sony / nnabla

Neural Network Libraries
https://nnabla.org/
Apache License 2.0
2.73k stars 334 forks source link

increate acceptable threshold for interpolate test on Windows #1142

Closed TomonobuTsujikawa closed 1 year ago

TomonobuTsujikawa commented 1 year ago

We sometime face the failure on interpolate test:

        np.testing.assert_allclose(actual, desired, rtol=rtol, atol=atol,
>                                  equal_nan=equal_nan, err_msg=err_msg, verbose=verbose)
E       AssertionError: 
E       Not equal to tolerance rtol=1e-05, atol=0.01
E       Backward (accum) of the backward function (interpolate_backward) wrt 0-th / 2 input fails.
E       Mismatched elements: 53 / 288 (18.4%)
E       Max absolute difference: 0.01009312
E       Max relative difference: 0.00629964
E        x: array([1.612266, 1.612266, 1.612266, 1.612266, 1.612266, 1.612266,
E              1.612266, 1.612266, 1.612266, 1.612266, 1.612266, 1.612266,
E              1.612266, 1.612266, 1.612266, 1.612266, 1.612266, 1.612266,...
E        y: array([1.605988, 1.61171 , 1.607895, 1.602173, 1.60408 , 1.602173,
E              1.607895, 1.607895, 1.605988, 1.602173, 1.605988, 1.607895,
E              1.60408 , 1.61171 , 1.609802, 1.607895, 1.60408 , 1.60408 ,...

This failure only happens on Windows environment, so there might be due to Windows specific implementation in python or cuda layer, so we will increase threshold value only for Windows.