Type safe deepl learning framework for OCaml written from scratch

Currently supporting

• Fully connected
• Convolutional 3D
• Pooling 3D
• Flatten 3D

layers

The main idea behind

is to check the Neural Network architecture at compile time, making code like:

make_input3 [| 28; 28; 1|]
|> make_conv3d (* some params*)
|> make_fully_connected (* fully connected layer after the convolutional 3D one does not make any sense *)

absolutely impossible. This is something that makes the hole codebase much safer and something that is not done by Owl deep learning library.

Initially I was using Lapack linear algebra bindings for OCaml but it makes the operations like matrix reshape or flatten extremely complicated and expensive so the matrix library for this framework is also written from scratch.

This repo is an experimental project and a work in progress so sometimes it doesn’t even compile