黑马程序员技术交流社区
标题: 【上海校区】使用自己的数据集训练GoogLenet InceptionNet V1 V2 ... [打印本页]
作者: 不二晨 时间: 2018-9-9 10:53
标题: 【上海校区】使用自己的数据集训练GoogLenet InceptionNet V1 V2 ...
InceptionNet,是 Google 的研究人员提出的网络结构(所以也叫做`GoogLeNet`),在当时取得了非常大的影响,因为网络的结构变得前所未有,它颠覆了大家对卷积网络的串联的印象和固定做法,采用了一种非常有效的 inception 模块,得到了比 VGG 更深的网络结构,但是却比 VGG 的参数更少,因为其去掉了后面的全连接层,所以参数大大减少,同时有了很高的计算效率。
1、googlenet 的网络示意图:
2、Inception 模块 在上面的网络中,我们看到了多个四个并行卷积的层,这些四个卷积并行的层就是 inception 模块,可视化如下
一个 inception 模块的四个并行线路如下:
1.一个 1 x 1 的卷积,一个小的感受野进行卷积提取特征
2.一个 1 x 1 的卷积加上一个 3 x 3 的卷积,1 x 1 的卷积降低输入的特征通道,减少参数计算量,然后接一个 3 x 3 的卷积做一个较大感受野的卷积
3.一个 1 x 1 的卷积加上一个 5 x 5 的卷积,作用和第二个一样
4.一个 3 x 3 的最大池化加上 1 x 1 的卷积,最大池化改变输入的特征排列,1 x 1 的卷积进行特征提取
最后将四个并行线路得到的特征在通道这个维度上拼接在一起
二、项目文件结构说明tensorflow_models_nets:
|__dataset #数据文件
|__record #里面存放record文件
|__train #train原始图片
|__val #val原始图片
|__models #保存训练的模型
|__slim #这个是拷贝自slim模块:https://github.com/tensorflow/models/tree/master/research/slim
|__test_image #存放测试的图片
|__create_labels_files.py #制作trian和val TXT的文件
|__create_tf_record.py #制作tfrecord文件
|__inception_v1_train_val.py #inception V1的训练文件
|__inception_v3_train_val.py # inception V3训练文件
|__predict.py # 模型预测文件
三、训练模型过程1、训练和测试的图片数据集 下面是我下载的数据集,共有五类图片,分别是:flower、guitar、animal、houses和plane,每组数据集大概有800张左右。为了照顾网友,下面的数据集,都已经放在Github项目的文件夹dataset上了,记得给个“star”哈
animal:http://www.robots.ox.ac.uk/~vgg/data/pets/
flower:http://www.robots.ox.ac.uk/~vgg/data/flowers/
plane:http://www.robots.ox.ac.uk/~vgg/ ... /airplanes_side.tar
house:http://www.robots.ox.ac.uk/~vgg/data/houses/houses.tar
guitar:http://www.robots.ox.ac.uk/~vgg/data/guitars/guitars.tar
下载图片数据集后,需要划分为train和val数据集,前者用于训练模型的数据,后者主要用于验证模型。这里提供一个create_labels_files.py脚本,可以直接生成训练train和验证val的数据集txt文件。
#-*-coding:utf-8-*-
"""
@Project: googlenet_classification
@File : create_labels_files.py
@Author : panjq
@E-mail : pan_jinquan@163.com
@Date : 2018-08-11 10:15:28
"""
import os
import os.path
def write_txt(content, filename, mode='w'):
"""保存txt数据
:param content:需要保存的数据,type->list
:param filename:文件名
:param mode:读写模式:'w' or 'a'
:return: void
"""
with open(filename, mode) as f:
for line in content:
str_line = ""
for col, data in enumerate(line):
if not col == len(line) - 1:
# 以空格作为分隔符
str_line = str_line + str(data) + " "
else:
# 每行最后一个数据用换行符“\n”
str_line = str_line + str(data) + "\n"
f.write(str_line)
def get_files_list(dir):
'''
实现遍历dir目录下,所有文件(包含子文件夹的文件)
:param dir:指定文件夹目录
:return:包含所有文件的列表->list
'''
# parent:父目录, filenames:该目录下所有文件夹,filenames:该目录下的文件名
files_list = []
for parent, dirnames, filenames in os.walk(dir):
for filename in filenames:
# print("parent is: " + parent)
# print("filename is: " + filename)
# print(os.path.join(parent, filename)) # 输出rootdir路径下所有文件(包含子文件)信息
curr_file=parent.split(os.sep)[-1]
if curr_file=='flower':
labels=0
elif curr_file=='guitar':
labels=1
elif curr_file=='animal':
labels=2
elif curr_file=='houses':
labels=3
elif curr_file=='plane':
labels=4
files_list.append([os.path.join(curr_file, filename),labels])
return files_list
if __name__ == '__main__':
train_dir = 'dataset/train'
train_txt='dataset/train.txt'
train_data = get_files_list(train_dir)
write_txt(train_data,train_txt,mode='w')
val_dir = 'dataset/val'
val_txt='dataset/val.txt'
val_data = get_files_list(val_dir)
write_txt(val_data,val_txt,mode='w')
注意,上面Python代码,已经定义每组图片对应的标签labels:
flower ->labels=0
guitar ->labels=1
animal ->labels=2
houses ->labels=3
plane ->labels=4
2、制作tfrecords数据格式 有了 train.txt和val.txt数据集,我们就可以制作train.tfrecords和val.tfrecords文件了,项目提供一个用于制作tfrecords数据格式的Python文件:create_tf_record.py,鄙人已经把代码放在另一篇博客:《Tensorflow生成自己的图片数据集TFrecords》https://blog.csdn.net/guyuealian/article/details/80857228 ,代码有详细注释了,所以这里不贴出来了.
