Raise NotImplementedError on missing backend functionality

[fritzo]

Addresses #2053

Note I've been having mkl errors when using .summary(). It would be nice to find a non-mkl workaround if not torch.backends.mkl.is_available().

Tested

• added a unit test

[neerajprad]

if not torch.backends.mkl.is_available()

I think the diagnostics should just have nan if mkl isn't available (that's probably required by the fft function), and IIRC we made a change to do that sometime back. This looks like a regression to me. cc. @fehiepsi

[neerajprad]

LGTM! I'll fix the mkl issue in another PR.

[neerajprad]

@fritzo - could you paste the error trace? I think the only diagnostic function that is likely to throw this is effective_sample_size but that is wrapped in a _safe block. I think the issue is that perhaps pytest is treating that warning as an error? If that's the case, we could just add @pytest.mark.filterwarnings("ignore:..") to the test.

def _safe(fn):
    """
    Safe version of utilities in the :mod:`pyro.ops.stats` module. Wrapped
    functions return `NaN` tensors instead of throwing exceptions.

    :param fn: stats function from :mod:`pyro.ops.stats` module.
    """
    @functools.wraps(fn)
    def wrapped(sample, *args, **kwargs):
        try:
            val = fn(sample, *args, **kwargs)
        except Exception:
            warnings.warn(tb.format_exc())
            val = torch.full(sample.shape[2:], float("nan"),
                             dtype=sample.dtype, device=sample.device)
        return val

    return wrapped

[fritzo]

Thanks, @neerajprad your suggestion of @pytest.mark.filterwarnings fixed my issue.