tsalib VS functorch

The one thing I really *really* wish got more attention was named tensors and the tensor type system. Tensor misalignment errors are a constant source of silently-failing bugs. While 3rd party libraries have attempted to fill this gap, it really needs better native support. In particular it seems like bad form to me for programmers to have to remember the specific alignment and broadcasting rules, and then have to apply them to an often poorly documented order of tensor indices. I'd really like to see something like tsalib's warp operator made part of the main library and generalized to arbitrary function application, like a named-tensor version of fold. But preferably using notation closer to that of torchtyping.

I would suggest checking https://pytorch.org/functorch/ and https://github.com/metaopt/torchopt for efficient inference and training with ensembles (e.g., t be independent actors in a multi-agent setting or multiple critics).

There were some bugs still with advanced indexing in an older release of functorch, I believe they should be fixed now though: https://github.com/pytorch/functorch/pull/862

Tangentially related but there is an effort to get some of the features of JAX into PyTorch: https://pytorch.org/functorch/

Fwiw, composable vmap and stuff like that have also been implemented in PyTorch now - see functorch :) https://github.com/pytorch/functorch

Fwiw, it’s not like Pytorch’s design prevents function transformations from being implemented. See functorch for an example of grad/vmap function transforms: https://github.com/pytorch/functorch

You may find vmap from functorch to be quite useful: https://github.com/pytorch/functorch

If you're interested in function transformations in PyTorch, try out functorch :) https://github.com/pytorch/functorch

The pytorch programming model is just really hard to adapt to an XLA-like compiler. Imperative python code doesn't translate to an ML graph compiler particularly well; Jax's API is functional, so it's easier to translate to the XLA API. By contrast, torch/xla uses "lazy tensors" that record the computation graph and compile when needed. The trouble is, if the compute graph changes from run to run, you end up recompiling a lot.

I guess in Jax you'd just only apply `jax.jit` to the parts where the compute graph is static? I'd be curious to see examples of how this works in practice. Fwiw, there's an offshoot of pytorch that is aiming to provide this sort of API (see https://github.com/pytorch/functorch and look at eager_compilation.py).

(Disclaimer: I worked on this until quite recently.)