feature: new tests added for tsne to expand test coverage

sklearnex/manifold/tests/test_tsne.py

@@ -15,12 +15,236 @@
 # ===============================================================================
 
 import numpy as np
+import pytest
 from numpy.testing import assert_allclose
+from sklearn.metrics.pairwise import pairwise_distances
 
+# Note: n_components must be 2 for now
+from onedal.tests.utils._dataframes_support import (
+    _as_numpy,
+    _convert_to_dataframe,
+    get_dataframes_and_queues,
+)
 
-def test_sklearnex_import():
+
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
+def test_sklearnex_import(dataframe, queue):
+    """Test TSNE compatibility with different backends and queues, and validate sklearnex module."""
     from sklearnex.manifold import TSNE
 
     X = np.array([[0, 0, 0], [0, 1, 1], [1, 0, 1], [1, 1, 1]])
-    tsne = TSNE(n_components=2, perplexity=2.0).fit(X)
+    X_df = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
+    tsne = TSNE(n_components=2, perplexity=2.0, random_state=42, init="pca").fit(X_df)
+    embedding = tsne.fit_transform(X_df)
+    embedding = _as_numpy(embedding)
     assert "daal4py" in tsne.__module__
+    assert tsne.n_components == 2
+    assert tsne.perplexity == 2.0
+    assert tsne.random_state == 42
+    assert tsne.init == "pca"
+
+
+@pytest.mark.parametrize(
+    "X_generator,n_components,perplexity,expected_shape,should_raise",
+    [
+        pytest.param(
+            lambda rng: np.array([[0, 0, 0], [0, 1, 1], [1, 0, 1], [1, 1, 1]]),
+            2,
+            2.0,
+            (4, 2),
+            False,
+            id="Basic functionality",
+        ),
+        pytest.param(
+            lambda rng: rng.random((100, 10)),
+            2,
+            30.0,
+            (100, 2),
+            False,
+            id="Random data",
+        ),
+        pytest.param(
+            lambda rng: np.array([[0, 0], [1, 1], [2, 2]]),
+            2,
+            2.0,
+            (3, 2),
+            False,
+            id="Valid minimal data",
+        ),
+        pytest.param(
+            lambda rng: np.empty((0, 10)),
+            2,
+            5.0,
+            None,
+            True,
+            id="Empty data",
+        ),
+        pytest.param(
+            lambda rng: np.array([[0, 0], [1, np.nan], [2, np.inf]]),
+            2,
+            5.0,
+            None,
+            True,
+            id="Data with NaN/Inf",
+        ),
+        pytest.param(
+            lambda rng: rng.random((50, 500)) * (rng.random((50, 500)) > 0.99),
+            2,
+            30.0,
+            (50, 2),
+            False,
+            id="Sparse-like high-dimensional data",
+        ),
+        pytest.param(
+            lambda rng: np.hstack(
+                [
+                    np.ones((50, 1)),  # First column is 1
+                    rng.random((50, 499)) * (rng.random((50, 499)) > 0.99),
+                ]
+            ),
+            2,
+            30.0,
+            (50, 2),
+            False,
+            id="Sparse-like data with constant column",
+        ),
+        pytest.param(
+            lambda rng: np.where(
+                np.arange(50 * 500).reshape(50, 500) % 10 == 0, 0, rng.random((50, 500))
+            ),
+            2,
+            30.0,
+            (50, 2),
+            False,
+            id="Sparse-like data with every tenth element zero",
+        ),
+        pytest.param(
+            lambda rng: rng.random((10, 5)),
+            2,
+            0.5,
+            (10, 2),
+            False,
+            id="Extremely low perplexity",
+        ),
+    ],
+)
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_tsne_functionality_and_edge_cases(
+    X_generator,
+    n_components,
+    perplexity,
+    expected_shape,
+    should_raise,
+    dataframe,
+    queue,
+    dtype,
+):
+    from sklearnex.manifold import TSNE
+
+    rng = np.random.default_rng(
+        seed=42
+    )  # Use generator to ensure independent dataset per test
+    X = X_generator(rng)
