2.19 The high-level API end to end
A compact smoke test of the everyday xgboost surface, using plain Racket lists for the data: build a DMatrix, train, predict, read a metric with eval-one-iter / parse-eval-line, and confirm a save-model / load-model round-trip reproduces the predictions.
<r13-require> ::=
(require racket/file xgboost)
<r13-provide> ::=
(provide run-example)
<r13-run> ::=
(define (run-example) (define features '((1.0 2.0 0.5) (2.0 1.0 1.5) (3.0 0.5 0.0) (0.5 3.0 2.0) (4.0 2.0 1.0) (1.5 1.5 0.5) (2.5 3.5 1.5) (0.0 1.0 0.0))) (define labels '(3.5 3.5 6.5 2.0 9.0 4.0 7.0 1.0)) (define dtrain (make-dmatrix features #:labels labels)) (define booster (train dtrain #:objective "reg:squarederror" #:eval-metric "rmse" #:max-depth 3 #:eta 0.1 #:verbosity 0 #:rounds 20)) (define preds (predict booster dtrain)) (define metrics (parse-eval-line (eval-one-iter booster 19 (list (cons "train" dtrain))))) (define tmp (make-temporary-file "xgboost-e2e-~a.json")) (define loaded-preds (dynamic-wind void (lambda () (save-model booster tmp) (predict (load-model tmp) dtrain)) (lambda () (when (file-exists? tmp) (delete-file tmp))))) (hash 'predictions preds 'loaded-predictions loaded-preds 'metrics metrics))
The harness "test/13-high-level-root-api.rkt" prints the prediction count and train RMSE, and asserts the predictions are finite and survive the save/load round-trip unchanged.
<*> ::=