TensorRT5/6 FC Layer not support Int8 quantization.

TensorRT5/6 FC Layer not support Int8 quantization.

Hi,

Can you provide the following information so we can better help?

Provide details on the platforms you are using:
Linux distro and version
GPU type
Nvidia driver version
CUDA version
CUDNN version
Python version [if using python]
Tensorflow version
TensorRT version
If Jetson, OS, hw versions

Files

Include any logs, source, models (.uff, .pb, etc.) that would be helpful to diagnose the problem.

If relevant, please include the full traceback.


Reproducibility

Please provide a minimal test case that reproduces your error.

Ubuntu16.04
P40
396.26
9.0
7.6.3
3.6.8
1.12
6.0.1.5

Case 1: Just use any *.engine with a fully connected layer and quantize it, the following error occurs.

TensorRT] WARNING: No implementation of layer FC obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.

Case 2: If I use a 1x1 convolutional layer instead of a fully connected layer with the same weight, the quantization process is ok.

And the *.engine8 size of Case 2 is much small Case 1, saying that fully connected layer is not supported for quantization.

Hi,

I can help you more quickly if you provide a minimal script that reproduces your error similar to the one you provided here: https://devtalk.nvidia.com/default/topic/1064929/tensorrt/resizebilinear/post/5392890/#5392890

Please provide (1) a simple script failing to quantize a network with FC layer and (2) a simple script successfully quantizing a network with a 1x1 conv layer.

Thanks,
NVIDIA Enterprise Support

how to check my pytorch layers contain INT64?