>>13631516What exactly can you do in Jax that you cannot do in Pytorch? As far as I am aware, Jax main focus is its autograd that can take pure python and numpy code. It is also not faster in all cases apparently, Pytorch is still slightly faster for linear/dense layers
> while jax is more of a researcher thingThat is also the same thing that was said about Pytorch when it was starting to gain momentum.
>>13631497Like
>>13631519 said Pytorch is 'easier' and more intuitive. Simply because its API is well written and the documentation is clean
Tensorflow was a mess in 1.x, multiple ways of doing the same thing, tensorflow.contrib was full of deprecated shit you shouldn't touch etc. Tensorflow 2.x is copying everything Pytorch did, at which point why not just jump to Pytorch?
Tensorflow still has a few nice things but they are mainly for deployment in devices (tflite and others) but now you can just convert your models to ONNX and deploy them through any library worth a shit
