scikit-learn
diff --git a/‎doc/whats_new/v1.5.rst
Lines changed: 4 additions & 0 deletions b/‎doc/whats_new/v1.5.rst
Lines changed: 4 additions & 0 deletions
diff --git a/‎sklearn/feature_extraction/text.py
Lines changed: 17 additions & 10 deletions b/‎sklearn/feature_extraction/text.py
Lines changed: 17 additions & 10 deletions
@@ -47,6 +47,10 @@ Changelog
   for storing the inverse document frequency.
   :pr:`18843` by :user:`Paolo Montesel <thebabush>`.
 
+- |Enhancement| :class:`feature_extraction.text.TfidfTransformer` now preserves
+  the data type of the input matrix if it is `np.float64` or `np.float32`.
+  :pr:`28136` by :user:`Guillaume Lemaitre <glemaitre>`.
+
 :mod:`sklearn.impute`
 .....................
 
 
@@ -1666,23 +1666,21 @@ def fit(self, X, y=None):
         )
         if not sp.issparse(X):
             X = sp.csr_matrix(X)
-        dtype = X.dtype if X.dtype in FLOAT_DTYPES else np.float64
+        dtype = X.dtype if X.dtype in (np.float64, np.float32) else np.float64
 
         if self.use_idf:
-            n_samples, n_features = X.shape
+            n_samples, _ = X.shape
             df = _document_frequency(X)
             df = df.astype(dtype, copy=False)
 
             # perform idf smoothing if required
-            df += int(self.smooth_idf)
+            df += float(self.smooth_idf)
             n_samples += int(self.smooth_idf)
 
             # log+1 instead of log makes sure terms with zero idf don't get
             # suppressed entirely.
+            # `np.log` preserves the dtype of `df` and thus `dtype`.
             self.idf_ = np.log(n_samples / df) + 1.0
-            # FIXME: for backward compatibility, we force idf_ to be np.float64
-            # In the future, we should preserve the `dtype` of `X`.
-            self.idf_ = self.idf_.astype(np.float64, copy=False)
 
         return self
 
@@ -1705,14 +1703,18 @@ def transform(self, X, copy=True):
         """
         check_is_fitted(self)
         X = self._validate_data(
-            X, accept_sparse="csr", dtype=FLOAT_DTYPES, copy=copy, reset=False
+            X,
+            accept_sparse="csr",
+            dtype=[np.float64, np.float32],
+            copy=copy,
+            reset=False,
         )
         if not sp.issparse(X):
-            X = sp.csr_matrix(X, dtype=np.float64)
+            X = sp.csr_matrix(X, dtype=X.dtype)
 
         if self.sublinear_tf:
             np.log(X.data, X.data)
-            X.data += 1
+            X.data += 1.0
 
         if hasattr(self, "idf_"):
             # the columns of X (CSR matrix) can be accessed with `X.indices `and
@@ -1725,7 +1727,12 @@ def transform(self, X, copy=True):
         return X
 
     def _more_tags(self):
-        return {"X_types": ["2darray", "sparse"]}
+        return {
+            "X_types": ["2darray", "sparse"],
+            # FIXME: np.float16 could be preserved if _inplace_csr_row_normalize_l2
+            # accepted it.
+            "preserves_dtype": [np.float64, np.float32],
+        }
 
 
 class TfidfVectorizer(CountVectorizer):