how to use AC loss or ACE loss

[guava1427]

Hello,

I tried to use AC or ACE losses instead of CE loss for the binary segmentation. Though I have used a certain network for the CE loss many times, the network does not work for AC/ACE losses.

In the meantime, I used a label array whose shape is equal to the prediction. Also, its channels have zero or one values like the code below.

from aceloss import ACLossV2
criterion = ACLossV2(classes=2)
outputs = model(inputs)
masks2 = torch.zeros_like(outputs)
masks2[:, 0, :, :] = (masks == 0).squeeze(1) # shape: [batch size, channel size, width, height]
masks2[:, 1, :, :] = (masks == 1).squeeze(1)
loss = criterion(outputs, masks2)

The loss value is larger than 1e4, and the output for the prediction looks meaningless. Did I miss something? I didn't change any code for the ACELoss class. Thank you.

[Luoxd1996]

Hi，
I don't know what's your problem you have met. In my understanding, the output needs to be seed into an activation function (softmax or sigmoid) and then calculate loss functions.
Best,
Xiangde.

[YangBing-9527]

Hello,

I tried to use AC or ACE losses instead of CE loss for the binary segmentation. Though I have used a certain network for the CE loss many times, the network does not work for AC/ACE losses.

In the meantime, I used a label array whose shape is equal to the prediction. Also, its channels have zero or one values like the code below.

from aceloss import ACLossV2 criterion = ACLossV2(classes=2) outputs = model(inputs) masks2 = torch.zeros_like(outputs) masks2[:, 0, :, :] = (masks == 0).squeeze(1) # shape: [batch size, channel size, width, height] masks2[:, 1, :, :] = (masks == 1).squeeze(1) loss = criterion(outputs, masks2)

The loss value is larger than 1e4, and the output for the prediction looks meaningless. Did I miss something? I didn't change any code for the ACELoss class. Thank you.

I also encountered the same problem. UNET + MSELoss works normally, but ACLoss does not converge, even when I used the pre-trained model.

[PhuongNDVN]

I got same issue, the loss reduces slowly and IoU does not increase. I trained with BCELoss/CrossEntropyLoss before and my Network works well. Here is my code for Active Contour Loss

class ActiveContourLoss(torch.nn.Module):
    def __init__(self, miu=1.0, numClasses=1):
        super(ActiveContourLoss, self).__init__()
        self.miu = miu
        self.numClasses = numClasses


    def forward(self, pred, mask):
        '''
        pred: prediction shape (B, numClasses, W, H)
        mask: ground truth (B, W, H)
        '''
        if self.numClasses == 1:
            pred = torch.sigmoid(pred)
        else:
            pred = torch.nn.functional.softmax(pred, dim=1)
        min_pool_x = torch.nn.functional.max_pool2d(pred * -1, (3, 3), 1, 1) * -1
        contour = torch.relu(torch.nn.functional.max_pool2d(
            min_pool_x, (3, 3), 1, 1) - min_pool_x)
        
        # length
        length = torch.sum(torch.abs(contour))
        
        # regions
        label = torch.zeros_like(pred)
        for k in range(self.numClasses):
            value = k
            if(self.numClasses == 1):
                value = 1
            label[:, k, :, :] = (mask == value)

        label = label.float()
        c_in = torch.ones_like(pred)
        c_out = torch.zeros_like(pred)
        region_in = torch.abs(torch.sum(pred * ((label - c_in) ** 2)))
        region_out = torch.abs(torch.sum((1 - pred) * ((label - c_out) ** 2)))
        region = self.miu * region_in + region_out

        return (region + length)