Difficulties upgrading from v0.2.1 to v0.3.1

[jblindsay]

Hi again, I recently updated from v0.2.1 and have had to deal with several breaking changes in the API. Most of them were things that I could work around, however, there is one that has caused me much issue and I cannot figure out how to resolve it. RandomForestClassification now requires that the target variable satify Ord, meaning that floating-point variables are not allowed. I get that, being a classification-type analysis, the target should be nominal/ordinal in nature. However, what I cannot figure out is that ClassificationMetrics now has a trait that requires the RealNumber + FloatNumber traits. This causes a compilation error for me. What am I not understanding here? Also, out of curiosity, why has the accuracy() function in ClassificationMetrics been removed? Thanks again.

[Mec-iS]

Hi!

As per previous question, please check first the tests module in the relative module, in this case ensemble/random_forest_classifier.rs. There should be also an example how to use the accuracy metrics, it has been moved up into metrics/accuracy.rs, not removed.

Consider providing code examples and the compilation error to facilitate debugging when you run into a issue.

Cheers

[jblindsay]

Hi @Mec-iS I appreciate you taking the time to respond so quickly. Perhaps I'm being daft here (likely the case!) but I don't see any examples of precision, recall, or f1 in the tests for the random forest classifier. And I cannot seem to use these functions from the Classificationmetrics struct because they require the RealNumber and FloatNumber traits, which the RandomForestClassification struct cannot use, because it itself requires the Ord trait which f32 and f64 do not implement. I just can't seem to figure out how to rectify this and unfortunately, unless I'm missing it, I don't see those methods used in the tests. I see accuracy, and I can get it to work, but not the other metrics. Thanks once again.

[Mec-iS]

Hi,
no problem, it is usually better to provide the code snippet and the compiler error as it is much easier to understand the context that trying to explain it in words.

From what I see you are trying to pass a Vec<usize> to the first argument of precision(...) while the score can only be computed on f32 or f64. In the new version types are a little stricter so the compiler complains. The example in the precision.rs module uses this starting data:

// predicted data
let y_pred: Vec<f64> = vec![0., 1., 1., 0.];
// ground truth
let y_true: Vec<f64> = vec![0., 0., 1., 1.];

As you can see y_pred needs now to be a floating point value.
You should modify your code to use f32 as data type, so how you should modify your code depends on how you compute/set the y_train value.

[jblindsay]

I suppose my issue though is that I can't call:

let classifier = RandomForestClassifier::fit(
        &x_train, 
        &y_train, 
        params
    ).expect("Error while developing the model.");

With y_train as a Vec<f32> with the changes in v0.3.1. Because it must implement Ord it must be an integer type. Ultimately, with v0.3.1, the classification models must be fit with integer type independent variables, but the metrics require floating-point values.

[Mec-iS]

yes the first ones are labels (usize that is what Rust uses for indices in arrays) but the second one needs floating point numbers to compute the score.