Just to make sure people aren't scared off by this: jax provides a lot more than just low level linear algebra. It has some fundamental NN functions in its lax submodule, and the numpy API itself goes way way past linear algebra. Numpy plus autodiff, plus automatic vectorization, plus automatic parallelization, plus some core NN functions, plus a bunch more stuff.
Jax plus optax (for common optimizers and easily making new optimizers) is plenty sufficient for a lot of NN needs. After that, the other libraries are really just useful for initialization and state management (which is still very useful; I use haiku myself).
I'm a pretty big fan of moving away from thinking about ML/Stats/etc specifically and people should more generally embrace the idea of differentiable programming as just a way to program and solve a range of problems.
JAX means that the average python programmer just needs to understand the basics of derivatives and their use (not how to compute them, just what they are and why they're useful) and suddenly has an amazing amount of power they can add to normal code.
The real power of JAX, for me at least, is that you can write the solution to your problem, whatever that problem may be, and use derivatives and gradient descent to find an answer. Sometimes this solution might be essentially a neural network, other times the generalized linear model, but sometimes it might not fit obviously into either of these paradigms.
Jax is definitely the right direction for the python ecosystem, but it can't solve all your problems. At some point you still need a fast language.
In particular i want to be able to measure feature importance on both inputs and internal layers on a sample by sample basis. This is the only thing currently holding me back from using JAX right now.
Alternatively a simle to read/understand/port implementation of DeepLIFT would work too.
thanks
Is that big industry lab Google or Deepmind? haha
Haiku is really cool - I haven't used Flax. It'll be really interested to see the development of JAX as time goes on. I also saw some benchmarks that show its neck-and-neck with PyTorch as the fastest of the three, but I think with more optimization its ceiling is higher than PyTorch's.
Of course. It's the only library that can be explained from first principles: https://jax.readthedocs.io/en/latest/autodidax.html
Even still, do you think researchers will want to take the time to learn all of that when PyTorch gives them no real reason to switch? Every day spent learning JAX is another day spent not reviewing literature, writing papers, or developing new models.
pytorch historically hasn't really focused on forward mode auto differentiation: https://github.com/pytorch/pytorch/issues/10223
this definitely limits its generality relative to jax, which makes it less than ideal for anything other than 'typical' deep neural networks
this is especially true when the research in question is related to things like physics or combining physical models and machine learning, which imho is very interesting. those are use cases that pytorch just isn't good at.
Every day spent learning JAX is also another day spent not trying to fit a round peg into a square hole of other libraries. I made the leap when I was doing things that were painful in pytorch. In terms of time, I think I came out ahead.
Not everything is a nail, and pytorch is better for some things, an jax is better for others. "Every day spent learning the screwdriver is a day spent not using your hammer."
To get started JAX is just knowing Python and adding `grad`, `jit` and `vmap` to the mix, it takes about 5 minutes to get going.
To me this is the real power of JAX, it can be viewed as a few functions that make it easy to take any python code you've written and work with derivatives using that. This gives it tremendous flexibility in helping you solve problems.
As an example, I mostly do statistical work with it, rather than NN focused work. It took probably a few minutes to implement a GLM with custom priors over all the parameters, and the use then Hessian for the Laplace approximation of parameter uncertainty. The proper way to solve this would have been using PyMC but this worked good enough for me, and building the model in scratch in JAX took less time than refreshing the PyMC api for me.
For example, a long chain of pmaps, each with some sort of device partitioning logic, not JIT compiling is extremely hard to understand. I basically had to binary search code until the compile errors disappeared.