masahi edited a comment on pull request #7137: URL: https://github.com/apache/tvm/pull/7137#issuecomment-750438126
@kevinthesun @zhiics I asked torchvision people about boxes with negative scores in https://github.com/pytorch/vision/issues/3198 and now I fully understand the issue. See in particular this great answer https://github.com/pytorch/vision/issues/3198#issuecomment-750346278 My conclusion is that TVM's conversion rule for PyTorch NMS is definitely wrong and needs fixing. Here is my take away from the above discussion: * PyTorch detection model has two use of NMS - one in `ROIHead` and another in `RegionProposalNetwork`. * NMS scores in `ROIHead` are probability. There, they do score thresholding with user-chosen threshold before NMS. This NMS doesn't send boxes with negative scores to TVM. * NMS scores in `RegionProposalNetwork` correspond to "objectness" and they don't apply softmax or sigmoid to the output from objectness network. The scores are estimate of [logit function](https://en.wikipedia.org/wiki/Logit) and negative logit totally makes sense - it just mean probability is < 0.5. So a totally reasonable box can end up having a negative score. We **shouldn't** arbitrarily cut negative boxes from RPN. It's highly possible that one of the reasons you didn't get the same output from pytorch detection models after compiling to TVM is this incorrect assumption we've been making about negative boxes, because RPN output in PyTorch/TVM are totally different - we are only considering only about half of them. The good news is, I found a way to reduce the number of boxes that are sent to NMS. See these parameters https://github.com/pytorch/vision/blob/master/torchvision/models/detection/mask_rcnn.py#L68-L71. They are something like `topk` parameter for each classes separately. The default is 1000, and there are five classes/levels in RPN NMS. So that explain why we are getting about 4500 boxes to RPN NMS. If we set the `rpn_post_nms_top_n_test` to 200, we will get at most 1000 boxes, 200 boxes for each level in the feature pyramid. That will significantly make detection model go faster and still consider a lot of boxes to keep accuracy high. So please take a look at the above issue in torchvision and my comment carefully and let's merge this ASAP. ---------------------------------------------------------------- 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]
