pointpillar（OpenPCDet）原创改进(0)——重构神经网络部分的代码逻辑

1.说明

想要对pointpillar神经网络部分大改，但个人感觉PCdet架构下源代码多层嵌套的循环形式代码的可读性低，用习惯了YOLO的代码看这个很难受，不容易在源代码的基础上进行改进。因此花时间重构了一下网络结构，增加可读性和可编程性。让网络代码可以像YOLO代码一样一层一层的添加新的模块，随意改进新的网络结构，重构代码放在下面。

2. 使用方法

将源代码路径OpenPCDet/pcdet/models/backbones_2d/base_bev_backbone.py文件中的BaseBEVBackbone类全部替换成下列代码(整个类全部替换)

添加图片注释，不超过 140 字（可选）

3.代码逻辑

定义ConvModule类进行3*3卷集。定义deConvModule进行上采样。定义ConvModule2进行多层卷集网络。

在网络class BaseBEVBackbone(nn.Module)中逐层定义网络结构

 

在前向传播中逐层传播

4.附录代码

class ConvModule(nn.Module):       #可变层卷集
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, use_bn=True, use_relu=True):
        super(ConvModule, self).__init__()
        layers = [nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=not use_bn)]
        if use_bn:
            layers.append(nn.BatchNorm2d(out_channels, eps=1e-3, momentum=0.01))
        if use_relu:
            layers.append(nn.ReLU(inplace=True))
        self.conv = nn.Sequential(*layers)
    def forward(self, x):
        return self.conv(x)
class deConvModule(nn.Module):     #上采样层
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, use_bn=True, use_relu=True):
        super(deConvModule, self).__init__()
        layers = [nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride, padding, bias=not use_bn)]
        if use_bn:
            layers.append(nn.BatchNorm2d(out_channels, eps=1e-3, momentum=0.01))
        if use_relu:
            layers.append(nn.ReLU(inplace=True))
        self.deconv = nn.Sequential(*layers)
    def forward(self, x):
        return self.deconv(x)
class ConvModule2(nn.Module):          #可变层卷集
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, use_bn=True, use_relu=True, num_1x1_layers=0):
        super(ConvModule2, self).__init__()
        layers = []
        # 添加 3x3 卷积层
        layers.append(nn.ZeroPad2d(1))
        layers.append(nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding=0, bias=not use_bn))
        if use_bn:
            layers.append(nn.BatchNorm2d(out_channels, eps=1e-3, momentum=0.01))
        if use_relu:
            layers.append(nn.ReLU(inplace=True))
        in_channels = out_channels  # 保证下一层输入通道匹配
        # 添加 1x1 卷积层
        for _ in range(num_1x1_layers):
            layers.append(nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=not use_bn))
            if use_bn:
                layers.append(nn.BatchNorm2d(out_channels, eps=1e-3, momentum=0.01))
            if use_relu:
                layers.append(nn.ReLU(inplace=True))
            in_channels = out_channels
        self.conv = nn.Sequential(*layers)
    def forward(self, x):
        return self.conv(x)

class BaseBEVBackbone(nn.Module):     #定义网络
    def __init__(self, model_cfg, input_channels):
        super().__init__()
        self.model_cfg = model_cfg

        if self.model_cfg.get('LAYER_NUMS', None) is not None:
            assert len(self.model_cfg.LAYER_NUMS) == len(self.model_cfg.LAYER_STRIDES) == len(self.model_cfg.NUM_FILTERS)
            layer_nums = self.model_cfg.LAYER_NUMS
            layer_strides = self.model_cfg.LAYER_STRIDES
            num_filters = self.model_cfg.NUM_FILTERS
        else:
            layer_nums = layer_strides = num_filters = []

        if self.model_cfg.get('UPSAMPLE_STRIDES', None) is not None:
            assert len(self.model_cfg.UPSAMPLE_STRIDES) == len(self.model_cfg.NUM_UPSAMPLE_FILTERS)
            num_upsample_filters = self.model_cfg.NUM_UPSAMPLE_FILTERS
            upsample_strides = self.model_cfg.UPSAMPLE_STRIDES
        else:
            upsample_strides = num_upsample_filters = []
        c_in_list = [input_channels, *num_filters[:-1]]
        
        idx=0                           #网络结构p0
        self.c1=ConvModule2(c_in_list[idx],num_filters[idx],stride=layer_strides[idx],padding=0,num_1x1_layers=layer_nums[idx])      #0
        self.d1=deConvModule(num_filters[idx], num_upsample_filters[idx],upsample_strides[idx],stride=upsample_strides[idx])         #1
        idx=1                           #网络结构p1
        self.c2=ConvModule2(c_in_list[idx],num_filters[idx],stride=layer_strides[idx],padding=0,num_1x1_layers=layer_nums[idx])      #2
        self.d2=deConvModule(num_filters[idx], num_upsample_filters[idx],upsample_strides[idx],stride=upsample_strides[idx])         #3
        idx=2                           #网络结构p3
        self.c3=ConvModule2(c_in_list[idx],num_filters[idx],stride=layer_strides[idx],padding=0,num_1x1_layers=layer_nums[idx])      #4
        self.d3=deConvModule(num_filters[idx], num_upsample_filters[idx],upsample_strides[idx],stride=upsample_strides[idx])         #5
        c_in = sum(num_upsample_filters)
        self.num_bev_features = c_in
    def forward(self, data_dict):
        spatial_features = data_dict['spatial_features']
        ups = []
        ret_dict = {}
        x = spatial_features
         #网络结构p0
        x=self.c1(x)    #0
        y=self.d1    #1
        ups.append(y)   
        #网络结构p0
        x=self.c2(x)    #2
        y=self.d2(x)    #3
        ups.append(y) 
        #网络结构p0
        x=self.c3(x)    #4
        y=self.d3(x)    #5
        ups.append(y)
        for i, u in enumerate(ups):
            print(f"ups[{i}] shape:", u.shape)
        if len(ups) > 1:
            x = torch.cat(ups, dim=1)  #6
        elif len(ups) == 1:
            x = ups[0]
        data_dict['spatial_features_2d'] = x
        return data_dict