keras 实现轻量级网络ShuffleNet教程


Posted in Python onJune 19, 2020

ShuffleNet是由旷世发表的一个计算效率极高的CNN架构,它是专门为计算能力非常有限的移动设备(例如,10-150 MFLOPs)而设计的。该结构利用组卷积和信道混洗两种新的运算方法,在保证计算精度的同时,大大降低了计算成本。ImageNet分类和MS COCO对象检测实验表明,在40 MFLOPs的计算预算下,ShuffleNet的性能优于其他结构,例如,在ImageNet分类任务上,ShuffleNet的top-1 error 7.8%比最近的MobileNet低。在基于arm的移动设备上,ShuffleNet比AlexNet实际加速了13倍,同时保持了相当的准确性。

Paper:ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile

Github:https://github.com/zjn-ai/ShuffleNet-keras

网络架构

组卷积

组卷积其实早在AlexNet中就用过了,当时因为GPU的显存不足因而利用组卷积分配到两个GPU上训练。简单来讲,组卷积就是将输入特征图按照通道方向均分成多个大小一致的特征图,如下图所示左面是输入特征图右面是均分后的特征图,然后对得到的每一个特征图进行正常的卷积操作,最后将输出特征图按照通道方向拼接起来就可以了。

keras 实现轻量级网络ShuffleNet教程

目前很多框架都支持组卷积,但是tensorflow真的不知道在想什么,到现在还是不支持组卷积,只能自己写,因此效率肯定不及其他框架原生支持的方法。组卷积层的代码编写思路就与上面所说的原理完全一致,代码如下。

def _group_conv(x, filters, kernel, stride, groups):
 """
 Group convolution
 # Arguments
  x: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  filters: Integer, number of output channels
  kernel: An integer or tuple/list of 2 integers, specifying the
   width and height of the 2D convolution window.
  strides: An integer or tuple/list of 2 integers,
   specifying the strides of the convolution along the width and height.
   Can be a single integer to specify the same value for
   all spatial dimensions.
  groups: Integer, number of groups per channel
  
 # Returns
  Output tensor
 """
 channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
 in_channels = K.int_shape(x)[channel_axis]
 
 # number of input channels per group
 nb_ig = in_channels // groups
 # number of output channels per group
 nb_og = filters // groups
 
 gc_list = []
 # Determine whether the number of filters is divisible by the number of groups
 assert filters % groups == 0
 
 for i in range(groups):
  if channel_axis == -1:
   x_group = Lambda(lambda z: z[:, :, :, i * nb_ig: (i + 1) * nb_ig])(x)
  else:
   x_group = Lambda(lambda z: z[:, i * nb_ig: (i + 1) * nb_ig, :, :])(x)
  gc_list.append(Conv2D(filters=nb_og, kernel_size=kernel, strides=stride, 
        padding='same', use_bias=False)(x_group))
  
 return Concatenate(axis=channel_axis)(gc_list)

通道混洗

通道混洗是这篇paper的重点,尽管组卷积大量减少了计算量和参数,但是通道之间的信息交流也受到了限制因而模型精度肯定会受到影响,因此作者提出通道混洗,在不增加参数量和计算量的基础上加强通道之间的信息交流,如下图所示。

keras 实现轻量级网络ShuffleNet教程

通道混洗层的代码实现很巧妙参考了别人的实现方法。通过下面的代码说明,d代表特征图的通道序号,x是经过通道混洗后的通道顺序。

>>> d = np.array([0,1,2,3,4,5,6,7,8]) 
>>> x = np.reshape(d, (3,3)) 
>>> x = np.transpose(x, [1,0]) # 转置
>>> x = np.reshape(x, (9,)) # 平铺
'[0 1 2 3 4 5 6 7 8] --> [0 3 6 1 4 7 2 5 8]'

