svm-techml-smile

Bridging tools between tech.ml.dataset and Smile, for SVM/SVR use cases.

Why

tech.ml used to ship this interop itself, but dropped it when Smile changed its licence. This library carries the pieces back for the workflows we actually use: dataset ↔ DataFrame / double-array conversion, RBF kernel estimation, std-scaling, and SVM/SVR fit + predict.

It is also the only place Smile is imported. Consumers depend on this library and get the Smile version pinned here (smile-core 7.0.0) — no Smile coordinate in their own deps.edn / project.clj, no version skew across projects.

Install

[net.clojars.alex314159/svm-techml-smile "0.1.1"]

smile-core 7.0.0 is compiled for Java 25+, so point JAVA_HOME at a 25/26 JDK and pass --enable-native-access=ALL-UNNAMED (already in this project's :jvm-opts).

Namespaces

ns	what's in it
svm-techml-smile.core	re-exports everything below — require this one ns if you don't care about the split
svm-techml-smile.bridge	dataset/column → Smile DataFrame, DoubleVector, double arrays
svm-techml-smile.kernel	RBF sigma estimation, std-scaling, fit-RBF-SVM / fit-RBF-SVR, predict, rsq
svm-techml-smile.features	small spec DSL for deriving dataset columns
svm-techml-smile.preprocessing	cross-sectional SVR model spec + data prep (one-hot, scaling, splits)
svm-techml-smile.classification	time-series SVM model spec, label conversion, evaluation

Note core aliases via a macro: REPL-driven tooling (Calva, CIDER) resolves the docs, purely static analysis (clj-kondo, offline Cursive index) doesn't — require the defining ns there.

Development

lein test
lein bench    # criterium, :bench profile only — never in the released jar