How can I setup TensorRT 5.1.5.0 with Cudnn 7.6

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

Hi, all,

My environment is Ubuntu 16.04, with tensorflow-gpu 1.13.X, cuda 10.1, cudnn 7.6. I have already configured tensorRT 5.1.5.0 well. In python, all the import package are fine, such as:

import tensorflow as tf
import tensorrt
import uff
import tensorflow.contrib.tensorrt

But when I try to save TensorRT model, there is the INFO: tensorflow : Running against TensorRT version 0.0.0. I searched the 5.X document which lists that cudnn and tf version should be 7.5 and 1.12.

Could you give me some advises to solve this problem? or should I have to degrade my tensorflow and cudnn?

Thanks, all

2

Hi wyhcjdtc,

Have you tried our NGC containers using nvidia-docker for your tasks before? They save a lot of hassle in trying to get all of the framework versions to match, and make work easily reproducible.

nvidia-docker setup instructions are here: https://github.com/NVIDIA/nvidia-docker#ubuntu-140416041804-debian-jessiestretch

Check out the latest TensorRT container here: https://ngc.nvidia.com/catalog/containers/nvidia:tensorrt

If you follow through with using Python in the TensorRT container, make sure you run

/opt/tensorrt/python/python_setup.sh

as mentioned on the NGC link above.

Thanks,
NVIDIA Enterprise Support

3

Thanks, I’ll try to see if the docker is suitable for my project. But I still wonder if the problem I met is caused by tf and cudnn version.

Cheers,

Yhan