George Wigley a5051a9521 Update torch.masked.mean to upcast dtype for bool tensors (#139999) ... When calling `torch.masked.mean(...)` with a boolean tensor, the dtype is inferred to be bool. When the mean is being computed, the sum operator is used. When the sum operator is used with dtype=torch.bool, the result is clamped to True (1) leading to an incorrect mean being calculated. The below example shows how the incorrect result occurs: ``` a = torch.tensor([True, True]) count = torch.sum(torch.ones(a.shape, dtype=torch.int64)) # 2 total = torch.sum(a, dtype=torch.bool) # True (1) mean = total / count # 0.5 ``` This PR upcasts the dtype used for the sumation to int32 in the case of bool tensors allowing for the correct result to be computed. Pull Request resolved: https://github.com/pytorch/pytorch/pull/139999 Approved by: https://github.com/cpuhrsch		2025-01-08 10:35:19 +00:00
..
_internal	Update torch.masked.mean to upcast dtype for bool tensors (#139999)	2025-01-08 10:35:19 +00:00
__init__.py	[BE][Easy][19/19] enforce style for empty lines in import segments in torch/[o-z]*/ (#129771)	2024-08-01 17:07:14 +00:00
_comparison.py	Strict shape checking for NJTs with TestCase.assertEqual() (#131898)	2024-07-30 20:05:48 +00:00
_creation.py	Ensure noncontiguous tensor creation tests offsetting (#136396)	2024-10-02 00:40:43 +00:00
_utils.py	[BE][Easy][19/19] enforce style for empty lines in import segments in torch/[o-z]*/ (#129771)	2024-08-01 17:07:14 +00:00