注意:
(1)create_tf_record.py将train和val数据分别保存为单个record文件,当图片数据很多时候,会导致单个record文件超级巨大的情况,解决方法就是,将数据分成多个record文件保存,读取时,只需要将多个record文件的路径列表交给“tf.train.string_input_producer”即可。
(2)如何将数据保存为多个record文件呢?请参考鄙人的博客:《Tensorflow生成自己的图片数据集TFrecords》https://blog.csdn.net/guyuealian/article/details/80857228
create_tf_record.py提供几个重要的函数:
- create_records():用于制作records数据的函数,
- read_records():用于读取records数据的函数,
- get_batch_images():用于生成批训练数据的函数
- get_example_nums:统计tf_records图像的个数(example个数)
- disp_records(): 解析record文件,并显示图片,主要用于验证生成record文件是否成功
3、GoogLenet网络结构 GoogLenet InceptionNet有很多的变体, 比如`InceptionV1`,`V2`, `V3`, `V4`版本,网上已经有很多使用TensorFlow实现的,但很尴尬的是,代码基本上都是你抄我,我复制你。原型代码貌似都是来自黄文坚著作《TensorFlow实战》-第六章的《6.3TensorFlow 实现 GooglelnceptionNet》。要想改动为实际可用的、可训练、可测试的图像分类模型,还是要花很大的力气的。
本人一开始就想把黄文坚第六章的《6.3TensorFlow 实现 GooglelnceptionNet》的源码封装成可训练的过程,但训练过程发现模型一直不收敛,识别率一直在0.2-0.3%浮动,不知道是我参数设置问题,还是我训练代码出错了,反正就是不行!!!
官网TensorFlow已经提供了使用TF-slim实现的InceptionNet V1,V2,V3,V4模型。TF-Slim是tensorflow中定义、训练和评估复杂模型的轻量级库。tf-slim中的组件可以轻易地和原生tensorflow框架以及例如tf.contrib.learn这样的框架进行整合。
1、官网模型地址:https://github.com/tensorflow/models/tree/master/research/slim/nets
2、slim/nets下的模型都是用TF-slim实现的网络结构,关系TF-slim的用法,可参考:
《tensorflow中slim模块api介绍》:https://blog.csdn.net/guvcolie/article/details/77686555
4、训练方法实现过程 inception_v3要求训练数据height, width = 299, 299(亲测用224也是可以的),项目使用create_tf_record.py制作了训练train299.tfrecords和验证val299.tfrecords数据,下面是inception_v3_train_val.py文件代码说明:
from datetime import datetime
import slim.nets.inception_v3 as inception_v3
from create_tf_record import *
import tensorflow.contrib.slim as slim
labels_nums = 5 # 类别个数
batch_size = 16 #
resize_height = 299 # 指定存储图片高度
resize_width = 299 # 指定存储图片宽度
depths = 3
data_shape = [batch_size, resize_height, resize_width, depths]
下面就一步一步实现训练过程:
(1)先占几个坑用来填数据:
tf.placeholder()是TensorFlow的占位符节点,由placeholder方法创建,其也是一种常量,但是由用户在调用run方法时传递的,可以简单理解为形参,用于定义过程,在执行的时候再赋具体的值。利用tf.placeholder(),代码就可以很方便的实现:is_training=True时,填充train数据进行训练过程,is_training=False时,填充val数据进行验证过程
# 定义input_images为图片数据
input_images = tf.placeholder(dtype=tf.float32, shape=[None, resize_height, resize_width, depths], name='input')
# 定义input_labels为labels数据
# input_labels = tf.placeholder(dtype=tf.int32, shape=[None], name='label')
input_labels = tf.placeholder(dtype=tf.int32, shape=[None, labels_nums], name='label')
# 定义dropout的概率
keep_prob = tf.placeholder(tf.float32,name='keep_prob')
is_training = tf.placeholder(tf.bool, name='is_training')
(2)定义一个训练函数train():
def train(train_record_file,
train_log_step,
train_param,
val_record_file,
val_log_step,
labels_nums,
data_shape,
snapshot,
snapshot_prefix):
'''
:param train_record_file: 训练的tfrecord文件
:param train_log_step: 显示训练过程log信息间隔
:param train_param: train参数
:param val_record_file: 验证的tfrecord文件
:param val_log_step: 显示验证过程log信息间隔
:param val_param: val参数
:param labels_nums: labels数
:param data_shape: 输入数据shape
:param snapshot: 保存模型间隔
:param snapshot_prefix: 保存模型文件的前缀名
:return:
'''
(3)读取训练和验证数据
create_tf_record.py文件提供了读取records数据的函数read_records(),以及获得批训练的数据get_batch_images()。一般而言,训练时需要打乱输入数据,因此函数get_batch_images()提供了shuffle=True参数可以打乱输入数据,但该函数仅仅对一个批次的数据进行打乱的,并未达到随机打乱所有训练数据的要求,鄙人建议是:在制作records数据时就打乱训练数据,即设置create_records()函数的参数shuffle=True,而对val数据,不要求打乱数据。
[base_lr,max_steps]=train_param
