Yolo v3调优记录

Yolo v3调优记录

1.使用专门针对背景的Focal Loss

def focal_loss(out_put,y_true):
    # loss = ((torch.sigmoid(out_put))**2*(1-y_true)+ (1-torch.sigmoid(out_put))**2*y_true)\
    #         *F.binary_cross_entropy_with_logits((out_put), y_true, reduction='none')
    loss = 0.8*(torch.sigmoid(out_put)) ** 2 * F.binary_cross_entropy_with_logits(1-out_put, 1-y_true, reduction='none')+F.binary_cross_entropy_with_logits((out_put), y_true, reduction='none')
    return loss

针对一阶段网络前背景分类不平衡问题，使用只对背景加入loss调节因子的focal loss（实验中直接使用原版的focal loss导致ar下降，和yolo作者在论文中讲述一样），使得ar提升。

γ = 2，求导：

修改 yolo_layer.c 中的 l.output[obj_index] 即可

2.使用Senet通道加权层

class SELayer(nn.Module):
    def __init__(self, channel, reduction=16):
        super(SELayer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(channel, channel // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(channel // reduction, channel, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y.expand_as(x)

 

 

3.使用Giou Loss

        pxmin = px_y[..., 0, :, :] - pw_h[..., 0, :, :] / 2
        pxmax = px_y[..., 0, :, :] + pw_h[..., 0, :, :] / 2
        pymin = px_y[..., 1, :, :] - pw_h[..., 1, :, :] / 2
        pymax = px_y[..., 1, :, :] + pw_h[..., 1, :, :] / 2
        pxmin = torch.min(pxmin,pxmax)
        pxmax = torch.max(pxmin, pxmax)
        pymin = torch.min(pymin, pymax)
        pymax = torch.max(pymin, pymax)
        gxmin =Y[..., 0, :, :]
        gxmax = Y[..., 2, :, :]
        gymin = Y[..., 1, :, :]
        gymax = Y[..., 3, :, :]
        AG=(gxmax-gxmin)*(gymax-gymin)
        AP=(pxmax-pxmin)*(pymax-pymin)
        AC=torch.max(torch.max(pxmax,gxmax)-torch.min(pxmin,gxmin), torch.zeros_like(pymin))*torch.max(torch.max(pymax,gymax)-torch.min(pymin,gymin),torch.zeros_like(pymin))
        I = torch.max((torch.min(pymin, gymin) - torch.max(pymax, gymax)), torch.zeros_like(pymin)) * torch.max((torch.min(pxmin, gxmin) - torch.max(pxmax, gxmax)), torch.zeros_like(pymin))
        area=AP+AG-I
        GIOU=I/area-torch.div(AC-area, AC+0.000001)
        objs= torch.nonzero(object_mask)
        GIOU_loss+=((1-GIOU)*obj_mask.squeeze(2)).sum()

遇到的nan的问题，AC+0.000001后解决

GHMC Loss

def GHMC_loss(out_put,y_true):
    bins = 10
    edges = [float(x) / bins for x in range(bins + 1)]
    edges[-1] += 1e-6
    weights = torch.zeros_like(out_put)
    g = torch.abs(torch.sigmoid(out_put) - y_true)
    tot = 1
    for k in out_put.size():
        tot *= k
    for i in range(bins):
        inds = (g >= edges[i]) & (g < edges[i + 1])
        num_in_bin = inds.sum().item()
        if num_in_bin > 0:
            weights[inds] = tot / num_in_bin
    weights /= 10
    loss = F.binary_cross_entropy_with_logits(out_put,y_true,weight=weights,reduction='none')
    return loss