ankane/torch.rb

Torch.rb – Deep Learning for Ruby

What it is

A Ruby gem that wraps Facebook’s LibTorch (the C++ core of PyTorch) so you can write and run PyTorch‑style deep‑learning code in Ruby.
Mirrors the PyTorch Python API closely, with Ruby‑idiomatic tweaks (e.g., add! for in‑place ops, tensor? for boolean checks, and Numo arrays instead of NumPy).

Key components

Tensor operations – creation (Torch.rand, Torch.zeros, etc.), arithmetic, indexing, conversion to/from Numo::NArray.
Autograd – automatic differentiation with requires_grad, backward, and gradient access via .grad.
Neural‑network module – define models by subclassing Torch::NN::Module, use layers like Conv2d, Linear, and functional helpers under Torch::NN::F.
Optimizers – e.g., Torch::Optim::SGD with typical zero_grad, step workflow.
Saving/loading – Torch.save / Torch.load for model state dictionaries (compatible with PyTorch files, with a small Python‑side conversion note).
Device support – CPU, CUDA GPUs (Torch::CUDA.available?, net.cuda), Apple‑silicon Metal (Torch::Backends::MPS).

Installation

Download a matching LibTorch build for your platform (CPU‑only or CUDA‑enabled). Example for macOS arm64:

curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.14.0.zip > libtorch.zip
unzip -q libtorch.zip

Tell the gem where LibTorch lives and add it to your Gemfile:

bundle config set build.torch-rb --with-torch-dir=/path/to/libtorch

gem "torch-rb"

Run bundle install. Compilation takes ~5‑10 minutes; Windows is not supported.

Getting started

Follow the 60‑minute blitz tutorial in tutorials/blitz/README.md.
Additional tutorials cover transfer learning, sequence models, and word embeddings.
Example projects include MNIST image classification, MovieLens collaborative filtering, and GANs.

Typical workflow (Ruby code)

# Define a simple CNN
class MyNet < Torch::NN::Module
  def initialize
    super()
    @conv1 = Torch::NN::Conv2d.new(1, 6, 3)
    @conv2 = Torch::NN::Conv2d.new(6, 16, 3)
    @fc1   = Torch::NN::Linear.new(16*6*6, 120)
    @fc2   = Torch::NN::Linear.new(120, 84)
    @fc3   = Torch::NN::Linear.new(84, 10)
  end

  def forward(x)
    x = Torch::NN::F.max_pool2d(Torch::NN::F.relu(@conv1.call(x)), [2,2])
    x = Torch::NN::F.max_pool2d(Torch::NN::F.relu(@conv2.call(x)), 2)
    x = Torch.flatten(x, 1)
    x = Torch::NN::F.relu(@fc1.call(x))
    x = Torch::NN::F.relu(@fc2.call(x))
    @fc3.call(x)
  end
end

net = MyNet.new
input = Torch.randn(1,1,32,32)
output = net.call(input)

criterion = Torch::NN::MSELoss.new
target = Torch.randn(10).view(1,-1)
loss = criterion.call(output, target)

optimizer = Torch::Optim::SGD.new(net.parameters, lr: 0.01)
optimizer.zero_grad
loss.backward
optimizer.step

Device handling

if Torch::CUDA.available?
  net.cuda               # move model to GPU
  input = input.cuda
end

Or on Apple silicon:

if Torch::Backends::MPS.available?
  device = Torch.device('mps')
  net.to(device)
end

Ecosystem

Companion gems for specific domains: torchvision-ruby, torchtext-ruby, torchaudio-ruby, torchcodec-ruby, torchrec-ruby, torchdata-ruby.
Higher‑level tools: transformers-ruby (transformer models) and safetensors-ruby (efficient tensor storage).

Why it matters

Enables Ruby developers to experiment with modern deep‑learning models without switching languages.
Leverages the same performant C++ backend used by PyTorch, so models run at native speed on CPU or GPU.
Keeps the Ruby community aligned with the rapidly evolving PyTorch ecosystem through a familiar API.

Resources

Build status badge shows CI testing on multiple platforms.
Detailed changelog, contribution guide, and scripts for GPU testing (AWS Deep Learning AMI) are provided.

Torch.rb brings the power of PyTorch to Ruby, offering a full‑featured, GPU‑accelerated deep‑learning stack that feels native to Ruby programmers.

ankane/torch.rb

Torch.rb – Deep Learning for Ruby

Related