[batch_size,resize_height,resize_width,depths]=data_shape
# 获得训练和测试的样本数
train_nums=get_example_nums(train_record_file)
val_nums=get_example_nums(val_record_file)
print('train nums:%d,val nums:%d'%(train_nums,val_nums))
# 从record中读取图片和labels数据
# train数据,训练数据一般要求打乱顺序shuffle=True
train_images, train_labels = read_records(train_record_file, resize_height, resize_width, type='normalization')
train_images_batch, train_labels_batch = get_batch_images(train_images, train_labels,
batch_size=batch_size, labels_nums=labels_nums,
one_hot=True, shuffle=True)
# val数据,验证数据可以不需要打乱数据
val_images, val_labels = read_records(val_record_file, resize_height, resize_width, type='normalization')
val_images_batch, val_labels_batch = get_batch_images(val_images, val_labels,
batch_size=batch_size, labels_nums=labels_nums,
one_hot=True, shuffle=False)
(4)定义inception_v3网络模型
TensorFlow的inception_v3是用tf.contrib.slim写的。slim.arg_scope用于为tensorflow里的layer函数提供默认参数值,以使构建模型的代码更加紧凑苗条(slim)。inception_v3已经定义好了默认的参数值:inception_v3_arg_scope(),因此,需要在定义inception_v3模型之前,设置默认参数。
# Define the model:
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
out, end_points = inception_v3.inception_v3(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=keep_prob, is_training=is_training)
(5)指定损失函数和准确率(重点)
inception_v3模型的默认参数inception_v3.inception_v3_arg_scope()已经定义了L2正则化项(你可以看一下源码:weights_regularizer=slim.l2_regularizer(weight_decay)),因此定义损失函数时,需要把L2正则化项的损失也加进来优化。这里就需要特别特别……说明一下:
- 若使用 tf.losses自带的loss函数,则都会自动添加到loss集合中,不需要add_loss()了:如:tf.losses.softmax_cross_entropy()
- 如使用tf.nn的自带的损失函数,则必须手动添加,如: tf.nn.sparse_softmax_cross_entropy_with_logits()和 tf.nn.softmax_cross_entropy_with_logits()
- 特别的,若自定义myloss损失函数,若myloss损失函数中使用了tf.losses中的loss函数,并将该loss添加到slim.losses.add_loss()中, 这时使用tf.losses.get_total_loss()函数时相当于累加两次myloss,因为tf.losses中的loss值都会自动添加到slim.losses接合中。因此若使用tf.losses中自带的loss函数,则不需要add_loss()了,否则相当于重复添加了
本项目源码使用交叉熵损失函数tf.losses.softmax_cross_entropy()和L2正则项add_regularization_losses=True!仅仅两条语句就Ok了,简单了吧,不得不惊叹tf.contrib.slim库的强大,大大简化网络代码的定义。若你使用原生的tf定义损失函数,你会发现计算L2正则项的损失,特别麻烦。
# Specify the loss function: tf.losses定义的loss函数都会自动添加到loss函数,不需要add_loss()了
tf.losses.softmax_cross_entropy(onehot_labels=input_labels, logits=out)#添加交叉熵损失loss=1.6
# slim.losses.add_loss(my_loss)
loss = tf.losses.get_total_loss(add_regularization_losses=True)#添加正则化损失loss=2.2
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(out, 1), tf.argmax(input_labels, 1)), tf.float32))
(6)定义优化方案
这里使用GradientDescentOptimizer梯度下降法,当然也可以选择MomentumOptimizer或者AdadeltaOptimizer其他优化方法。由于inception_v3使用了batch_norm层,需要更新每一层的`average`和`variance`参数, 更新的过程不包含在正常的训练过程中, 需要我们去手动更新,并通过`tf.get_collection`获得所有需要更新的`op`。
一个不确定的说明:
(1)若使用train_op = optimizer.minimize(loss)函数时,则需要手动更新每一层的`average`和`variance`参数,并通过`tf.get_collection`获得所有需要更新的`op`
(2)若使用slim.learning.create_train_op()产生训练的op,貌似会自动更新每一层的参数,这个不确定!主要是我发现即使没有tf.get_collection,使用slim.learning.create_train_op()时,训练也是收敛的。
# Specify the optimization scheme:
optimizer = tf.train.GradientDescentOptimizer(learning_rate=base_lr)
# 在定义训练的时候, 注意到我们使用了`batch_norm`层时,需要更新每一层的`average`和`variance`参数,
# 更新的过程不包含在正常的训练过程中, 需要我们去手动像下面这样更新
# 通过`tf.get_collection`获得所有需要更新的`op`
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# 使用`tensorflow`的控制流, 先执行更新算子, 再执行训练
