int8 PTQ using random

[Luo__Andy]

Hi, 

  I want to do some quick perf benchmark for int8 quantized models but it is not time consuming to find and prepare all the calibration datasets. The accuracy is not a concern.

  Is there any way to generate int8 PTQ models with random input data? I did check the NNCF documentation and not find out how to do it. 

Regards,

Andy

[Aznie_Intel]

Hi Luo_Andy,

Thanks for reaching out.

In general, you may use a random input dataset to apply 8-bit post-training quantization. NNCF API has two main capabilities, Basic quantization and Quantization with accuracy control where you need to prepare calibration datasets and validation datasets on your own if you are using random datasets. The minimum dataset should be 300 samples.

You may refer to this Dataset Preparation for Object Detection Sample to prepare your dataset.

Regards,

Aznie

[Aznie_Intel]

Hi Luo_Andy,

This thread will no longer be monitored since we have provided information. If you need any additional information from Intel, please submit a new question. 

Regards,

Aznie