package torch

ocaml-torch

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 differentiation.

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 v1.13.0.

On Linux note that you will need the PyTorch version using the appropriate cxx11 abi depending on your g++ version. cpu version, cuda 11.6 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 ocaml-jupyter.

Here is a sample utop session.

Build a Simple Example

To build a first torch program, create a file example.ml 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:

(executables
  (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).

Tutorials and Examples

• MNIST tutorial.

• Finetuning a ResNet-18 model.

• Generative Adversarial Networks.

• Running some Python model.

Some more advanced applications from external repos:

• An OCaml port of mini-dalle by Arulselvan Madhavan.

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

Sample Code

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

  (* 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
    in

    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 ~dim:(-1) (model test_images) = test_labels)
      |> Tensor.to_kind ~kind:(T Float)
      |> Tensor.sum
      |> Tensor.float_value
      |> fun sum -> sum /. test_samples
    in
    printf "%d %f %.2f%%\n%!" index (Tensor.float_value loss) (100. *. test_accuracy);
  done

• Some ResNet examples on CIFAR-10.

• A simplified version of char-rnn illustrating character level language modeling using Recurrent Neural Networks.

• Neural Style Transfer applies the style of an image to the content of another image. This uses some deep Convolutional Neural Network.

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:

• ResNet-18 weights.

• ResNet-34 weights.

• ResNet-50 weights.

• ResNet-101 weights.

• ResNet-152 weights.

• DenseNet-121 weights.

• DenseNet-161 weights.

• DenseNet-169 weights.

• SqueezeNet 1.0 weights.

• SqueezeNet 1.1 weights.

• VGG-13 weights.

• VGG-16 weights.

• AlexNet weights.

• Inception-v3 weights.

• MobileNet-v2 weights.

• EfficientNet b0 weights, b1 weights, b2 weights, b3 weights, b4 weights.

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 images/tiger.jpg

Acknowledgements

Many thanks to @LaurentMazare for the original work of ocaml-torch.