erksch
commented
2 years ago Hey, thank you for your interest!
Yes, we are still early in development and are planning to rewrite the API to abstract the IValue concept, thus making it more powerful and allowing more inference use cases. Currently, you might have to change your module structure if you don't only use float or long tensors as input and output, this constraint would vanish with the new IValue API.
But nonetheless, if your modules are already so simple that the current API is sufficient or if changing your module is no big deal, I would not hesitate to give it a try! The breaking changes that will follow should be fairly easy to migrate.
And as soon as we have IValue support we can think about hitting a stable version. Until then, I think we will not change the current API much, we already use it in production for our simpler models.
Hi guys, Thank you for creating this valuable library. I know you are at an early stage of developement, but I would like to use this ASAP so I would like to hear from you if you think you have a stable-enough version for doing some inference on pretty standard nets, nothing too fancy.