Closed soeaver closed 6 years ago
Sure. I have already implemented a simplified version of ORN on Caffe/pyTorch (supports 1x1 or 3x3 ARFs only). If you need it, I will find some time next week to clean it up and release. Thanks for your interest!
@ZhouYanzhao That' s very awesome! Looking forward your future work.
Waiting for the release for months! I've tried the torch version. It's much better than my caffe implementation. There must be something I misunderstood. Hope the official release will come out soon.
@samson-wang Thank you for your patience! The alpha version of Caffe/PyTorch implementation (PR #6) is released for your reference.
@ZhouYanzhao I've tried it. Thank you!
In my test case, ORAlign is much better than the ORPooling. However, there is no ORAlign in the caffe implementation.
An interesting observation. I tried to use pad:2
in conv1
instead of resizing image to 32x32
. The test accuracy drops a lot. pad:1
is OK.
@samson-wang ORAlign is not supported in the alpha version of caffe implementation, but it should be easy for you to port it from the torch/pytorch implementation :) As for your observation, my little guess is that inappropriate padding could cause misalignment of hierarchical receptive fields.
I'll close this issue since the implementations are released. If you have more questions, feel free to:
Thanks!