Description

Proof of concept repo to run the scikit-learn tests in Pyodide.

You can run all tests by module like this:

./run-tests-by-modules.sh

You can run selected tests like this:

node --experimental-fetch scikit-learn-pytest.js --pyargs 'sklearn.tree'

scikit-learn-pytest.js is strongly inspired from numpy/numpy#21895.

Manually curated list of issues

Test failures

• sklearn.ensemble.tests.test_forest::test_poisson_vs_mse
• sklearn.experimental.tests.test_enable_iterative_imputer.py::test_imports_strategies
• sklearn.experimental.tests.test_enable_successive_halving.py::test_imports_strategies
• sklearn.feature_extraction.tests.test_text.py::test_tfidf_no_smoothing
• plenty of errors in sklearn._loss.tests.test_loss::test_loss_dtype
• sklearn.svm.tests.test_bounds::test_newrand_set_seed[None-81]
• sklearn.tests.test_build::test_openmp_parallelism_enabled
• sklearn.tree.tests.test_tree::test_poisson_vs_mse
• sklearn.utils.tests.test_estimator_checks::test_all_estimators_all_public

Fatal errors

Two different kind of errors: "null function or function signature mismatch" and "memory access out of bounds".

• sklearn.decomposition.tests
• sklearn.ensemble.tests
• sklearn.inspection.tests
• sklearn.linear_model.tests
• sklearn.tests
• sklearn.utils.tests

TODO: This would be nice to try to pinpoint whether some particular tests causes the issue. In my experience the pytest output can not be trusted for this since it can vary from run to run. collected-tests.txt could be used to run tests by smaller chunks in separate node instances.