jumpy validation accuracy graph

helo,
i have a very weird problem.
my validation accuracy graph is very jumpy, and i dont know how to fix it
this is the graph:

this is a multi label problem. i calculate accuracy with 0.5 threshold
i think that the problem related to my loss function
(i needed to minimize even one class success in the model prediction, and not only when he success in all of them)

i have a costume loss function:

    def loss_function(predictions, labels, batch_size):
        eps = 0.000000000000000000000000001
        losses = 0
        k = len(lebels[0])
        ones = torch.ones(1, 1).expand(batch_size, k).cuda()
        loss1 = -((ones-labels)*(((ones-predictions)+(eps)).log())).sum(dim=1)
        prod = (ones-labels)*k - labels*((predictions+ eps).log())
        loss2 = torch.min(prod, dim=1)[0]
        losses = (loss1 + loss2).sum()
        return losses / batch_size

i hope someone can understand the problem
thank you very much.