tensorRT int8

My data directory don’t have int8 data file as well.
It can be inferred from code that the INT8 sample reads batches of data for calibration first, then reading other batches for INT8 inference and calculate the accuracy Top1 and Top5 scores. So I followed this to use image data for the calibrator. It can be woked for FLOAT. In mnist, my retrained model get 98.88% top-1 accuracy, but INT8 is only 8% top-1 accuracy. I don’t know where is wrong and how to select calibration data?