邳州城乡建设局网站,wordpress 自定义注册页面,专门做财经的网站,美食网站开发前期准备空间通道重组卷积SCConv
空间通道重组卷积SCConv#xff0c;全称Spatial and Channel Reconstruction Convolution#xff0c;CPR2023年提出#xff0c;可以即插即用#xff0c;能够在减少参数的同时提升性能的模块。其核心思想是希望能够实现减少特征冗余从而提高算法的效…空间通道重组卷积SCConv
空间通道重组卷积SCConv全称Spatial and Channel Reconstruction ConvolutionCPR2023年提出可以即插即用能够在减少参数的同时提升性能的模块。其核心思想是希望能够实现减少特征冗余从而提高算法的效率。一般压缩模型的方法分为三种分别是network pruning, weight quantization, low-rank factorization以及knowledge distillation虽然这些方法能够达到减少参数的效果但是往往都会导致模型性能的衰减。另一种方法就是在构建模型时利用特殊的模块或操作减少模型参数获得轻量级的网络模型这种方法能够在保证性能的同时达到参数减少的效果。
原文地址SCConv: Spatial and Channel Reconstruction Convolution for Feature Redundancy
作者提出的SCConv包含两部分分别是Spatial Reconstruction Unit (SRU)和Channel Reconstruction Unit (CRU)下面是SSConv的总体结构。 可以看出SCConv模块设计对于输入的特征图先利用1x1的卷积改变为适合的通道数之后便分别是SRU和CRU两个模块对于特征图进行处理最后在通过1x1的卷积将特征通道数恢复并进行残差操作。
代码实现如下
import torch
import torch.nn.functional as F
import torch.nn as nn class GroupBatchnorm2d(nn.Module):def __init__(self, c_num:int, group_num:int 16, eps:float 1e-10):super(GroupBatchnorm2d,self).__init__()assert c_num group_numself.group_num group_numself.gamma nn.Parameter( torch.randn(c_num, 1, 1) )self.beta nn.Parameter( torch.zeros(c_num, 1, 1) )self.eps epsdef forward(self, x):N, C, H, W x.size()x x.view( N, self.group_num, -1 )mean x.mean( dim 2, keepdim True )std x.std ( dim 2, keepdim True )x (x - mean) / (stdself.eps)x x.view(N, C, H, W)return x * self.gamma self.betaclass SRU(nn.Module):def __init__(self,oup_channels:int, group_num:int 16,gate_treshold:float 0.5 ):super().__init__()self.gn GroupBatchnorm2d( oup_channels, group_num group_num )self.gate_treshold gate_tresholdself.sigomid nn.Sigmoid()def forward(self,x):gn_x self.gn(x)w_gamma self.gn.gamma/sum(self.gn.gamma)reweigts self.sigomid( gn_x * w_gamma )# Gateinfo_mask reweigtsself.gate_tresholdnoninfo_mask reweigtsself.gate_tresholdx_1 info_mask * xx_2 noninfo_mask * xx self.reconstruct(x_1,x_2)return xdef reconstruct(self,x_1,x_2):x_11,x_12 torch.split(x_1, x_1.size(1)//2, dim1)x_21,x_22 torch.split(x_2, x_2.size(1)//2, dim1)return torch.cat([ x_11x_22, x_12x_21 ],dim1)class CRU(nn.Module):alpha: 0alpha1def __init__(self, op_channel:int,alpha:float 1/2,squeeze_radio:int 2 ,group_size:int 2,group_kernel_size:int 3,):super().__init__()self.up_channel up_channel int(alpha*op_channel)self.low_channel low_channel op_channel-up_channelself.squeeze1 nn.Conv2d(up_channel,up_channel//squeeze_radio,kernel_size1,biasFalse)self.squeeze2 nn.Conv2d(low_channel,low_channel//squeeze_radio,kernel_size1,biasFalse)#upself.GWC nn.Conv2d(up_channel//squeeze_radio, op_channel,kernel_sizegroup_kernel_size, stride1,paddinggroup_kernel_size//2, groups group_size)self.PWC1 nn.Conv2d(up_channel//squeeze_radio, op_channel,kernel_size1, biasFalse)#lowself.PWC2 nn.Conv2d(low_channel//squeeze_radio, op_channel-low_channel//squeeze_radio,kernel_size1, biasFalse)self.advavg nn.AdaptiveAvgPool2d(1)def forward(self,x):# Splitup,low torch.split(x,[self.up_channel,self.low_channel],dim1)up,low self.squeeze1(up),self.squeeze2(low)# TransformY1 self.GWC(up) self.PWC1(up)Y2 torch.cat( [self.PWC2(low), low], dim 1 )# Fuseout torch.cat( [Y1,Y2], dim 1 )out F.softmax( self.advavg(out), dim1 ) * outout1,out2 torch.split(out,out.size(1)//2,dim1)return out1out2class ScConv(nn.Module):def __init__(self,op_channel:int,group_num:int 16,gate_treshold:float 0.5,alpha:float 1/2,squeeze_radio:int 2 ,group_size:int 2,group_kernel_size:int 3,):super().__init__()self.SRU SRU( op_channel, group_num group_num, gate_treshold gate_treshold )self.CRU CRU( op_channel, alpha alpha, squeeze_radio squeeze_radio ,group_size group_size ,group_kernel_size group_kernel_size )def forward(self,x):x self.SRU(x)x self.CRU(x)return xif __name__ __main__:x torch.randn(1,32,16,16)model ScConv(32)print(model(x).shape)
