i.MX ProcessorsのトピックCustom Convolution Quantization to run on NPU

Custom Convolution Quantization to run on NPU

taklause — Fri, 21 Oct 2022 14:05:26 GMT

Hello, I ll try to convert certain networks to the NPU with the best possible performance.

Therefore I saw in the NXP IMX.8 ML Guide that the NPU provides faster inference for "per-tensor" quanized models.

I saw in some example (ex. Posenet) that all the conv. layers have been quantized Layerwise ,not Channel-wise as a Tensorflow Default.

I have yet not been able to quantize the convolution in a simple example in a Layer wise manner. Do you might have an example how to achieve this?

Thanks Daniel

Bio_TICFSL — Mon, 24 Oct 2022 13:42:38 GMT

Hello,

Do you see that examples with the NPU? It appears that do not work with ARM architectures.

Regards

taklause — Thu, 01 Dec 2022 06:42:58 GMT

It can be done by using the eIQ converter tool. Thanks