scikit-learn · ogrisel · Dec 21, 2021 · Oct 21, 2021 · Oct 21, 2021 · Oct 22, 2021
diff --git a/doc/whats_new/v1.1.rst b/doc/whats_new/v1.1.rst
@@ -266,6 +266,13 @@ Changelog
   Setting a transformer to "passthrough" will pass the features unchanged.
   :pr:`20860` by :user:`Shubhraneel Pal <shubhraneel>`.
 
+:mod:`sklearn.svm`
+...................
+
+- |Enhancement| :class:`svm.OneClassSVM`, :class:`svm.NuSVC`,
+  :class:`svm.NuSVR`, :class:`svm.SVC` and :class:`svm.SVR` now expose
+  `n_iter_`, the number of iterations of the libsvm optimization routine.
+  :pr:`21408` by :user:`Juan Martín Loyola <jmloyola>`.
 - |Fix| :class: `pipeline.Pipeline` now does not validate hyper-parameters in
   `__init__` but in `.fit()`.
   :pr:`21888` by :user:`iofall <iofall>` and :user: `Arisa Y. <arisayosh>`.

diff --git a/sklearn/svm/_base.py b/sklearn/svm/_base.py
@@ -274,6 +274,17 @@ def fit(self, X, y, sample_weight=None):
             self.intercept_ *= -1
             self.dual_coef_ = -self.dual_coef_
 
+        # Since, in the case of SVC and NuSVC, the number of models optimized by
+        # libSVM could be greater than one (depending on the input), `n_iter_`
+        # stores an ndarray.
+        # For the other sub-classes (SVR, NuSVR, and OneClassSVM), the number of
+        # models optimized by libSVM is always one, so `n_iter_` stores an
+        # integer.
+        if self._impl in ["c_svc", "nu_svc"]:
+            self.n_iter_ = self._num_iter
+        else:
+            self.n_iter_ = self._num_iter.item()
+
         return self
 
     def _validate_targets(self, y):
@@ -320,6 +331,7 @@ def _dense_fit(self, X, y, sample_weight, solver_type, kernel, random_seed):
             self._probA,
             self._probB,
             self.fit_status_,
+            self._num_iter,
         ) = libsvm.fit(
             X,
             y,
@@ -360,6 +372,7 @@ def _sparse_fit(self, X, y, sample_weight, solver_type, kernel, random_seed):
             self._probA,
             self._probB,
             self.fit_status_,
+            self._num_iter,
         ) = libsvm_sparse.libsvm_sparse_train(
             X.shape[1],
             X.data,

diff --git a/sklearn/svm/_classes.py b/sklearn/svm/_classes.py
@@ -670,6 +670,13 @@ class SVC(BaseSVC):
 
         .. versionadded:: 1.0
 
+    n_iter_ : ndarray of shape (n_classes * (n_classes - 1) // 2,)
+        Number of iterations run by the optimization routine to fit the model.
+        The shape of this attribute depends on the number of models optimized
+        which in turn depends on the number of classes.
+
+        .. versionadded:: 1.1
+
     support_ : ndarray of shape (n_SV)
         Indices of support vectors.
 
@@ -925,6 +932,13 @@ class NuSVC(BaseSVC):
 
         .. versionadded:: 1.0
 
+    n_iter_ : ndarray of shape (n_classes * (n_classes - 1) // 2,)
+        Number of iterations run by the optimization routine to fit the model.
+        The shape of this attribute depends on the number of models optimized
+        which in turn depends on the number of classes.
+
+        .. versionadded:: 1.1
+
     support_ : ndarray of shape (n_SV,)
         Indices of support vectors.
 
@@ -1140,6 +1154,11 @@ class SVR(RegressorMixin, BaseLibSVM):
 
         .. versionadded:: 1.0
 
+    n_iter_ : int
+        Number of iterations run by the optimization routine to fit the model.
+
+        .. versionadded:: 1.1
+
     n_support_ : ndarray of shape (n_classes,), dtype=int32
         Number of support vectors for each class.
 
@@ -1328,6 +1347,11 @@ class NuSVR(RegressorMixin, BaseLibSVM):
 
         .. versionadded:: 1.0
 
+    n_iter_ : int
+        Number of iterations run by the optimization routine to fit the model.
+
+        .. versionadded:: 1.1
+
     n_support_ : ndarray of shape (n_classes,), dtype=int32
         Number of support vectors for each class.
 