利用keras后端实现代码:

def _channel_shuffle(x, groups):
 """
 Channel shuffle layer
 
 # Arguments
  x: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  groups: Integer, number of groups per channel
  
 # Returns
  Shuffled tensor
 """
 
 if K.image_data_format() == 'channels_last':
  height, width, in_channels = K.int_shape(x)[1:]
  channels_per_group = in_channels // groups
  pre_shape = [-1, height, width, groups, channels_per_group]
  dim = (0, 1, 2, 4, 3)
  later_shape = [-1, height, width, in_channels]
 else:
  in_channels, height, width = K.int_shape(x)[1:]
  channels_per_group = in_channels // groups
  pre_shape = [-1, groups, channels_per_group, height, width]
  dim = (0, 2, 1, 3, 4)
  later_shape = [-1, in_channels, height, width]
 
 x = Lambda(lambda z: K.reshape(z, pre_shape))(x)
 x = Lambda(lambda z: K.permute_dimensions(z, dim))(x) 
 x = Lambda(lambda z: K.reshape(z, later_shape))(x)
 
 return x

ShuffleNet Unit

ShuffleNet的主要构成单元。下图中,a图为深度可分离卷积的基本架构,b图为1步长时用的单元,c图为2步长时用的单元。

keras 实现轻量级网络ShuffleNet教程

ShuffleNet架构

注意,对于第二阶段(Stage2),作者没有在第一个1×1卷积上应用组卷积,因为输入通道的数量相对较少。

keras 实现轻量级网络ShuffleNet教程

环境

Python 3.6

Tensorlow 1.13.1

Keras 2.2.4

实现

支持channel first或channel last

# -*- coding: utf-8 -*-
"""
Created on Thu Apr 25 18:26:41 2019
@author: zjn
"""
import numpy as np
from keras.callbacks import LearningRateScheduler
from keras.models import Model
from keras.layers import Input, Conv2D, Dropout, Dense, GlobalAveragePooling2D, Concatenate, AveragePooling2D
from keras.layers import Activation, BatchNormalization, add, Reshape, ReLU, DepthwiseConv2D, MaxPooling2D, Lambda
from keras.utils.vis_utils import plot_model
from keras import backend as K
from keras.optimizers import SGD
 
def _group_conv(x, filters, kernel, stride, groups):
 """
 Group convolution
 
 # Arguments
  x: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  filters: Integer, number of output channels
  kernel: An integer or tuple/list of 2 integers, specifying the
   width and height of the 2D convolution window.
  strides: An integer or tuple/list of 2 integers,
   specifying the strides of the convolution along the width and height.
   Can be a single integer to specify the same value for
   all spatial dimensions.
  groups: Integer, number of groups per channel
  
 # Returns
  Output tensor
 """
 
 channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
 in_channels = K.int_shape(x)[channel_axis]
 
 # number of input channels per group
 nb_ig = in_channels // groups
 # number of output channels per group
 nb_og = filters // groups
 
 gc_list = []
 # Determine whether the number of filters is divisible by the number of groups
 assert filters % groups == 0
 
 for i in range(groups):
  if channel_axis == -1:
   x_group = Lambda(lambda z: z[:, :, :, i * nb_ig: (i + 1) * nb_ig])(x)
  else:
   x_group = Lambda(lambda z: z[:, i * nb_ig: (i + 1) * nb_ig, :, :])(x)
  gc_list.append(Conv2D(filters=nb_og, kernel_size=kernel, strides=stride, 
        padding='same', use_bias=False)(x_group))
  
 return Concatenate(axis=channel_axis)(gc_list)
def _channel_shuffle(x, groups):
 """
 Channel shuffle layer
 
 # Arguments
  x: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  groups: Integer, number of groups per channel
  
 # Returns
  Shuffled tensor
 """
 if K.image_data_format() == 'channels_last':
  height, width, in_channels = K.int_shape(x)[1:]
  channels_per_group = in_channels // groups
  pre_shape = [-1, height, width, groups, channels_per_group]
  dim = (0, 1, 2, 4, 3)
  later_shape = [-1, height, width, in_channels]
 else:
  in_channels, height, width = K.int_shape(x)[1:]
  channels_per_group = in_channels // groups
  pre_shape = [-1, groups, channels_per_group, height, width]
  dim = (0, 2, 1, 3, 4)
  later_shape = [-1, in_channels, height, width]
 
 x = Lambda(lambda z: K.reshape(z, pre_shape))(x)
 x = Lambda(lambda z: K.permute_dimensions(z, dim))(x) 
 x = Lambda(lambda z: K.reshape(z, later_shape))(x)
 
 return x
 
def _shufflenet_unit(inputs, filters, kernel, stride, groups, stage, bottleneck_ratio=0.25):
 """
 ShuffleNet unit
 
