This is just a question, Currently, it is not possible to convert code directly between different ML frameworks, such as JAX and Pytorch. what do you think are the biggest challenges?

CodePudding user response：

I think there is no practical challenge to this issue. It is just the competition between companies that who can build a better framework. For that reason, they have specific politics for their framework which causes this problem that you've mentioned. Therefore, there wouldn't be any collaboration between companies to make these frameworks compatible with each other.

CodePudding user response：

It is in fact possible to automatically convert code between some frameworks; for example, you can convert JAX code to tensorflow using the experimental jax2tf module. In principle you could create a similar approach to convert JAX code to pytorch, but I don't know of any efforts to do that.

Fundamentally JAX is well-suited to be the source of this kind of conversion because it relies on a small-ish set of fundamental primitives, and is built around an extensible transformation framework on these primitives. I would think it would be more difficult to use tensorflow or pytorch as the source implementation of an automated transformation because they lack these features.