Fix deprecation

[kahaaga]

In #427, I fixed the error disallowing different-type outcome containers for Probabilities. However, that introduced the deprecation warning again. This fixes the resulting deprecations warning.

[Datseris]

Increment patch, merge, and tag, please.

[Datseris]

AH NO! Test fail!!!!!!!!

[kahaaga]

Test fails are unrelated. Seems like a random tolerance thing related to the sum of probabilities. We can safely merge this without affecting anything else.

[Datseris]

Test have been restarted 5 times by now but they still fail. I think the failure should be fixed. It isn't even a test failure but an @assert statement failing. It isn't failing on main branch.

[kahaaga]

Test have been restarted 5 times by now but they still fail. I think the failure should be fixed. It isn't even a test failure but an @Assert statement failing. It isn't failing on main branch.

The error occurred because the sums of the counts computed using the CosineSimilarityBinning outcome space didn't match the expected number of outcomes obtained from encoded_space_cardinality (see attached definition below).

The actual counts were missing an observation. The reason is that the cosine similarity is supposed to be bounded on [-1, 1], but sometimes, due to imprecision errors, it sometimes got a value of 1.00000000...2. The FixedRectangularBinning used to encode the cosine similarity distances is defined on [-1, 1 + eps()). But this didn't catch these points that fall juuuust outside the right-hand range of the binning. I fixed it by restricting the computed cosine similarity value to 1-eps().

Note: another alternative would be to use the precise argument to the FixedRectangularBinning, but that would make binning four times slower, so I opted for just adjusting the value of the cosine similarity so that it is within the bin edges.

Why this suddenly occurs I have not been able to track down. My best guess is that it has something to do with random number generation during testing.

Anyways: it is fixed now, so I'm merging and tagging.

relevant code

function probs_and_outs_from_histogram(est::AddConstant, outcomemodel::OutcomeSpace,
        cts::Counts{T, 1}, outs, x) where T

    c = est.c
    m = length(cts)
    n = encoded_space_cardinality(outcomemodel, x) # Normalize based on *encoded* data.
    probs = zeros(m)
    for (k, nₖ) in enumerate(cts)
        probs[k] = (nₖ + c) / (n + (c * m))
    end
    @assert sum(probs) ≈ 1
    
    return Probabilities(probs, outs,)
end

[Datseris]

thanks!