EEGClassifier classifier not supporting criterion=torch.nn.CrossEntropyLoss()

[bokey007]

Hi, I want to pass custom criterion to the EEGClassifier.

but its seems to support only torch.nn.NLLLoss.

I even tried criterion=torch.nn.CrossEntropyLoss() but it does not work.

Is their any way to pass our custom loss instead of NLLLoss.

looking forward to your answer,
Bokey

[gemeinl]

If you attempt to do trialwise decoding, criterion=CrossEntropyLoss will work.
If you attempt to do cropped decoding, set criterion=CroppedLoss and criterion__loss_function=torch.nn.functional.cross_entropy).

[bokey007]

Hi thanka for the reply. I am trying to create custom criterion as follows :

class Cosine(nn.modules.loss._WeightedLoss):
    def __init__(self, s=64.0, m=0.40):
        super(CosFace, self).__init__()
        self.s = s
        self.m = m


    def forward(self, cosine: torch.Tensor, label: torch.Tensor):
        index = torch.where(label != -1)[0]
        m_hot = torch.zeros(index.size()[0], cosine.size()[1], device=cosine.device)
        m_hot.scatter_(1, label[index, None], self.m)
        cosine[index] -= m_hot
        ret = cosine * self.s
        return ret

But am getting following error 👍

RuntimeError: grad can be implicitly created only for scalar outputs

Can u please help me understand ?

[robintibor]

Hi,

as the error indicates, your loss function needs to return a single value (hence the "scalar outputs" error). May be your loss function creates a value per example? then you still need to call torch.mean on it before returning, e.g., maybe
return torch.mean(ret) would work?