with tf.control_dependencies(update_ops):
# create_train_op that ensures that when we evaluate it to get the loss,
# the update_ops are done and the gradient updates are computed.
# train_op = slim.learning.create_train_op(total_loss=loss,optimizer=optimizer)
train_op = slim.learning.create_train_op(total_loss=loss, optimizer=optimizer)
(7)训练迭代训练
TF-Slim自带了非常强大的训练工具 slim.learning.train()函数,下面是该函数的简单用法
g = tf.Graph()
# Create the model and specify the losses...
...
total_loss = slim.losses.get_total_loss()
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
# create_train_op ensures that each time we ask for the loss, the update_ops
# are run and the gradients being computed are applied too.
train_op = slim.learning.create_train_op(total_loss, optimizer)
logdir = ... # Where checkpoints are stored.
slim.learning.train(
train_op,
logdir,
number_of_steps=1000,
save_summaries_secs=300,
save_interval_secs=600):
不过啦~本人在循环迭代过程并未使用 slim.learning.train()函数,而是使用原生普通的tensorflow代码。主要是因为我想根据自己的需要控制迭代过程,显示log信息和保存模型:
说明:
1、step_train()函数可以实现测试trian的准确率(这里仅测试训练集的一个batch),而val的数据数据是全部都需要测试的
2、可以设置模型保存间隔,源码还实现保存val准确率最高的模型
def step_train(train_op,loss,accuracy,
train_images_batch,train_labels_batch,train_nums,train_log_step,
val_images_batch,val_labels_batch,val_nums,val_log_step,
snapshot_prefix,snapshot):
'''
循环迭代训练过程
:param train_op: 训练op
:param loss: loss函数
:param accuracy: 准确率函数
:param train_images_batch: 训练images数据
:param train_labels_batch: 训练labels数据
:param train_nums: 总训练数据
:param train_log_step: 训练log显示间隔
:param val_images_batch: 验证images数据
:param val_labels_batch: 验证labels数据
:param val_nums: 总验证数据
:param val_log_step: 验证log显示间隔
:param snapshot_prefix: 模型保存的路径
:param snapshot: 模型保存间隔
:return: None
'''
saver = tf.train.Saver()
max_acc = 0.0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(max_steps + 1):
batch_input_images, batch_input_labels = sess.run([train_images_batch, train_labels_batch])
_, train_loss = sess.run([train_op, loss], feed_dict={input_images: batch_input_images,
input_labels: batch_input_labels,
keep_prob: 0.5, is_training: True})
# train测试(这里仅测试训练集的一个batch)
if i % train_log_step == 0:
train_acc = sess.run(accuracy, feed_dict={input_images: batch_input_images,
input_labels: batch_input_labels,
keep_prob: 1.0, is_training: False})
print "%s: Step [%d] train Loss : %f, training accuracy : %g" % (
datetime.now(), i, train_loss, train_acc)
# val测试(测试全部val数据)
if i % val_log_step == 0:
mean_loss, mean_acc = net_evaluation(sess, loss, accuracy, val_images_batch, val_labels_batch, val_nums)
print "%s: Step [%d] val Loss : %f, val accuracy : %g" % (datetime.now(), i, mean_loss, mean_acc)
# 模型保存:每迭代snapshot次或者最后一次保存模型
if (i % snapshot == 0 and i > 0) or i == max_steps:
print('-----save:{}-{}'.format(snapshot_prefix, i))
saver.save(sess, snapshot_prefix, global_step=i)
# 保存val准确率最高的模型
if mean_acc > max_acc and mean_acc > 0.7:
max_acc = mean_acc
path = os.path.dirname(snapshot_prefix)
best_models = os.path.join(path, 'best_models_{}_{:.4f}.ckpt'.format(i, max_acc))
print('------save:{}'.format(best_models))
saver.save(sess, best_models)
coord.request_stop()
coord.join(threads)
有木有觉得step_train()函数,跟Caffe的solver.prototxt的参数设置很像,比如snapshot_prefix、snapshot等参数,没错,我就是学Caffe转学TensorFlow的。因为习惯了Caffe的训练方式,所以啦,我把循环迭代改成类似于Caffe的形式,娃哈哈!
