t-vi commented on pull request #5834:
URL: https://github.com/apache/incubator-tvm/pull/5834#issuecomment-646561327
I have to ask about the test criterion for
`qnn_test.py::test_serialized_modules`.
It seems to pass here locally (with PyTorch 1.5 cpu wheel).
It gets ~70% identical, but so I would think that a more natural test is
```
num_identical = np.sum(np.abs(tvm_result - pt_result) < 1e-2)
```
for me the outputs are 0.5ish and so 1e-2 is about (2**-6), i.e. to 2x-4x of
quantization accuracy. Then I get 100% match for the criterion.
This might not work well for complex models, but for this simple one it
should work. What do you think?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
