amorgun · February 27, 2017 11:24 · Feb 27, 2017
diff --git a/convert_svm.py b/convert_svm.py
@@ -0,0 +1,21 @@
+from scipy.sparse import csr_matrix
+import numpy as np
+
+def make_sparse(clf):
+    """
+    Make sklearn.svm.SVC trained on dense data work on sparse features without fitting if again.
+    """
+    clf._sparse = True
+    clf.support_vectors_ = csr_matrix(clf.support_vectors_)
+
+    # From https://github.com/scikit-learn/scikit-learn/blob/14031f65d144e3966113d3daec836e443c6d7a5b/sklearn/svm/base.py#L288-L293
+    n_class = len(clf._label) - 1
+    n_SV = clf.support_vectors_.shape[0]
+
+    dual_coef_indices = np.tile(np.arange(n_SV), n_class)
+    dual_coef_indptr = np.arange(0, dual_coef_indices.size + 1,
+                                 dual_coef_indices.size / n_class)
+    clf.dual_coef_ = csr_matrix(
+        (clf.dual_coef_.ravel(), dual_coef_indices, dual_coef_indptr),
+        (n_class, n_SV))
+    return clf