其中net_evaluation为评价函数:
def net_evaluation(sess,loss,accuracy,val_images_batch,val_labels_batch,val_nums):
val_max_steps = int(val_nums / batch_size)
val_losses = []
val_accs = []
for _ in xrange(val_max_steps):
val_x, val_y = sess.run([val_images_batch, val_labels_batch])
# print('labels:',val_y)
# val_loss = sess.run(loss, feed_dict={x: val_x, y: val_y, keep_prob: 1.0})
# val_acc = sess.run(accuracy,feed_dict={x: val_x, y: val_y, keep_prob: 1.0})
val_loss,val_acc = sess.run([loss,accuracy], feed_dict={input_images: val_x, input_labels: val_y, keep_prob:1.0, is_training: False})
val_losses.append(val_loss)
val_accs.append(val_acc)
mean_loss = np.array(val_losses, dtype=np.float32).mean()
mean_acc = np.array(val_accs, dtype=np.float32).mean()
return mean_loss, mean_acc
完整的inception_v3_train_val.py代码:
#coding=utf-8
import tensorflow as tf
import numpy as np
import pdb
import os
from datetime import datetime
import slim.nets.inception_v3 as inception_v3
from create_tf_record import *
import tensorflow.contrib.slim as slim
labels_nums = 5 # 类别个数
batch_size = 16 #
resize_height = 299 # 指定存储图片高度
resize_width = 299 # 指定存储图片宽度
depths = 3
data_shape = [batch_size, resize_height, resize_width, depths]
# 定义input_images为图片数据
input_images = tf.placeholder(dtype=tf.float32, shape=[None, resize_height, resize_width, depths], name='input')
# 定义input_labels为labels数据
# input_labels = tf.placeholder(dtype=tf.int32, shape=[None], name='label')
input_labels = tf.placeholder(dtype=tf.int32, shape=[None, labels_nums], name='label')
# 定义dropout的概率
keep_prob = tf.placeholder(tf.float32,name='keep_prob')
is_training = tf.placeholder(tf.bool, name='is_training')
def net_evaluation(sess,loss,accuracy,val_images_batch,val_labels_batch,val_nums):
val_max_steps = int(val_nums / batch_size)
val_losses = []
val_accs = []
for _ in xrange(val_max_steps):
val_x, val_y = sess.run([val_images_batch, val_labels_batch])
# print('labels:',val_y)
# val_loss = sess.run(loss, feed_dict={x: val_x, y: val_y, keep_prob: 1.0})
# val_acc = sess.run(accuracy,feed_dict={x: val_x, y: val_y, keep_prob: 1.0})
val_loss,val_acc = sess.run([loss,accuracy], feed_dict={input_images: val_x, input_labels: val_y, keep_prob:1.0, is_training: False})
val_losses.append(val_loss)
val_accs.append(val_acc)
mean_loss = np.array(val_losses, dtype=np.float32).mean()
mean_acc = np.array(val_accs, dtype=np.float32).mean()
return mean_loss, mean_acc
def step_train(train_op,loss,accuracy,
train_images_batch,train_labels_batch,train_nums,train_log_step,
val_images_batch,val_labels_batch,val_nums,val_log_step,
snapshot_prefix,snapshot):
'''
循环迭代训练过程
:param train_op: 训练op
:param loss: loss函数
:param accuracy: 准确率函数
:param train_images_batch: 训练images数据
:param train_labels_batch: 训练labels数据
:param train_nums: 总训练数据
:param train_log_step: 训练log显示间隔
:param val_images_batch: 验证images数据
:param val_labels_batch: 验证labels数据
:param val_nums: 总验证数据
:param val_log_step: 验证log显示间隔
:param snapshot_prefix: 模型保存的路径
:param snapshot: 模型保存间隔
:return: None
'''
saver = tf.train.Saver()
max_acc = 0.0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(max_steps + 1):
batch_input_images, batch_input_labels = sess.run([train_images_batch, train_labels_batch])
_, train_loss = sess.run([train_op, loss], feed_dict={input_images: batch_input_images,
input_labels: batch_input_labels,
keep_prob: 0.5, is_training: True})
# train测试(这里仅测试训练集的一个batch)
if i % train_log_step == 0:
train_acc = sess.run(accuracy, feed_dict={input_images: batch_input_images,
input_labels: batch_input_labels,
