Interpret the content of calibration table

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hi,
My environments are Xeon E5-2620 + Nvidia T4,
I run sample_int8 mnist successfully within tensorrt:19.10-py3 container,
the code same as TensorRT/sampleINT8.cpp at release/6.0 · NVIDIA/TensorRT · GitHub

After the execution, it produced CalibrationTablemnist and its contain had many “ : value”,
according to README: “value” corresponds to the floating point activation scales determined during calibration for each tensor in the network. The real “ : value” were listed below:
TRT-6001-EntropyCalibration2
data: 3c008912
conv1: 3c88edfc
pool1: 3c88edfc
…

  1. Each value is so huge and much great than 127(the max value of int8), how the values can be calibrated for int8 processing? Can you give me an explanation or example how to convert these values to the range [-127, 127]?

2.My input dataset are download from MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges, but the images parsing by sample_int8 are 12800 not 40000(README said), why?

2

Hi,

Similar to below Dev Talk topic. Request you to follow below topic to updates:
https://devtalk.nvidia.com/default/topic/1068669/tensorrt/calibrationtable-and-executable-engine

Thanks