Problem of TensorRT3.0 under int8 mode

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

When we use TensorRT2.1, the gap between int8 mode and float32 mode is very small.

But when we switched to TensorR3.0, the result of int8 mode is very poor when compared with float32 mode.

Is there any bug of TensorRT3.0 int8 mode or the method of using TensorRT3.0 is different from using TensorRT2.1?

2

@sjones@nvdia if you have answer, you may replay me on this forum or send me a email rkangaa@connect.ust.hk

3

I got a similar problem like yours. When i use TensorRT3, the gap between int8 mode and float32 mode is very small. But when i switched to TensorR4.0.1.6,the result of int8 mode is very poor when compared with float32 mode. and i check the generated CalibrationTable, i found much difference between CalibrationTable generated under TensorRT3 and TensorRT4.

Do you fix the problem? how to fix it?
if you got any solution, could you please let me know. thanks.

4

Nvidia fix this bug in tensorrt3.0.4, you may ask nvidia to solve this problem for TR4. If you have other quanestions about tensorrt, you may send me an email, rkangaa@connect.ust.hk.