keep_prob: 1.0, is_training: False})
print "%s: Step [%d] train Loss : %f, training accuracy : %g" % (
datetime.now(), i, train_loss, train_acc)
# val测试(测试全部val数据)
if i % val_log_step == 0:
mean_loss, mean_acc = net_evaluation(sess, loss, accuracy, val_images_batch, val_labels_batch, val_nums)
print "%s: Step [%d] val Loss : %f, val accuracy : %g" % (datetime.now(), i, mean_loss, mean_acc)
# 模型保存:每迭代snapshot次或者最后一次保存模型
if (i % snapshot == 0 and i > 0) or i == max_steps:
print('-----save:{}-{}'.format(snapshot_prefix, i))
saver.save(sess, snapshot_prefix, global_step=i)
# 保存val准确率最高的模型
if mean_acc > max_acc and mean_acc > 0.7:
max_acc = mean_acc
path = os.path.dirname(snapshot_prefix)
best_models = os.path.join(path, 'best_models_{}_{:.4f}.ckpt'.format(i, max_acc))
print('------save:{}'.format(best_models))
saver.save(sess, best_models)
coord.request_stop()
coord.join(threads)
def train(train_record_file,
train_log_step,
train_param,
val_record_file,
val_log_step,
labels_nums,
data_shape,
snapshot,
snapshot_prefix):
'''
:param train_record_file: 训练的tfrecord文件
:param train_log_step: 显示训练过程log信息间隔
:param train_param: train参数
:param val_record_file: 验证的tfrecord文件
:param val_log_step: 显示验证过程log信息间隔
:param val_param: val参数
:param labels_nums: labels数
:param data_shape: 输入数据shape
:param snapshot: 保存模型间隔
:param snapshot_prefix: 保存模型文件的前缀名
:return:
'''
[base_lr,max_steps]=train_param
[batch_size,resize_height,resize_width,depths]=data_shape
# 获得训练和测试的样本数
train_nums=get_example_nums(train_record_file)
val_nums=get_example_nums(val_record_file)
print('train nums:%d,val nums:%d'%(train_nums,val_nums))
# 从record中读取图片和labels数据
# train数据,训练数据一般要求打乱顺序shuffle=True
train_images, train_labels = read_records(train_record_file, resize_height, resize_width, type='normalization')
train_images_batch, train_labels_batch = get_batch_images(train_images, train_labels,
batch_size=batch_size, labels_nums=labels_nums,
one_hot=True, shuffle=True)
# val数据,验证数据可以不需要打乱数据
val_images, val_labels = read_records(val_record_file, resize_height, resize_width, type='normalization')
val_images_batch, val_labels_batch = get_batch_images(val_images, val_labels,
batch_size=batch_size, labels_nums=labels_nums,
one_hot=True, shuffle=False)
# Define the model:
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
out, end_points = inception_v3.inception_v3(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=keep_prob, is_training=is_training)
# Specify the loss function: tf.losses定义的loss函数都会自动添加到loss函数,不需要add_loss()了
tf.losses.softmax_cross_entropy(onehot_labels=input_labels, logits=out)#添加交叉熵损失loss=1.6
# slim.losses.add_loss(my_loss)
loss = tf.losses.get_total_loss(add_regularization_losses=True)#添加正则化损失loss=2.2
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(out, 1), tf.argmax(input_labels, 1)), tf.float32))
# Specify the optimization scheme:
optimizer = tf.train.GradientDescentOptimizer(learning_rate=base_lr)
# global_step = tf.Variable(0, trainable=False)
# learning_rate = tf.train.exponential_decay(0.05, global_step, 150, 0.9)
#
# optimizer = tf.train.MomentumOptimizer(learning_rate, 0.9)
# # train_tensor = optimizer.minimize(loss, global_step)
# train_op = slim.learning.create_train_op(loss, optimizer,global_step=global_step)
# 在定义训练的时候, 注意到我们使用了`batch_norm`层时,需要更新每一层的`average`和`variance`参数,
# 更新的过程不包含在正常的训练过程中, 需要我们去手动像下面这样更新
# 通过`tf.get_collection`获得所有需要更新的`op`
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# 使用`tensorflow`的控制流, 先执行更新算子, 再执行训练
with tf.control_dependencies(update_ops):
