Search code examples
dockertensorflow

Where is the Dockerfile (or build) for the official tensorflow/tensorflow:2.xx-gpu Docker images?

I am trying to build a GPU-enabled Docker image for my ML application, but I have some fairly strict requirements:

  • tensorflow==2.13
  • python>=3.10,<3.12

The tensorflow/tensorflow:2.13.0-gpu image on Docker hub is based on Ubuntu 20.04 so use a quite old Python 3.8.

There is also tensorflow/build:2.13-python3.11 that technically works great, but is a whopping 11GB. I guess I am hoping to get an image closer to the 6.6GB that tensorflow/tensorflow:2.13.0-gpu clocks in at, but with a newer Python. (I acknowledge even this is huge, and most of that is from GPU support)

Is there an extant public (and trusted) Docker image that meets my needs?

Otherwise I could try to build my own. In this case I would like to replicate and slightly modify the same build behind tensorflow/tensorflow:2.13.0-gpu, but I am not sure where that lives. (It did not have luck just reproducing the commands shown in docker history.) Where is the Dockerfile (or build script) for the official tensorflow/tensorflow:2.13-gpu Docker image?

Solution

  • I found the code to build tensorflow/tensorflow runtime images here: https://github.com/tensorflow/build/tree/fdd023e/tensorflow_runtime_dockerfiles

    I was able to successfully build a new image with my own choice of Python version and Tensorflow as follows:

    1. Clone the repo
    2. docker build -f gpu.Dockerfile --target=base --build-arg="TENSORFLOW_PACKAGE=tensorflow~=2.13.0" . -t tensorflow/custom-runtime:2.13-python3.11