当前位置:网站首页>tensorflow2.0 cnn(layerwise)
tensorflow2.0 cnn(layerwise)
2022-07-31 15:37:00 【全栈程序员站长】
大家好,又见面了,我是你们的朋友全栈君。
实验环境:tensorflow版本1.2.0,python2.7
介绍
depthwise_conv2d来源于深度可分离卷积:
Xception: Deep Learning with Depthwise Separable Convolutions
tf.nn.depthwise_conv2d(input,filter,strides,padding,rate=None,name=None,data_format=None)除去name参数用以指定该操作的name,data_format指定数据格式,与方法有关的一共五个参数:
- input: 指需要做卷积的输入图像,要求是一个4维Tensor,具有
[batch, height, width, in_channels]这样的shape,具体含义是[训练时一个batch的图片数量, 图片高度, 图片宽度, 图像通道数] - filter: 相当于CNN中的卷积核,要求是一个4维Tensor,具有
[filter_height, filter_width, in_channels, channel_multiplier]这样的shape,具体含义是[卷积核的高度,卷积核的宽度,输入通道数,输出卷积乘子],同理这里第三维in_channels,就是参数value的第四维 - strides: 卷积的滑动步长。
- padding: string类型的量,只能是”SAME”,”VALID”其中之一,这个值决定了不同边缘填充方式。
- rate: 这个参数的详细解释见【Tensorflow】tf.nn.atrous_conv2d如何实现空洞卷积?
结果返回一个Tensor,shape为[batch, out_height, out_width, in_channels * channel_multiplier],注意这里输出通道变成了in_channels * channel_multiplier
实验
为了形象的展示depthwise_conv2d,我们必须要建立自定义的输入图像和卷积核
img1 = tf.constant(value=[[[[1],[2],[3],[4]],[[1],[2],[3],[4]],[[1],[2],[3],[4]],[[1],[2],[3],[4]]]],dtype=tf.float32)
img2 = tf.constant(value=[[[[1],[1],[1],[1]],[[1],[1],[1],[1]],[[1],[1],[1],[1]],[[1],[1],[1],[1]]]],dtype=tf.float32)
img = tf.concat(values=[img1,img2],axis=3)filter1 = tf.constant(value=0, shape=[3,3,1,1],dtype=tf.float32)
filter2 = tf.constant(value=1, shape=[3,3,1,1],dtype=tf.float32)
filter3 = tf.constant(value=2, shape=[3,3,1,1],dtype=tf.float32)
filter4 = tf.constant(value=3, shape=[3,3,1,1],dtype=tf.float32)
filter_out1 = tf.concat(values=[filter1,filter2],axis=2)
filter_out2 = tf.concat(values=[filter3,filter4],axis=2)
filter = tf.concat(values=[filter_out1,filter_out2],axis=3)建立好了img和filter,就可以做卷积了
out_img = tf.nn.conv2d(input=img, filter=filter, strides=[1,1,1,1], padding='VALID')好了,用一张图来详细展示这个过程
这是普通的卷积过程,我们再来看深度卷积。
out_img = tf.nn.depthwise_conv2d(input=img, filter=filter, strides=[1,1,1,1], rate=[1,1], padding='VALID')现在我们可以形象的解释一下depthwise_conv2d卷积了。看普通的卷积,我们对卷积核每一个out_channel的两个通道分别和输入的两个通道做卷积相加,得到feature map的一个channel,而depthwise_conv2d卷积,我们对每一个对应的in_channel,分别卷积生成两个out_channel,所以获得的feature map的通道数量可以用in_channel* channel_multiplier来表达,这个channel_multiplier,就可以理解为卷积核的第四维。
代码清单
import tensorflow as tf
img1 = tf.constant(value=[[[[1],[2],[3],[4]],[[1],[2],[3],[4]],[[1],[2],[3],[4]],[[1],[2],[3],[4]]]],dtype=tf.float32)
img2 = tf.constant(value=[[[[1],[1],[1],[1]],[[1],[1],[1],[1]],[[1],[1],[1],[1]],[[1],[1],[1],[1]]]],dtype=tf.float32)
img = tf.concat(values=[img1,img2],axis=3)
filter1 = tf.constant(value=0, shape=[3,3,1,1],dtype=tf.float32)
filter2 = tf.constant(value=1, shape=[3,3,1,1],dtype=tf.float32)
filter3 = tf.constant(value=2, shape=[3,3,1,1],dtype=tf.float32)
filter4 = tf.constant(value=3, shape=[3,3,1,1],dtype=tf.float32)
filter_out1 = tf.concat(values=[filter1,filter2],axis=2)
filter_out2 = tf.concat(values=[filter3,filter4],axis=2)
filter = tf.concat(values=[filter_out1,filter_out2],axis=3)
out_img = tf.nn.depthwise_conv2d(input=img, filter=filter, strides=[1,1,1,1], rate=[1,1], padding='VALID')输出:
rate=1, VALID mode result:
[[[[ 0. 36. 9. 27.] [ 0. 54. 9. 27.]] [[ 0. 36. 9. 27.] [ 0. 54. 9. 27.]]]]发布者:全栈程序员栈长,转载请注明出处:https://javaforall.cn/127979.html原文链接:https://javaforall.cn
边栏推荐
猜你喜欢

After Grafana is installed, the web opens and reports an error

基于ABP实现DDD

女性服务社群产品设计

.NET 20周年专访 - 张善友:.NET 技术是如何赋能并改变世界的

BGP综合实验(建立对等体、路由反射器、联邦、路由宣告及聚合)

mongo enters error

工程流体力学复习

Female service community product design

TRACE32 - SNOOPer-based variable logging
![[Meetup Preview] OpenMLDB+OneFlow: Link feature engineering to model training to accelerate machine learning model development](/img/f6/311d5a4c70993df6291250d2025d3f.jpg)
[Meetup Preview] OpenMLDB+OneFlow: Link feature engineering to model training to accelerate machine learning model development
随机推荐
json到底是什么(c# json)
The use of border controls
Excel quickly aligns the middle name of the table (two-word name and three-word name alignment)
四象限时间管理有多好用?
Destruction order of thread_local variables
【MySQL】Mysql范式及外键作用
Deployment应用生命周期与Pod健康检查
Linux查看redis版本(查看mongodb版本)
Tencent Cloud Deployment----DevOps
The principle of hough transform detection of straight lines (opencv hough straight line detection)
update data table update
[Meetup Preview] OpenMLDB+OneFlow: Link feature engineering to model training to accelerate machine learning model development
Female service community product design
11 pinia使用
Qt实战案例(54)——利用QPixmap设计图片透明度
Bilateral filtering acceleration "recommended collection"
TRACE32——常用操作
多主复制的适用场景(1)-多IDC
Snake Project (Simple)
Kubernetes principle analysis and practical application manual, too complete