machine learning - MLJ prediction type deterministic Error when using evaluate!()

Question

I have an issue running the evaluate!() command from MLJ. I am using the Multinomial Naive Bayes Classifier to classify sentiment in a bunch of tweets. When I run for single case like this it runs all fine,

using MLJ
    
X = coerce(wordCountVec,Count)
y = coerce(data.sentiment_labels, Multiclass)
train_idx,test_idx = partition(eachindex(y), 0.7,shuffle = true)

nb_m = @load MultinomialNBClassifier pkg = "NaiveBayes"
mach = machine(nb_m,X,y)
MLJ.fit!(mach, rows = train_idx)
    
yhat = MLJ.predict_mode(mach, rows = test_idx)
micro_f1score(yhat,y[test_idx])

#out > 0.68

But however when I try to use evaluate!() to use cross validation

evaluate!(
    mach,
    resampling = CV(nfolds = 3),
    measure = micro_f1score
)

I get the following error

ArgumentError:

[34mMultinomialNBClassifier @358[39m <: Probabilistic but prediction_type([34mMulticlassFScore{Float64,…} @139[39m) = :deterministic.

To override measure checks, set check_measure=false.

I am going to take a guess here and say it might be because when I predict in the first case (not using evaluate!()) I use predict_mode rather than predict, so I end up with deterministic values rather than probabilities. Any ideas what I can do it fix this when using the evaluate() function or is this some other error?

question from:https://stackoverflow.com/questions/65937483/mlj-prediction-type-deterministic-error-when-using-evaluate

Answer 1

You can specify the operation argument to evaluate!, like this:

evaluate!(
    mach,
    resampling = CV(nfolds = 3),
    measure = micro_f1score,
    operation = predict_mode
)