devbox
我家自动化
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

YOLOv9独家原创改进|使用可改变核卷积AKConv改进RepNCSPELAN4

作者：我家自动化 | 2024-04-09 15:21:02

赞

踩

repncspelan4

专栏介绍：YOLOv9改进系列 | 包含深度学习最新创新，主力高效涨点！！！

一、改进点介绍

        AKConv是一种具有任意数量的参数和任意采样形状的可变卷积核，对不规则特征有更好的提取效果。

        RepNCSPELAN4是YOLOv9中的特征提取模块，类似YOLOv5和v8中的C2f与C3模块。

二、RepNCSPELAN4-AKConv模块详解

 2.1 模块简介

        RepNCSPELAN4-AKConv的主要思想：  使用AKConv替换RepNCSPELAN4中的Conv模块。

三、 RepNCSPELAN4-AKConv模块使用教程

3.1 RepNCSPELAN4-AKConv模块的代码

  1. class RepNCSP_AKConv(RepNCSP):
  2.     # CSP Bottleneck with 3 convolutions
  3.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
  4.         super().__init__(c1, c2, n, shortcut, g, e)
  5.         c_ = int(c2 * e)  # hidden channels
  6.         self.cv1 = AKConv(c1, c_)
  7.         self.cv2 = AKConv(c1, c_)
  8.         self.cv3 = AKConv(2 * c_, c2)  # optional act=FReLU(c2)
  9.         self.m = nn.Sequential(*(RepNBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
  10.  
  11.  
  12. class RepNCSPELAN4AKConv1(RepNCSPELAN4):
  13.     def __init__(self, c1, c2, c3, c4, c5=1):  # ch_in, ch_out, number, shortcut, groups, expansion
  14.         super().__init__(c1, c2, c3, c4, c5)
  15.         self.cv1 = Conv(c1, c3, 1, 1)
  16.         self.cv2 = nn.Sequential(RepNCSP_AKConv(c3//2, c4, c5), Conv(c4, c4, 3, 1))
  17.         self.cv3 = nn.Sequential(RepNCSP_AKConv(c4, c4, c5), Conv(c4, c4, 3, 1))
  18.         self.cv4 = AKConv(c3+(2*c4), c2, 1, 1)
  19.  
  20. from einops import rearrange
  21.  
  22. class AKConv(nn.Module):
  23.     def __init__(self, inc, outc, num_param=5, stride=1):
  24.         """
  25.         初始化参数说明：
  26.             inc: 输入通道数, outc: 输出通道数, num_param:（卷积核）参数量, stride = 1:卷积步长默认为1, bias = None：默认无偏执
  27.             """
  28.         super(AKConv, self).__init__()
  29.         self.num_param = num_param
  30.         self.stride = stride
  31.         self.conv = Conv(inc, outc, k=(num_param, 1), s=(num_param, 1) )
  32.         self.p_conv = nn.Conv2d(inc, 2 * num_param, kernel_size=3, padding=1, stride=stride)
  33.         nn.init.constant_(self.p_conv.weight, 0)
  34.         self.p_conv.register_full_backward_hook(self._set_lr)
  35.  
  36.     @staticmethod
  37.     def _set_lr(module, grad_input, grad_output):
  38.         grad_input = (grad_input[i] * 0.1 for i in range(len(grad_input)))
  39.         grad_output = (grad_output[i] * 0.1 for i in range(len(grad_output)))
  40.  
  41.     def forward(self, x):
  42.         # N is num_param.
  43.         offset = self.p_conv(x)
  44.         dtype = offset.data.type()
  45.         N = offset.size(1) // 2
  46.         # (b, 2N, h, w)
  47.         p = self._get_p(offset, dtype)
  48.  
  49.         # (b, h, w, 2N)
  50.         p = p.contiguous().permute(0, 2, 3, 1)
  51.         q_lt = p.detach().floor()
  52.         q_rb = q_lt + 1
  53.  
  54.         q_lt = torch.cat([torch.clamp(q_lt[..., :N], 0, x.size(2) - 1), torch.clamp(q_lt[..., N:], 0, x.size(3) - 1)],
  55.                          dim=-1).long()
  56.         q_rb = torch.cat([torch.clamp(q_rb[..., :N], 0, x.size(2) - 1), torch.clamp(q_rb[..., N:], 0, x.size(3) - 1)],
  57.                          dim=-1).long()
  58.         q_lb = torch.cat([q_lt[..., :N], q_rb[..., N:]], dim=-1)
  59.         q_rt = torch.cat([q_rb[..., :N], q_lt[..., N:]], dim=-1)
  60.  
  61.         # clip p
  62.         p = torch.cat([torch.clamp(p[..., :N], 0, x.size(2) - 1), torch.clamp(p[..., N:], 0, x.size(3) - 1)], dim=-1)
  63.  
  64.         # bilinear kernel (b, h, w, N)
  65.         g_lt = (1 + (q_lt[..., :N].type_as(p) - p[..., :N])) * (1 + (q_lt[..., N:].type_as(p) - p[..., N:]))
  66.         g_rb = (1 - (q_rb[..., :N].type_as(p) - p[..., :N])) * (1 - (q_rb[..., N:].type_as(p) - p[..., N:]))
