VarifocalNet
VarifocalNet copied to clipboard
hyz-xmaster
VFocalLoss in yolov5
hi，i test VFocalLoss in yolov5，but not getting improvment.
did you have done some test about yolov5 ? or any suggestion ?
thank your ~
class VFocalLoss(nn.Module):
def __init__(self, loss_fcn, gamma=2.0, alpha=0.75): #runs/train/exp28
super(VFocalLoss, 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
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
Hi, I haven't yet played YOLOv5 with Varifocal Loss. I think you may try the methods below to see if you can achieve some improvements.
- Try different values of
gammaandalpha. - Predict the IACS (as the final detection score) instead of the classification score with Varifocal Loss. Please refer to our paper for what is the IACS.
Hello I tried to implement this in YOLOv5 but why do I get an error saying "AttributeError: 'Model' object has no attribute 'reduction'" Help me pleaseee thank youu
This is the full error I get
the FocalLoss there is actually a VFocalLoss, I just named it focal loss.
hi，i test VFocalLoss in yolov5，but not getting improvment. did you have done some test about yolov5 ? or any suggestion ? thank your ~
class VFocalLoss(nn.Module):
def __init__(self, loss_fcn, gamma=2.0, alpha=0.75): #runs/train/exp28 super(VFocalLoss, 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 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