@@ -1512,6 +1536,11 @@ class OneClassSVM(OutlierMixin, BaseLibSVM):
 
         .. versionadded:: 1.0
 
+    n_iter_ : int
+        Number of iterations run by the optimization routine to fit the model.
+
+        .. versionadded:: 1.1
+
     n_support_ : ndarray of shape (n_classes,), dtype=int32
         Number of support vectors for each class.
 

diff --git a/sklearn/svm/_libsvm.pxi b/sklearn/svm/_libsvm.pxi
@@ -56,6 +56,7 @@ cdef extern from "libsvm_helper.c":
                          char *, char *, char *, char *)
 
     void copy_sv_coef   (char *, svm_model *)
+    void copy_n_iter  (char *, svm_model *)
     void copy_intercept (char *, svm_model *, np.npy_intp *)
     void copy_SV        (char *, svm_model *, np.npy_intp *)
     int copy_support (char *data, svm_model *model)

diff --git a/sklearn/svm/_libsvm.pyx b/sklearn/svm/_libsvm.pyx
@@ -152,6 +152,9 @@ def fit(
 
     probA, probB : array of shape (n_class*(n_class-1)/2,)
         Probability estimates, empty array for probability=False.
+
+    n_iter : ndarray of shape (max(1, (n_class * (n_class - 1) // 2)),)
+        Number of iterations run by the optimization routine to fit the model.
     """
 
     cdef svm_parameter param
@@ -199,6 +202,10 @@ def fit(
     SV_len  = get_l(model)
     n_class = get_nr(model)
 
+    cdef np.ndarray[int, ndim=1, mode='c'] n_iter
+    n_iter = np.empty(max(1, n_class * (n_class - 1) // 2), dtype=np.intc)
+    copy_n_iter(n_iter.data, model)
+
     cdef np.ndarray[np.float64_t, ndim=2, mode='c'] sv_coef
     sv_coef = np.empty((n_class-1, SV_len), dtype=np.float64)
     copy_sv_coef (sv_coef.data, model)
@@ -248,7 +255,7 @@ def fit(
     free(problem.x)
 
     return (support, support_vectors, n_class_SV, sv_coef, intercept,
-           probA, probB, fit_status)
+           probA, probB, fit_status, n_iter)
 
 
 cdef void set_predict_params(

diff --git a/sklearn/svm/_libsvm_sparse.pyx b/sklearn/svm/_libsvm_sparse.pyx
@@ -41,6 +41,7 @@ cdef extern from "libsvm_sparse_helper.c":
                                   double, int, int, int, char *, char *, int,
                                   int)
     void copy_sv_coef   (char *, svm_csr_model *)
+    void copy_n_iter  (char *, svm_csr_model *)
     void copy_support   (char *, svm_csr_model *)
     void copy_intercept (char *, svm_csr_model *, np.npy_intp *)
     int copy_predict (char *, svm_csr_model *, np.npy_intp *, char *, BlasFunctions *)
@@ -159,6 +160,10 @@ def libsvm_sparse_train ( int n_features,
     cdef np.npy_intp SV_len = get_l(model)
     cdef np.npy_intp n_class = get_nr(model)
 
+    cdef np.ndarray[int, ndim=1, mode='c'] n_iter
+    n_iter = np.empty(max(1, n_class * (n_class - 1) // 2), dtype=np.intc)
+    copy_n_iter(n_iter.data, model)
+
     # copy model.sv_coef
     # we create a new array instead of resizing, otherwise
     # it would not erase previous information
@@ -217,7 +222,7 @@ def libsvm_sparse_train ( int n_features,
     free_param(param)
 
     return (support, support_vectors_, sv_coef_data, intercept, n_class_SV,
-            probA, probB, fit_status)
+            probA, probB, fit_status, n_iter)
 
