search

ARM-software / CMSIS_5

CMSIS Version 5 Development Repository
http://arm-software.github.io/CMSIS_5/index.html
Apache License 2.0
1.33k stars 1.08k forks source link

A suggestions #575

Closed lswzjuer closed 5 years ago

lswzjuer commented 5 years ago

Hope to be able to join the support of binarization convolution which is based on Xnor and bitcountering operations

JonatanAntoni commented 5 years ago

Hi @lswzjuer,

Can you update this issue to be slightly more expressive, please? What exactly do you want to suggest? -> Update the title accordingly, should be the rough summary instead of just "A suggestion". -> Rephrase the description to clearly what you want to achieve, why, and how if feasible.

Thanks for getting in touch, Jonatan

majianjia commented 5 years ago

I think op is referencing this work: https://pjreddie.com/media/files/papers/xnor.pdf

lswzjuer commented 5 years ago

Hi @lswzjuer, Can you update this issue to be slightly more expressive, please? What exactly do you want to suggest? -> Update the title accordingly, should be the rough summary instead of just "A suggestion". -> Rephrase the description to clearly what you want to achieve, why, and how if feasible. Thanks for getting in touch, Jonatan We can binarize the activation values and weights to constrain them to {-1, +1}，If all of the operands of the convolutions are binary, then the convolutions can be estimated by XNOR and bitcounting operations without Multiplication and addition. XorNet has so much potential, I suggest you in the Source/ConvolutionFunctions provides support for binary convolution (just like arm_convolve_HWC_q2. c)

JonatanAntoni commented 5 years ago

I tend to remove this issue because it's still not comprehensible for me what the request is about.

On the first glance it sounds like some enhancement to the CMSIS-DSP library. But reading "activation values" makes me think you are talking about CMSIS-NN functionality. Please bare in mind not everybody reading here is an expert in DSP or NN. CMSIS covers a broad range of aspects. I doubt "a suggestion" tells anything to a wider audience.

Please consider creating a new issue, stating clearly what you want to achieve, why and preferable how.