zero question

[MrChenDingding]

It is a binary classification problem. After using the boundary loss proposed by you, all categories are judged as the background in the output result of the network, that is, the output result is 0, and no segmentation area is found. What is the problem causing this?

This code is to find the loss function and distance graph

label = Image.open(os.path.join(os.path.join(self.label_path, image_name)))
label = np.array(label)
seg_label = np.eye(self.num_classes)[label.reshape([-1])]
seg_label = seg_label.reshape((int(self.input_shape[0]), int(self.input_shape[1]), self.num_classes))
seg_label = np.transpose(seg_label, [2, 0, 1])
dist_map = self.one_hot2dist(seg_label)

def one_hot2dist(self, seg_label: np.ndarray) -> np.ndarray:
"""
根据独热编码分割图像计算距离图，用于后续损失计算
:param seg:
:return:
"""
C: int = len(seg_label)

res = np.zeros_like(seg_label)
for c in range(C):
    posmask = seg_label[c].astype(np.bool)

    if posmask.any():
        negmask = ~posmask
        # print('negmask:', negmask)
        # print('distance(negmask):', distance(negmask))
        res[c] = distance(negmask) * negmask - (distance(posmask) - 1) * posmask
        # print('res[c]', res[c])
return res

out = model(img)

----------------------#

损失计算

----------------------#

if ce_loss:
ce = CE_Loss(out, seg_labels, weights)
if focal_loss:
focal = Focal_Loss(out, seg_labels, weights)
if dice_loss:
dice = DiceLoss(out, seg_labels, idc=[0, 1])
if boundary_loss:
boundary = SurfaceLoss(out, dist_maps, idc=idc)

loss = (ce + focal + dice) * (1 - alpha) + alpha * boundary

def SurfaceLoss(inputs, dist_maps, idc=[1]):
temp_inputs = torch.softmax(inputs, dim=1)

temp_inputs = torch.softmax(inputs, dim=1)[:, idc, ...]

temp_dist_maps = dist_maps[:, idc, ...]

multipled = temp_inputs * temp_dist_maps

loss = multipled.mean()

return loss