def smooth_l1_loss_modify(predict, target, mask, sigma=3):
# predict: bx2x96x128
# target : bx2x96x128
# mask : bx2x96x128
num_object = mask.sum().item() / mask.size(1)
sigma2 = sigma * sigma
diff = predict[mask] - target[mask]
diff_abs = diff.abs()
near = (diff_abs < 1 / sigma2).float()
far = 1 - near
return (near * 0.5 * sigma2 * torch.pow(diff, 2) + far * (diff_abs - 0.5 / sigma2)).sum() / num_object
def l2_loss_modify(predict, target, mask):
# predict: bx2x96x128
# target : bx2x96x128
# mask : bx2x96x128
num_object = mask.sum().item() / mask.size(1)
if num_object == 0 : num_object = 1
masked_predict = predict[mask]
masked_target = target[mask]
return torch.pow(masked_predict - masked_target, 2).sum() / num_object
def l1_loss_modify(predict, target, mask):
# predict: bx2x96x128
# target : bx2x96x128
# mask : bx2x96x128
num_object = mask.sum().item() / mask.size(1)
if num_object == 0 : num_object = 1
masked_predict = predict[mask]
masked_target = target[mask]
return torch.abs(masked_predict - masked_target).sum() / num_object
标签：bx2x96x128,num,target,predict,object,mask,soomthl1,l2,l1 From： https://www.cnblogs.com/xiaoruirui/p/16851858.html