【TensorFlow】-LeNet-5
1.难点说明
1.输出尺寸的计算公式
不是整数下取整
输 出 = [ n + 2 p − f s + 1 ] 输出 = [\frac{n+2p-f}{s}+1] 输出=[sn+2p−f+1]
2.drop
防止过拟合,只在训练过程中使用
dropout一般只在全连接层而不是卷积层或者池化层使用
3.tf.argmax(vector, axis=1)
其中axis:0表示按列,1表示按行。返回的是vector中的最大值的索引号
4.get_collection()
返回一个列表,这个列表包含所有这个losses集合中的元素,这些元素就是损失函数的不同部分
5.tf.add_n()
tf.add_n([p1, p2, p3…])函数是实现一个列表的元素的相加。输入的对象是一个列表,列表里的元素可以是向量、矩阵等
6.saver.save()
saver.save(sess,'./model/model.ckp',
global_step = global_step)
'''
@global_step 这样可以让每个被保存模型的文件名末尾加上训练的轮数
比如“model.ckpt-1000” 表示训练1000轮之后得到的模型
'''
每次保存会生成3个文件
- model.ckpt.meta——保存了tensorflow计算图的结构
- model.ckpt——tensorflow程序中每一个变量的取值
- checkpoint——文件中保存了一个目录下所有的模型文件列表
7.tf.nn.conv2d()
tf.nn.conv2d(input, # Tensor,具有[batch, in_height, in_width, in_channels]
filter, # [filter_height, filter_width, in_channels, out_channels]
strides,
padding, # "SAME","VALID"
use_cudnn_on_gpu=True, # 是否使用cudnn加速
name=None)
对于 VALID
n
e
w
−
height
=
n
e
w
−
w
i
d
t
h
=
⌈
(
W
−
F
+
1
)
S
⌉
n e w_{-} \text {height}=n e w_{-} w i d t h=\left\lceil\frac{(W-F+1)}{S}\right\rceil
new−height=new−width=⌈S(W−F+1)⌉
对于 SAME
new
−
height
=
new
−
width
=
⌈
W
S
⌉
\text {new}_{-} \text {height}=\text {new}_{-} \text {width}=\left\lceil\frac{W}{S}\right\rceil
new−height=new−width=⌈SW⌉
- W-为输入size,F为filer的size
- 向上取整
参考:ensorFlow中CNN的两种padding方式“SAME”和“VALID”
8.tf.nn.bias_add和tf.add()
tf.nn.bias_add(x,y,name=None)
这个函数的作用是将偏差bias加到value上面
可以看作是tf.add的一个特例。其中bias必须是一维的
import tensorflow as tf
a=tf.constant([[1,1],[2,2],[3,3]],dtype=tf.float32)
b=tf.constant([1,-1],dtype=tf.float32)
c=tf.constant([1],dtype=tf.float32)
with tf.Session() as sess:
print('bias_add:')
print(sess.run(tf.nn.bias_add(a, b)))
#执行下面语句错误
#print(sess.run(tf.nn.bias_add(a, c)))
print('add:')
print(sess.run(tf.add(a, c)))
9.正则项
只有全连接层的权重需要加入正则
10.tf.Print()
Print(
input_, # 通过这个操作的张量。 (流入的数据流)
data,
message=None, # 一个字符串,错误消息的前缀
first_n=None,
summarize=None, # 只打印每个张量的固定数目的条目
name=None
)
2.mnist_train.py
'''
输入x - [batch, 28,28,1] - 四维向量
输出y_ - [None, 10]
'''
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
import mnist_inference
BATCH_SIZE = 100
LEARNING_RATE_BASE = 0.01
LEARNING_RATE_DECAY =0.99
REGULARAZTION_RATE = 0.0001 #正则化项的权重
MOVING_AVERAGE_DECAY = 0.99 #滑动平均模型的衰减率
TRAINING_STEPS = 8000
def train( mnist ):
'''定义输入输出placeholder'''
x = tf.placeholder(tf.float32, [BATCH_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.NUM_CHANNELS], name = 'x-input1')
y_ = tf.placeholder(tf.float32, [None, mnist_inference.OUTPUT_NODE] , name='y-input')
regularizer = tf.contrib.layers.l2_regularizer( REGULARAZTION_RATE ) #返回一个可以计算l2正则化项的函数
'''前向传播过程'''
y = mnist_inference.inference(x,True,regularizer)
'''滑动平均更新参数'''
global_step = tf.Variable(0, trainable=False) # 迭代轮数变量,控制衰减率
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply( tf.trainable_variables()) # 更新列表中的变量
'''损失函数'''
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_,1))
cross_entropy_mean = tf.reduce_mean(cross_entropy)
loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses'))
'''
通过exponential_decay函数生成学习率,使用呈指数衰减的学习率
在minimize函数中传入global_step将自动更新global_step参数,从而使得学习率learning_rate也得到相应更新
'''
learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE,
global_step,
mnist.train.num_examples/BATCH_SIZE,
LEARNING_RATE_DECAY,
