[BUG] correctly set tags of Hurdle and Truncated depending on inner distribution

[fkiraly]

This PR correctly sets the tag of Hurdle and Truncated depending on the inner distribution.

Related to #566

[tingiskhan]

Hmm, this seems really strange:

Why is it testing the PMF for a truncated Normal? I think this was the issue I had earlier as well, when I used the wrong "measure" tag.

[tingiskhan]

I think that this might be the issue:

I think we need to remove the method names that aren't relevant when changing measuretype.

[tingiskhan]

@fkiraly , see #571

[fkiraly]

see #571, continuous distributions should return 0 for pmf - this should actually be the default?

[fkiraly]

The "one-off" error from #566 is still present. Any ideas?

[tingiskhan]

Hm, very strange - I see that it tests Poisson for Truncated, I thought I changed all Poisson references to Negative Binomial. What happens if you replace it with NegativeBinomial instead?

[fkiraly]

that would not fix the error, just "make the tests pass"?

[tingiskhan]

I'll debug and see what's wrong!

[tingiskhan]

Can't seem to reproduce the error:

import tqdm

from skpro.distributions import Normal, TruncatedDistribution, Hurdle
import numpy as np


def ppf_and_cdf(d):
    capabilities_exact = d.get_tags()["capabilities:exact"]

    if "ppf" not in capabilities_exact or "cdf" not in capabilities_exact:
        return
    x = d.sample()
    x_approx = d.ppf(d.cdf(x))
    if d.ndim > 0:
        assert np.allclose(x.values, x_approx.values)
    else:
        assert np.allclose(x, x_approx)


cases = TruncatedDistribution.get_test_params()

for i in tqdm.tqdm(range(1_000)):
    for case in cases:
        dist = TruncatedDistribution(**case)
        ppf_and_cdf(dist)

Tried bumping to 5,000, still no issue. I'm on scipy==1.16.0 and Python=3.12. Will try to update scipy if see if that's the issue.

Edit: No issue

[tingiskhan]

Perhaps we should move the eps fix to truncated instead?

[tingiskhan]

@fkiraly, what happens if you merge the changes of #576?

[fkiraly]

ok, I will try to merge #576 into this and then we see

[fkiraly]

I think there is a genuine bug in #576, unrelated to the fix. What I will do, to make debugging easier: I will merge this without #576. and then you can run check_estimator locally to fix the distribution - does this make sense?

[tingiskhan]

Sounds good. I had messed up which variable to look at when selecting the eps since I assumed that shifted_p would be an array, hence the failing tests