303217b34070dc47a86622b62764098999b0d7f5,gpytorch/lazy/lazy_tensor.py,LazyTensor,_quad_form_derivative,#LazyTensor#Any#Any#,378

Before Change

        toggled = [False] * len(args)

        for i, arg in enumerate(args):
            if not arg.requires_grad:
                arg.requires_grad = True
                toggled[i] = True

        loss = (left_vecs * self._matmul(right_vecs)).sum()
        loss.requires_grad_(True)
        grads = torch.autograd.grad(loss, args, allow_unused=True)

After Change

        from collections import deque

        args = tuple(self.representation())
        args_with_grads = tuple(arg for arg in args if arg.requires_grad)

        // Easy case: if we don"t require any gradients, then just return!
        if not len(args_with_grads):
            return tuple(None for _ in args)

        // Normal case: we"ll use the autograd to get us a derivative
        with torch.autograd.enable_grad():
            loss = (left_vecs * self._matmul(right_vecs)).sum()
            loss.requires_grad_(True)
            actual_grads = deque(torch.autograd.grad(loss, args_with_grads, allow_unused=True))

        // Now make sure that the object we return has one entry for every item in args
        grads = []
        for arg in args:
            if arg.requires_grad:
                grads.append(actual_grads.popleft())
            else:
                grads.append(None)

Instances