Issue with INT8 quantization of PCB Defect detection model on EII 3.0.1

[sagi-scalers-ai]

System information (version)

EII => 3.0.1

Operating System => Ubuntu 20.04

Problem classification => INFO

We are trying to use EII on 4th Gen Intel Xeon(Sapphire Rapids). We are able to custom build EII with latest OpenVINO to leverage AMX optimizations using the steps from Issue with current EII 3.0.1 in leveraging AMX optimization on 4th Gen Intel Xeon(Sapphire Rapids) .

But we see that the PCB defect detection demo on EII is not using INT8 quantized model and this might not effectively use the capability of AMX optimization on Sapphire Rapids.

Please guide us on the steps for INT8 quantization of PCB defect detection model.

[JesusE_Intel · Accepted solution]

Hi sagi-scalers-ai,

Thanks for reaching out, I am in the process of re-installing EII to investigate further. I believe the models for the PCB Defect detection are provided in FP16 and FP32. You could try to quantize the model to INT8 using the OpenVINO toolkit using the Post-Training Optimization Tool (POT) default quantization as this may not require an annotated dataset.

https://docs.openvino.ai/latest/pot_default_quantization_usage.html

Regards,

Jesus

[JesusE_Intel]

Hi sagi-scalers-ai,

I can confirm that all the IR models included in Edge Insights for Industrial are either in FP16 or FP32 format.

Where you able to quantize the model using OpenVINO?

Regards,

Jesus

[JesusE_Intel]

If you need any additional information, please submit a new question as this thread will no longer be monitored.

[FaizanK]

Hi, 

where can i find older version of EII like 2.6.3, or reference implementation of textile defect classifcation?