Lenet神经网络是YannLeCun等人在1998年提出的,该神经网络充分考虑图像的相关性。
Lenet神经网络结构为:
1),输入为32X32X1的图片大小,为单通道;
2),进行卷积,卷积核大小为5X5X1,个数为6,步长为1,非0填充;
3)将卷积结果通过非线性激活函数;
4)进行池化,池化大小为2X2,步长为1,全0填充;
5)进行卷积,卷积核大小为5X5X6,个数16,步长1,非全0填充;
6)将卷积结果通过非线性激活函数;
7)进行池化,池化大小为2X2,步长1,全0填充;
8)全连接层进行10分类。
根据Lenet神经网络的结构可得,Lenet具有以下特点:
1)卷积(conv),池化(ave-pooling),非线性激活函数(sigmoid)相互交替;
2)层与层之间稀疏连接,减少计算复杂度。
对Lenet进行微调整,使其适应Mnist数据集:
由于Mnist数据集图片大小为28X28X1的灰度图片,而Lenet神经网络的输入为32X32X1的大小,所以要进行微调;
①输入为28*28*1的图片大小,为单通道的输入;
②进行卷积,卷积核大小为5*5*1,个数为32,步长为1,全零填充模式;
③将卷积结果通过非线性激活函数;
④进行池化,池化大小为2*2,步长为2,全零填充模式;
⑤进行卷积,卷积核大小为5*5*32,个数为64,步长为1,全零填充模式;
⑥将卷积结果通过非线性激活函数;
⑦进行池化,池化大小为2*2,步长为2,全零填充模式;
⑧全连接层,进行10分类。
Lenet进行微调后的结构如下所示:
Lenet神经网络在Mnist数据集上的实现,主要分为三个部分:前向传播过程(mnist_lenet5_forward.py)、反向传播过程(mnist_lenet5_backword.py)、测试过程(mnist_lenet5_test.py)。
第一,前向传播过程(mnist_lenet5_forward.py)实现对网络中参数和偏置的初始化、定义卷积结构和池化结构、定义前向传播过程。具体代码如下所示:
# coding:utf-8
import tensorflow as tf
# 每张图片分辨率为28*28
IMAGE_SIZE = 28
# Mnist数据集为灰度图,故输入图片通道数NUM_CHANNELS取值为1
NUM_CHANNELS = 1
# 第一层卷积核大小为5
CONV1_SIZE = 5
# 卷积核个数为32
CONV1_KERNEL_NUM = 32
# 第二层卷积核大小为5
CONV2_SIZE = 5
# 卷积核个数为64
CONV2_KERNEL_NUM = 64
# 全连接层第一层为 512 个神经元
FC_SIZE = 512
# 全连接层第二层为 10 个神经元
OUTPUT_NODE = 10
# 权重w计算
def get_weight(shape, regularizer):
w = tf.Variable(tf.truncated_normal(shape, stddev=0.1))
if regularizer != None: tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(regularizer)(w))
return w
# 偏置b计算
def get_bias(shape):
b = tf.Variable(tf.zeros(shape))
return b
# 卷积层计算
def conv2d(x, w):
return tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding='SAME')
# 最大池化层计算
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def forward(x, train, regularizer):
# 实现第一层卷积
conv1_w = get_weight([CONV1_SIZE, CONV1_SIZE, NUM_CHANNELS, CONV1_KERNEL_NUM], regularizer)
conv1_b = get_bias([CONV1_KERNEL_NUM])
conv1 = conv2d(x, conv1_w)
# 非线性激活
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_b))
# 最大池化
pool1 = max_pool_2x2(relu1)
# 实现第二层卷积
conv2_w = get_weight([CONV2_SIZE, CONV2_SIZE, CONV1_KERNEL_NUM, CONV2_KERNEL_NUM], regularizer)
conv2_b = get_bias([CONV2_KERNEL_NUM])
conv2 = conv2d(pool1, conv2_w)
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_b))
pool2 = max_pool_2x2(relu2)
# 获取一个张量的维度
pool_shape = pool2.get_shape().as_list()
# pool_shape[1] 为长 pool_shape[2] 为宽 pool_shape[3]为高
nodes = pool_shape[1] * pool_shape[2] * pool_shape[3]
# 得到矩阵被拉长后的长度,pool_shape[0]为batch值
reshaped = tf.reshape(pool2, [pool_shape[0], nodes])
# 实现第三层全连接层
fc1_w = get_weight([nodes, FC_SIZE], regularizer)
fc1_b = get_bias([FC_SIZE])
fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_w) + fc1_b)
# 如果是训练阶段,则对该层输出使用dropout
if train: fc1 = tf.nn.dropout(fc1, 0.5)
# 实现第四层全连接层
fc2_w = get_weight([FC_SIZE, OUTPUT_NODE], regularizer)
fc2_b = get_bias([OUTPUT_NODE])
y = tf.matmul(fc1, fc2_w) + fc2_b
return y
第二,反向传播过程(mnist_lenet5_backward.py),完成训练神经网络的参数。具体代码如下所示:
# coding:utf-8
import tensorflow as tf
import mnist_lenet5_forward
import numpy as np
import os
from tensorflow.examples.tutorials.mnist import input_data
# batch的数量
BATCH_SIZE = 100
# 初始学习率
LEARNING_RATE_BASE = 0.005
# 学习率衰减率
LEARNING_RATE_DECAY = 0.99
# 正则化
REGULARIZER = 0.0001
# 滑动平均衰减率
MOVING_AVERAGE_DECAY = 0.99
# 模型保存路径
MODEL_SAVE_PATH = "./model/"
# 模型名称
MODEL_NAME = "mnist_model"
def backward(mnist):
# 卷积层输入为四阶张量
# 第一阶表示每轮喂入的图片数量,第二阶和第三阶分别表示图片的行分辨率和列分辨率,第四阶表示通道数
x = tf.placeholder(tf.float32, [
BATCH_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.NUM_CHANNELS
