Scorch: a minimalist neural net framework in Scala inspired by PyTorch

Discussion
1

Hi all,
I created Scorch: a minimalist neural net framework in Scala inspired by PyTorch.
You can find it here: https://github.com/botkop/scorch
Under the hood it uses the n-dimensional arrays of nd4j (https://nd4j.org/).
I am thinking of making a version that uses mxnet’s ndarray.
Do you think this makes sense?
(To be honest, I find the Scala interface of mxnet rather bloated.)

Thank you.

2

The last few releases of MXNet (1.2 and 1.3) have greatly improved the Scala API and resolved memory management issues. I’d recommend giving it a fresh look.

1 Like
3

@simonco they are still weak, doesn’t even have autodiff
I’m wondering if scorch can be set to use mxnet C++ component as backend, this can take the best of both.