Currently, we take the first 500 images of an ImageNet dataset for calibration. This is 1% of the full dataset, or 100% of the "min" dataset.
In fact, we should create a package that creates a calibration dataset from an ImageNet dataset. It can still take 100% of the "min" dataset and 1% of the full dataset, but should have three variations e.g.:
Currently, we take the first 500 images of an ImageNet dataset for calibration. This is 1% of the full dataset, or 100% of the "min" dataset.
In fact, we should create a package that creates a calibration dataset from an ImageNet dataset. It can still take 100% of the "min" dataset and 1% of the full dataset, but should have three variations e.g.:
first.500
: the current behaviour;mlperf.option1
: use cal_image_list_option_1.txt (used by Intel);mlperf.option2
: use cal_image_list_option_2.txt.If we use the calibrated dataset, not the original one, we can probably prevent https://github.com/ctuning/ck-openvino/issues/12.