masahi edited a comment on pull request #7137: URL: https://github.com/apache/tvm/pull/7137#issuecomment-750769271
Also I find that gluoncv MaskRCNN mostly follows the design of PyTorch MaskRCNN. But they apply sigmoid to objectness network outputs, so in their case there are no negative scores and all inputs to RPN NMS, which could be thousand of boxes, are valid, even if "by valid" we mean the previous wrong definition of having positive score. So without this fix, if you compare MaskRCNN performance of PyTorch and GluonCV after we compile them to TVM, PyTorch model would run much faster because our cheat would work to ignore negative boxes, while there is no negative boxes in GluonCV MaskRCNN. So equivalent models give inconsistent result after they get compiled to TVM. This is non sense. So we have to accept the fact that we need to deal with lots of boxes in MaskRCNN. ---------------------------------------------------------------- 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: [email protected]
