Building an older Tensorflow 2.x

Question

I have been trying to do this for a while now. I wasn't able to get a build on my local machine so I went the Docker route. I was able to successfully the use docker image devel-gpu to build Tensorflow. The problem is that it built the latest and greatest (2.5). I have searched and searched for a way to build an older 2.x version but have had no luck.

How do you build, say TensorFlow 2.1 using docker?

I seemed like they quit making docker images for individual versions that contained the source code to build.

Answer 1

If you are fine with using already built images there are older versions available on Docker Hub. Here is a link for 2.1: https://hub.docker.com/r/tensorflow/tensorflow/tags?page=1&name=2.1 This is the easiest way to obtain an older version.

If you want to build it yourself:

1. Clone the repository:

git clone https://github.com/tensorflow/tensorflow.git && cd tensorflow

2. Switch to the desired release:

git checkout v2.1.0

3. You'll find Dockerfiles in tensorflow/tools/dockerfiles/dockerfiles/ but before building I recommend you to check out tensorflow/tools/dockerfiles/README.md for exact building instructions. For version 2.1.0 it states that you have to run build command setting context to the directory with the README.md . That is:

cd tensorflow/tools/dockerfiles
docker build -f ./dockerfiles/cpu.Dockerfile -t tf .

And that's pretty much it, when building is finished you'll have an image named tf:latest (customize with flag -t in build command).