  67.         g_lb = (1 + (q_lb[..., :N].type_as(p) - p[..., :N])) * (1 - (q_lb[..., N:].type_as(p) - p[..., N:]))
  68.         g_rt = (1 - (q_rt[..., :N].type_as(p) - p[..., :N])) * (1 + (q_rt[..., N:].type_as(p) - p[..., N:]))
  69.  
  70.         # resampling the features based on the modified coordinates.
  71.         x_q_lt = self._get_x_q(x, q_lt, N)
  72.         x_q_rb = self._get_x_q(x, q_rb, N)
  73.         x_q_lb = self._get_x_q(x, q_lb, N)
  74.         x_q_rt = self._get_x_q(x, q_rt, N)
  75.  
  76.         # bilinear
  77.         x_offset = g_lt.unsqueeze(dim=1) * x_q_lt + \
  78.                    g_rb.unsqueeze(dim=1) * x_q_rb + \
  79.                    g_lb.unsqueeze(dim=1) * x_q_lb + \
  80.                    g_rt.unsqueeze(dim=1) * x_q_rt
  81.  
  82.         x_offset = self._reshape_x_offset(x_offset, self.num_param)
  83.         out = self.conv(x_offset)
  84.  
  85.         return out
  86.  
  87.     # generating the inital sampled shapes for the AKConv with different sizes.
  88.     def _get_p_n(self, N, dtype):
  89.         base_int = round(math.sqrt(self.num_param))
  90.         row_number = self.num_param // base_int
  91.         mod_number = self.num_param % base_int
  92.         p_n_x, p_n_y = torch.meshgrid(
  93.             torch.arange(0, row_number),
  94.             torch.arange(0, base_int), indexing='xy')
  95.         p_n_x = torch.flatten(p_n_x)
  96.         p_n_y = torch.flatten(p_n_y)
  97.         if mod_number > 0:
  98.             mod_p_n_x, mod_p_n_y = torch.meshgrid(
  99.                 torch.arange(row_number, row_number + 1),
  100.                 torch.arange(0, mod_number), indexing='xy')
  101.  
  102.             mod_p_n_x = torch.flatten(mod_p_n_x)
  103.             mod_p_n_y = torch.flatten(mod_p_n_y)
  104.             p_n_x, p_n_y = torch.cat((p_n_x, mod_p_n_x)), torch.cat((p_n_y, mod_p_n_y))
  105.         p_n = torch.cat([p_n_x, p_n_y], 0)
  106.         p_n = p_n.view(1, 2 * N, 1, 1).type(dtype)
  107.         return p_n
  108.  
  109.     # no zero-padding
  110.     def _get_p_0(self, h, w, N, dtype):
  111.         p_0_x, p_0_y = torch.meshgrid(
  112.             torch.arange(0, h * self.stride, self.stride),
  113.             torch.arange(0, w * self.stride, self.stride), indexing='xy')
  114.  
  115.         p_0_x = torch.flatten(p_0_x).view(1, 1, h, w).repeat(1, N, 1, 1)
  116.         p_0_y = torch.flatten(p_0_y).view(1, 1, h, w).repeat(1, N, 1, 1)
  117.         p_0 = torch.cat([p_0_x, p_0_y], 1).type(dtype)
  118.  
  119.         return p_0
  120.  
  121.     def _get_p(self, offset, dtype):
  122.         N, h, w = offset.size(1) // 2, offset.size(2), offset.size(3)
  123.  
  124.         # (1, 2N, 1, 1)
  125.         p_n = self._get_p_n(N, dtype)
  126.         # (1, 2N, h, w)
  127.         p_0 = self._get_p_0(h, w, N, dtype)
  128.         p = p_0 + p_n + offset
  129.         return p
  130.  
  131.     def _get_x_q(self, x, q, N):
  132.         b, h, w, _ = q.size()
  133.         padded_w = x.size(3)
  134.         c = x.size(1)
  135.         # (b, c, h*w)
  136.         x = x.contiguous().view(b, c, -1)
  137.  
  138.         # (b, h, w, N)
  139.         index = q[..., :N] * padded_w + q[..., N:]  # offset_x*w + offset_y
  140.         # (b, c, h*w*N)
  141.         index = index.contiguous().unsqueeze(dim=1).expand(-1, c, -1, -1, -1).contiguous().view(b, c, -1)
  142.  
  143.         x_offset = x.gather(dim=-1, index=index).contiguous().view(b, c, h, w, N)
  144.  
  145.         return x_offset
  146.  
  147.     #  Stacking resampled features in the row direction.
  148.     @staticmethod
  149.     def _reshape_x_offset(x_offset, num_param):
  150.         b, c, h, w, n = x_offset.size()
  151.         x_offset = rearrange(x_offset, 'b c h w n -> b c (h n) w')
  152.         return x_offset

