Nimit
January 6, 2022, 6:50pm
1
Hi there,
I am trying to convert a BERT model to ONNX. However, I think there is some discrepancy in the ONNX conversion module. I ran the sample conversion presented here on the website: (optional) Exporting a Model from PyTorch to ONNX and Running it using ONNX Runtime — PyTorch Tutorials 1.10.1+cu102 documentation
I did not find the test to succeed and there is a significant numerical error between the PyTorch model and the ONNX version of the model generated by the conversion.
Python (3.8.10)
Nimit
January 7, 2022, 3:17am
2
Update: I have the latest branch of PyTorch (built from source) and ONNX (built from source).
Here’s a sample output when I run the code:
Traceback (most recent call last):
File "sample_convert.py", line 96, in <module>
np.testing.assert_allclose(to_numpy(torch_out), ort_outs[0], rtol=1e-03, atol=1e-05)
File "/usr/local/lib/python3.7/site-packages/numpy/testing/_private/utils.py", line 1531, in assert_allclose
verbose=verbose, header=header, equal_nan=equal_nan)
File "/usr/local/lib/python3.7/site-packages/numpy/testing/_private/utils.py", line 844, in assert_array_compare
raise AssertionError(msg)
AssertionError:
Not equal to tolerance rtol=0.001, atol=1e-05
Mismatched elements: 451416 / 451584 (100%)
Max absolute difference: 7.256867
Max relative difference: 317452.75
x: array([[[[ 0.178333, 0.240012, 0.491639, ..., 0.368835, 0.33173 ,
0.327067],
[ 0.153973, 0.266167, 0.592167, ..., 0.480333, 0.449961,...
y: array([[[[ 0.620332, 0.832479, 1.062318, ..., -0.410807, -0.471132,
-0.290673],
[ 0.496837, 0.832677, 1.17443 , ..., -0.337473, -0.433076,...