1.使用 Focal loss

在util/loss.py中，computeloss类用于计算损失函数

# Focal loss
        g = h['fl_gamma']  # focal loss gamma
        if g > 0:
            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
1
2
3
4

其中这一段就是开启Focal loss的关键！！！

parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
 
1
2

使用的data/hyps/hyp.scratch-low.yaml为参数配置文件，进去修改fl_gamma即可

fl_gamma实际上就是公式中红色椭圆的部分
看看代码更易于理解：

 def forward(self, pred, true):
        loss = self.loss_fcn(pred, true)
        # p_t = torch.exp(-loss)
        # loss *= self.alpha * (1.000001 - p_t) ** self.gamma  # non-zero power for gradient stability

        # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
        pred_prob = torch.sigmoid(pred)  # prob from logits
        p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
        modulating_factor = (1.0 - p_t) ** self.gamma
        loss *= alpha_factor * modulating_factor

        if self.reduction == 'mean':
            return loss.mean()
        elif self.reduction == 'sum':
            return loss.sum()
        else:  # 'none'
            return loss
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

调参上的技巧

1.1 增加alpha

focalloss其实是两个参数，一个参数就是我们前述的fl_gamma，同样的道理我们也可以增加fl_alpha来调节alpha参数
（1）进入参数配置文件

增加

fl_alpha: 0.95     # my focal loss alpha：nagetive example rate
1

（2）然后回到核心代码那里替换这一段

        # Focal loss
        g = h['fl_gamma']  # focal loss gamma
        if g > 0:
            a=h['fl_alpha']
            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)

            # ————————————————使用Varifocal Loss损失函数———————————————————————————————————
            #BCEcls, BCEobj = VFLoss(BCEcls, g,a), VFLoss(BCEobj, g,a)
            # print(BCEcls)
            # print
            # ————————————————使用Varifocal Loss损失函数———————————————————————————————————
1
2
3
4
5
6
7
8
9
10
11

Varifocal 和foacl loss二选一，另一个注释掉就行

（2）使用Varifocal Loss

Varifocal Loss

p输入为前景类的预测概率；q为ground-truth

class VFLoss(nn.Module):
    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
        super(VFLoss, self).__init__()
        # 传递 nn.BCEWithLogitsLoss() 损失函数  must be nn.BCEWithLogitsLoss()
        self.loss_fcn = loss_fcn  #
        self.gamma = gamma
        self.alpha = alpha
        self.reduction = loss_fcn.reduction
        self.loss_fcn.reduction = 'mean'  # required to apply VFL to each element

    def forward(self, pred, true):

        loss = self.loss_fcn(pred, true)

        pred_prob = torch.sigmoid(pred)  # prob from logits
                                                                    #p
        focal_weight = true * (true > 0.0).float() + self.alpha * (pred_prob - true).abs().pow(self.gamma) * (
                    true <= 0.0).float()
        loss *= focal_weight

        if self.reduction == 'mean':
            return loss.mean()
        elif self.reduction == 'sum':
            return loss.sum()
        else:
            return loss
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

true:q，即为ground-truth
(pred_prob - true):p，即前景类的预测概率

直接使用代码会报这个错

后面self.loss_fcn.reduction = 'mean'修改为self.loss_fcn.reduction = 'none'就没问题了

Focal loss和Varifocal Loss始终是不如原先的效果，可能很大一部分是参数问题