Docker images

[mathpluscode]

The benchmark will be performed using dockers to ensure reproducibilities.

Therefore we will need to build a docker image supporting both packages.
The main idea would be using conda inside docker and having two separate virtual environments so that they can have different package dependencies.

It would be also nice to push the image into docker hub.

[fepegar]

What's the benefit of using Docker instead of only conda?

[mathpluscode]

What's the benefit of using Docker instead of only conda?

I'm not sure I understand it correctly, but my understanding is:

• Conda mostly only handle py packages but docker can handle system-wide packages. Especially with TensorFlow, we have specific TensorFlow docker images that you know it handles the system dependencies for you. For instance, on remote servers, you may not have the right to install additional system packages, and you can not modify some system settings. With docker, it is doable and reproducible. A particular case with TensorFlow is when we want to do profiling to measure/optimize the code, we will need root access which is impossible on clusters.
• As I'm suggesting using conda inside docker, it means that using docker is not necessary. I suggest docker more from the point of the view that "this is the most reliable way to make sure everything is reproducible".

[fepegar]

using docker is not necessary

Shall we close this issue then? :D JK

Conda mostly only handle py packages

I believe conda supports all kind of binaries, not just Python stuff.

It's still not very clear to me that conda plus a requirements file to pip-install everything wouldn't be enough. But I'm looking forward to being convinced of the opposite!

[mathpluscode]

I consider docker as an extra feature, as DeepReg until now only use conda as well LOL.

It is not urgent, but with docker, you can literally do not bother anything regarding the installation requirement as it will be all done automatically.