Question about Calibration Dataset

Hello everyone,

In the slides of this presentation http://on-demand.gputechconf.com/gtc/2017/presentation/s7310-8-bit-inference-with-tensorrt.pdf, they point that the calibration dataset should be a subset of the validation dataset.
Can you please explain why it can’t be a subset of the training dataset?

Thanks!

Hi,

In presentation, it’s just a recommendation/suggestion that ideally calibration dataset should be a subset of validation dataset.

Calibration dataset can be from training/test data set as well. Just make sure the calibration files are representative of the overall inference data files.

Thanks