摘要
本文介绍我自研的下采样模块。本次改进的下采样模块是一种通用的改进方法，你可以用分类任务的主干网络中，也可以用在分割和超分的任务中。已经有粉丝用来改进ConvNext模型，取得了非常好的效果，配合一些其他的改进，发一篇CVPR、ECCV之类的顶会完全没有问题。
本次我将这个模块用来改进YoloV9，实现大幅度涨点。
自研下采样模块及其变种
第一种改进方法
将输入分成两个分支，一个分支用卷积，一个分支分成两部分，一部分用MaxPool，一部分用AvgPool。然后，在最后合并起来。代码如下：
import torch
import torch.nn as nn
def autopad(k, p=None, d=1): # kernel, padding, dilation
"""Pad to 'same' shape outputs."""
if d > 1:
k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k] # actual kernel-size
if p is None:
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
return p
class Conv(nn.Module):
"""Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""
default_act = nn.SiLU() # default activation
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
"""Initialize Conv layer with given arguments including activation."""
super().__init__()
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
self.bn = nn.BatchNorm2d(c2)
self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
def forward(self, x):
"""Apply convolution, batch normalization and activation to input tensor."""
return self.act(self.bn(self.conv(x)))
def forward_fuse(self, x):
"""Perform transposed convolution of 2D data."""
return self.act(self.conv(x))
class DownSimper(nn.Module):
"""DownSimper."""
def __init__(self, c1, c2):
super().__init__()
self.c = c2 // 2
self.cv1 = Conv(c1, self.c, 3, 2, d=3)
self.cv2 = Conv(c1, self.c, 1, 1, 0)
def forward(self, x):
x1 = self.cv1(x)
x = self.cv2(x)
x2, x3 = x.chunk(2, 1)
x2 = torch.nn.functional.max_pool2d(x2, 3, 2, 1)
x3 = torch.nn.functional.avg_pool2d(x3, 3, 2, 1)
return torch.cat((x1, x2, x3), 1)
结构图：
左侧卷积中d=3，代表使用空洞卷积或者是膨胀卷积，可以带来更大的感受野。d=3，k=3等同卷积核为9.
YoloV9官方测试结果
yolov9 summary: 580 layers, 60567520 parameters, 0 gradients, 264.3 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 15/15 00:02
all 230 1412 0.878 0.991 0.989 0.732
c17 230 131 0.92 0.992 0.994 0.797
c5 230 68 0.828 1 0.992 0.807
helicopter 230 43 0.895 0.977 0.969 0.634
c130 230 85 0.955 0.999 0.994 0.684
f16 230 57 0.839 0.965 0.966 0.689
b2 230 2 1 0.978 0.995 0.647
other 230 86 0.91 0.942 0.957 0.525
b52 230 70 0.917 0.971 0.979 0.806
kc10 230 62 0.958 0.984 0.987 0.826
command 230 40 0.964 1 0.995 0.815
f15 230 123 0.939 0.995 0.995 0.702
kc135 230 91 0.949 0.989 0.978 0.691
a10 230 27 0.863 0.963 0.982 0.458
b1 230 20 0.926 1 0.995 0.712
aew 230 25 0.929 1 0.993 0.812
f22 230 17 0.835 1 0.995 0.706
p3 230 105 0.97 1 0.995 0.804
p8 230 1 0.566 1 0.995 0.697
f35 230 32 0.908 1 0.995 0.547
f18 230 125 0.956 0.992 0.993 0.828
v22 230 41 0.921 1 0.995 0.682
su-27 230 31 0.925 1 0.994 0.832
il-38 230 27 0.899 1 0.995 0.816
tu-134 230 1 0.346 1 0.995 0.895
su-33 230 2 0.96 1 0.995 0.747
an-70 230 2 0.718 1 0.995 0.796
tu-22 230 98 0.912 1 0.995 0.804
改进方法
测试结果
yolov9 summary: 595 layers, 58708576 parameters, 0 gradients, 274.0 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 15/15 00:35
all 230 1412 0.952 0.974 0.99 0.738
c17 230 131 0.981 0.992 0.995 0.832
c5 230 68 0.963 0.985 0.995 0.847
helicopter 230 43 0.968 0.93 0.972 0.635
c130 230 85 0.988 0.996 0.995 0.669
f16 230 57 0.976 0.947 0.975 0.687
b2 230 2 0.767 1 0.995 0.516
other 230 86 0.981 0.907 0.968 0.573
b52 230 70 0.969 0.971 0.985 0.812
kc10 230 62 0.986 0.984 0.989 0.835
command 230 40 0.988 1 0.995 0.82
f15 230 123 0.965 0.992 0.989 0.697
kc135 230 91 0.984 0.989 0.981 0.725
a10 230 27 1 0.794 0.976 0.495
b1 230 20 0.979 1 0.995 0.682
aew 230 25 0.944 1 0.995 0.802
f22 230 17 1 0.881 0.992 0.717
p3 230 105 0.981 0.992 0.995 0.81
p8 230 1 0.756 1 0.995 0.697
f35 230 32 0.99 0.938 0.982 0.55
f18 230 125 0.981 0.992 0.991 0.829
v22 230 41 0.99 1 0.995 0.684
su-27 230 31 0.985 1 0.995 0.849
il-38 230 27 0.987 1 0.995 0.84
tu-134 230 1 0.756 1 0.995 0.895
su-33 230 2 0.99 1 0.995 0.747
an-70 230 2 0.838 1 0.995 0.848
tu-22 230 98 1 1 0.995 0.832