def test_vsi_loss(input_tensors: Tuple[torch.Tensor, torch.Tensor], device: str) -> None:
x, y = input_tensors
x.requires_grad_()
loss = VSILoss(data_range=1.)(x.to(device), y.to(device))
loss.backward()
assert torch.isfinite(x.grad).all(), \
f'Expected finite gradient values after back propagation, got {x.grad}'
def test_vsi_loss_zero_for_equal_input(input_tensors: Tuple[torch.Tensor, torch.Tensor], device: str) -> None:
x, _ = input_tensors
y = x.clone()
x.requires_grad_()
loss = VSILoss(data_range=1.)(x.to(device), y.to(device))
assert torch.isclose(loss, torch.zeros_like(loss)), \
f'Expected loss equals zero for identical inputs, got {loss}'