mmdetection 选用自定义 backbone训练

学习率

lr = (base_lr / 8) x num_gpus x (img_per_gpu/2)

其中 base_lr 是源码中给定的 学习率，这是在 8张GPU上面，每张GPU上面放2张图片得到的学习率，所以当对应的GPU数目和每张GPU上的图片数目发生改变时，学习率成正比例变化。

选用新的backbone来训练

这里选用 res2net_26w_6s 作为backbone;
以cascade_rcnn_r50_fpn_1x.py 作为config文件

• 在 mmdetection-master\mmdet\models\backbones中新建一个 res2net.py文件，并将模型代码复制进去 https://github.com/Res2Net/Res2Net-PretrainedModels/blob/master/res2net.py

• 由于这里需要多尺度特征图，可参考 原有的 resnet.py，返回的是stage1,stage2,stage3,stage4组成的list，因此新增的 res2net.py的代码需要进行修改

  • 1). 源码：

    class Res2Net(nn.Module):
       def __init__(self, block, layers, baseWidth = 26, scale = 4, num_classes=1000):
  

  修改：添加一个register；需要再import from ..registry import BACKBONES; 同时把block中的参数先写上默认参数，不然后面的config文件中传 block名字时比较麻烦

    @BACKBONES.register_module
    class Res2Net(nn.Module):
       def __init__(self, block=Bottle2neck, layers=[3,4,6,3], baseWidth=26, scale=6, num_classes=1000):
  

  • 2). 源码：class Res2Net下面的 forward

        def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
    
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
    
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
    
        return x
    

    修改：返回stage1, stage2,stage3,stage4

        def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
    
        x1 = self.layer1(x)
        x2 = self.layer2(x1)
        x3 = self.layer3(x2)
        x4 = self.layer4(x3)
    
        #x = self.avgpool(x)
        #x = x.view(x.size(0), -1)
        #x = self.fc(x)
    
        return [x1,x2,x3,x4]
    

  • 3). 增加一个 init_weights函数(这边是不添加就报错，所以就直接从 resnet.py 中复制过来了)，如resnet.py中一样，修改对应的block名称

      def init_weights(self, pretrained=None):
        if isinstance(pretrained, str):
            from mmdet.apis import get_root_logger
            logger = get_root_logger()
            load_checkpoint(self, pretrained, strict=False, logger=logger)
        elif pretrained is None:
            for m in self.modules():
                if isinstance(m, nn.Conv2d):
                    kaiming_init(m)
                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
                    constant_init(m, 1)
  
            if self.dcn is not None:
                for m in self.modules():
                    if isinstance(m, Bottle2neck) and hasattr(
                            m, 'conv2_offset'):
                        constant_init(m.conv2_offset, 0)
  
            if self.zero_init_residual:
                for m in self.modules():
                    if isinstance(m, Bottle2neck):
                        constant_init(m.norm3, 0)
                    # elif isinstance(m, BasicBlock):
                    #     constant_init(m.norm2, 0)
        else:
            raise TypeError('pretrained must be a str or None')
  

• 修改 config 文件,cascade_rcnn_r50_fpn_1x.py
  源码：

# model settings
model = dict(
    type='CascadeRCNN',
    num_stages=3,
    pretrained='torchvision://resnet50',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        style='pytorch'),

修改：

model = dict(
    type='CascadeRCNN',
    num_stages=3,
    #pretrained='torchvision://resnet50',
    pretrained='/media/mtc/dac70756-6922-445c-9cbe-58847ebdc56f/zql/mmdetection/pretrained_models/res2net50_26w_6s-19041792.pth',
    backbone=dict(
        type='Res2Net',
        layers=[3,4,6,3], baseWidth=26, scale=6, num_classes=1000,
        ),