staircase=True)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
'''
tf.control_dependencies机制更新反向传播参数和每一个参数的滑动平均值
'''
with tf.control_dependencies([train_step, variables_averages_op]):
train_op = tf.no_op(name='train')
saver = tf.train.Saver()
with tf.Session() as sess:
# 初始化所有变量
init_op = tf.global_variables_initializer()
sess.run(init_op)
print("------------开始训练--------------")
for i in range(TRAINING_STEPS):
xs, ys = mnist.train.next_batch( BATCH_SIZE )
# 类似地将输入的训练数据格式调整为一个四维矩阵,并将这个调整后的数据传入sess.run过程
reshaped_xs = np.reshape(xs, (BATCH_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.NUM_CHANNELS))
train_op_renew,loss_value, step = sess.run([train_op, loss, global_step],feed_dict={x: reshaped_xs, y_: ys})
if i % 1000 == 0:
print ( "After %d training step (s) , loss on training batch is %g." % (step, loss_value))
saver.save(sess,'./model/model.ckp',
global_step = global_step)
print("------------------训练结束-----------------")
# 主程序入口
def main(argv=None):
'''
主程序入口
声明处理MNIST数据集的类,这个类在初始化时会自动下载数据
'''
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
if mnist != None:
print("-------------数据加载完毕------------------")
train(mnist)
if __name__ == '__main__':
tf.app.run ()
3.mnist_inference.py
'''
1.-----------conv1--------------------------------
输入: 28*28*1
f:(5*5*32) s:1 padding='same'
输出:(28*28*32)
2.-----------pool1---------------------------------
f:(2*2) s:2 padding='same'
输出:(14*14*32)
3.-----------conv2----------------------------------
f:(5*5*64) s:1 padding='same'
输出:(14*14*64)
4.-----------pool2-----------------------------------
f:(2*2) s:1 padding='same'
输出:(7*7*64) >>>> reshape 成一个(batch_sizes,7*7*64)
5.-----------fc1-------------------------------------
输入:(batch_sizes,7*7*64)
6.-----------fc2-------------------------------------
输入:(,512)
输出:(,10)
'''
import tensorflow as tf
INPUT_NODE = 784
OUTPUT_NODE =10
IMAGE_SIZE =28
NUM_CHANNELS = 1
NUM_LABELS = 10
'''第一层卷积层的尺寸和深度'''
C0NV1_DEEP = 32
C0NV1_SIZE = 5
'''第二层卷积层的尺寸和深度'''
CONV2_DEEP = 64
CONV2_SIZE = 5
'''全连接层的节点个数'''
FC_SIZE = 512
def inference(input_tensor, train, regularizer):
'''
卷积神经网络的前向传播过程
@ train 用于区分训练和测试过程
@ input_tensor 输入变量 四维
'''
'''
conv1
输入: 28*28*1 [batch_size,:]
f:(5*5*32) s:1 padding='same'
输出:(28*28*32)
'''
with tf.variable_scope('layer1-conv1'):
conv1_weights = tf.get_variable("weight",
[C0NV1_SIZE, C0NV1_SIZE, NUM_CHANNELS, C0NV1_DEEP],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv1_biases = tf.get_variable("bias", [C0NV1_DEEP], initializer=tf.constant_initializer(0.0))
conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
'''pool1'''
with tf.name_scope("layer2-pool1"):
pool1 = tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
with tf.variable_scope("layer3-conv2"):
conv2_weights = tf.get_variable("weight",
[CONV2_SIZE,CONV2_SIZE,C0NV1_DEEP,CONV2_DEEP],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv2_biases = tf.get_variable("bias", [CONV2_DEEP], initializer=tf.constant_initializer(0.0))
conv2 = tf.nn.conv2d(pool1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
with tf.name_scope("layer4-pool2"):
pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
pool_shape = pool2.get_shape().as_list() # 获取pool2的输出
'''获取的pool_shape包含batch_size层'''
nodes = pool_shape[1] * pool_shape[2] * pool_shape[3]
