FocalTversky loss function expected input/output size?

1

I am currently working on a semantic segmentation model and I am trying out a different loss function is this case. The loss function i am using is FocalTversky Loss function.

‘’’
class FocalTverskyLoss(nn.Module):
def init(self):
super().init()

def forward(self, preds, target, alpha=0.7, beta = 0.3, epsilon=1e-6, gamma=3):
    
    preds = torch.sigmoid(preds) 
    
    #flatten label and preds tensors
    preds = preds.reshape(-1)
    target = target.reshape(-1)
    
    #True Positives, False Positives & False Negatives 
    TP = (preds * target).sum()
    FP = ((1-target) * preds).sum()
    FN = (target * (1-preds)).sum()
    Tversky = (TP + epsilon)/(TP + alpha*FP + beta*FN + epsilon)
    FocalTversky = (1 - Tversky)**gamma
    
    return FocalTversky

‘’’

However when I run it in my code like this

outputs = self.model(images)
loss = self.loss_function(preds=outputs,target=labels).to(device)
train_loss += loss.item()
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()

it gives this error,

RuntimeError Traceback (most recent call last)
in
1 imgdir = ‘Z:\HuaSheng\datasets\sen12msgrss\DFC_Public_Dataset’
----> 2 trainer(imgdir= imgdir, classes = list(range(0,10)) ,fsave=‘Rip_chkpt_FTL.pth’, reloadmode=‘same’, checkpoint=None, num_epochs = 110)

in init(self, imgdir, classes, num_epochs, fsave, reloadmode, checkpoint, bs, report)
59 for self.epoch in range(self.num_epochs):
60 print(’\n’+’*'6+‘TRAIN FOR ONE EPOCH’+’'6)
—> 61 train_loss = self.train()
62
63 print(’\n’+’‘6+‘EVAL FOR ONE EPOCH’+’’*6)

in train(self)
163 self.writer.flush()
164
–> 165 loss = self.loss_function(preds=outputs,target=labels.view(1, -1)).to(device)
166 #weight=torch.FloatTensor([0.,2.5,0.,1.5,2.1,0.3,4.5,0.,4.5,0.]).to(device)
167 train_loss += loss.item()

~\anaconda3\envs\pytLocal38\lib\site-packages\torch\nn\modules\module.py in _call_impl(self, *input, **kwargs)
725 result = self._slow_forward(*input, **kwargs)
726 else:
–> 727 result = self.forward(*input, **kwargs)
728 for hook in itertools.chain(
729 _global_forward_hooks.values(),

in forward(self, preds, target, alpha, beta, epsilon, gamma)
57
58 #True Positives, False Positives & False Negatives
—> 59 TP = (preds * target).sum()
60 FP = ((1-target) * preds).sum()
61 FN = (target * (1-preds)).sum()

RuntimeError: The size of tensor a (5017600) must match the size of tensor b (501760) at non-singleton dimension 0

My input size for my preds and target is torch.Size([10, 10, 224, 224]), torch.Size([1, 501760])

Is there a way I can make my target dim 0 into 10? Or is that even right for me to do when I am training it? Thanks alot for your help

2

Could you explain what the shapes of preds and target represent?
Currently it seems preds uses a batch size of 10, while you have a single target, which will then raise this error after flattening these tensors.

PS: you can post code snippets by wrapping them into three backticks ```, which makes debugging easier. :wink:

3

Hi @ptrblck, sorry for the poor posting format! haha The target represent the labels of the image and the prediction is the output after fitting in the model. The image I am working on right now consist of 13 channel images with 10 classes inside. The chip size of the image is 224. Where every pixel in the image contains a classes used for semantic segmantation modelling.

However, I have come across that using this loss function is not suitable as it is a multiclassifcation problem where as this focaltverskyloss is based off binarycrossentropy loss. So I think it may not be suitable to use this loss

4

Thanks for the update. While your use case makes sense, it’s still unclear why the batch size of the output images is 10 while the target has only a single sample.
If your target is supposed to contain 10 values, then I guess a reshaping operation might be wrong on this tensor.