TensorFlow 2 PyTorch?

[innat]

No description provided.

[AlexanderLutsenko]

@innat But why?

[innat]

A few months ago, there was a Kaggle competition where a TF model was needed to submit. And nobuco did well. Now, a new competition will be organized where the PyTorch model will be needed to submit.

[AlexanderLutsenko]

@innat I actually have the reverse converter written, although It's not nearly as sophisticated as Nobuco. Problem is, I don't think it's worth the effort. Hardly anyone uses Tensorflow as a research framework nowadays. What I believe has much potential is JAX. Perhaps we should focus on developing a JAX->Pytorch converter instead.

[innat]

Hardly anyone uses Tensorflow as a research framework nowadays.

We are considering general tools, not particular ones for the research area. And IMO, TensorFlow wouldn't be gone any time soon.

What I believe has much potential is JAX. Perhaps we should focus on developing a JAX->Pytorch converter instead.

Good stuff.
This may interest you, https://twitter.com/fchollet/status/1678777783848419330

[AlexanderLutsenko]

This may interest you, https://twitter.com/fchollet/status/1678777783848419330

@innat Wow, that's huge! Never seriously considered making a unified API since no one would trust my determination to support it for a number of years. But the Keras team may actually pull it off. Hope they don't change too many things in the upcoming release so I could add Keras 3 in Nobuco with just a handful of import replacements.

[innat]

@AlexanderLutsenko
Most probably, you can collaborate with keras-core and add functionalities of nobuco directly to keras-core as part of its APIs.

check: https://twitter.com/kornia_foss/status/1678837500083568643
check: keras-team/keras-core#450
check: keras-team/tf-keras#262

[AlexanderLutsenko]

@innat Thanks, I'll look into it.