comaniac commented on pull request #8712: URL: https://github.com/apache/tvm/pull/8712#issuecomment-896431270
> I think we can change everything to 1e-4, given that we do expect precision issues if the compute order is different I'm a bit concern to use 1e-4 everywhere even compute order is indeed an issue, because the error may propagate and accumulate. If we allow every op to have 1e-4 then it's possible to hurt the accuracy eventually especially when the graph is large. That's why I would prefer to know the specific case that we really have no way but can only use 1e-4. The flaky issue can be workaround by other approaches like fixing the random seed IMHO. -- 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]
