Use torch.testing.assert_close in test_functional_tensor #3876
Conversation
|
|
||
| if scripted_fn_atol >= 0: | ||
| scripted_fn = torch.jit.script(fn) | ||
| # scriptable function test | ||
| s_transformed_batch = scripted_fn(batch_tensors, **fn_kwargs) | ||
| self.assertTrue(transformed_batch.allclose(s_transformed_batch, atol=scripted_fn_atol)) | ||
| torch.testing.assert_close(transformed_batch, s_transformed_batch, rtol=1e-5, atol=scripted_fn_atol) |
There was a problem hiding this comment.
For ref the rtol=1e-5 comes from the current default of https://pytorch.org/docs/stable/generated/torch.allclose.html
I'd prefer to leave rtol to its default in assert_close if possible, but rtol must be set if atol is set. Would you know the reason @pmeier ? np.testing.assert_allclose doesn't have this constraint it seems
There was a problem hiding this comment.
We went for this logic, since we have non-zero defaults for rtol and atol. Imagine setting atol=0 and the tensors still match because rtol > 0. See https://github.com/pytorch/pytorch/blob/74c12da4517c789bea737dc947d6adc755f63176/torch/testing/_asserts.py#L391-L396.
| torch.max(true_out.float() - out.float()), | ||
| 1.0, | ||
| torch.testing.assert_close( | ||
| out, true_out, rtol=0.0, atol=1.0, check_stride=False, |
There was a problem hiding this comment.
we're comparing the max abs difference here instead of the max difference in master, but it's probably more correct with the proposed changes
…on into assert_close_func_tensor
) Summary: Co-authored-by: Philip Meier <[email protected]> Reviewed By: vincentqb, cpuhrsch Differential Revision: D28679977 fbshipit-source-id: 59cd7c52bd5a94a75141c719b85424bf075c65a6
part of #3865