Spaces:
Sleeping
Sleeping
File size: 3,337 Bytes
28c6826 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 |
""" CBAM (sort-of) Attention
Experimental impl of CBAM: Convolutional Block Attention Module: https://arxiv.org/abs/1807.06521
WARNING: Results with these attention layers have been mixed. They can significantly reduce performance on
some tasks, especially fine-grained it seems. I may end up removing this impl.
Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
from torch import nn as nn
import torch.nn.functional as F
from .conv_bn_act import ConvBnAct
class ChannelAttn(nn.Module):
""" Original CBAM channel attention module, currently avg + max pool variant only.
"""
def __init__(self, channels, reduction=16, act_layer=nn.ReLU):
super(ChannelAttn, self).__init__()
self.fc1 = nn.Conv2d(channels, channels // reduction, 1, bias=False)
self.act = act_layer(inplace=True)
self.fc2 = nn.Conv2d(channels // reduction, channels, 1, bias=False)
def forward(self, x):
x_avg = x.mean((2, 3), keepdim=True)
x_max = F.adaptive_max_pool2d(x, 1)
x_avg = self.fc2(self.act(self.fc1(x_avg)))
x_max = self.fc2(self.act(self.fc1(x_max)))
x_attn = x_avg + x_max
return x * x_attn.sigmoid()
class LightChannelAttn(ChannelAttn):
"""An experimental 'lightweight' that sums avg + max pool first
"""
def __init__(self, channels, reduction=16):
super(LightChannelAttn, self).__init__(channels, reduction)
def forward(self, x):
x_pool = 0.5 * x.mean((2, 3), keepdim=True) + 0.5 * F.adaptive_max_pool2d(x, 1)
x_attn = self.fc2(self.act(self.fc1(x_pool)))
return x * x_attn.sigmoid()
class SpatialAttn(nn.Module):
""" Original CBAM spatial attention module
"""
def __init__(self, kernel_size=7):
super(SpatialAttn, self).__init__()
self.conv = ConvBnAct(2, 1, kernel_size, act_layer=None)
def forward(self, x):
x_avg = torch.mean(x, dim=1, keepdim=True)
x_max = torch.max(x, dim=1, keepdim=True)[0]
x_attn = torch.cat([x_avg, x_max], dim=1)
x_attn = self.conv(x_attn)
return x * x_attn.sigmoid()
class LightSpatialAttn(nn.Module):
"""An experimental 'lightweight' variant that sums avg_pool and max_pool results.
"""
def __init__(self, kernel_size=7):
super(LightSpatialAttn, self).__init__()
self.conv = ConvBnAct(1, 1, kernel_size, act_layer=None)
def forward(self, x):
x_avg = torch.mean(x, dim=1, keepdim=True)
x_max = torch.max(x, dim=1, keepdim=True)[0]
x_attn = 0.5 * x_avg + 0.5 * x_max
x_attn = self.conv(x_attn)
return x * x_attn.sigmoid()
class CbamModule(nn.Module):
def __init__(self, channels, spatial_kernel_size=7):
super(CbamModule, self).__init__()
self.channel = ChannelAttn(channels)
self.spatial = SpatialAttn(spatial_kernel_size)
def forward(self, x):
x = self.channel(x)
x = self.spatial(x)
return x
class LightCbamModule(nn.Module):
def __init__(self, channels, spatial_kernel_size=7):
super(LightCbamModule, self).__init__()
self.channel = LightChannelAttn(channels)
self.spatial = LightSpatialAttn(spatial_kernel_size)
def forward(self, x):
x = self.channel(x)
x = self.spatial(x)
return x
|