+    X = X.astype(dtype) if X.size > 0 else X
+    X_df = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
+
+    if should_raise:
+        with pytest.raises(ValueError):
+            TSNE(n_components=n_components, perplexity=perplexity).fit_transform(X_df)
+    else:
+        tsne = TSNE(n_components=n_components, perplexity=perplexity, random_state=42)
+        embedding = tsne.fit_transform(X_df)
+        embedding = _as_numpy(embedding)
+        assert embedding.shape == expected_shape
+        assert np.all(np.isfinite(embedding))
+        assert np.any(embedding != 0)
+
+
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
+@pytest.mark.parametrize("init", ["pca", "random"])
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_tsne_constant_data(init, dataframe, queue, dtype):
+    from sklearnex.manifold import TSNE
+
+    X = np.ones((10, 10), dtype=dtype)
+    X_df = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
+    tsne = TSNE(n_components=2, init=init, perplexity=5, random_state=42)
+    embedding = tsne.fit_transform(X_df)
+    embedding = _as_numpy(embedding)
+    assert embedding.shape == (10, 2)
+    if init == "pca":
+        assert np.isclose(embedding[:, 0].std(), 0, atol=1e-6)  # Constant first dimension
+        assert np.allclose(embedding[:, 1], 0, atol=1e-6)  # Zero second dimension
+    elif init == "random":
+        assert np.all(np.isfinite(embedding))
+
+
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_tsne_reproducibility(dataframe, queue, dtype):
+    from sklearnex.manifold import TSNE
+
+    rng = np.random.default_rng(seed=42)
+    X = rng.random((50, 10)).astype(dtype)
+    X_df = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
+    tsne_1 = TSNE(n_components=2, random_state=42).fit_transform(X_df)
+    tsne_2 = TSNE(n_components=2, random_state=42).fit_transform(X_df)
+    # in case of dpctl.tensor.usm_ndarray covert to numpy array
+    tsne_1 = _as_numpy(tsne_1)
+    tsne_2 = _as_numpy(tsne_2)
+    assert_allclose(tsne_1, tsne_2, rtol=1e-5)
+
+
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_tsne_complex_and_gpu_validation(dataframe, queue, dtype):
+    from sklearnex.manifold import TSNE
+
+    X = np.array(
+        [
+            [1, 1, 1, 1],
+            [1.1, 1.1, 1.1, 1.1],
+            [0.9, 0.9, 0.9, 0.9],
+            [2e9, 2e-9, -2e9, -2e-9],
+            [5e-5, 5e5, -5e-5, -5e5],
+            [9e-7, -9e7, 9e-7, -9e7],
+            [1, -1, 1, -1],
+            [-1e-9, 1e-9, -1e-9, 1e-9],
+            [42, 42, 42, 42],
+            [8, -8, 8e8, -8e-8],
+            [1e-3, 1e3, -1e3, -1e-3],
+            [0, 1e9, -1e-9, 1],
+            [0, 0, 1, -1],
+            [0, 0, 0, 0],
+            [-1e5, 0, 1e5, -1],
+            [1, 0, -1e8, 1e8],
+        ]
+    )
+    n_components = 2
+    perplexity = 3.0
+    expected_shape = (16, 2)
+
+    X = X.astype(dtype)
+    X_df = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
+    tsne = TSNE(n_components=n_components, perplexity=perplexity, random_state=42)
+    embedding = tsne.fit_transform(X_df)
+
+    # Validate results
+    assert embedding.shape == expected_shape
+    embedding = _as_numpy(embedding)
+    assert np.all(np.isfinite(embedding))
+    assert np.any(embedding != 0)
+
+    # Ensure close points in original space remain close in embedding
+    group_a_indices = [0, 1, 2]  # Hardcoded index of similar points
+    group_b_indices = [3, 4, 5]  # Hardcoded index of dissimilar points from a
+    embedding_distances = pairwise_distances(
+        X, metric="euclidean"
+    )  # Get an array of distance where [i, j] is distance b/t i and j
+    # Check for distance b/t two points in group A < distance of this point and any point in group B
+    for i in group_a_indices:
+        for j in group_a_indices:
+            assert (
+                embedding_distances[i, j] < embedding_distances[i, group_b_indices].min()
+            ), f"Point {i} in Group A is closer to a point in Group B than to another point in Group A."