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

#修改前
# self.base_network = resnet.resnet101(pretrained=pretrained)
# self.dec_c2 = CombinationModule(512, 256, batch_norm=True)
# self.dec_c3 = CombinationModule(1024, 512, batch_norm=True)
# self.dec_c4 = CombinationModule(2048, 1024, batch_norm=True)

#修改后
self.base_network = resnet.resnet18(pretrained=pretrained)
self.dec_c2 = CombinationModule(128, 64, batch_norm=True)
self.dec_c3 = CombinationModule(256, 128, batch_norm=True)
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。

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

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

import math
import os

import torch
import torch.nn as nn
import torch.utils.model_zoo as model_zoo

BatchNorm2d = nn.BatchNorm2d

def conv_bn(inp, oup, stride):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        BatchNorm2d(oup),
        nn.ReLU6(inplace=True)
    )

def conv_1x1_bn(inp, oup):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        BatchNorm2d(oup),
        nn.ReLU6(inplace=True)
    )

class InvertedResidual(nn.Module):
    def __init__(self, inp, oup, stride, expand_ratio):
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]

        hidden_dim = round(inp * expand_ratio)
        self.use_res_connect = self.stride == 1 and inp == oup

        if expand_ratio == 1:
            self.conv = nn.Sequential(
                # dw
                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
                BatchNorm2d(hidden_dim),
                nn.ReLU6(inplace=True),
                # pw-linear
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                BatchNorm2d(oup),
            )
        else:
            self.conv = nn.Sequential(
                # pw
                nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
                BatchNorm2d(hidden_dim),
                nn.ReLU6(inplace=True),
                # dw
                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
                BatchNorm2d(hidden_dim),
                nn.ReLU6(inplace=True),
                # pw-linear
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                BatchNorm2d(oup),
            )

    def forward(self, x):
        if self.use_res_connect:
            return x + self.conv(x)
        else:
            return self.conv(x)

class MobileNetV2(nn.Module):
    def __init__(self, n_class=1000, input_size=224, width_mult=1.):
        super(MobileNetV2, self).__init__()
        block = InvertedResidual
        input_channel = 32
        last_channel = 1280
        interverted_residual_setting = [
            # t, c, n, s
            [1, 16, 1, 1],
            [6, 24, 2, 2],
            [6, 32, 3, 2],
            [6, 64, 4, 2],
            [6, 96, 3, 1],
            [6, 160, 3, 2],
            [6, 320, 1, 1],
        ]

        # building first layer
        assert input_size % 32 == 0
        input_channel = int(input_channel * width_mult)
        self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
        self.features = [conv_bn(3, input_channel, 2)]
        # building inverted residual blocks
        for t, c, n, s in interverted_residual_setting:
            output_channel = int(c * width_mult)
            for i in range(n):
                if i == 0:
                    self.features.append(block(input_channel, output_channel, s, expand_ratio=t))
                else:
                    self.features.append(block(input_channel, output_channel, 1, expand_ratio=t))
                input_channel = output_channel
        # building last several layers
        self.features.append(conv_1x1_bn(input_channel, self.last_channel))
        # make it nn.Sequential
        self.features = nn.Sequential(*self.features)

        # building classifier
        self.classifier = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(self.last_channel, n_class),
        )

        self._initialize_weights()

    def forward(self, x):
        x = self.features(x)
        x = x.mean(3).mean(2)
        x = self.classifier(x)
        return x

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                n = m.weight.size(1)
                m.weight.data.normal_(0, 0.01)
                m.bias.data.zero_()


def load_url(url, model_dir='./model_data', map_location=None):
    if not os.path.exists(model_dir):
        os.makedirs(model_dir)
    filename = url.split('/')[-1]
    cached_file = os.path.join(model_dir, filename)
    if os.path.exists(cached_file):
        return torch.load(cached_file, map_location=map_location)
    else:
        return model_zoo.load_url(url,model_dir=model_dir)

def mobilenetv2(pretrained=False, **kwargs):
    model = MobileNetV2(n_class=1000, **kwargs)
    if pretrained:
        model.load_state_dict(load_url('http://sceneparsing.csail.mit.edu/model/pretrained_resnet/mobilenet_v2.pth.tar'), strict=False)
    return model

class Backbone(nn.Module):
    def __init__(self,pretrained=False):
        super(Backbone, self).__init__()
        model=mobilenetv2(pretrained)
        self.features = model.features[:-1]
        
    def forward(self, x):
        
        C1=self.features[:4](x)
        C2=self.features[4:7](C1)
        C3=self.features[7:14](C2)
        C4=self.features[14:](C3)

        return C1,C2,C3,C4

if __name__ == '__main__':
    import numpy as np

    device='cpu'
    model=Backbone().to(device)
    input=np.ones((1,3,512,512)).astype(np.float32)

    dummy_input = torch.from_numpy(input).to(device)
    logit=model(dummy_input)
    print(logit[-1].size())
    print(logit[-2].size())
    print(logit[-3].size())
    print(logit[-4].size())
 

4个featuremap的大小分别是

torch.Size([1, 320, 16, 16])
torch.Size([1, 96, 32, 32])
torch.Size([1, 32, 64, 64])
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

# channels = [3, 64, 256, 512, 1024, 2048]  # 当下面采用resnet101的时候用这个
# self.base_network = resnet.resnet101(pretrained=pretrained)
# self.dec_c2 = CombinationModule(512, 256, batch_norm=True)
# self.dec_c3 = CombinationModule(1024, 512, batch_norm=True)
# self.dec_c4 = CombinationModule(2048, 1024, batch_norm=True)

# channels = [3, 64, 64, 512, 1024, 2048]  # 用resnet18的时候是这个
# self.base_network = resnet.resnet18(pretrained=pretrained)
# self.dec_c2 = CombinationModule(128, 64, batch_norm=True)
# self.dec_c3 = CombinationModule(256, 128, batch_norm=True)
# self.dec_c4 = CombinationModule(512, 256, batch_norm=True)

channels = [3, 64, 24, 512, 1024, 2048]  # mobilenetv2的时候是这个
self.base_network = mobilenet.Backbone(pretrained=pretrained)
self.dec_c2 = CombinationModule(32, 24, batch_norm=True)
self.dec_c3 = CombinationModule(96, 32, batch_norm=True)
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。