devbox
很楠不爱3
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

YOLOv8改进RepNCSPELAN4（结合YOLOv9二创）_repncspelan4改进yolov8

作者：很楠不爱3 | 2024-06-13 02:51:02

赞

踩

repncspelan4改进yolov8
  1. import torch
  2. import torch.nn as nn
  3. import torch.nn.functional as F
  4. import numpy as np
  5.  
  6. class RepConvN(nn.Module):
  7.     """RepConv is a basic rep-style block, including training and deploy status
  8.     This code is based on https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py
  9.     """
  10.     default_act = nn.SiLU()  # default activation
  11.  
  12.     def __init__(self, c1, c2, k=3, s=1, p=1, g=1, d=1, act=True, bn=False, deploy=False):
  13.         super().__init__()
  14.         assert k == 3 and p == 1
  15.         self.g = g
  16.         self.c1 = c1
  17.         self.c2 = c2
  18.         self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
  19.  
  20.         self.bn = None
  21.         self.conv1 = Conv(c1, c2, k, s, p=p, g=g, act=False)
  22.         self.conv2 = Conv(c1, c2, 1, s, p=(p - k // 2), g=g, act=False)
  23.  
  24.     def forward_fuse(self, x):
  25.         """Forward process"""
  26.         return self.act(self.conv(x))
  27.  
  28.     def forward(self, x):
  29.         """Forward process"""
  30.         if hasattr(self, 'conv'):
  31.             return self.forward_fuse(x)
  32.         id_out = 0 if self.bn is None else self.bn(x)
  33.         return self.act(self.conv1(x) + self.conv2(x) + id_out)
  34.  
  35.     def get_equivalent_kernel_bias(self):
  36.         kernel3x3, bias3x3 = self._fuse_bn_tensor(self.conv1)
  37.         kernel1x1, bias1x1 = self._fuse_bn_tensor(self.conv2)
  38.         kernelid, biasid = self._fuse_bn_tensor(self.bn)
  39.         return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
  40.  
  41.     def _avg_to_3x3_tensor(self, avgp):
  42.         channels = self.c1
  43.         groups = self.g
  44.         kernel_size = avgp.kernel_size
  45.         input_dim = channels // groups
  46.         k = torch.zeros((channels, input_dim, kernel_size, kernel_size))
  47.         k[np.arange(channels), np.tile(np.arange(input_dim), groups), :, :] = 1.0 / kernel_size ** 2
  48.         return k
  49.  
  50.     def _pad_1x1_to_3x3_tensor(self, kernel1x1):
  51.         if kernel1x1 is None:
  52.             return 0
  53.         else:
  54.             return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1])
  55.  
  56.     def _fuse_bn_tensor(self, branch):
  57.         if branch is None:
  58.             return 0, 0
  59.         if isinstance(branch, Conv):
  60.             kernel = branch.conv.weight
  61.             running_mean = branch.bn.running_mean
  62.             running_var = branch.bn.running_var
  63.             gamma = branch.bn.weight
  64.             beta = branch.bn.bias
  65.             eps = branch.bn.eps
  66.         elif isinstance(branch, nn.BatchNorm2d):
  67.             if not hasattr(self, 'id_tensor'):
  68.                 input_dim = self.c1 // self.g
  69.                 kernel_value = np.zeros((self.c1, input_dim, 3, 3), dtype=np.float32)
  70.                 for i in range(self.c1):
  71.                     kernel_value[i, i % input_dim, 1, 1] = 1
  72.                 self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
  73.             kernel = self.id_tensor
  74.             running_mean = branch.running_mean
  75.             running_var = branch.running_var
  76.             gamma = branch.weight
  77.             beta = branch.bias
  78.             eps = branch.eps
  79.         std = (running_var + eps).sqrt()
  80.         t = (gamma / std).reshape(-1, 1, 1, 1)
  81.         return kernel * t, beta - running_mean * gamma / std
  82.  
  83.     def switch_to_deploy(self):
  84.         if hasattr(self, 'conv'):
  85.             return
  86.         kernel, bias = self.get_equivalent_kernel_bias()
  87.         self.conv = nn.Conv2d(in_channels=self.conv1.conv.in_channels,
  88.                               out_channels=self.conv1.conv.out_channels,
  89.                               kernel_size=self.conv1.conv.kernel_size,
  90.                               stride=self.conv1.conv.stride,
  91.                               padding=self.conv1.conv.padding,
  92.                               dilation=self.conv1.conv.dilation,
  93.                               groups=self.conv1.conv.groups,
  94.                               bias=True).requires_grad_(False)
  95.         self.conv.weight.data = kernel
  96.         self.conv.bias.data = bias
  97.         for para in self.parameters():
  98.             para.detach_()
  99.         self.__delattr__('conv1')
  100.         self.__delattr__('conv2')
  101.         if hasattr(self, 'nm'):
  102.             self.__delattr__('nm')
  103.         if hasattr(self, 'bn'):
  104.             self.__delattr__('bn')
  105.         if hasattr(self, 'id_tensor'):
  106.             self.__delattr__('id_tensor')
  107.  
  108. class RepNBottleneck(nn.Module):
  109.     # Standard bottleneck
  110.     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):  # ch_in, ch_out, shortcut, kernels, groups, expand
  111.         super().__init__()
  112.         c_ = int(c2 * e)  # hidden channels
  113.         self.cv1 = RepConvN(c1, c_, k[0], 1)
  114.         self.cv2 = Conv(c_, c2, k[1], 1, g=g)
  115.         self.add = shortcut and c1 == c2
  116.  
  117.     def forward(self, x):
  118.         return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
  119.  
  120. class RepNCSP(nn.Module):
  121.     # CSP Bottleneck with 3 convolutions
  122.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
  123.         super().__init__()
  124.         c_ = int(c2 * e)  # hidden channels
  125.         self.cv1 = Conv(c1, c_, 1, 1)
  126.         self.cv2 = Conv(c1, c_, 1, 1)
  127.         self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
  128.         self.m = nn.Sequential(*(RepNBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
  129.  
  130.     def forward(self, x):
  131.         return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
  132.  
  133. class RepNCSPELAN4(nn.Module):
  134.     # csp-elan
  135.     def __init__(self, c1, c2, c3, c4, c5=1):  # ch_in, ch_out, number, shortcut, groups, expansion
  136.         super().__init__()
  137.         self.c = c3//2
  138.         self.cv1 = Conv(c1, c3, 1, 1)
  139.         self.cv2 = nn.Sequential(RepNCSP(c3//2, c4, c5), Conv(c4, c4, 3, 1))
  140.         self.cv3 = nn.Sequential(RepNCSP(c4, c4, c5), Conv(c4, c4, 3, 1))
  141.         self.cv4 = Conv(c3+(2*c4), c2, 1, 1)
  142.  
  143.     def forward(self, x):
  144.         y = list(self.cv1(x).chunk(2, 1))
  145.         y.extend((m(y[-1])) for m in [self.cv2, self.cv3])
  146.         return self.cv4(torch.cat(y, 1))
  147.  
  148.     def forward_split(self, x):
  149.         y = list(self.cv1(x).split((self.c, self.c), 1))
  150.         y.extend(m(y[-1]) for m in [self.cv2, self.cv3])
  151.         return self.cv4(torch.cat(y, 1))
  1. # Ultralytics YOLO 
    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/很楠不爱3/article/detail/710609
    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。