Sparse Model Benchmark about xnnpack HOT 5 CLOSED

The layout of the saved models is the layout that was used during training. We modify the layout when loading for inference.

There does exist C++ code that runs the sparse models that used to generate the times in the paper, we do not consider it suitable for release due to its current lack of generality. We're working on a more general solution. Please reach out to the authors via email if you have a specific research use case that is being blocked by this.

from xnnpack.

Hi @ekelsen,

The layout of the saved models is the layout that was used during training. We modify the layout when loading for inference.

As I started to read tensorflow lite's code, I noticed that convolution operations are replaced by gemm kernels (in convolution-nhwc.c) rather then spmm kernels (in convolution-nchw.c). Is it because that the spmm are not integrated into the tf lite yet?

My main purpose for now is to see how is spmm used during tf lite inference when xnnpack is applied. It would be really helpful if you can recommond some documents/readings regards this.

Thank you very much!

from xnnpack.

It is not integrated into TF-lite. We will release a delegate that won't change the TF-lite code, but will essentially make TF-lite call XNNPACK to run a sparse model.

from xnnpack.

Hi @ekelsen,
Is there a roughly expected date for the release?

Thank you!

from xnnpack.

We hope to release it by CVPR.

from xnnpack.