16x8 quantization support

[andy-aic]

Hi,

I'm getting started with the xmos.ai platform and I read in the documentation that 16x8 quantization has "partial support". What does "partial support" mean?

I read about this here: https://github.com/xmos/ai_tools/blob/d ... or-xcoreai

Thanks for you help.

[albertoisorna]

Hello,

Fully int8 quantization is supported, I would recommend first running it in int8 to get the base precission.
Regarding 16x8, so 16-bit activation, and 8-bit weights are currently a work in progress. I would guess not all operators are supported at this stage, so depending on wich layers is using your model you will be able to run it or not.

You can give a try the conversion using: https://www.tensorflow.org/lite/perform ... perimental
And then export the model and try to run it on the board.

[deepak.panickal]

Hi andy-aic ,

We have a 16x8 audio-based example at https://github.com/xmos/ai_tools/tree/d ... io_network .
16x8 quantization is meant to be used for audio-based networks, such as for denoising, dereverb etc, so that you can have higher precision.
We support most operators that you would encounter for such a model.
Please try out the tools and let us know if you run into any issues.

Thank you,
Deepak Panickal

[andy-aic]

Thank you for your replies!