 
 def libsvm_sparse_predict (np.ndarray[np.float64_t, ndim=1, mode='c'] T_data,

diff --git a/sklearn/svm/src/libsvm/LIBSVM_CHANGES b/sklearn/svm/src/libsvm/LIBSVM_CHANGES
@@ -7,4 +7,5 @@ This is here mainly as checklist for incorporation of new versions of libsvm.
   * Improved random number generator (fix on windows, enhancement on other
     platforms). See <https://github.com/scikit-learn/scikit-learn/pull/13511#issuecomment-481729756>
   * invoke scipy blas api for svm kernel function to improve performance with speedup rate of 1.5X to 2X for dense data only. See <https://github.com/scikit-learn/scikit-learn/pull/16530>
+  * Expose the number of iterations run in optimization. See <https://github.com/scikit-learn/scikit-learn/pull/21408>
 The changes made with respect to upstream are detailed in the heading of svm.cpp
diff --git a/sklearn/svm/src/libsvm/libsvm_helper.c b/sklearn/svm/src/libsvm/libsvm_helper.c
@@ -2,6 +2,13 @@
 #include <numpy/arrayobject.h>
 #include "svm.h"
 #include "_svm_cython_blas_helpers.h"
+
+
+#ifndef MAX
+    #define MAX(x, y) (((x) > (y)) ? (x) : (y))
+#endif
+
+
 /*
  * Some helper methods for libsvm bindings.
  *
@@ -128,6 +135,9 @@ struct svm_model *set_model(struct svm_parameter *param, int nr_class,
     if ((model->rho = malloc( m * sizeof(double))) == NULL)
         goto rho_error;
 
+    // This is only allocated in dynamic memory while training.
+    model->n_iter = NULL;
+
     model->nr_class = nr_class;
     model->param = *param;
     model->l = (int) support_dims[0];
@@ -218,6 +228,15 @@ npy_intp get_nr(struct svm_model *model)
     return (npy_intp) model->nr_class;
 }
 
+/*
+ * Get the number of iterations run in optimization
+ */
+void copy_n_iter(char *data, struct svm_model *model)
+{
+    const int n_models = MAX(1, model->nr_class * (model->nr_class-1) / 2);
+    memcpy(data, model->n_iter, n_models * sizeof(int));
+}
+
 /*
  * Some helpers to convert from libsvm sparse data structures
  * model->sv_coef is a double **, whereas data is just a double *,
@@ -363,9 +382,11 @@ int free_model(struct svm_model *model)
     if (model == NULL) return -1;
     free(model->SV);
 
-    /* We don't free sv_ind, since we did not create them in
+    /* We don't free sv_ind and n_iter, since we did not create them in
        set_model */
-    /* free(model->sv_ind); */
+    /* free(model->sv_ind);
+     * free(model->n_iter);
+     */
     free(model->sv_coef);
     free(model->rho);
     free(model->label);

diff --git a/sklearn/svm/src/libsvm/libsvm_sparse_helper.c b/sklearn/svm/src/libsvm/libsvm_sparse_helper.c
@@ -3,6 +3,12 @@
 #include "svm.h"
 #include "_svm_cython_blas_helpers.h"
 
+
+#ifndef MAX
+    #define MAX(x, y) (((x) > (y)) ? (x) : (y))
+#endif
+
+
 /*
  * Convert scipy.sparse.csr to libsvm's sparse data structure
  */
@@ -122,6 +128,9 @@ struct svm_csr_model *csr_set_model(struct svm_parameter *param, int nr_class,
     if ((model->rho = malloc( m * sizeof(double))) == NULL)
         goto rho_error;
 
+    // This is only allocated in dynamic memory while training.
+    model->n_iter = NULL;
+
     /* in the case of precomputed kernels we do not use
        dense_to_precomputed because we don't want the leading 0. As
        indices start at 1 (not at 0) this will work */
@@ -348,6 +357,15 @@ void copy_sv_coef(char *data, struct svm_csr_model *model)
     }
 }
 
+/*
+ * Get the number of iterations run in optimization
+ */
+void copy_n_iter(char *data, struct svm_csr_model *model)
+{
+    const int n_models = MAX(1, model->nr_class * (model->nr_class-1) / 2);
+    memcpy(data, model->n_iter, n_models * sizeof(int));
+}
+
 /*
  * Get the number of support vectors in a model.
  */
@@ -402,6 +420,7 @@ int free_problem(struct svm_csr_problem *problem)
 int free_model(struct svm_csr_model *model)
 {
     /* like svm_free_and_destroy_model, but does not free sv_coef[i] */
+    /* We don't free n_iter, since we did not create them in set_model. */
     if (model == NULL) return -1;
     free(model->SV);
     free(model->sv_coef);

diff --git a/sklearn/svm/src/libsvm/svm.cpp b/sklearn/svm/src/libsvm/svm.cpp
@@ -55,6 +55,10 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
      Sylvain Marie, Schneider Electric
      see <https://github.com/scikit-learn/scikit-learn/pull/13511#issuecomment-481729756>
 
