白云岛资源网

keras 实现轻量级网络ShuffleNet教程
站长资源 2024/12/26 佚名
52 1538
白云岛资源网 Design By www.pvray.com

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。

白云岛资源网 Design By www.pvray.com
广告合作：本站广告合作请联系QQ：858582 申请时备注：广告合作（否则不回）
免责声明：本站资源来自互联网收集,仅供用于学习和交流,请遵循相关法律法规,本站一切资源不代表本站立场,如有侵权、后门、不妥请联系本站删除！
白云岛资源网 Design By www.pvray.com
www.pvray.com 白云岛资源网
39,976影音资源
44,792技术资源
21,817软件资源
651,128站长资源

P70系列延期，华为新旗舰将在下月发布

3月20日消息，近期博主@数码闲聊站 透露，原定三月份发布的华为新旗舰P70系列延期发布，预计4月份上市。

而博主@定焦数码 爆料，华为的P70系列在定位上已经超过了Mate60，成为了重要的旗舰系列之一。它肩负着重返影像领域顶尖的使命。那么这次P70会带来哪些令人惊艳的创新呢？

根据目前爆料的消息来看，华为P70系列将推出三个版本，其中P70和P70 Pro采用了三角形的摄像头模组设计，而P70 Art则采用了与上一代P60 Art相似的不规则形状设计。这样的外观是否好看见仁见智，但辨识度绝对拉满。

友情链接

杰晶网络 DDR爱好者之家 桃源资源网 杰网资源 富贵资源网 南强小屋 铁雪资源网 幽灵资源网 万梅资源网 狼山资源网 白云岛资源网 昆仑资源网 相思资源网 明霞山资源网 内蒙古资源网 黑松山资源网 茶园资源网 饿虎岗资源网 大旗谷资源网 常春岛资源网 岱庙资源网 兴国资源网 快活林资源网 蝙蝠岛资源网 帝王谷资源网 白云城资源网 伏龙阁资源网 清风细雨楼 天枫庄资源网 圆月山庄资源网 无争山庄资源网 神水资源网 移花宫资源网 神剑山庄资源网 无为清净楼资源网 金钱帮资源网 丐帮资源网 华山资源网 极乐门资源网 小李飞刀资源网 凤求凰客栈 风云阁资源网 金狮镖局 鸳鸯亭资源网 千金楼资源网 更多链接
白云岛资源网 Design By www.pvray.com