# create_train_op that ensures that when we evaluate it to get the loss,
# the update_ops are done and the gradient updates are computed.
# train_op = slim.learning.create_train_op(total_loss=loss,optimizer=optimizer)
train_op = slim.learning.create_train_op(total_loss=loss, optimizer=optimizer)
# 循环迭代过程
step_train(train_op, loss, accuracy,
train_images_batch, train_labels_batch, train_nums, train_log_step,
val_images_batch, val_labels_batch, val_nums, val_log_step,
snapshot_prefix, snapshot)
if __name__ == '__main__':
train_record_file='dataset/record/train299.tfrecords'
val_record_file='dataset/record/val299.tfrecords'
train_log_step=100
base_lr = 0.01 # 学习率
max_steps = 10000 # 迭代次数
train_param=[base_lr,max_steps]
val_log_step=200
snapshot=2000#保存文件间隔
snapshot_prefix='models/model.ckpt'
train(train_record_file=train_record_file,
train_log_step=train_log_step,
train_param=train_param,
val_record_file=val_record_file,
val_log_step=val_log_step,
labels_nums=labels_nums,
data_shape=data_shape,
snapshot=snapshot,
snapshot_prefix=snapshot_prefix)
OK啦,运行启动训练看看log信息:
2018-08-16 10:08:57.107124: Step [0] train Loss : 2.762746, training accuracy : 0.3125
2018-08-16 10:09:19.281263: Step [0] val Loss : 2.931877, val accuracy : 0.215726
2018-08-16 10:10:33.807865: Step [100] train Loss : 1.849171, training accuracy : 0.375
2018-08-16 10:11:56.997064: Step [200] train Loss : 2.248142, training accuracy : 0.0625
2018-08-16 10:12:22.684584: Step [200] val Loss : 163.246002, val accuracy : 0.200941
2018-08-16 10:13:44.102429: Step [300] train Loss : 1.785683, training accuracy : 0.25
……
2018-08-16 10:48:24.938470: Step [2500] train Loss : 0.732916, training accuracy : 0.3125
2018-08-16 10:49:45.958701: Step [2600] train Loss : 0.749750, training accuracy : 0.25
2018-08-16 10:50:10.845769: Step [2600] val Loss : 9.741004, val accuracy : 0.387769
2018-08-16 10:51:31.777861: Step [2700] train Loss : 1.074746, training accuracy : 0.4375
2018-08-16 10:52:52.909256: Step [2800] train Loss : 0.814188, training accuracy : 0.125
2018-08-16 10:53:17.725089: Step [2800] val Loss : 9.216277, val accuracy : 0.368952
2018-08-16 10:54:38.721697: Step [2900] train Loss : 0.762590, training accuracy : 0.375
2018-08-16 10:55:59.860650: Step [3000] train Loss : 0.733000, training accuracy : 0.1875
2018-08-16 10:56:24.746242: Step [3000] val Loss : 13.225746, val accuracy : 0.237903
2018-08-16 10:57:45.828758: Step [3100] train Loss : 0.833523, training accuracy : 0.5625
2018-08-16 10:59:06.822897: Step [3200] train Loss : 0.710151, training accuracy : 0.625
……
2018-08-16 12:40:31.923101: Step [9500] train Loss : 0.700521, training accuracy : 1
2018-08-16 12:41:53.206480: Step [9600] train Loss : 0.782273, training accuracy : 1
2018-08-16 12:42:17.757492: Step [9600] val Loss : 1.299307, val accuracy : 0.860887
2018-08-16 12:43:38.987012: Step [9700] train Loss : 0.700440, training accuracy : 0.9375
2018-08-16 12:45:00.040759: Step [9800] train Loss : 0.702021, training accuracy : 0.75
2018-08-16 12:45:25.000334: Step [9800] val Loss : 1.402472, val accuracy : 0.836021
2018-08-16 12:46:46.077850: Step [9900] train Loss : 0.701689, training accuracy : 1
2018-08-16 12:48:07.302272: Step [10000] train Loss : 0.703496, training accuracy : 1
2018-08-16 12:48:32.193206: Step [10000] val Loss : 1.131343, val accuracy : 0.889113
-----save:models/model.ckpt-10000
------save:models/best_models_10000_0.8891.ckpt
可以看到,前2000次迭代,不管是train还是val的识别率都很低,徘徊在20%-30%,但到了2000次以上,识别率一步一步往上蹭,到10000步时,train的识别率达到100%啦,而val的识别率稳定在80%以上,对于只训练了10000次,其效果还是不错的~有需要的话,可自行训练十万次以上哈
