MAINT: Update TSNE for sklearn1.8 (#2793)

david-cortes-intel · web-flow · commit be23e7bd866b · 2025-11-26T15:54:56.000+01:00
* update tsne for sklearn1.8

* fix test

* more corrections

* more fixes for older sklearn

* missing else

* more fixes for older sklearn

* correction

* remove redundant check

* more clear conditions
diff --git a/daal4py/sklearn/manifold/_t_sne.py b/daal4py/sklearn/manifold/_t_sne.py
@@ -122,36 +122,89 @@ def _daal_tsne(self, P, n_samples, X_embedded):
 
         return X_embedded
 
+    # Comment 2025-11-24: This appears to be a copy-paste from an earlier version of the original
+    # scikit-learn with some modifications to make calls to oneDAL under a narrow subset of
+    # allowed input parameters, copy-pastying the rest of the sklearn code when oneDAL is not
+    # called. Note that the conditions checked here are out of synch with the latest sklearn by now.
+    # An early 'is supported' check that offloads to stock sklearn was added later on, which results
+    # in having a lot of dead code paths in this function that can be safely removed.
+    # Note: this method is called from inside 'fit' from the base class in stock scikit-learn.
+    # Hence, the offloading logic is different than in other classes, as falling back to 'fit'
+    # from the base class would lead to a circular loop.
     def _fit(self, X, skip_num_points=0):
         """Private function to fit the model using X as training data."""
-        if isinstance(self.init, str) and self.init == "warn":
-            warnings.warn(
-                "The default initialization in TSNE will change "
-                "from 'random' to 'pca' in 1.2.",
-                FutureWarning,
-            )
-            self._init = "random"
+
+        _patching_status = PatchingConditionsChain("sklearn.manifold.TSNE._tsne")
+        _patching_status.and_conditions(
+            [
+                (
+                    self.method == "barnes_hut",
+                    'Used t-SNE method is not "barnes_hut" which is the only supported.',
+                ),
+                (self.n_components == 2, "Number of components != 2."),
+                (self.verbose == 0, "Verbose mode is set."),
+                (
+                    daal_check_version((2021, "P", 600)),
+                    "oneDAL version is lower than 2021.6.",
+                ),
+                (
+                    not (
+                        isinstance(self.init, str) and self.init == "pca" and issparse(X)
+                    ),
+                    "PCA initialization is not supported with sparse input matrices.",
+                ),
+                # Note: these conditions below should result in errors, but stock scikit-learn
+                # does not check for errors at this exact point. Hence, this offloads the erroring
+                # out to the base class, wherever in the process they might be encountered.
+                (
+                    np.isscalar(self.angle) and self.angle > 0.0 and self.angle < 1.0,
+                    "'angle' must be between 0.0 - 1.0",
+                ),
+                (self.early_exaggeration >= 1.0, "early_exaggeration must be at least 1"),
+                (
+                    (
+                        isinstance(self.init, str)
+                        and self.init
+                        in ["random", "pca"]
+                        + (
+                            ["warn"]
+                            if sklearn_check_version("1.0")
+                            and not sklearn_check_version("1.2")
+                            else []
+                        )
+                    )
+                    or isinstance(self.init, np.ndarray),
+                    "'init' must be 'exact', 'pca', or a numpy array.",
+                ),
+            ]
+        )
+        _dal_ready = _patching_status.get_status(logs=True)
+        if not _dal_ready:
+            return super()._fit(X, skip_num_points)
+
+        if sklearn_check_version("1.0") and not sklearn_check_version("1.2"):
+            if isinstance(self.init, str) and self.init == "warn":
+                warnings.warn(
+                    "The default initialization in TSNE will change "
+                    "from 'random' to 'pca' in 1.2.",
+                    FutureWarning,
+                )
+                self._init = "random"
+            else:
+                self._init = self.init
         else:
             self._init = self.init
 
