• BBAVectors斜框检测网络更换轻量型主干网络

            本文将会手把手教会帅比看官如何更换BBAVectors斜框检测模型骨干网络。

            本博客假设各位看官帅比已经能利用BBAVectors训练自己的数据，但是不知道如何更换其主干网络。原始代码中采用的resnet101做为骨干网络，这个网络训练的时候对资源要求较高，那么如何换为更轻量的resnet18、mobilnet呢？

    一、主干网络修改为resnet18模型

    1、修改ctrbox_net.py中主干网络参数

    1. #修改前
    2. # self.base_network = resnet.resnet101(pretrained=pretrained)
    3. # self.dec_c2 = CombinationModule(512, 256, batch_norm=True)
    4. # self.dec_c3 = CombinationModule(1024, 512, batch_norm=True)
    5. # self.dec_c4 = CombinationModule(2048, 1024, batch_norm=True)
    6.  
    7. #修改后
    8. self.base_network = resnet.resnet18(pretrained=pretrained)
    9. self.dec_c2 = CombinationModule(128, 64, batch_norm=True)
    10. self.dec_c3 = CombinationModule(256, 128, batch_norm=True)
    11. self.dec_c4 = CombinationModule(512, 256, batch_norm=True)

            首先我们要明白上面几个256、512、1024、2048其实分别是resnet101主干网络中4个featuremap的通道大小。而分析resnet.py的resnet18我们可知，他的4个featuremap的通道数分别是64、128、256、512，你只需要在resnet.py末尾加上下面几句代码，然后运行resnet.py即可知道，仅仅是通道数不同，所以我们直接修改上述参数即可。

    1. if __name__ == '__main__':
    2.     device='cpu'
    3.     input=np.ones((1,3,512,512)).astype(np.float32)
    4.     dummy_input = torch.from_numpy(input).to(device)
    5.     
    6.     model=resnet18().to(device)
    7.     logit=model(dummy_input)
    8.     print(logit[-1].size()) #torch.Size([1, 512, 16, 16])
    9.     print(logit[-2].size()) #torch.Size([1, 256, 32, 32])
    10.     print(logit[-3].size()) #torch.Size([1, 128, 64, 64])
    11.     print(logit[-4].size()) #torch.Size([1, 64, 128, 128])

    2.修改 ctrbox_net.py中的 channels

            当 down_ratio=4的时候，c2_combine输出的通道数是256，这也是后面4个head的输入通道数， ctrbox_net .py的计算代码如下

    1. class CTRBOX(nn.Module):
    2.     def __init__(self, heads, pretrained, down_ratio, final_kernel, head_conv,export=False):
    3.         super(CTRBOX, self).__init__()
    4.         channels = [3, 64, 256, 512, 1024, 2048]
    5.         assert down_ratio in [2, 4, 8, 16]
    6.         self.l1 = int(np.log2(down_ratio))

            head的输入通道数就是channels[self.l1]=256。但是改成resnet18之后，c2_combine输出的通道数变成了64。这个时候你就会自然而然将main.py中的down_ratio改成2，这样channels[self.l1]=64。

            如果你是通过修改down_ratio=2，就大错特错了。因为ctrbox_net最终的输出大小都是输入大小4倍下采样，即输入512x512，4个head的输出都是128*128，计算loss的时候，gt的大小也应该是128x128。而在dataset/base.py中也会用到down_ration，如果改成2，gt的大小是计算出来就是256x256，这样就会报错。        

    **解决方法**

            我们不改down_ratio，在完成1中的内容修改之后，直接把 ctrbox_net.py中的channels = [3, 64, 256, 512, 1024, 2048]中的256改成64，即channels = [3, 64, 64, 512, 1024, 2048]即可开始训练，简单暴力。

