Zheng-Bicheng commented on PR #16653:
URL: https://github.com/apache/tvm/pull/16653#issuecomment-1985423565

   > If that's what achievable, I'd say we should add the tests, merge this PR, 
and in parallel try to understand why this discrepancy is quite high, being 5%. 
Maybe that understanding will involve the PaddlePaddle community as well, so 
it's a longer process.
   
   I fully endorse your point of view. I also believe that testing for the 
reasons behind errors is crucial, as it helps make TVM's code more 'robust.' 
After merging the PR, I will create a separate issue to discuss this matter, as 
it not only concerns the PaddlePaddle community but also the ONNX community.
   
   I've conducted some tests regarding the error. I pruned the model to only 
leave a quantized convolutional operator and tested it with input data of the 
same shape but different values. I found that this error doesn't exist for 
every input data; in most cases, it doesn't occur.
   
   Subsequently, I analyzed each element of the output. I found that when the 
error occurs, the majority of elements are the same, with only a small portion 
being different. Based on my work experience, I speculate that this is likely 
due to improper handling of overflow during computation.


-- 
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.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to