3.2 在YOlO v9中的添加教程

阅读YOLOv9添加模块教程或使用下文操作

        1. 将YOLOv9工程中models下common.py文件中的最下行（否则可能因类继承报错）增加模块的代码。

​

         2. 将YOLOv9工程中models下yolo.py文件中的第681行（可能因版本变化而变化）增加以下代码。

​

            RepNCSPELAN4, SPPELAN, RepNCSPELAN4AKConv1}:

3.3 运行配置文件

  1. # YOLOv9
  2. # Powered bu https://blog.csdn.net/StopAndGoyyy
  3.  
  4. # parameters
  5. nc: 80  # number of classes
  6. #depth_multiple: 0.33  # model depth multiple
  7. depth_multiple: 1  # model depth multiple
  8. #width_multiple: 0.25  # layer channel multiple
  9. width_multiple: 1  # layer channel multiple
  10. #activation: nn.LeakyReLU(0.1)
  11. #activation: nn.ReLU()
  12.  
  13. # anchors
  14. anchors: 3
  15.  
  16. # YOLOv9 backbone
  17. backbone:
  18.   [
  19.    [-1, 1, Silence, []],  
  20.    
  21.    # conv down
  22.    [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
  23.  
  24.    # conv down
  25.    [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
  26.  
  27.    # elan-1 block
  28.    [-1, 1, RepNCSPELAN4AKConv1, [256, 128, 64, 1]],  # 3
  29.  
  30.    # avg-conv down
  31.    [-1, 1, ADown, [256]],  # 4-P3/8
  32.  
  33.    # elan-2 block
  34.    [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 5
  35.  
  36.    # avg-conv down
  37.    [-1, 1, ADown, [512]],  # 6-P4/16
  38.  
  39.    # elan-2 block
  40.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 7
  41.  
  42.    # avg-conv down
  43.    [-1, 1, ADown, [512]],  # 8-P5/32
  44.  
  45.    # elan-2 block
  46.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 9
  47.   ]
  48.  
  49. # YOLOv9 head
  50. head:
  51.   [
  52.    # elan-spp block
  53.    [-1, 1, SPPELAN, [512, 256]],  # 10
  54.  
  55.    # up-concat merge
  56.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  57.    [[-1, 7], 1, Concat, [1]],  # cat backbone P4
  58.  
  59.    # elan-2 block
  60.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 13
  61.  
  62.    # up-concat merge
  63.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
  64.    [[-1, 5], 1, Concat, [1]],  # cat backbone P3
  65.  
  66.    # elan-2 block
  67.    [-1, 1, RepNCSPELAN4, [256, 256, 128, 1]],  # 16 (P3/8-small)
  68.  
  69.    # avg-conv-down merge
  70.    [-1, 1, ADown, [256]],
  71.    [[-1, 13], 1, Concat, [1]],  # cat head P4
  72.  
  73.    # elan-2 block
  74.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 19 (P4/16-medium)
  75.  
  76.    # avg-conv-down merge
  77.    [-1, 1, ADown, [512]],
  78.    [[-1, 10], 1, Concat, [1]],  # cat head P5
  79.  
  80.    # elan-2 block
  81.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 22 (P5/32-large)
  82.    
  83.    
  84.    # multi-level reversible auxiliary branch
  85.    
  86.    # routing
  87.    [5, 1, CBLinear, [[256]]], # 23
  88.    [7, 1, CBLinear, [[256, 512]]], # 24
  89.    [9, 1, CBLinear, [[256, 512, 512]]], # 25
  90.    
  91.    # conv down
  92.    [0, 1, Conv, [64, 3, 2]],  # 26-P1/2
  93.  
  94.    # conv down
  95.    [-1, 1, Conv, [128, 3, 2]],  # 27-P2/4
  96.  
  97.    # elan-1 block
  98.    [-1, 1, RepNCSPELAN4, [256, 128, 64, 1]],  # 28
  99.  
  100.    # avg-conv down fuse
  101.    [-1, 1, ADown, [256]],  # 29-P3/8
  102.    [[23, 24, 25, -1], 1, CBFuse, [[0, 0, 0]]], # 30  
  103.  
  104.    # elan-2 block
  105.    [-1, 1, RepNCSPELAN4, [512, 256, 128, 1]],  # 31
  106.  
  107.    # avg-conv down fuse
  108.    [-1, 1, ADown, [512]],  # 32-P4/16
  109.    [[24, 25, -1], 1, CBFuse, [[1, 1]]], # 33 
  110.  
  111.    # elan-2 block
  112.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 34
  113.  
  114.    # avg-conv down fuse
  115.    [-1, 1, ADown, [512]],  # 35-P5/32
  116.    [[25, -1], 1, CBFuse, [[2]]], # 36
  117.  
  118.    # elan-2 block
  119.    [-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
  120.    
  121.    
  122.    
  123.    # detection head
  124.  
  125.    # detect
  126.    [[31, 34, 37, 16, 19, 22], 1, DualDDetect, [nc]],  # DualDDetect(A3, A4, A5, P3, P4, P5)
  127.   ]

3.4 训练过程

​

欢迎关注!

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/我家自动化/article/detail/393603
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号