[grahamannett]

seem's basically the same (didn't check every framework but seemed like most big names were on both) but being built on the cuda base image might make it more easy to attach a GPU?

Also I just noticed this is about the craziest dockerfile. It's all one RUN which good luck trying to troubleshoot if theres a non obvious error during building

[paulfurtado]

Every RUN statement creates a new layer in the image. Best practice for image size and io performance is to use the lowest number of layers possible.

That said, recent versions of docker do have the ability to squash layers at the end of a build which could give the best of both worlds.