π Geometric Computer Vision Library for Spatial AI
Python artificial-intelligencecomputer-visiondeep-learninghacktoberfestimage-processingmachine-learningneural-networkpythonpytorchroboticsspatial-ai
Preserve KMeans accumulation precision under autocast (#6597)
A Aswanth Rajf50e332913e6
diff --git a/changelog.d/6597.fixed.md b/changelog.d/6597.fixed.mdnew file mode 100644index 000000000..2183b2e1b--- /dev/null+++ b/changelog.d/6597.fixed.md@@ -0,0 +1 @@+`KMeans.fit` preserves its promoted accumulation dtype under autocast, preventing infinite or biased cluster centers for unit-range data.diff --git a/kornia/contrib/kmeans.py b/kornia/contrib/kmeans.pyindex 94beeebd0..fe32e3708 100644--- a/kornia/contrib/kmeans.py+++ b/kornia/contrib/kmeans.py@@ -24,6 +24,7 @@ import operatorimport torchfrom kornia.core.check import KORNIA_CHECK, KORNIA_CHECK_SAME_SIZE, KORNIA_CHECK_SHAPE+from kornia.core.utils import _autocast_disabledfrom kornia.geometry.linalg import euclidean_distance@@ -259,7 +260,8 @@ class KMeans:one_hot = torch.nn.functional.one_hot(cluster_assignment, num_classes=self.num_clusters)cluster_counts = one_hot.sum(0)if work_dtype.is_floating_point:- cluster_sums = one_hot.to(work_dtype).t() @ X_work+ with _autocast_disabled(X.device.type):+ cluster_sums = one_hot.to(work_dtype).t() @ X_workelse:# integer sums are exact with scatter_add_; MPS lacks integer matmul and index_add_ on torch 2.5index = cluster_assignment[:, None].expand(-1, X_work.shape[1])diff --git a/tests/contrib/test_kmeans.py b/tests/contrib/test_kmeans.pyindex 0ea64edd3..bd25158b0 100644--- a/tests/contrib/test_kmeans.py+++ b/tests/contrib/test_kmeans.py@@ -231,6 +231,21 @@ class TestKMeans(BaseTester):# pre-fix bfloat16 error (0.25, well under BaseTester's default bfloat16 tolerance) passself.assert_close(kmeans.cluster_centers, expected.to(device=device, dtype=dtype), rtol=0.0, atol=0.0)+ @pytest.mark.parametrize("autocast_dtype", [torch.float16, torch.bfloat16])+ def test_autocast_preserves_cluster_sum_precision(self, device, dtype, autocast_dtype):+ """Keep centroid accumulation in the work dtype under autocast."""+ if device.type not in ("cpu", "cuda"):+ pytest.skip("autocast precision regression is exercised on CPU and CUDA")+ if device.type == "cuda" and autocast_dtype == torch.bfloat16 and not torch.cuda.is_bf16_supported():+ pytest.skip("CUDA device does not support bfloat16 autocast")+ x = torch.full((100000, 3), 0.75, device=device, dtype=dtype)+ kmeans = kornia.contrib.KMeans(1, torch.zeros_like(x[:1]), tolerance=None, max_iterations=1)++ with torch.autocast(device_type=device.type, dtype=autocast_dtype):+ kmeans.fit(x)++ self.assert_close(kmeans.cluster_centers, x[:1], rtol=0.0, atol=0.0)+@staticmethoddef _create_data(device, dtype):# create example dataset