+   Modified 2021:
+
+   - Exposed number of iterations run in optimization, Juan Martín Loyola.
+     See <https://github.com/scikit-learn/scikit-learn/pull/21408/>
  */
 
 #include <math.h>
@@ -553,6 +557,7 @@ class Solver {
                 double *upper_bound;
 		double r;	// for Solver_NU
                 bool solve_timed_out;
+		int n_iter;
 	};
 
 	void Solve(int l, const QMatrix& Q, const double *p_, const schar *y_,
@@ -919,6 +924,9 @@ void Solver::Solve(int l, const QMatrix& Q, const double *p_, const schar *y_,
 	for(int i=0;i<l;i++)
 		si->upper_bound[i] = C[i];
 
+	// store number of iterations
+	si->n_iter = iter;
+
 	info("\noptimization finished, #iter = %d\n",iter);
 
 	delete[] p;
@@ -1837,6 +1845,7 @@ struct decision_function
 {
 	double *alpha;
 	double rho;	
+	int n_iter;
 };
 
 static decision_function svm_train_one(
@@ -1902,6 +1911,7 @@ static decision_function svm_train_one(
 	decision_function f;
 	f.alpha = alpha;
 	f.rho = si.rho;
+	f.n_iter = si.n_iter;
 	return f;
 }
 
@@ -2387,6 +2397,8 @@ PREFIX(model) *PREFIX(train)(const PREFIX(problem) *prob, const svm_parameter *p
                 NAMESPACE::decision_function f = NAMESPACE::svm_train_one(prob,param,0,0, status,blas_functions);
 		model->rho = Malloc(double,1);
 		model->rho[0] = f.rho;
+		model->n_iter = Malloc(int,1);
+		model->n_iter[0] = f.n_iter;
 
 		int nSV = 0;
 		int i;
@@ -2523,8 +2535,12 @@ PREFIX(model) *PREFIX(train)(const PREFIX(problem) *prob, const svm_parameter *p
 			model->label[i] = label[i];
 
 		model->rho = Malloc(double,nr_class*(nr_class-1)/2);
+		model->n_iter = Malloc(int,nr_class*(nr_class-1)/2);
 		for(i=0;i<nr_class*(nr_class-1)/2;i++)
+		{
 			model->rho[i] = f[i].rho;
+			model->n_iter[i] = f[i].n_iter;
+		}
 
 		if(param->probability)
 		{
@@ -2978,6 +2994,9 @@ void PREFIX(free_model_content)(PREFIX(model)* model_ptr)
 
 	free(model_ptr->nSV);
 	model_ptr->nSV = NULL;
+
+	free(model_ptr->n_iter);
+	model_ptr->n_iter = NULL;
 }
 
 void PREFIX(free_and_destroy_model)(PREFIX(model)** model_ptr_ptr)

diff --git a/sklearn/svm/src/libsvm/svm.h b/sklearn/svm/src/libsvm/svm.h
@@ -76,6 +76,7 @@ struct svm_model
 	int l;			/* total #SV */
 	struct svm_node *SV;		/* SVs (SV[l]) */
 	double **sv_coef;	/* coefficients for SVs in decision functions (sv_coef[k-1][l]) */
+	int *n_iter;		/* number of iterations run by the optimization routine to fit the model */
 
 	int *sv_ind;            /* index of support vectors */
 
@@ -101,6 +102,7 @@ struct svm_csr_model
 	int l;			/* total #SV */
 	struct svm_csr_node **SV;		/* SVs (SV[l]) */
 	double **sv_coef;	/* coefficients for SVs in decision functions (sv_coef[k-1][l]) */
+	int *n_iter;		/* number of iterations run by the optimization routine to fit the model */
 
         int *sv_ind;            /* index of support vectors */