-        if isinstance(self._init, str) and self._init == "pca" and issparse(X):
-            raise TypeError(
-                "PCA initialization is currently not supported "
-                "with the sparse input matrix. Use "
-                'init="random" instead.'
-            )
-
-        if self.method not in ["barnes_hut", "exact"]:
-            raise ValueError("'method' must be 'barnes_hut' or 'exact'")
-        if self.angle < 0.0 or self.angle > 1.0:
-            raise ValueError("'angle' must be between 0.0 - 1.0")
-        if self.learning_rate == "warn":
-            warnings.warn(
-                "The default learning rate in TSNE will change "
-                "from 200.0 to 'auto' in 1.2.",
-                FutureWarning,
-            )
-            self._learning_rate = 200.0
+        if sklearn_check_version("1.0") and not sklearn_check_version("1.2"):
+            if self.learning_rate == "warn":
+                warnings.warn(
+                    "The default learning rate in TSNE will change "
+                    "from 200.0 to 'auto' in 1.2.",
+                    FutureWarning,
+                )
+                self._learning_rate = 200.0
+            else:
+                self._learning_rate = self.learning_rate
         else:
             self._learning_rate = self.learning_rate
         if self._learning_rate == "auto":
@@ -227,28 +280,15 @@ def _fit(self, X, skip_num_points=0):
                     "or provide the dense distance matrix."
                 )
 
-        if self.method == "barnes_hut" and self.n_components > 3:
-            raise ValueError(
-                "'n_components' should be inferior to 4 for the "
-                "barnes_hut algorithm as it relies on "
-                "quad-tree or oct-tree."
-            )
         random_state = check_random_state(self.random_state)
 
-        if self.early_exaggeration < 1.0:
-            raise ValueError(
-                "early_exaggeration must be at least 1, but is {}".format(
-                    self.early_exaggeration
-                )
-            )
-
         if not sklearn_check_version("1.2"):
             if self.n_iter < 250:
                 raise ValueError("n_iter should be at least 250")
 
         n_samples = X.shape[0]
 
-        neighbors_nn = None
+        # neighbors_nn = None # <- unused variable in stock sklearn, commented out due to coverity
         if self.method == "exact":
             # Retrieve the distance matrix, either using the precomputed one or
             # computing it.
@@ -278,9 +318,8 @@ def _fit(self, X, skip_num_points=0):
                     "All distances should be positive, the " "metric given is not correct"
                 )
 
-            if (
-                self.metric != "euclidean"
-                and getattr(self, "square_distances", True) is True
+            if self.metric != "euclidean" and (
+                sklearn_check_version("1.2") or self.square_distances is True
             ):
                 distances **= 2
 
@@ -339,15 +378,14 @@ def _fit(self, X, skip_num_points=0):
             # Free the memory used by the ball_tree
             del knn
 
-            if (
-                getattr(self, "square_distances", True) is True
-                or self.metric == "euclidean"
+            # knn return the euclidean distance but we need it squared
+            # to be consistent with the 'exact' method. Note that the
+            # the method was derived using the euclidean method as in the
+            # input space. Not sure of the implication of using a different
+            # metric.
+            if sklearn_check_version("1.2") or (
+                self.square_distances is True or self.metric == "euclidean"
             ):
-                # knn return the euclidean distance but we need it squared
-                # to be consistent with the 'exact' method. Note that the
-                # the method was derived using the euclidean method as in the
-                # input space. Not sure of the implication of using a different
-                # metric.
                 distances_nn.data **= 2
 
             # compute the joint probability distribution for the input space
@@ -358,16 +396,23 @@ def _fit(self, X, skip_num_points=0):
         elif self._init == "pca":
             pca = PCA(
                 n_components=self.n_components,
-                svd_solver="randomized",
                 random_state=random_state,
             )
+            if sklearn_check_version("1.2"):
+                # Always output a numpy array, no matter what is configured globally
+                pca.set_output(transform="default")
             X_embedded = pca.fit_transform(X).astype(np.float32, copy=False)
-            warnings.warn(
-                "The PCA initialization in TSNE will change to "
-                "have the standard deviation of PC1 equal to 1e-4 "
-                "in 1.2. This will ensure better convergence.",
-                FutureWarning,
-            )
+            if sklearn_check_version("1.0") and not sklearn_check_version("1.2"):
+                warnings.warn(
+                    "The PCA initialization in TSNE will change to "
+                    "have the standard deviation of PC1 equal to 1e-4 "
+                    "in 1.2. This will ensure better convergence.",
+                    FutureWarning,
+                )
+            if sklearn_check_version("1.2"):
+                # PCA is rescaled so that PC1 has standard deviation 1e-4 which is
+                # the default value for random initialization. See issue #18018.
+                X_embedded = X_embedded / np.std(X_embedded[:, 0]) * 1e-4
         elif self._init == "random":
             # The embedding is initialized with iid samples from Gaussians with
             # standard deviation 1e-4.
@@ -377,40 +422,11 @@ def _fit(self, X, skip_num_points=0):
         else:
             raise ValueError("'init' must be 'pca', 'random', or " "a numpy array")
 
