when compiling .pb model, the lager dataset to used for quantification is better?



  • I notice the nncase compiling model need a dataset to quantification, the dataset should be larger?

    ncc -i tflite -o k210model --dataset ./images ./mbnetv1.tflite ./mbnetv1.kmodel
    

  • Staff

    We call the dataset as Calibrate Dataset
    larger is better, but the result would not be too much different between large or small scale calibrate dataset.