WIP Fix 0.7/1.0 Incompatiabilities

[fundamental]

I'm pretty sure I used this project back in the 0.4 era or so and I was trying to use it again yesterday. It failed to load on the build of 0.7 I had due to some dependencies. I updated to 1.0 (being hopeful) and it looks like there's been a fair amount of bitrot due to all of the changes in the core Julia libs/API.

While I certainly haven't fixed all the problems, this set of commits should be a stepping stone in getting GassianMixtures to run again. I'm not sure if I'll have the time to finish updating the code, so if anyone is interested feel free to take over where I left off. The commits should be atomic enough that there aren't any surprises.

[davidavdav]

Thanks!

Yes, I haven't used this package in a long time myself. If I had to write it again, I'd probably do it differently this time.

There still are some travis failures. DO you think you can fix these?

[fundamental]

This PR is now 1 assert away from passing tests (which isn't to say that everything is working quite right, but it is a start) on 0.7. If the prediction assert at test/scikitlearn.jl:29 is disabled, it 'successfully' runs locally.

[cstjean]

Wow, that @assert should have been a @test.

It would be interesting to see what value sum(predict(gmm, X_train) .== 1) has. Since there's an srand in there, and Julia's RNG has changed since I wrote that code, it's possible that there's some outlier which is misclassified under srand(42). You might try srand(43) 🙂. Or you could @test_broken it, and someone else will look at it some day. ScikitLearn compatibility shouldn't prevent this package from being updated; thank you for your efforts.

[fundamental]

Tests 'pass' on 0.7, 1.0, and nightly now. There's plenty of smaller issues to be addressed, but it may be acceptable to merge this at this stage.

I'll make no guarantee of correctness, though it is less broken (on 0.7/1.0) than it was before.

[cstjean]

I think that in 0.7/1.0, this line should just be getindex.(argmax(ScikitLearnBase.predict_proba(gmm, X), dims=2), 2). Then the test might work.

[fundamental]

Fixed and the test works, though the intent isn't entirely clear given my quick read. Half of the return value is 1 and half is 2, and the assert now passes.

[cstjean]

The training set has 300 points in gaussian #1 and 600 in gaussian #2. They are cleanly separated. After performing clustering, how many points should be in cluster #1? Either 300 or 600, since we don't know which cluster corresponds to which gaussian.