Predictions are not consistent

rikhuijzer commented 2 years ago

As kindly pointed out by @ablaom on Slack and copied here to avoid the information being lost.

In general, integrating PPL into MLJ is hard based on experience with SossMLJ.jl. For the first model in the test suite:

julia> predictions1 = predict(tm, chns, X)[3:4]
2-element Vector{Float64}:
 -0.3010857674840958
  0.0005393573897319209

julia> predictions2 = predict(tm, chns, MLJBase.selectrows(X, 3:4))
2-element Vector{Float64}:
 -0.28599208954026617
  0.005011192524686957

These should be the same because of the basic MLJ assumptions outlined in https://alan-turing-institute.github.io/MLJ.jl/dev/adding_models_for_general_use/#Mathematical-assumptions.

Related discussed at

cscherrer commented 2 years ago

Is this a SossMLJ issue?

ablaom commented 2 years ago

Nope. I'm pretty sure we resolved that one!

rikhuijzer / MLJTuringInterface.jl

Predictions are not consistent #3