YOLOv5结合BiFPN

现在yolov5的neck用的是PANet，在EfficientDet论文中提出了BiFPN结构，还有更加不错的性能。所以就尝试将yolov5中的PANet层改为BiFPN。

需要修改的地方

• 主要是修改yaml配置文件
  我修改的是yolov5x.yaml，将concat层连接不同的layer。有三层BiFPN，最后输出为了和yolov5对应，我没有用在论文里写的p3-p7五个节点，我只用了三个。（这里应该要用relu使w>0，我训练报错就没有用relu）
• 修改common.py

class Concat(nn.Module):
    # Concatenate a list of tensors along dimension
    def __init__(self, c1, c2):
        super(Concat, self).__init__()
        # self.relu = nn.ReLU()
        self.w1 = nn.Parameter(torch.ones(2, dtype=torch.float32), requires_grad=True)
        self.w2 = nn.Parameter(torch.ones(3, dtype=torch.float32), requires_grad=True)
        self.epsilon = 0.0001
        self.conv = nn.Conv2d(c1, c2, kernel_size=1, stride=1, padding=0)
        self.swish = MemoryEfficientSwish()

    def forward(self, x):
        outs = self._forward(x)
        return outs

    def _forward(self, x): # intermediate result
        if len(x) == 2:
            # w = self.relu(self.w1)
            w = self.w1
            weight = w / (torch.sum(w, dim=0) + self.epsilon)
            x = self.conv(self.swish(weight[0] * x[0] + weight[1] * x[1]))
        elif len(x) == 3: # final result
            # w = self.relu(self.w2)
            w = self.w2
            weight = w / (torch.sum(w, dim=0) + self.epsilon)
            x = self.conv(self.swish(weight[0] * x[0] + weight[1] * x[1] + weight[2] * x[2]))

• 修改yolo.py
  channel现在不是直接concat了，而是进行pairwise add操作，所以out-channel不再是sum。

 # elif m is Concat:
 #    c2 = sum([ch[x] for x in f])
   elif m is Concat:
      c2 = max([ch[x] for x in f])

配置文件的修改在https://github.com/XingZeng307/YOLOv5_with_BiFPN/blob/main/models/yolov5x.yaml
现在算是一个初步的结合吧。还在训练当中，有结果了再贴图。

参考资料

https://github.com/ultralytics/yolov5

https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/tree/15403b5371a64defb2a7c74e162c6e880a7f462c

Mingxing Tan, Ruoming Pang, and Quoc V Le. EfficientDet: Scalable and efficient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020.