bhack
bhack
Any news on this? We are adding a Gaussian filter again in Keras-cv: https://github.com/keras-team/keras-cv/pull/143
No, I meant these kind of optimizations could be handled by XLA/MLIR
/cc @innat
/cc @ianstenbit @tanzhenyu
There was an old compositional implementation of a 2d deformable conv: https://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers/convolution/deformable_conv.html But I have not benchmarked it under XLA or the new bridge to understand how it is efficiently...
> I believe that implementation is a copy of the following: Yes probably it is derived but with some refactoring. It could be interesting if you can try to jit_compile...
I still see np code there that instead in the tensorlayer flavour was removed.
We had a V2 version at https://github.com/smallsunsun1/deformable_conv_v2 but I've not verfied the impl.
Yes but it is just for benchmarking so I am not worried about the license while it is not used in a PR: https://github.com/RuaHU/keras_DCNv2
> maybe that's something the XLA can help. If XLA cannot speed up the operator enough or it cannot compile we could always check with the upstream why it cannot...