 # Arguments
  inputs: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  filters: Integer, number of output channels
  kernel: An integer or tuple/list of 2 integers, specifying the
   width and height of the 2D convolution window.
  strides: An integer or tuple/list of 2 integers,
   specifying the strides of the convolution along the width and height.
   Can be a single integer to specify the same value for
   all spatial dimensions.
  groups: Integer, number of groups per channel
  stage: Integer, stage number of ShuffleNet
  bottleneck_channels: Float, bottleneck ratio implies the ratio of bottleneck channels to output channels
   
 # Returns
  Output tensor
  
 # Note
  For Stage 2, we(authors of shufflenet) do not apply group convolution on the first pointwise layer 
  because the number of input channels is relatively small.
 """
 channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
 in_channels = K.int_shape(inputs)[channel_axis]
 bottleneck_channels = int(filters * bottleneck_ratio)
 
 if stage == 2:
  x = Conv2D(filters=bottleneck_channels, kernel_size=kernel, strides=1,
     padding='same', use_bias=False)(inputs)
 else:
  x = _group_conv(inputs, bottleneck_channels, (1, 1), 1, groups)
 x = BatchNormalization(axis=channel_axis)(x)
 x = ReLU()(x)
 
 x = _channel_shuffle(x, groups)
 x = DepthwiseConv2D(kernel_size=kernel, strides=stride, depth_multiplier=1, 
      padding='same', use_bias=False)(x)
 x = BatchNormalization(axis=channel_axis)(x)
  
 if stride == 2:
  x = _group_conv(x, filters - in_channels, (1, 1), 1, groups)
  x = BatchNormalization(axis=channel_axis)(x)
  avg = AveragePooling2D(pool_size=(3, 3), strides=2, padding='same')(inputs)
  x = Concatenate(axis=channel_axis)([x, avg])
 else:
  x = _group_conv(x, filters, (1, 1), 1, groups)
  x = BatchNormalization(axis=channel_axis)(x)
  x = add([x, inputs])
 return x
 
def _stage(x, filters, kernel, groups, repeat, stage):
 """
 Stage of ShuffleNet
 
 # Arguments
  x: Tensor, input tensor of with `channels_last` or 'channels_first' data format
  filters: Integer, number of output channels
  kernel: An integer or tuple/list of 2 integers, specifying the
   width and height of the 2D convolution window.
  strides: An integer or tuple/list of 2 integers,
   specifying the strides of the convolution along the width and height.
   Can be a single integer to specify the same value for
   all spatial dimensions.
  groups: Integer, number of groups per channel
  repeat: Integer, total number of repetitions for a shuffle unit in every stage
  stage: Integer, stage number of ShuffleNet
  
 # Returns
  Output tensor
 """
 x = _shufflenet_unit(x, filters, kernel, 2, groups, stage)
 
 for i in range(1, repeat):
  x = _shufflenet_unit(x, filters, kernel, 1, groups, stage)
 return x
 
def ShuffleNet(input_shape, classes):
 """
 ShuffleNet architectures
 
 # Arguments
  input_shape: An integer or tuple/list of 3 integers, shape
   of input tensor
  k: Integer, number of classes to predict
  
 # Returns
  A keras model
 """
 inputs = Input(shape=input_shape)
 
 x = Conv2D(24, (3, 3), strides=2, padding='same', use_bias=True, activation='relu')(inputs)
 x = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(x)
 
 x = _stage(x, filters=384, kernel=(3, 3), groups=8, repeat=4, stage=2)
 x = _stage(x, filters=768, kernel=(3, 3), groups=8, repeat=8, stage=3)
 x = _stage(x, filters=1536, kernel=(3, 3), groups=8, repeat=4, stage=4)
 
 x = GlobalAveragePooling2D()(x)
 
 x = Dense(classes)(x)
 predicts = Activation('softmax')(x)
 model = Model(inputs, predicts)
 return model
 
if __name__ == '__main__':
 model = ShuffleNet((224, 224, 3), 1000)
 #plot_model(model, to_file='ShuffleNet.png', show_shapes=True)

以上这篇keras 实现轻量级网络ShuffleNet教程就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持三水点靠木。

