Jetson TX2 上的 Tensorflow 导入导致分段错误（核心转储）

Question

I followed this installation tutorial ( https://devtalk.nvidia.com/default/topic/1038957/tensorflow-for-jetson-tx2-/ ) on a Jetson TX2 right after flashing it with Jetpack 4.2.2. I'm using the default python 3.6.8.

When I open a python3 terminal and import tensorflow, the terminals waits for a few seconds then prints "Segmentation fault (core dumped)".

There were no error messages during the install. Any help would be greatly appreciated, thank you.

Notes: I noticed looking here ( Which TensorFlow and CUDA version combinations are compatible? ) that tensorflow 1.14 will only work with cuDNN 7.4, but by default the sdkmanager installs cuDNN 7.5.

Answer 1

If you are using docker on 4.2.2 and have this issue:

Since you have seemed to check all the compatibility issues, I would suggest checking if docker is running the correct runtime. If your /etc/docker/daemon.json file looks like this:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

change it to this to force docker to use the nvidia-runtime. For some strange reason it doesn't seem to use this if not specified, even though it is the only runtime apparently present.

{
    "default-runtime": "nvidia",
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

Finally, for Jetpack 4.2.2, your docker version with docker --version should be:

Docker version 18.09.7, build 2d0083d