Consider moving away from TensorFlow
adamltyson opened this issue · comments
TF is causing us some headaches, namely lack of support for Windows, and not being on conda-forge. There are other options, but PyTorch seems to be the leading candidate currently. It's on conda-forge, seems fairly easy to install, and it's now part of the Linux Foundation (so can't be killed off by Meta).
I don't know how easy the TF->PyTorch transition would be, but the 3D ResNet here isn't that complicated. A re-write may be a way of speeding things up.
I remember someone had started looking into this (and reported some speedups), but I can't remember who!
I remember someone had started looking into this (and reported some speedups), but I can't remember who!
It was @noisysky! - https://github.com/noisysky/fastai-examples
I wonder if keras-core (soon to be keras 3.0) is a potential solution? keras is on conda-forge, and if the tensorflow backend requirement was removed, we could use pytorch without changing much of the classification code.