])
y_ = tf.placeholder(tf.float32, [None, mnist_lenet5_forward.OUTPUT_NODE])
y = mnist_lenet5_forward.forward(x, True, REGULARIZER)
# 声明一个全局计数器
global_step = tf.Variable(0, trainable=False)
# 对网络最后一层的输出y做softmax,求取输出属于某一类的概率
ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
# 向量求均值
cem = tf.reduce_mean(ce)
# 正则化的损失值
loss = cem + tf.add_n(tf.get_collection('losses'))
# 指数衰减学习率
learning_rate = tf.train.exponential_decay(
LEARNING_RATE_BASE,
global_step,
mnist.train.num_examples / BATCH_SIZE,
LEARNING_RATE_DECAY,
staircase=True
)
# 梯度下降算法优化器
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
# 采用滑动平均的方法更新参数
ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
ema_op = ema.apply(tf.trainable_variables())
# 将train_step和ema两个训练参数绑定train_op上
with tf.control_dependencies([train_step, ema_op]):
train_op = tf.no_op(name='train')
# 实例化一个保存和恢复变量的saver
saver = tf.train.Saver()
with tf.Session() as sess:
init_op=tf.global_variables_initializer()
sess.run(init_op)
#通过checkpoint文件定位到最新保存的模型,若文件存在,则加载最新模型
ckpt=tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess,ckpt.model_checkpoint_path)
for i in range(5000):
#读取一个batch数据,将输入数据xs转成网络输入相同形状的矩阵
xs,ys=mnist.train.next_batch(BATCH_SIZE)
reshaped_xs=np.reshape(xs,(
BATCH_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.NUM_CHANNELS
))
_,loss_value,step=sess.run([train_step,loss,global_step],feed_dict={x:reshaped_xs,y_:ys})
if i%100==0:
print("after %d training step,loss is %g"%(step,loss_value))
saver.save(sess,os.path.join(MODEL_SAVE_PATH,MODEL_NAME),global_step=global_step)
def main():
mnist=input_data.read_data_sets("./data/",one_hot=True)
backward(mnist)
if __name__=='__main__':
main()
第三,测试过程(mnist_lenet5_test.py),对Mnist数据集中的测试数据进行预测,测试模型准确率。具体代码如下所示:
# coding:utf-8
import time
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import mnist_lenet5_forward
import mnist_lenet5_backward
import numpy as np
# 创建一个默认图,在该图中执行以下操作
def test(mnist):
with tf.Graph().as_default() as g:
x = tf.placeholder(tf.float32, [
mnist.test.num_examples,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.NUM_CHANNELS])
y_ = tf.placeholder(tf.float32, [None, mnist_lenet5_forward.OUTPUT_NODE])
# 训练好的网络,故不使用 dropout
y = mnist_lenet5_forward.forward(x, False, None)
ema = tf.train.ExponentialMovingAverage(mnist_lenet5_backward.MOVING_AVERAGE_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
# 判断预测值和实际值是否相同
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
# 求平均得到准确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
while True:
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(mnist_lenet5_backward.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
# 根据读入的模型名字切分出该模型是属于迭代了多少次保存的
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
reshaped_x = np.reshape(mnist.test.images, (
mnist.test.num_examples,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.IMAGE_SIZE,
mnist_lenet5_forward.NUM_CHANNELS))
# 利用多线程提高图片和标签的批获取效率
#coord = tf.train.Coordinator() # 3
#threads = tf.train.start_queue_runners(sess=sess, coord=coord) # 4
accuracy_score = sess.run(accuracy, feed_dict={x: reshaped_x, y_: mnist.test.labels})
print("After %s training step(s), test accuracy = %g" % (global_step, accuracy_score))
# 关闭线程协调器
#coord.request_stop() # 6
#coord.join(threads) # 7
else:
print('No checkpoint file found')
return
time.sleep(5)
def main():
mnist = input_data.read_data_sets("./data/", one_hot=True)
test(mnist)
if __name__ == '__main__':
main()