This sounds eminently sensible but it never pans out that way in practice for me:
How do you understand the ML algorithms if you haven't implemented them yourself?
And if you have implemented them for the sake of understanding, in whatever language you fancied, why not just keep using that implementation?
The reason I'd choose Python in practice would be as a social compromise with collaborators.
(I say this having recently switched from Julia to Common Lisp because I didn't feel the ecosystem was giving me much practical benefit.)