# 通过tf.reshape函数将第四层的输出变成一个batch的向量。
reshaped = tf.reshape(pool2, [pool_shape[0], nodes])
with tf.variable_scope("layer5-fc1"):
fc1_weights = tf.get_variable("weight", [nodes, FC_SIZE],
initializer=tf.truncated_normal_initializer(stddev=0.1))
'''add_to_collection函数将一个张量加入一个集合'losses' '''
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc1_weights))
fc1_biases = tf.get_variable('bias', [FC_SIZE], initializer=tf.constant_initializer(0.1))
fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_weights) + fc1_biases)
if train:
fc1 = tf.nn.dropout(fc1, 0.5)
with tf.variable_scope('layer6-fc2'):
fc2_weights = tf.get_variable("weight", [FC_SIZE, NUM_LABELS],
initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc2_weights))
fc2_biases = tf.get_variable("bias", [NUM_LABELS], initializer=tf.constant_initializer(0.1))
logit = tf.matmul(fc1, fc2_weights) + fc2_biases
return logit
4.mnist_eval.py
'''
测试集数量:5000
@ minst.validation.images.shape (5000, 784)
'''
import time
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
import mnist_inference
import mnist_train
'''每10秒加载一次模型,并测试最新的正确率'''
EVAL_INTERVAL_SECS = 10
def evaluate( mnist ):
with tf.Graph().as_default() as g: # 将默认图设为g
'''定义输入输出的格式'''
x = tf.placeholder(tf.float32, [mnist.validation.images.shape[0],
mnist_inference.IMAGE_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.NUM_CHANNELS], name='x-input1')
y_ = tf.placeholder(tf.float32, [None, mnist_inference.OUTPUT_NODE], name='y-input')
xs = mnist.validation.images
# 类似地将输入的测试数据格式调整为一个四维矩阵
reshaped_xs = np.reshape(xs, (mnist.validation.images.shape[0],
mnist_inference.IMAGE_SIZE,
mnist_inference.IMAGE_SIZE,
mnist_inference.NUM_CHANNELS))
validate_feed = {x: reshaped_xs, y_: mnist.validation.labels}
'''前向传播测试,不需要正则项'''
y = mnist_inference.inference(x,None, None)
#使用tf.argmax(y, 1)就可以得到输入样例的预测类别
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
# 首先将一个布尔型的数组转换为实数,然后计算平均值
#True为1,False为0
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
#通过变量重命名的方式来加载模型
variable_averages = tf.train.ExponentialMovingAverage(mnist_train.MOVING_AVERAGE_DECAY)
variable_to_restore = variable_averages.variables_to_restore()
# 所有滑动平均的值组成的字典,处在/ExponentialMovingAverage下的值
# 为了方便加载时重命名滑动平均量,tf.train.ExponentialMovingAverage类
# 提供了variables_to_store函数来生成tf.train.Saver类所需要的变量
saver = tf.train.Saver(variable_to_restore) #这些值要从模型中提取
#每隔EVAL_INTERVAL_SECS秒调用一次计算正确率的过程以检测训练过程中正确率的变化
#while True:
for i in range(2): # 为了降低个人电脑的压力,此处只利用最后生成的模型对测试数据集做测试
with tf.Session() as sess:
# 会通过checkpoint文件自动找到目录中最新模型的文件名
ckpt = tf.train.get_checkpoint_state( "./model")
if ckpt and ckpt.model_checkpoint_path:
#加载模型
saver.restore(sess, ckpt.model_checkpoint_path)
#得到所有的滑动平均值
#通过文件名得到模型保存时迭代的轮数
global_step = ckpt.model_checkpoint_path.split('-')[-1]
accuracy_score = sess.run(accuracy, feed_dict = validate_feed) #使用此模型检验
#没有初始化滑动平均值,只是调用模型的值,inference只是提供了一个变量的接口,完全没有赋值
print("After %s training steps, validation accuracy = %g" %(global_step, accuracy_score))
else:
print("No checkpoint file found")
return
time.sleep(EVAL_INTERVAL_SECS)
# time sleep()函数推迟调用线程的运行,可通过参数secs指秒数,表示进程挂起的时间。
def main( argv=None ):
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
evaluate(mnist)
if __name__=='__main__':
tf.app.run()