Getting a trait bound error when trying to use cross_validate from smartcore crate

help
1

It seems smartcore is fairly "popular" as far as Rust machine learning crates go, but I'm struggling to find more than a couple of bare examples. Like the title says, I'm trying to perform grid search parameter tuning with cross-validation on a random forest classifier, but I'm getting an error in the cross_validate() function. Here's the excerpt of the function I'm having an issue with (ignoring the return type and actual search values since I'm planning on updating those):

use smartcore::linalg::basic::arrays::{MutArray, Array2};
use smartcore::linalg::basic::matrix::DenseMatrix;
use smartcore::metrics::{mean_squared_error, accuracy};
use smartcore::model_selection::{train_test_split, cross_validate, KFold};
use smartcore::ensemble::random_forest_classifier::*;


fn grid_search(x_train: &DenseMatrix<f64>, y_train: &Vec<u32>) -> Option<(i32, u16, usize, usize)> {
    let mut best_score = 0.;
    let mut best_params = None;

    let n_trees = vec![10, 50, 100, 200, 500];
    let max_depth = vec![2, 4, 6, 8, 10];
    let min_samples_split = vec![2, 4, 6, 8, 10];
    let min_samples_leaf = vec![1, 2, 4, 6, 8, 10];
    let max_features = vec![2, 4, 6, 8, 10];

    for n_tree in n_trees {
        for m_depth in &max_depth {
            for m_split in &min_samples_split {
                for m_leaf in &min_samples_leaf {
                    for m_feat in &max_features {
                        let cv_score = cross_validate(
                            RandomForestClassifier::fit,
                            x_train,
                            y_train,
                            Default::default()
                                .with_n_trees(*m_feat)
                                .with_max_depth(*m_depth)
                                .with_min_samples_split(*m_split)
                                .with_min_samples_leaf(*m_leaf),
                        &KFold::default().with_n_splits(10),
                        &accuracy
                        ).unwrap();
                        
                        if cv_score.mean_test_score() > best_score {
                            best_score = cv_score.mean_test_score();
                            best_params = Some((n_tree, m_depth, m_split, m_leaf));
                        }
                    }
                }
            }
        }
    }

    println!("Best score: {}", best_score);
    println!("Best params: {:?}", best_params);

    Some(
        (
            best_params.unwrap().0,
            *best_params.unwrap().1,
            *best_params.unwrap().2,
            *best_params.unwrap().3
        )
    )
}

I'm having a hard time interpreting the error message that it's giving me since it says that the fit function within the cross_validate function needs to have the trait SupervisedEstimator , which, as far as I can tell, it does.

the trait bound `for<'r, 's> fn(&'r _, &'s _, smartcore::ensemble::random_forest_classifier::RandomForestClassifierParameters) -> std::result::Result<smartcore::ensemble::random_forest_classifier::RandomForestClassifier<_, _, _, _>, smartcore::error::Failed> {smartcore::ensemble::random_forest_classifier::RandomForestClassifier::<_, _, _, _>::fit}: smartcore::api::SupervisedEstimator<_, _, _>` is not satisfied
the following other types implement trait `smartcore::api::SupervisedEstimator<X, Y, P>`:
  <smartcore::ensemble::random_forest_classifier::RandomForestClassifier<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::ensemble::random_forest_classifier::RandomForestClassifierParameters>>
  <smartcore::ensemble::random_forest_regressor::RandomForestRegressor<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::ensemble::random_forest_regressor::RandomForestRegressorParameters>>
  <smartcore::linear::elastic_net::ElasticNet<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::linear::elastic_net::ElasticNetParameters>>
  <smartcore::linear::lasso::Lasso<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::linear::lasso::LassoParameters>>
  <smartcore::linear::linear_regression::LinearRegression<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::linear::linear_regression::LinearRegressionParameters>>
  <smartcore::linear::logistic_regression::LogisticRegression<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::linear::logistic_regression::LogisticRegressionParameters<TX>>>
  <smartcore::linear::ridge_regression::RidgeRegression<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::linear::ridge_regression::RidgeRegressionParameters<TX>>>
  <smartcore::naive_bayes::bernoulli::BernoulliNB<TX, TY, X, Y> as smartcore::api::SupervisedEstimator<X, Y, smartcore::naive_bayes::bernoulli::BernoulliNBParameters<TX>>>
and 7 others

main.rs(102, 40): required by a bound introduced by this call
mod.rs(239, 8): required by a bound in `smartcore::model_selection::cross_validate`

I realize this is probably more specific to smartcore than Rust, but I figured I'd post here in the hopes that someone might be able to help. If this is too specific, I'll create a GitHub issue on the repo.

2

I don't know smartcore, but judging from your error message, the function you provide as estimator does not implement the right trait (SupervisedEstimator). Looking at the implementors of SupervisedEstimator, RandomForestClassifier is amonst them. Could you try providing RandomForestClassifier, instead of RandomForestClassifier::fit:

3

I tried that as well but end up getting

expected value, found struct `RandomForestClassifier`
random_forest_classifier.rs(101, 1): `RandomForestClassifier` defined here
main.rs(103, 29): use struct literal syntax instead: `RandomForestClassifier { trees: val, classes: val, samples: val }`

From the description of the cross_validate function arguments:

fit_estimator - a fit function of an estimator
4

Ok, my bad. SupervisedEstimator provides the new method. So I guess

let cv_score = cross_validate(
    RandomForestClassifier::new(),
    ...
)

should do the trick.

5

I think they forgot to update the documentation. In version v0.2.1 there was the fit_estimator parameter. If you look at the function signature in version v0.3.0 there is the _estimator argument, but no fit_estimator any more.

6

Thank you so much, this was honestly driving me crazy. I don't see any errors anymore though, so I'm very hopeful that it'll actually run smoothly. It doesn't look like smartcore's random forest has any built-in parallelization, so it'll probably be a while before I find out if something breaks once I start running it.

1 Like
7

This topic was automatically closed 90 days after the last reply. We invite you to open a new topic if you have further questions or comments.