tensorflow2.0 cnn(layerwise)

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_conv2dDerived from depthwise separable convolution:

Xception: Deep Learning with Depthwise Separable Convolutions

tf.nn.depthwise_conv2d(input,filter,strides,padding,rate=None,name=None,data_format=None)

复制

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,output convolution multiplier],The same goes for the third dimension herein_channels,就是参数value的第四维
• strides： The sliding step size of the convolution.
• padding： string类型的量,只能是”SAME”,”VALID”其中之一,This value determines how different edges are filled.
• rate： For a detailed explanation of this parameter, see 【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],Note that here the output channel becomesin_channels * channel_multiplier

实验

for image displaydepthwise_conv2d,We have to create custom input images and convolution kernels

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)

复制

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)

复制

建立好了img和filter,You can do convolution

out_img = tf.nn.conv2d(input=img, filter=filter, strides=[1,1,1,1], padding='VALID')

复制

好了,Use a diagram to illustrate this process in detail

This is the normal convolution process,Let's look at depthwise convolution again.

out_img = tf.nn.depthwise_conv2d(input=img, filter=filter, strides=[1,1,1,1], rate=[1,1], padding='VALID')

复制

Now we can explain it visuallydepthwise_conv2d卷积了.See ordinary convolution,We apply convolution kernels to eachout_channelThe two channels of , respectively, are convolved and added to the two channels of the input,得到feature map的一个channel,而depthwise_conv2d卷积,We correspond to each onein_channel,Convolution to generate two respectivelyout_channel,所以获得的feature mapnumber of channels availablein_channel* channel_multiplier来表达,这个channel_multiplier,It can be understood as the fourth dimension of the convolution kernel.

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')

复制

输出：

rate=1, VALID mode result:
[[[[ 0. 36. 9. 27.] [ 0. 54. 9. 27.]] [[ 0. 36. 9. 27.] [ 0. 54. 9. 27.]]]]

复制

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