TensorRT INT8 plugin layer

[dvbr]

Hello,

I would like to know if there is a way to develop an INT8 plugin layer ?

If it can be done, there is a sample available to described the way ?

If not, tensorRT 5 will be updated soon to allow that ?

David

[NVES]

Currently, Plugin layers can support four data formats and layouts:
NCHW single (FP32), half precision (FP16) and integer (INT32) tensors
NC/2HW2 and NHWC8 half precision (FP16) tensors

we are always gathering feedback from the community and adding features to future TRT versions.

[Leoli96]

@NVES

Currently, Plugin layers can support four data formats and layouts:
NCHW single (FP32), half precision (FP16) and integer (INT32) tensors
NC/2HW2 and NHWC8 half precision (FP16) tensors

we are always gathering feedback from the community and adding features to future TRT versions.

Hi NVES,

Do you mean that currently we cannot do INT8 inference with customized plugin layer?

[xxdyx110]

@NVES

Currently, Plugin layers can support four data formats and layouts:
NCHW single (FP32), half precision (FP16) and integer (INT32) tensors
NC/2HW2 and NHWC8 half precision (FP16) tensors

we are always gathering feedback from the community and adding features to future TRT versions.

Hi NVES,

Do you mean that currently we cannot do INT8 inference with customized plugin layer?

Hi, I have test TRT int8 inference, but the accuracy is totally wrong! I wander how about yours?