5、模型预测 模型预测代码,比较简单,这里直接贴出源码:
#coding=utf-8
import tensorflow as tf
import numpy as np
import pdb
import cv2
import os
import glob
import slim.nets.inception_v3 as inception_v3
from create_tf_record import *
import tensorflow.contrib.slim as slim
def predict(models_path,image_dir,labels_filename,labels_nums, data_format):
[batch_size, resize_height, resize_width, depths] = data_format
labels = np.loadtxt(labels_filename, str, delimiter='\t')
input_images = tf.placeholder(dtype=tf.float32, shape=[None, resize_height, resize_width, depths], name='input')
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
out, end_points = inception_v3.inception_v3(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=1.0, is_training=False)
# 将输出结果进行softmax分布,再求最大概率所属类别
score = tf.nn.softmax(out)
class_id = tf.argmax(score, 1)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, models_path)
images_list=glob.glob(os.path.join(image_dir,'*.jpg'))
for image_path in images_list:
im=read_image(image_path,resize_height,resize_width,normalization=True)
im=im[np.newaxis,:]
#pred = sess.run(f_cls, feed_dict={x:im, keep_prob:1.0})
pre_score,pre_label = sess.run([score,class_id], feed_dict={input_images:im})
max_score=pre_score[0,pre_label]
print "{} is: pre labels:{},name:{} score: {}".format(image_path,pre_label,labels[pre_label], max_score)
sess.close()
if __name__ == '__main__':
class_nums=5
image_dir='test_image'
labels_filename='dataset/label.txt'
models_path='models/model.ckpt-10000'
batch_size = 1 #
resize_height = 299 # 指定存储图片高度
resize_width = 299 # 指定存储图片宽度
depths=3
data_format=[batch_size,resize_height,resize_width,depths]
predict(models_path,image_dir, labels_filename, class_nums, data_format)
预测结果:
test_image/flower.jpg is: pre labels:[0],name:['flower'] score: [ 0.99865556]
test_image/houses.jpg is: pre labels:[3],name:['houses'] score: [ 0.99899763]
test_image/animal.jpg is: pre labels:[2],name:['animal'] score: [ 0.96808302]
test_image/guitar.jpg is: pre labels:[1],name:['guitar'] score: [ 0.99999511]
四、其他模型训练方法 上面的程序inception_v3_train_val.py是实现googLenet inception V3训练的过程,实质上,稍微改改就可以支持训练 inception V1,V2 啦,改动方法也很简单,以 inception V1为例:
(1)import 改为:
# 将
import slim.nets.inception_v3 as inception_v3
# 改为
import slim.nets.inception_v1 as inception_v1
(2)record数据
inception V1要求输入的数据是224*224,因此制作record数据时,需要设置:
resize_height = 224 # 指定存储图片高度
resize_width = 224 # 指定存储图片宽度
(3)定义模型和默认参数修改:
# 将
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
out, end_points = inception_v3.inception_v3(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=keep_prob, is_training=is_training)
# 改为
with slim.arg_scope(inception_v1.inception_v1_arg_scope()):
out, end_points = inception_v1.inception_v1(inputs=input_images, num_classes=labels_nums, dropout_keep_prob=keep_prob, is_training=is_training)
(4)修改优化方案
inception V3中使用的优化方案是: optimizer = tf.train.GradientDescentOptimizer(learning_rate=base_lr),但在V1中,我发现训练不收敛,后来我改为 optimizer = tf.train.MomentumOptimizer(learning_rate=base_lr,momentum= 0.9),又可以正常收敛了。总之,要是发现训练不收敛,请尝试修改几个参数:
1、增大或减小学习率参数:base_lr
2、改变优化方案:如使用MomentumOptimizer或者AdadeltaOptimizer等优化方法
3、是否有设置默认的模型参数:如slim.arg_scope(inception_v1.inception_v1_arg_scope())
4、计算损失时,增加或去掉正则项:tf.losses.get_total_loss(add_regularization_losses=False)
……最后,就可以Train了!是的,就是那么简单~
五、将ckpt转pb文件: tensorflow实现将ckpt转pb文件的方法,请参考:
https://blog.csdn.net/guyuealian/article/details/82218092
如果你觉得该帖子帮到你,还望贵人多多支持,鄙人会再接再厉,继续努力的~
作者: 不二晨 时间: 2018-9-13 16:26
很不错,受教了
作者: 魔都黑马少年梦 时间: 2018-11-1 16:47
欢迎光临 黑马程序员技术交流社区 (http://bbs.itheima.com/) |
黑马程序员IT技术论坛 X3.2 |