1.2.9 Convergence and errors🔗ℹ

Coordinate descent cycles over the coefficients until no update moves the objective by more than #:thresh (default 1e-7) times the null deviance, or until #:max-iters passes (default 100000). ols defaults to a tighter 1e-10, because the unpenalized solution is approached slowly. Every result records its passes; a tighter threshold takes more:

> (for/list ([thresh '(1e-7 1e-12)])     (elnet-result-num-passes (lasso X y #:lambda 0.1 #:thresh thresh)))

'(19 34)

Problems are reported in three ways:

• Before the solver runs, an argument of the wrong type, a ragged or empty matrix, a non-finite entry, or a response of the wrong length raises exn:fail:contract (see Data).

• Data that glmnet cannot fit raises exn:fail with a readable message: for example when every predictor is constant, when a class probability collapses under perfect separation (a larger #:lambda usually fixes that), or when a Gaussian response is constant, as R reports too.

• When glmnet fits no λ at all, for example because the solver hits #:max-iters before converging, the fit raises exn:fail with glmnet’s reason, naming the procedure called. R warns and returns an empty model instead. A regularization path that fails after fitting some λ keeps those and logs a warning with the topic 'glmnet, as R warns. The warning names the procedure called, and for cross-validation what it was fitting: all the data, or the training data of a held-out fold.

> (ols (matrix [[1.0 2.0] [1.0 2.0] [1.0 2.0]]) '(1.0 2.0 3.0))

ols: all used predictors have zero variance

> (lasso X y #:lambda 0.01 #:max-iters 1)

lasso: glmnet fitted no lambda: convergence was not reached

at the first lambda within #:max-iters passes; try a larger

#:max-iters (jerr=-1)