I was wondering if this was possible at all?
The reason I am asking is because there has been some attempts at doing this but from Python to OCaml like languages. Outlined here: https://coq.discourse.group/t/machine-learning-and-hammers-for-coq/303
Yes it is totally possible, see these resources:
- OCaml Owl scientific computing library, in particular Neural Networks and Algorithmic Differentiation chapters of their manual
- Deep Learning with OCaml blog post from Jane Street
- OCaml in Data Science discussion at this forum
- Will OCaml be a good choice for writing high performance, parallelizable machine learning libraries discussion at this forum
If you are interested in improving the OCaml capabilites in this area, I think joining Owl development is the way to go, you can see their repositories here: https://github.com/owlbarn/
They have some ideas on possible improvements of the library at Proposed Projects page. I personally think, that saving and loading ONNX format is the most important improvement, to be able to load networks created in third-party tools, see the IR specification itself.
See the example on how to code the simple object-detection convolutional neural network (Mask R-CNN) at https://github.com/owlbarn/owl_mask_rcnn
5 Likes
There’s also a couple of followup posts to that first one:
- One on [reinforcement learning][https://blog.janestreet.com/playing-atari-games-with-ocaml-and-deep-rl/], using pytorch via OCaml (creating bots for playing Atari games!)
- One on transfer learning using pytorch as well.
2 Likes
Update: It seems that Owl project started to implement ONNX integration. Wish them luck, and more pull requests!
1 Like
An update about MLIR from Google - it was merged in the LLVM mainstream and now officially a part of the project - see https://mlir.llvm.org/
It should be released together with LLVM 10 and you can expect wider adoption after that move.
related: [ANN] PyTorch bindings
1 Like