It's a shame there isn't a Python TF implementation of this (apparently due to a different file structure between the tf.js model and TF SavedModel?), and it doesn't sound like it's a priority for the team.
TF 2.0 in general (SavedModel included) has been a giant clusterfuck so far. As we used to say at Google "the old thing is deprecated, the new thing is not ready yet" - internally Google is pretty much entirely like this. I get that they had to ship it sooner or later, but IMO they shipped 2.0 a bit too soon. Shipping sooner is fine if you fix the bugs quickly, but that's not really happening either - some rather potent bugs remain open for half a year or more.
So at the moment you have a choice of working with a version that's officially EOL (1.15) or using a version that has several critical features broken or missing, and on which TF Model Zoo doesn't work. Oh, and EOL version is also busted beyond belief. Worse yet, Google does not reward maintenance work, and since this is considered "launched", you can count on the bugs being there for an extended period of time until Google itself moves to 2.0, which I'd wager is unlikely to happen anytime soon. Facepalm.
Problem with Julia is that approximately nobody wants to learn yet another language just to do pretty much the same things that Python can do with the help of e.g. PyTorch or NumPy. Us programmers often overestimate the level of enthusiasm non-programmers have for learning programming languages. Heck, I'm kind of also running out of enthusiasm in this area myself.
Julia should be a much easier transition for a python programmer than Swift would be. If your argument is that swift can just be called from python so that the python programmers don't have to touch it directly, then I'd say:
1) That's not the most robust solution, especially for deep learning because auto-diff tools don't cross language barriers well
2) If you have an argument for why 1) isn't a big deal, then the next step would be demonstrating why that argument doesn't apply to julia just as well (if not better) than Swift. In fact, Julia has a great interoperability story with other languages, one good example is DiffEqPy and DiffEqR which are Python and R packages respectively for accessing julia's state of the art DifferentialEquations library.
I agree. I looked into it briefly (since I sorta already know Swift anyway and like the language) but I can't even run this stuff on Linux easily, so it's kind of useless for me so far. This doesn't in any way mean that Julia has a snowball's chance in hell of taking off though. Besides the language, Python had a decade of head start in terms of "batteries included". And people have gotten used to that particular brand of batteries.
Is Julia likely to become as massively popular as Python? Probably not. Is julia likely to become much bigger than it already is? Yes, I'd say so.
Julia's value proposition is mostly geared towards people trying to do things that are just too difficult, awkward or slow in other languages. It's not just about speed, dynamism and friendly syntax. It's also about solving the expression problem and providing unprecedented composability.
So while it doesn't have great appeal to end users, it does appeal to people who make the sort of things end users want so the community is currently swelling with very talented people making state of the art research libraries. Honestly, whether these libraries attract hordes of 'scripters' or not, doesn't really matter to julia's usefulness to the people who use it currently.
But I think that as julia's package ecosystem evolves and matures (We've only been at 1.0 for a year!), it's going to get more and more attractive to end users.
I'd like to add that it took the Julia community itself quite a bit to understand what this "value proposition" actually is, especially that it is not just a very nice high level language language geared towards linear algebra that is fast as C and allows some interaction with the type system!
I feel like a lot of performance could still be extracted from Python libs everyone is using. I.e. NumPy should probably use MKL on Intel out of the box - an easy 2x improvement in performance. On dense linear algebra MKL is pretty impressive. All the stuff that has to happen in Python will be slow as molasses, though, no question there. Parallelization is also currently much harder than it needs to be. I have a bit of code that does elastic deformation on images. It uses a single core for whatever reason (it does call out into numpy and scipy, obviously, this is not done in Python). All my options for parallelizing this suck mightily. I might end up doing exactly what you're opposed to: drop it down to C++ and use pybind, because currently this stuff is single handedly responsible for quadrupling the duration of training runs, which weren't super fast to begin with.
PyPy can't get you all the way there in many cases. Unlike Julia. The real problem with PyPy is Python's semantics. They preclude many essential optimizations to get those final few drops of performance.
ONNX is a turd in is present state as well. Too many unsupported ops and you can't really run ONNX on e.g. a phone or really anywhere. You could theoretically convert it into something that does run where you need it to run, but then you have to deal with the converter's supported ops as well, and design your net around the intersection of the two sets of ops and their parameters. If you're doing anything more complicated than basic classifiers, you're going to run into issues.
That said, I think in a few years with enough elbow grease ONNX could become pretty exciting.
As to what people should use for research: just use PyTorch. I switched to it when it first came out, and never looked back, _except_ when I need my stuff to run on e.g. a phone or something, which is something PyTorch won't help you with.
It's a shame there isn't a Python TF implementation of this (apparently due to a different file structure between the tf.js model and TF SavedModel?), and it doesn't sound like it's a priority for the team.