    3.结果

    二、修改为轻量mobilenetv2主干网络

    mobilenetv2的代码如下

    1. import math
    2. import os
    3.  
    4. import torch
    5. import torch.nn as nn
    6. import torch.utils.model_zoo as model_zoo
    7.  
    8. BatchNorm2d = nn.BatchNorm2d
    9.  
    10. def conv_bn(inp, oup, stride):
    11.     return nn.Sequential(
    12.         nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
    13.         BatchNorm2d(oup),
    14.         nn.ReLU6(inplace=True)
    15.     )
    16.  
    17. def conv_1x1_bn(inp, oup):
    18.     return nn.Sequential(
    19.         nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
    20.         BatchNorm2d(oup),
    21.         nn.ReLU6(inplace=True)
    22.     )
    23.  
    24. class InvertedResidual(nn.Module):
    25.     def __init__(self, inp, oup, stride, expand_ratio):
    26.         super(InvertedResidual, self).__init__()
    27.         self.stride = stride
    28.         assert stride in [1, 2]
    29.  
    30.         hidden_dim = round(inp * expand_ratio)
    31.         self.use_res_connect = self.stride == 1 and inp == oup
    32.  
    33.         if expand_ratio == 1:
    34.             self.conv = nn.Sequential(
    35.                 # dw
    36.                 nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
    37.                 BatchNorm2d(hidden_dim),
    38.                 nn.ReLU6(inplace=True),
    39.                 # pw-linear
    40.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
    41.                 BatchNorm2d(oup),
    42.             )
    43.         else:
    44.             self.conv = nn.Sequential(
    45.                 # pw
    46.                 nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
    47.                 BatchNorm2d(hidden_dim),
    48.                 nn.ReLU6(inplace=True),
    49.                 # dw
    50.                 nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
    51.                 BatchNorm2d(hidden_dim),
    52.                 nn.ReLU6(inplace=True),
    53.                 # pw-linear
    54.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
    55.                 BatchNorm2d(oup),
    56.             )
    57.  
    58.     def forward(self, x):
    59.         if self.use_res_connect:
    60.             return x + self.conv(x)
    61.         else:
    62.             return self.conv(x)
    63.  
    64. class MobileNetV2(nn.Module):
    65.     def __init__(self, n_class=1000, input_size=224, width_mult=1.):
    66.         super(MobileNetV2, self).__init__()
    67.         block = InvertedResidual
    68.         input_channel = 32
    69.         last_channel = 1280
    70.         interverted_residual_setting = [
    71.             # t, c, n, s
    72.             [1, 16, 1, 1],
    73.             [6, 24, 2, 2],
    74.             [6, 32, 3, 2],
    75.             [6, 64, 4, 2],
    76.             [6, 96, 3, 1],
    77.             [6, 160, 3, 2],
    78.             [6, 320, 1, 1],
    79.         ]
    80.  
    81.         # building first layer
    82.         assert input_size % 32 == 0
    83.         input_channel = int(input_channel * width_mult)
    84.         self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
    85.         self.features = [conv_bn(3, input_channel, 2)]
    86.         # building inverted residual blocks
    87.         for t, c, n, s in interverted_residual_setting:
    88.             output_channel = int(c * width_mult)
    89.             for i in range(n):
    90.                 if i == 0:
    91.                     self.features.append(block(input_channel, output_channel, s, expand_ratio=t))
    92.                 else:
    93.                     self.features.append(block(input_channel, output_channel, 1, expand_ratio=t))
    94.                 input_channel = output_channel
    95.         # building last several layers
    96.         self.features.append(conv_1x1_bn(input_channel, self.last_channel))
    97.         # make it nn.Sequential
    98.         self.features = nn.Sequential(*self.features)
    99.  
    100.         # building classifier
    101.         self.classifier = nn.Sequential(
    102.             nn.Dropout(0.2),
    103.             nn.Linear(self.last_channel, n_class),
    104.         )
    105.  
    106.         self._initialize_weights()
    107.  
    108.     def forward(self, x):
    109.         x = self.features(x)
    110.         x = x.mean(3).mean(2)
    111.         x = self.classifier(x)
    112.         return x
    113.  
    114.     def _initialize_weights(self):
    115.         for m in self.modules():
    116.             if isinstance(m, nn.Conv2d):
    117.                 n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
    118.                 m.weight.data.normal_(0, math.sqrt(2. / n))
    119.                 if m.bias is not None:
    120.                     m.bias.data.zero_()
    121.             elif isinstance(m, BatchNorm2d):
    122.                 m.weight.data.fill_(1)
    123.                 m.bias.data.zero_()
    124.             elif isinstance(m, nn.Linear):
    125.                 n = m.weight.size(1)
    126.                 m.weight.data.normal_(0, 0.01)
    127.                 m.bias.data.zero_()
    128.  
    129.  
    130. def load_url(url, model_dir='./model_data', map_location=None):
    131.     if not os.path.exists(model_dir):
    132.         os.makedirs(model_dir)
    133.     filename = url.split('/')[-1]
    134.     cached_file = os.path.join(model_dir, filename)
    135.     if os.path.exists(cached_file):
    136.         return torch.load(cached_file, map_location=map_location)
    137.     else:
    138.         return model_zoo.load_url(url,model_dir=model_dir)
    139.  
    140. def mobilenetv2(pretrained=False, **kwargs):
    141.     model = MobileNetV2(n_class=1000, **kwargs)
    142.     if pretrained:
    143.         model.load_state_dict(load_url('http://sceneparsing.csail.mit.edu/model/pretrained_resnet/mobilenet_v2.pth.tar'), strict=False)
    144.     return model
    145.  
    146. class Backbone(nn.Module):
    147.     def __init__(self,pretrained=False):
    148.         super(Backbone, self).__init__()
    149.         model=mobilenetv2(pretrained)
    150.         self.features = model.features[:-1]
    151.         
    152.     def forward(self, x):
    153.         
    154.         C1=self.features[:4](x)
    155.         C2=self.features[4:7](C1)
    156.         C3=self.features[7:14](C2)
    157.         C4=self.features[14:](C3)
    158.  
    159.         return C1,C2,C3,C4
    160.  
    161. if __name__ == '__main__':
    162.     import numpy as np
    163.  
    164.     device='cpu'
    165.     model=Backbone().to(device)
    166.     input=np.ones((1,3,512,512)).astype(np.float32)
    167.  
    168.     dummy_input = torch.from_numpy(input).to(device)
    169.     logit=model(dummy_input)
    170.     print(logit[-1].size())
    171.     print(logit[-2].size())
    172.     print(logit[-3].size())
    173.     print(logit[-4].size())

