post training optimization

[akshya]

Hi all,
I have a some questions on the post-training Optimization toolkit using the DefaultQuantization algorithm in simplified mode. As mentioned on the following links, it needs a calibration dataset of 300 samples to carry out the optimization process.
https://docs.openvinotoolkit.org/latest/pot_README.html

https://docs.openvinotoolkit.org/latest/pot_docs_BestPractices.html

Q1. Whether the number is fixed for all datasets?
Q2. Is there any way to figure out this number for custom dataset?
Q3. What to do, in case someone is doing post-training optimization on a pre-trained model, fine-tuned and trained with a fewer number of images(less than 300).

[Munesh_Intel · Accepted solution]

Hi Akshya

Thanks for contacting us.

We recommended minimum number of 300 representative samples from our experience, as we found that these are sufficient to get representative statistics in most cases. This is by no means a definite number.

As such, we leave it to you to determine the number of representative samples needed for your custom dataset.

Regards,

Munesh

[Munesh_Intel]

Hi Akshya,

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

Regards,

Munesh