How can I integrate this project API to TensorFlow or Caffe?

[ysh329]

Is there some tips or docs? Thanks a lot!

[AnthonyBarbier]

Hi,

We're working on it but it's still early stages. We'll publish some blog posts and examples in the coming weeks / months.

[mikyoreyes]

Thank you

[ysh329]

@AnthonyARM
Thank you very much for your work! 👍
We will continue to focus on this repo.

[Thunderbob]

The licensing terms allow you to do this if you intend to do it yourself. We intend to integrate our library with Caffe and TensorFlow however this is still work in progress, so watch this space!

[bhack]

I really hope that ARM can intergrate upstream like Intel.

[mikyoreyes]

Hi @Thunderbob! Will you be creating a repository for Caffe/Tensorflow with the ComputeLibrary?

[AnthonyBarbier]

We're only getting started, we haven't decided yet how we were going to contribute to these projects

[psyhtest]

As I contributed CLBlast support in Caffe, I see two options for optimising Caffe on ARM Mali GPUs when using the BLAS approach for computing convolutions:

1. Implement all BLAS routines needed by Caffe in a partial Mali-optimised library (based on this released compute library) and provide its integration with Caffe similar to ISAAC.
2. Provide a Mali-optimised plugin for CLBlast using a new plugin framework proposed by @intelfx. In fact, he already used the Mali-optimised SGEMM code that ARM released over a year ago for creating a Mali-optimised CLBlast overlay. (See this Jupyter notebook for performance improvements on Samsung Chromebook 2 with AlexNet and SqueezeNet 1.0; the overlay needed more work support GoogleNet and SqueezeNet 1.1.) The new framework should be more robust and allow to use Mali-optimised implementations on Android (where the LD_PRELOAD trick is not available).

I'm happy to provide further pointers and discuss.

[austingg]

@psyhtest seems openblas works best

[psyhtest]

@austingg Yes, but it was fairly optimised implementation on quad-core Cortex-A15 @ 1700 GHz vs unoptimised implementation on quad-core Mali-T628 @ 600 MHz...

[ManubARM]

Closing for inactivity. Anyway to answer the initial question:

There is a project made exactly to solve this kind of problems: ArmNN
Here is the link to the repo:
https://github.com/ARM-software/armnn