-        # Degrees of freedom of the Student's t-distribution. The suggestion
-        # degrees_of_freedom = n_components - 1 comes from
-        # "Learning a Parametric Embedding by Preserving Local Structure"
-        # Laurens van der Maaten, 2009.
-        degrees_of_freedom = max(self.n_components - 1, 1)
-
-        _patching_status = PatchingConditionsChain("sklearn.manifold.TSNE._tsne")
-        _patching_status.and_conditions(
-            [
-                (
-                    self.method == "barnes_hut",
-                    'Used t-SNE method is not "barnes_hut" which is the only supported.',
-                ),
-                (self.n_components == 2, "Number of components != 2."),
-                (self.verbose == 0, "Verbose mode is set."),
-                (
-                    daal_check_version((2021, "P", 600)),
-                    "oneDAL version is lower than 2021.6.",
-                ),
-            ]
-        )
-        _dal_ready = _patching_status.get_status(logs=True)
+        # Note: by this point, stock sklearn would calculate degrees of freedom, but oneDAL
+        # doesn't use them.
 
-        if _dal_ready:
-            X_embedded = check_array(X_embedded, dtype=[np.float32, np.float64])
-            return self._daal_tsne(P, n_samples, X_embedded=X_embedded)
-        return self._tsne(
-            P,
-            degrees_of_freedom,
-            n_samples,
-            X_embedded=X_embedded,
-            neighbors=neighbors_nn,
-            skip_num_points=skip_num_points,
-        )
+        X_embedded = check_array(X_embedded, dtype=[np.float32, np.float64])
+        return self._daal_tsne(P, n_samples, X_embedded=X_embedded)
 
     fit.__doc__ = BaseTSNE.fit.__doc__
     fit_transform.__doc__ = BaseTSNE.fit_transform.__doc__
diff --git a/doc/sources/algorithms.rst b/doc/sources/algorithms.rst
@@ -190,11 +190,11 @@ Dimensionality Reduction
      - All parameters are supported except:
 
        - ``metric`` != 'euclidean' or `'minkowski'` with ``p`` != `2`
-
        - ``n_components`` can only be `2`
-
+       - ``method`` != ``"barnes_hut"``
+       
        Refer to :ref:`TSNE acceleration details <acceleration_tsne>` to learn more.
-     - Sparse data is not supported
+     - Sparse data with ``init`` = ``"pca"`` is not supported
 
 Nearest Neighbors
 *****************
diff --git a/doc/sources/guide/acceleration.rst b/doc/sources/guide/acceleration.rst
@@ -34,9 +34,9 @@ The overall acceleration of TSNE depends on the acceleration of each of these al
   - ``metric`` != `'euclidean'` or `'minkowski'` with ``p`` != `2`
 - The Gradient Descent part of the algorithm supports all parameters except:
 
-  - ``n_components`` = `3`
-  - ``method`` = `'exact'`
-  - ``verbose`` != `0`
+  - ``n_components`` > ``2``
+  - ``method`` = ``'exact'``
+  - ``verbose`` != ``0``
 
 To get better performance, use parameters supported by both components.
 
diff --git a/sklearnex/manifold/tests/test_tsne.py b/sklearnex/manifold/tests/test_tsne.py
@@ -19,6 +19,8 @@
 from numpy.testing import assert_allclose
 from sklearn.metrics.pairwise import pairwise_distances
 
+from daal4py.sklearn._utils import sklearn_check_version
+
 # Note: n_components must be 2 for now
 from onedal.tests.utils._dataframes_support import (
     _as_numpy,
@@ -161,8 +163,12 @@ def test_tsne_functionality_and_edge_cases(
         assert np.any(embedding != 0)
 
 
+# Note: since sklearn1.2, the PCA initialization divides by standard deviations of components.
+# Since those will be zeros for constant data, it will end up producing NaNs, hence it's not tested.
 @pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
-@pytest.mark.parametrize("init", ["pca", "random"])
+@pytest.mark.parametrize(
+    "init", ["random"] + (["pca"] if not sklearn_check_version("1.2") else [])
+)
 @pytest.mark.parametrize("dtype", [np.float32, np.float64])
 def test_tsne_constant_data(init, dataframe, queue, dtype):
     from sklearnex.manifold import TSNE
