Open mbencer opened 3 years ago
I believe the model was from https://github.com/onnx/models/tree/master/text/machine_comprehension/bert-squad
If it is a test data issue, please help us report it to https://github.com/onnx/models/issues
Thank you!
@snnn Thanks for answer Reporting issue to model zoo was my first step some time ago: https://github.com/onnx/models/issues/415 My PR with fix expected data to model zoo was also merged: https://github.com/onnx/models/pull/418
But still very strange for me is that ./onnx_test_runner
PASS only for test data before my change. So there are some differences between onnx_test_runner
execution and my test script.
Have you maybe plans to update https://onnxruntimetestdata.blob.core.windows.net/models/20191107.zip package?
I will help take a look.
@snnn Great, thanks
Describe the bug I've tried to run BERT-Squad (opset10) model (https://onnxruntimetestdata.blob.core.windows.net/models/20191107.zip) using test data. Unfortunately, I see a big mismatch (and incorrect order) between expected values from sample data and values produced by ONNX Runtime.
System information
To Reproduce
Current behavior
Even if output order is not guaranteed (based on data shapes we can assume that expected outputs from .pb files are saved in order
unique_ids:0
,unstack:0
,unstack:1
) the mismatch is very huge. When I run it on other framework I got results the same as produced by ONNX Runtime which suggest problem with expected test data. The other problem can be missing "preprocessing" or "postprocessing" step which I am not aware of.What is important if I run this model with expected data using
./onnx_test_runner
app I see that everything succeeded.