Em Algorithm Speedup

[paschermayr]

Hi guys,

In your mle.jl file, on line 14 in the mle_step function, you have:

log_likelihoods = HMMBase.log_likelihoods(hmm, observations)

If you define the loglikelihoods function for both, univariate and multivariate, with a fixed type like:

function log_likelihoods(hmm::AbstractHMM{Univariate}, observations)::Array{Float64,2} 
    hcat(map(d -> logpdf.(d, observations), hmm.D)...)
end

you should archieve a drastic speed up. I had a similar problem and it did the trick. I am no expert, but I think the 'log_likelihoods' functions within the 'mle' function is not type stable.

Also, I have to say that your module is really well structured and intuitively written!

[maxmouchet]

Hi,

Thanks for the suggestion :-)

It does not seems to make a difference at a first glance (see the screenshot) but you're right that there is some Any types inside the function.

I'll look more into details.

[maxmouchet]

I don't know if we can do better in terms of time (2x17us ~= 35us), however we may be able to use slightly less memory (~40KiB instead of 79KiB):

@btime logpdf.(hmm.D[1], obs);
17.398 μs (5 allocations: 19.72 KiB)

[paschermayr]

Thanks for your answer!

I used the log_likelihood function in another function and the compiler wasn't able to correctly guess the output - so I thought you might get some speed up when you are using your function in the MLE part of your module.

On a related note, I saw the graph where you guys benchmarked your package with hmmlearn - why this drastic time increase for the likelihood calculation? This should be already be a part in the forward pass.