How to use correctly multilabel classification in pytorch

vision
1

Hi guys, I have some csv file with structure

  img |class0|class1|class2|class3|class4|class5...
|-----|------|------|------|------|------|------
|path1|0     |0     |1     |0     |0     |1     
|path2|0     |1     |1     |0     |1     |0     
|path3|0     |0     |0     |1     |0     |0     
.................................................
.................................................
|pathn|1     |1     |0     |1     |0     |0     
|-----|------|------|------|------|------|------
class mydataset(Dataset):
...
def __getitem__(self,idx):
...
labels=np.array(self.labels.iloc[idx,1:])    
    return {'img':img,'labels':labels}

I am trying to use

n_classes=df.shape[0]-1
model = EfficientNet.from_pretrained('efficientnet-b1',num_classes=n_classes).to(device)

with

loss = torch.nn.BCELoss()

but it doesn’t work, how to implement in my case multilabel classification?

2

What is the exact error you are getting in this case?

You might want to check that the datatype of labels is a torch tensor as expected.

The documentation shows a simple of example of input being passed to the loss function: BCELoss — PyTorch 1.9.0 documentation
although in your case you will likely also have a batch dimension.