数据参数
MODEL_DIR = 'inception_dec_2015' # inception-v3模型的文件夹
MODEL_FILE = 'tensorflow_inception_graph.pb' # inception-v3模型文件名
CACHE_DIR = 'datasets/tmp/bottleneck' # 图像的特征向量保存地址
INPUT_DATA = 'datasets/train' # 图片数据文件夹
VALIDATION_PERCENTAGE = 10 # 验证数据的百分比
TEST_PERCENTAGE = 10 # 测试数据的百分比
# inception-v3模型参数
BOTTLENECK_TENSOR_SIZE = 2048 # inception-v3模型瓶颈层的节点个数
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
# 神经网络的训练参数
LEARNING_RATE = 0.01 #学习率
STEPS = 100000
BATCH = 50
CHECKPOINT_EVERY = 100 #训练多久保存一次模型
NUM_CHECKPOINTS = 5 #最多保存的模型
- 数据处理：获取图片位置，图片输入获取瓶颈层向量
# 从数据文件夹中读取所有的图片列表并按训练、验证、测试分开
def create_image_lists(validation_percentage, test_percentage):
result = {} # 保存所有图像。key为类别名称。value也是字典，存储了所有的图片名称
sub_dirs = [x[0] for x in os.walk(INPUT_DATA)] # 获取所有子目录
is_root_dir = True # 第一个目录为当前目录，需要忽略
# 分别对每个子目录进行操作
for sub_dir in sub_dirs:
if is_root_dir:
is_root_dir = False
continue
# 获取当前目录下的所有有效图片
extensions = {'jpg', 'jpeg', 'JPG', 'JPEG'}
file_list = [] # 存储所有图像
dir_name = os.path.basename(sub_dir) # 获取路径的最后一个目录名字
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
# 将当前类别的图片随机分为训练数据集、测试数据集、验证数据集
label_name = dir_name.lower() # 通过目录名获取类别的名称
training_images = []
testing_images = []
validation_images = []
for file_name in file_list:
base_name = os.path.basename(file_name) # 获取该图片的名称
chance = np.random.randint(100) # 随机产生100个数代表百分比
if chance < validation_percentage:
validation_images.append(base_name)
elif chance < (validation_percentage + test_percentage):
testing_images.append(base_name)
else:
training_images.append(base_name)
# 将当前类别的数据集放入结果字典
result[label_name] = {
'dir': dir_name,
'training': training_images,
'testing': testing_images,
'validation': validation_images
}
# 返回整理好的所有数据
return result
# 通过类别名称、所属数据集、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] # 获取给定类别中的所有图片
category_list = label_lists[category] # 根据所属数据集的名称获取该集合中的全部图片
mod_index = index % len(category_list) # 规范图片的索引
base_name = category_list[mod_index] # 获取图片的文件名
sub_dir = label_lists['dir'] # 获取当前类别的目录名
full_path = os.path.join(image_dir, sub_dir, base_name) # 图片的绝对路径
return full_path
# 通过类别名称、所属数据集、图片编号获取特征向量值的地址
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inception-v3处理图片获取特征向量
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) # 将四维数组压缩成一维数组
return bottleneck_values
# 获取一张图片经过inception-v3模型处理后的特征向量
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的路径
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)
bottleneck_path = get_bottleneck_path(image_lists, label_name, index,
category)
# 如果该特征向量文件不存在，则通过inception-v3模型计算并保存
if not os.path.exists(bottleneck_path):
image_path = get_image_path(image_lists, INPUT_DATA, label_name, index,
category) # 获取图片原始路径
image_data = gfile.FastGFile(image_path, 'rb').read() # 获取图片内容
bottleneck_values = run_bottleneck_on_image(
sess, image_data, jpeg_data_tensor,
bottleneck_tensor) # 通过inception-v3计算特征向量
# 将特征向量存入文件
bottleneck_string = ','.join(str(x) for x in bottleneck_values)
with open(bottleneck_path, 'w') as bottleneck_file:
bottleneck_file.write(bottleneck_string)
else:
# 否则直接从文件中获取图片的特征向量
with open(bottleneck_path, 'r') as bottleneck_file:
bottleneck_string = bottleneck_file.read()
bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
# 返回得到的特征向量
return bottleneck_values
