前言:前面介绍了最基本的Lenet,下面介绍几种其他的网络结构
CNN-AlexNet
网络结构如下图:
从图中可以看出,采用双gpu训练
增加LRN归一化层:本质上,这个层也是为了防止激活函数的饱和的。
采用dropout防止过拟合
基于AlexNet进行微调 ,诞生了ZF-net
CNN-GoogleNet
GoogLeNet借鉴了NIN的特性,在原先的卷积过程中附加了11的 卷积核加上ReLU激活。
这不仅仅提升了网络的深度,提高了representation power,而且 文中还通过11的卷积来进行降维,减少了更新参数量。
-
NIN模型
Network-in-Network主要思想是,用全连接的多层感知机去代替 传统的卷积过程,以获取特征更加全面的表达,同时,因为前面已 经做了提升特征表达的过程,传统CNN最后的全连接层也被替换为 一个全局平均池化层,因为作者认为此时的map已经具备分类足够 的可信度了,它可以直接通过softmax来计算loss了。
NIN模型 -
在计算要求增加很多的地方应用维度缩减 和预测。即,在3x3和5x5的卷积前用一个1x1的卷积用于减少计算, 还用于修正线性激活。如下图所示,左边是加入维度缩减之前的, 右边是加入维度缩减之后的。
最终网络结构如下:
其他网络结构
还有VGG-net,RES-net等,结构都是在最基本的网络结构上进行改进,比如R esNet结构
使用了一种连接方式叫做“shortcut connection”,顾名思义, shortcut就是“抄近道”的意思。
参数更新用了RNN的思想
下面给一个CNN-VGGNet15层代码的示例,用来判断17种花的一个分类情况
def vgg_network(x, y):
net1_kernel_size = 32
net3_kernel_size = 64
net5_kernal_size_1 = 128
net5_kernal_size_2 = 128
net7_kernal_size_1 = 256
net7_kernal_size_2 = 256
net9_kernal_size_1 = 256
net9_kernal_size_2 = 256
net11_unit_size = 1000
net12_unit_size = 1000
net13_unit_size = 17
# cov3-64 lrn
with tf.variable_scope('net1'):
net = tf.nn.conv2d(x, filter=get_variable('w', [3, 3, 3, net1_kernel_size]), strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [net1_kernel_size]))
net = tf.nn.relu(net)
# lrn(input, depth_radius=5, bias=1, alpha=1, beta=0.5, name=None)
# 做一个局部响应归一化,是对卷积核的输出值做归一化
# depth_radius ==> 对应ppt公式上的n,bias => 对应ppt公式上的k, alpha => 对应ppt公式上的α, beta=>对应ppt公式上的β
net = tf.nn.lrn(net)
# maxpool
with tf.variable_scope('net2'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-128
with tf.variable_scope('net3'):
net = tf.nn.conv2d(net, filter=get_variable('w', [3, 3, net1_kernel_size, net3_kernel_size]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b', [net3_kernel_size]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net4'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-256 conv3-256
with tf.variable_scope('net5'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net3_kernel_size, net5_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net5_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net5_kernal_size_1, net5_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net5_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net6'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-512 conv3-512
with tf.variable_scope('net7'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net5_kernal_size_2, net7_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net7_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net7_kernal_size_1, net7_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net7_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net8'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# conv3-512 conv3-512
with tf.variable_scope('net9'):
net = tf.nn.conv2d(net, filter=get_variable('w1', [3, 3, net7_kernal_size_2, net9_kernal_size_1]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b1', [net9_kernal_size_1]))
net = tf.nn.relu(net)
net = tf.nn.conv2d(net, filter=get_variable('w2', [3, 3, net9_kernal_size_1, net9_kernal_size_2]),
strides=[1, 1, 1, 1],
padding='SAME')
net = tf.nn.bias_add(net, get_variable('b2', [net9_kernal_size_2]))
net = tf.nn.relu(net)
# maxpool
with tf.variable_scope('net10'):
net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# fc
with tf.variable_scope('net11'):
# 将四维的数据转换为两维的数据
shape = net.get_shape()
feature_number = shape[1] * shape[2] * shape[3]
net = tf.reshape(net, shape=[-1, feature_number])
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [feature_number, net11_unit_size])),
get_variable('b', [net11_unit_size]))
# fc
with tf.variable_scope('net12'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [net11_unit_size, net12_unit_size])),
get_variable('b', [net12_unit_size]))
# fc
with tf.variable_scope('net13'):
# 全连接
net = tf.add(tf.matmul(net, get_variable('w', [net12_unit_size, net13_unit_size])),
get_variable('b', [net13_unit_size]))
# softmax
with tf.variable_scope('net14'):
# softmax
act = tf.nn.softmax(net)
return act
网友评论