Paper https://arxiv.org/abs/1807.06521

https://github.com/Jongchan/attention-module

https://mp.weixin.qq.com/s?__biz=MzA4MjY4NTk0NQ==&mid=2247484531&idx=1&sn=625065862b28608428acb21da3330717&chksm=9f80bee5a8f737f399f0f564883337154dd8ca3ad5c246c85a86a88b0ac8ede7bf59ffc04554&token=897871599&lang=zh_CN&scene=21#wechat_redirect

image.png

代码

  1. import torch
  2. import math
  3. import torch.nn as nn
  4. import torch.nn.functional as F
  6. class BasicConv(nn.Module):
  7.     def __init__(self, in_planes, out_planes, kernel_size, stride=1, padding=0, dilation=1, groups=1, relu=True, bn=True, bias=False):
  8.         super(BasicConv, self).__init__()
  9.         self.out_channels = out_planes
  10.         self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
  11.         self.bn = nn.BatchNorm2d(out_planes,eps=1e-5, momentum=0.01, affine=True) if bn else None
  12.         self.relu = nn.ReLU() if relu else None
  14.     def forward(self, x):
  15.         x = self.conv(x)
  16.         if self.bn is not None:
  17.             x = self.bn(x)
  18.         if self.relu is not None:
  19.             x = self.relu(x)
  20.         return x
  22. class Flatten(nn.Module):
  23.     def forward(self, x):
  24.         return x.view(x.size(0), -1)
  26. class ChannelGate(nn.Module):
  27.     def __init__(self, gate_channels, reduction_ratio=16, pool_types=['avg', 'max']):
  28.         super(ChannelGate, self).__init__()
  29.         self.gate_channels = gate_channels
  30.         self.mlp = nn.Sequential(
  31.             Flatten(),
  32.             nn.Linear(gate_channels, gate_channels // reduction_ratio),
  33.             nn.ReLU(),
  34.             nn.Linear(gate_channels // reduction_ratio, gate_channels)
  35.             )
  36.         self.pool_types = pool_types
  37.     def forward(self, x):
  38.         channel_att_sum = None
  39.         for pool_type in self.pool_types:
  40.             if pool_type=='avg': # 平均池化
  41.                 avg_pool = F.avg_pool2d( x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
  42.                 channel_att_raw = self.mlp( avg_pool )
  43.             elif pool_type=='max': # 全局池化
  44.                 max_pool = F.max_pool2d( x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
  45.                 channel_att_raw = self.mlp( max_pool )
  46.             elif pool_type=='lp':
  47.                 lp_pool = F.lp_pool2d( x, 2, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
  48.                 channel_att_raw = self.mlp( lp_pool )
  49.             elif pool_type=='lse':
  50.                 # LSE pool only
  51.                 lse_pool = logsumexp_2d(x)
  52.                 channel_att_raw = self.mlp( lse_pool )
  54.             if channel_att_sum is None: # 首次计算
  55.                 channel_att_sum = channel_att_raw
  56.             else:
  57.                 channel_att_sum = channel_att_sum + channel_att_raw # 累加不同方式的池化
  59.         scale = torch.sigmoid( channel_att_sum ).unsqueeze(2).unsqueeze(3).expand_as(x)
  60.         return x * scale
  62. def logsumexp_2d(tensor):
  63.     """
  64.     tensor： (B, C, H, W)
  65.     """
  66.     tensor_flatten = tensor.view(tensor.size(0), tensor.size(1), -1) # (B, C, H×W)
  67.     s, _ = torch.max(tensor_flatten, dim=2, keepdim=True) # (B, C, 1×1) 每个通道中的最大值
  68.     outputs = s + (tensor_flatten - s).exp().sum(dim=2, keepdim=True).log()
  69.     return outputs
  71. class ChannelPool(nn.Module):
  72.     def forward(self, x):
  73.         # 找出 (H*W 平面每个点的在通道上最大值以及平均值)
  74.         return torch.cat( (torch.max(x,1)[0].unsqueeze(1), torch.mean(x,1).unsqueeze(1)), dim=1 )
  76. class SpatialGate(nn.Module):
  77.     def __init__(self, kernel_size):
  78.         super(SpatialGate, self).__init__()
  79.         self.kernel_size = kernel_size
  80.         self.compress = ChannelPool()
  81.         self.spatial = BasicConv(2, 1, self.kernel_size, stride=1, padding=(kernel_size-1) // 2, relu=False)
  82.     def forward(self, x):
  83.         x_compress = self.compress(x)
  84.         x_out = self.spatial(x_compress)
  85.         scale = torch.sigmoid(x_out) # broadcasting
  86.         return x * scale
  88. class CBAM(nn.Module):
  89.     def __init__(self, gate_channels, reduction_ratio=16, kernel_size=7, pool_types=['avg', 'max'], no_spatial=False):
  90.         super(CBAM, self).__init__()
  91.         self.ChannelGate = ChannelGate(gate_channels, reduction_ratio, pool_types)
  92.         self.no_spatial=no_spatial
  93.         if not no_spatial:
  94.             self.SpatialGate = SpatialGate(kernel_size)
  95.     def forward(self, x):
  96.         x_out = self.ChannelGate(x)
  97.         if not self.no_spatial:
  98.             x_out = self.SpatialGate(x_out)
  99.         return x_out