Closed stoicio closed 6 years ago
I have been able to work around this temporarily by using the offlicial tensorflow image as the base image instead of using nvidia's docker image.
I replaced the # FROM nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04
with FROM gcr.io/tensorflow/tensorflow:latest-gpu-py3
Sounds like you found a fix, so closing out.
I tried building an image from the Dockerfile.gpu file and the build process succeeded without any errors. However when I try to run some of the tensor flow scripts the scripts fails with the following error
F tensorflow/stream_executor/cuda/cuda_dnn.cc:221] Check failed: s.ok() could not find cudnnCreate in cudnn DSO; dlerror: /root/miniconda3/envs/carnd-term1/lib/python3.5/site-packages/tensorflow/python/_pywrap_tensorflow.so: undefined symbol: cudnnCreate Aborted (core dumped)