models
models copied to clipboard
ResNet101_DUC_HDC can have invalid test_data.
During testing models from model zoo I realized that onnxruntime and ngraph give similar results on ResNet101_DUC_HDC
but still have 80% missmatch.
Is it possible that test data for ResNet101_DUC_HDC
are invalid?
onnxruntime:
E AssertionError:
E Not equal to tolerance rtol=0.001, atol=1e-07
E
E Mismatch: 80.2%
E Max absolute difference: 0.9990995
E Max relative difference: 1.0197404e+09
E x: array([[[3.535608e-04, 1.008540e-03, 6.378655e-04, ..., 5.859119e-02,
E 3.697309e-02, 4.621094e-02],
E [1.310240e-04, 3.463278e-05, 2.679980e-05, ..., 2.111239e-02,...
E y: array([[[8.367564e-12, 4.105623e-11, 2.043494e-11, ..., 1.814237e-10,
E 6.354865e-11, 6.036995e-11],
E [3.100877e-12, 1.409852e-12, 8.585701e-13, ..., 6.537316e-11,...
nGraph:
E AssertionError:
E Not equal to tolerance rtol=0.001, atol=1e-07
E
E Mismatch: 80.2%
E Max absolute difference: 0.99909943
E Max relative difference: 1.0189351e+09
E x: array([[[3.535608e-04, 1.008540e-03, 6.378655e-04, ..., 5.859119e-02,
E 3.697309e-02, 4.621094e-02],
E [1.310240e-04, 3.463278e-05, 2.679980e-05, ..., 2.111239e-02,...
E y: array([[[8.374161e-12, 4.108864e-11, 2.045105e-11, ..., 1.815672e-10,
E 6.359892e-11, 6.041759e-11],
E [3.103328e-12, 1.410971e-12, 8.592453e-13, ..., 6.542469e-11,...
x - reference output
y - actual framework output
Is there a way to submit updated test data for review?
The issue is still present. Can you please revise the reference output data?