This issue helps towards making cuml.accel more compatible with scikit-learn. The TSNE estimator in scikit-learn stores the number of iterations actually run as a fitted attribute est.n_iter_.
The cuml version should do the same.
I think this is not as easy as simply creating the attribute, we need a way to thread the information back from the C++ implementation that does the work into an attribute.
This is being tested in sklearn.manifold.tests.test_t_sne::test_max_iter_used (or sklearn.manifold.tests.test_t_sne::test_n_iter_used in scikit-learn 1.4), which is currently marked as xfail. There are probably several other tests that will start passing once we have this attribute.
This issue helps towards making
cuml.accelmore compatible with scikit-learn. TheTSNEestimator in scikit-learn stores the number of iterations actually run as a fitted attributeest.n_iter_.The cuml version should do the same.
I think this is not as easy as simply creating the attribute, we need a way to thread the information back from the C++ implementation that does the work into an attribute.
This is being tested in
sklearn.manifold.tests.test_t_sne::test_max_iter_used(orsklearn.manifold.tests.test_t_sne::test_n_iter_usedin scikit-learn 1.4), which is currently marked as xfail. There are probably several other tests that will start passing once we have this attribute.