PyTorch bindings for OCaml

ocaml-torch provides some ocaml bindings for the PyTorch tensor library.
This brings to OCaml NumPy-like tensor computations with GPU acceleration and tape-based automatic

These bindings use the PyTorch C++ API and are
mostly automatically generated. The current GitHub tip and the opam package v0.7
corresponds to PyTorch v1.10.0.

On Linux note that you will need the PyTorch version using the cxx11 abi
cpu version,
cuda 10.2 version.

Opam Installation

The opam package can be installed using the following command.
This automatically installs the CPU version of libtorch.

opam install torch

You can then compile some sample code, see some instructions below.
ocaml-torch can also be used in interactive mode via
utop or

Here is a sample utop session.

Build a Simple Example

To build a first torch program, create a file with the
following content.

open Torch

let () =
  let tensor = Tensor.randn [ 4; 2 ] in
  Tensor.print tensor

Then create a dune file with the following content:

  (names example)
  (libraries torch))

Run dune exec example.exe to compile the program and run it!

Alternatively you can first compile the code via dune build example.exe then run the executable
_build/default/example.exe (note that building the bytecode target example.bc may
not work on macos).



Below is an example of a linear model trained on the MNIST dataset (full

  (* Create two tensors to store model weights. *)
  let ws = Tensor.zeros [image_dim; label_count] ~requires_grad:true in
  let bs = Tensor.zeros [label_count] ~requires_grad:true in

  let model xs = Tensor.(mm xs ws + bs) in
  for index = 1 to 100 do
    (* Compute the cross-entropy loss. *)
    let loss =
      Tensor.cross_entropy_for_logits (model train_images) ~targets:train_labels

    Tensor.backward loss;

    (* Apply gradient descent, disable gradient tracking for these. *)
    Tensor.(no_grad (fun () ->
        ws -= grad ws * f learning_rate;
        bs -= grad bs * f learning_rate));

    (* Compute the validation error. *)
    let test_accuracy =
      Tensor.(argmax (model test_images) = test_labels)
      |> Tensor.to_kind ~kind:(T Float)
      |> Tensor.sum
      |> Tensor.float_value
      |> fun sum -> sum /. test_samples
    printf "%d %f %.2f%%\n%!" index (Tensor.float_value loss) (100. *. test_accuracy);

Models and Weights

Various pre-trained computer vision models are implemented in the
vision library.
The weight files can be downloaded at the following links:

Running the pre-trained models on some sample images can the easily be done via the following commands.

dune exec examples/pretrained/predict.exe path/to/resnet18.ot tiger.jpg

Natural Language Processing models based on BERT can be found in the
ocaml-torch repo.

Alternative Installation Option

This alternative way to install ocaml-torch could be useful to run with GPU
acceleration enabled.

The libtorch library can be downloaded from the PyTorch
(1.10.0 cpu

Download and extract the libtorch library then to build all the examples run:

export LIBTORCH=/path/to/libtorch
git clone
cd ocaml-torch
make all
30 Oct 2021
Reverse Dependencies