Modify tfidf_transformer to enable custom vocabulary and approximate sublinear-tf scaling without sparse containers

skl2onnx/operator_converters/text_vectoriser.py

@@ -226,8 +226,11 @@ def convert_sklearn_text_vectorizer(scope: Scope, operator: Operator,
             "You may raise an issue at "
             "https://github.com/onnx/sklearn-onnx/issues.")
 
-    stop_words = op.stop_words_ | (
-        set(op.stop_words) if op.stop_words else set())
+    if hasattr(op, "stop_words_"):
+        stop_words = op.stop_words_ | (
+            set(op.stop_words) if op.stop_words else set())
+    else:
+        stop_words = set()
 
     if op.lowercase or stop_words:
         if len(operator.input_full_names) != 1:

skl2onnx/operator_converters/tfidf_transformer.py

@@ -8,7 +8,7 @@
 from ..common._topology import Scope, Operator
 from ..common._container import ModelComponentContainer
 from ..common._apply_operation import (
-    apply_mul, apply_identity, apply_normalizer)
+    apply_add, apply_log, apply_mul, apply_identity, apply_normalizer)
 
 
 def convert_sklearn_tfidf_transformer(scope: Scope, operator: Operator,
@@ -30,23 +30,23 @@ def convert_sklearn_tfidf_transformer(scope: Scope, operator: Operator,
         # code scikit-learn
         # np.log(X.data, X.data) --> does not apply on null coefficient
         # X.data += 1
-        raise RuntimeError(
-            "ONNX does not support sparse tensors before opset < 11, "
-            "sublinear_tf must be False.")
-
-        # In case sparse is enabled.
-        # C = operator.inputs[0].type.shape[1]
-        # logged = scope.get_unique_variable_name('logged')
-        # apply_log(scope, data, logged, container)
-        # if not op.use_idf and op.norm is None:
-        #     loggedplus1 = final
-        # else:
-        #     loggedplus1 = scope.get_unique_variable_name('loggedplus1')
-        # ones = scope.get_unique_variable_name('ones')
-        # cst = np.ones((C,), dtype=float_type)
-        # container.add_initializer(ones, proto_dtype, [C], cst.flatten())
-        # apply_add(scope, [logged, ones], loggedplus1, container, broadcast=1)
-        # data = [loggedplus1]
+        # ONNX does not support sparse tensors before opset < 11
+        # approximated by X.data += 1 --> np.log(X.data, X.data)
+        if operator.target_opset < 11:
+            plus1 = scope.get_unique_variable_name("plus1")
+            C = operator.inputs[0].type.shape[1]
+            ones = scope.get_unique_variable_name("ones")
+            cst = np.ones((C,), dtype=float_type)
+            container.add_initializer(ones, proto_dtype, [C], cst.flatten())
+            apply_add(scope, data + [ones], plus1, container, broadcast=1)
+            plus1logged = scope.get_unique_variable_name("plus1logged")
+            apply_log(scope, plus1, plus1logged, container)
+            data = [plus1logged]
+        else:
+            # sparse containers have not yet been implemented.
+            raise RuntimeError(
+                "ONNX does not support sparse tensors before opset < 11, "
+                "sublinear_tf must be False.")
 
     if op.use_idf:
         cst = op.idf_.astype(float_type)

tests/test_sklearn_tfidf_vectorizer_converter.py

@@ -577,6 +577,34 @@ def test_model_tfidf_vectorizer_nan(self):
         assert res.shape == (4, 9)
         assert numpy.isnan(res[0, 0])
 
+    @unittest.skipIf(
+        StrictVersion(onnx.__version__) <= StrictVersion("1.4.1"),
+        reason="Requires opset 9.")
+    def test_model_tfidf_vectorizer11_custom_vocabulary(self):
+        corpus = numpy.array([
+            "This is the first document.",
+            "This document is the second document.",
+            "And this is the third one.",
+            "Is this the first document?",
+        ]).reshape((4, 1))
+        vc = ["first", "second", "third", "document", "this"]
+        vect = TfidfVectorizer(ngram_range=(1, 1), norm=None, vocabulary=vc)
+        vect.fit(corpus.ravel())
+        self.assertFalse(hasattr(vect, "stop_words_"))
+        model_onnx = convert_sklearn(vect, "TfidfVectorizer",
+                                     [("input", StringTensorType())],
+                                     options=self.get_options(),
+                                     target_opset=TARGET_OPSET)
+        self.assertTrue(model_onnx is not None)
+        dump_data_and_model(
+            corpus,
+            vect,
+            model_onnx,
+            basename="SklearnTfidfVectorizer11CustomVocab-OneOff-SklCol",
+            allow_failure="StrictVersion(onnxruntime.__version__)"
+                          " <= StrictVersion('0.4.0')",
+        )
+
 
 if __name__ == "__main__":
     unittest.main()