# 随机获取一个batch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]
image_index = random.randrange(65535)
if not image_lists[label_name][category]:
continue
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
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor,
bottleneck_tensor):
bottlenecks = []
ground_truths = []
label_name_list = list(image_lists.keys())
# 枚举所有的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'
if not image_lists[label_name][category]:
continue
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_truths
- 模型迁移，构建最后一层fc分类器
def main(_):
# 读取所有的图片
image_lists = create_image_lists(VALIDATION_PERCENTAGE, TEST_PERCENTAGE)
n_classes = len(image_lists.keys())
with tf.Graph().as_default() as graph:
# 读取训练好的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())
# 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
])
# 定义新的神经网络输入
bottleneck_input = tf.placeholder(
tf.float32, [None, BOTTLENECK_TENSOR_SIZE],
name='BottleneckInputPlaceholder')
# 定义新的标准答案输入
ground_truth_input = tf.placeholder(
tf.float32, [None, n_classes], name='GroundTruthInput')
# 定义一层全连接层解决新的图片分类问题
with tf.name_scope('final_training_ops'):
weights = tf.Variable(
tf.truncated_normal(
[BOTTLENECK_TENSOR_SIZE, n_classes], stddev=0.1))
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))
# 训练过程
with tf.Session(graph=graph) as sess:
init = tf.global_variables_initializer().run()
# 模型和摘要的保存目录
import time
timestamp = str(int(time.time()))
out_dir = os.path.abspath(
os.path.join(os.path.curdir, 'runs', timestamp))
print('\nWriting to {}\n'.format(out_dir))
# 损失值和正确率的摘要
loss_summary = tf.summary.scalar('loss', cross_entropy_mean)
acc_summary = tf.summary.scalar('accuracy', evaluation_step)
# 训练摘要
train_summary_op = tf.summary.merge([loss_summary, acc_summary])
train_summary_dir = os.path.join(out_dir, 'summaries', 'train')
train_summary_writer = tf.summary.FileWriter(train_summary_dir,
sess.graph)
# 开发摘要
dev_summary_op = tf.summary.merge([loss_summary, acc_summary])
dev_summary_dir = os.path.join(out_dir, 'summaries', 'dev')
dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph)
# 保存检查点
checkpoint_dir = os.path.abspath(os.path.join(out_dir, 'checkpoints'))
checkpoint_prefix = os.path.join(checkpoint_dir, 'model')
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
saver = tf.train.Saver(
tf.global_variables(), max_to_keep=NUM_CHECKPOINTS)
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)
_, train_summaries = sess.run(
[train_step, train_summary_op],
feed_dict={
bottleneck_input: train_bottlenecks,
ground_truth_input: train_ground_truth
})
# 保存每步的摘要
train_summary_writer.add_summary(train_summaries, i)
# 在验证集上测试正确率
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, dev_summaries = sess.run(
[evaluation_step, dev_summary_op],
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))
# 每隔checkpoint_every保存一次模型和测试摘要
if i % CHECKPOINT_EVERY == 0:
dev_summary_writer.add_summary(dev_summaries, i)
path = saver.save(sess, checkpoint_prefix, global_step=i)
print('Saved model checkpoint to {}\n'.format(path))
# 最后在测试集上测试正确率
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))
# 保存标签
output_labels = os.path.join(out_dir, 'labels.txt')
with tf.gfile.FastGFile(output_labels, 'w') as f:
keys = list(image_lists.keys())
for i in range(len(keys)):
keys[i] = '%2d -> %s' % (i, keys[i])
f.write('\n'.join(keys) + '\n')
