Our custom loss does not converge with PyTorch

1

Here is the loss code:

    def our_loss(output, target):
         global S
         global batchSize
         global _EPSILON

         y_pred = torch.clamp(output, _EPSILON, 1.0-_EPSILON)
         out = -(target * torch.log(y_pred) + (1.0 - target) * torch.log(1.0 - y_pred))
         out = torch.flatten(out)
         tmp = torch.mv(S.float() , out.float())
         loss = torch.dot(out.t(), tmp)
         return loss

where S is a RBF generated Gram matrix of the input patterns.