    4个featuremap的大小分别是

    1. torch.Size([1, 320, 16, 16])
    2. torch.Size([1, 96, 32, 32])
    3. torch.Size([1, 32, 64, 64])
    4. torch.Size([1, 24, 128, 128])

    对比之前改成resnet18，你就知道channels和各通道应该改成下面这样

    1. # channels = [3, 64, 256, 512, 1024, 2048]  # 当下面采用resnet101的时候用这个
    2. # self.base_network = resnet.resnet101(pretrained=pretrained)
    3. # self.dec_c2 = CombinationModule(512, 256, batch_norm=True)
    4. # self.dec_c3 = CombinationModule(1024, 512, batch_norm=True)
    5. # self.dec_c4 = CombinationModule(2048, 1024, batch_norm=True)
    6.  
    7. # channels = [3, 64, 64, 512, 1024, 2048]  # 用resnet18的时候是这个
    8. # self.base_network = resnet.resnet18(pretrained=pretrained)
    9. # self.dec_c2 = CombinationModule(128, 64, batch_norm=True)
    10. # self.dec_c3 = CombinationModule(256, 128, batch_norm=True)
    11. # self.dec_c4 = CombinationModule(512, 256, batch_norm=True)
    12.  
    13. channels = [3, 64, 24, 512, 1024, 2048]  # mobilenetv2的时候是这个
    14. self.base_network = mobilenet.Backbone(pretrained=pretrained)
    15. self.dec_c2 = CombinationModule(32, 24, batch_norm=True)
    16. self.dec_c3 = CombinationModule(96, 32, batch_norm=True)
    17. self.dec_c4 = CombinationModule(320, 96, batch_norm=True)

    2.结果

     loss下降不如resnet18做骨干网络的loss，这是正常的情况。

    三、完整代码

    完整代码可以参考：见这里，这是我修改过的BBAVectors代码，含DOTA_Devkit， rolabelimg的xml转4点txt格式代码，划分数据集代码，以及导出含decoder的onnx模型代码，tensorrt模型转换与推理的代码，关于tensorrt推理可以看我其他博客。



     

  • 相关阅读:
    BlobDetector的使用与参数说明（OpenCV/C++）
    Redis分布式锁这样用，有坑？
    EdrawMax v12新鲜出炉,重新设计字体库UI
    CSAPP-BombLab详解
    项目实战——实现注册和登录模块
    鸿蒙开发实例 ｜ 可复用列表项的ListContainer
    常见分词算法综述
    vite打包部分页面不显示问题+图片不显示问题
    1.2 异步相关概念：深入了解
    AR空间音频能力，打造沉浸式声音体验
  • 原文地址：https://blog.csdn.net/qq_41043389/article/details/127561102