Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Is cuDNN not installed on the current Dockerfile?

See original GitHub issue

I think that cuDNN is not installed in the current Dockerfile/Docker image. It is necessary to modify Dockerfile as follows.

In dockerfile Now FROM nvidia/cuda New FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04 or FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04

https://hub.docker.com/r/nvidia/cuda/

Issue Analytics

State:
Created 6 years ago
Comments:8 (5 by maintainers)

Top GitHub Comments

1reaction

relutropycommented, Jun 22, 2017

In order to use Tensorflow-gpu, it seems that CUDA and cuDNN v 5.1 need to be installed. Even in Docker 's container, even in a pure computer.

It seems appropriate that CUDA and cuDNN are prepared in containers beforehand. Even if you use a pure computer, not only in the container, it seems necessary to have an environment where CUDA and cuDNN are installed first.

Therefore, it seems that you will use the Docker image which nvidia has released. Even the official tensorflow-gpu Docker image is as follows.

FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04 https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/docker/Dockerfile.gpu

In an environment where CUDA and cuDNN are installed, Tensorflow is installed. I think it would be nice to do the same.

0reactions

relutropycommented, Jun 22, 2017

Try the Docker build test with the Dockerfile that changed only the first line. FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04 We can install cuDNN no changes.

Requirement is important. The effect by installing cuDNN seems to have been better than ever.