I don't know if you are familiar with the (infamous) graveyard of projects of Google: https://killedbygoogle.com
A lot of Google's efforts are moving from tensorflow to jax jax. Going to jax-by-default would be an option, but I believe google will eventually kill jax as well when some middle manager comes up with whatever shiny new thing they have.
As much as I hate Meta and all they represent, they have a much better track record of keeping projects alive, so my vote would be for pytorch.
(just to clarify, we can use jax or pytorch already changing the backend, but a number of operations were optimized and benchmarked using tensorflow and there is room for improvement when using any of the other backends)
I don't know if you are familiar with the (infamous) graveyard of projects of Google: https://killedbygoogle.com
A lot of Google's efforts are moving from tensorflow to jax jax. Going to jax-by-default would be an option, but I believe google will eventually kill jax as well when some middle manager comes up with whatever shiny new thing they have.
As much as I hate Meta and all they represent, they have a much better track record of keeping projects alive, so my vote would be for
pytorch.(just to clarify, we can use jax or pytorch already changing the backend, but a number of operations were optimized and benchmarked using tensorflow and there is room for improvement when using any of the other backends)