[GitHub] sbodenstein commented on issue #8582: Yolo2 operator
sbodenstein commented on issue #8582: Yolo2 operator URL: https://github.com/apache/incubator-mxnet/pull/8582#issuecomment-343462534 One more comment on this PR: is there anyway of verifying that the gradient is correctly calculated? This is an automated message from the Apache Git Service. To respond to the message, please log on 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 With regards, Apache Git Services
[GitHub] sbodenstein commented on issue #8582: Yolo2 operator
sbodenstein commented on issue #8582: Yolo2 operator URL: https://github.com/apache/incubator-mxnet/pull/8582#issuecomment-343229939 > I am working on object detection operators decomposed and implemented using gluon blocks/nnvm operators that is highly modular and customizable. @zhreshold: OK great! That will be very nice for MXNet to have. Do you have a rough timeframe for those? This is an automated message from the Apache Git Service. To respond to the message, please log on 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 With regards, Apache Git Services
[GitHub] sbodenstein commented on issue #8582: Yolo2 operator
sbodenstein commented on issue #8582: Yolo2 operator URL: https://github.com/apache/incubator-mxnet/pull/8582#issuecomment-342901844 One issue for us: this layer is an 'output layer' and hence doesn't support out grads into the layer. This is an unfortunate restriction, as it rules out things like example weighting. In addition, it does not return the actual per-example loss, which is non-trivial to compute. I think (but stand to be corrected, @piiswrong) that MXNet is moving away from these `output layers`. One option: split this into a `yolo_region_layer` (based on DarkNet https://github.com/pjreddie/darknet/blob/master/src/region_layer.c), and then a separate loss layer that returns the loss. What are others thoughts on this? (could also have this 'output layer' and then also the separated region and loss layers, but this seems like a lot of code duplication). This is an automated message from the Apache Git Service. To respond to the message, please log on 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 With regards, Apache Git Services