Python 相关文章推荐
使用Python下载Bing图片(代码)
Nov 07 Python
python实现统计代码行数的方法
May 22 Python
json跨域调用python的方法详解
Jan 11 Python
搞清楚 Python traceback的具体使用方法
May 13 Python
Python3批量生成带logo的二维码方法
Jun 24 Python
python 随机生成10位数密码的实现代码
Jun 27 Python
Python 使用PyQt5 完成选择文件或目录的对话框方法
Jun 27 Python
Django admin.py 在修改/添加表单界面显示额外字段的方法
Aug 22 Python
python rsync服务器之间文件夹同步脚本
Aug 29 Python
python tkinter基本属性详解
Sep 16 Python
Python 批量读取文件中指定字符的实现
Mar 06 Python
深入了解python列表(LIST)
Jun 08 Python
Python爬虫实现HTTP网络请求多种实现方式
Jun 19 #Python
Keras设置以及获取权重的实现
Jun 19 #Python
Python包和模块的分发详细介绍
Jun 19 #Python
浅谈Keras中shuffle和validation_split的顺序
Jun 19 #Python
Python爬虫headers处理及网络超时问题解决方案
Jun 19 #Python
sklearn和keras的数据切分与交叉验证的实例详解
Jun 19 #Python
Python虚拟环境的创建和包下载过程分析
Jun 19 #Python
You might like
php miniBB中文乱码问题解决方法
2008/11/25 PHP
PHP 面向对象实现代码
2009/11/11 PHP
PHP 截取字符串函数整理(支持gb2312和utf-8)
2010/02/16 PHP
php.ini修改php上传文件大小限制的方法详解
2013/06/17 PHP
解析thinkphp的左右值无限分类
2013/06/20 PHP
PHP XML和数组互相转换详解
2016/10/26 PHP
php基于闭包实现函数的自调用(递归)实例分析
2016/11/11 PHP
Javascript valueOf 使用方法
2008/12/28 Javascript
网页整体变灰白色(兼容各浏览器)实例
2013/04/21 Javascript
zeroclipboard 单个复制按钮和多个复制按钮的实现方法
2014/06/14 Javascript
招聘网站基于jQuery实现自动刷新简历
2015/05/10 Javascript
基于jQuery+JSON的省市二三级联动效果
2015/06/05 Javascript
jQuery超精致图片轮播幻灯片特效代码分享
2015/09/10 Javascript
JS前端笔试题分析
2016/12/19 Javascript
微信小程序 使用canvas制作K线实例详解
2017/01/12 Javascript
JavaScript中三种常见的排序方法
2017/02/24 Javascript
基于node.js的fs核心模块读写文件操作(实例讲解)
2017/09/10 Javascript
vue.js vue-router如何实现无效路由(404)的友好提示
2017/12/20 Javascript
vue实现的上传图片到数据库并显示到页面功能示例
2018/03/17 Javascript
JS实现的判断方法、变量是否存在功能示例
2020/03/28 Javascript
JavaScript实现移动小精灵的案例代码
2020/12/12 Javascript
深入解析Python中的变量和赋值运算符
2015/10/12 Python
Python 爬虫之超链接 url中含有中文出错及解决办法
2017/08/03 Python
Python使用MD5加密算法对字符串进行加密操作示例
2018/03/30 Python
python 统计数组中元素出现次数并进行排序的实例
2018/07/02 Python
Pycharm 2019 破解激活方法图文详解
2019/10/11 Python
英国女性时尚鞋类的潮流制造者:Koi Footwear
2018/10/19 全球购物
写一个函数返回1+2+3+…+n的值(假定结果不会超过长整型变量的范围)
2014/09/05 面试题
自考生自我评价分享
2014/01/18 职场文书
百年校庆节目主持词
2014/03/27 职场文书
个人四风问题对照检查材料
2014/10/01 职场文书
2014教师评职称工作总结
2014/11/10 职场文书
Filebeat 采集 Nginx 日志的方法
2021/03/31 Servers
Unity连接MySQL并读取表格数据的实现代码
2021/06/20 MySQL
MongoDB连接数据库并创建数据等使用方法
2021/11/27 MongoDB
gojs实现蚂蚁线动画效果
2022/02/18 Javascript