import torch
  class MyModule(torch.nn.Module):

    def __init__(self, N, M):
      super(MyModule, self).__init__()
      self.weight = torch.nn.Parameter(torch.rand(N, M))

    def forward(self, input):
      if input.sum() > 0:
        output = self.weight.mv(input)
      else:
        output = self.weight + input
      return output

    # Compile the model code to a static representation
    my_script_module = torch.jit.script(MyModule(3, 4))

    # Save the compiled code and model data so it can be loaded elsewhere
    my_script_module.save("my_script_module.pt")

TorchScript

With TorchScript, PyTorch provides ease-of-use and flexibility in eager mode, while seamlessly transitioning to graph mode for speed, optimization, and functionality in C++ runtime environments.

Distributed Training

Optimize performance in both research and production by taking advantage of native support for asynchronous execution of collective operations and peer-to-peer communication that is accessible from Python and C++.

  import torch.distributed as dist
  from torch.nn.parallel import DistributedDataParallel
  
  dist.init_process_group(backend='gloo')
  model = DistributedDataParallel(model)

  ## Save your model
  torch.jit.script(model).save("my_mobile_model.pt")

  ## iOS prebuilt binary
  pod ‘LibTorch’
  ## Android prebuilt binary
  implementation 'org.pytorch:pytorch_android:1.3.0'

  ## Run your model (Android example)
  Tensor input = Tensor.fromBlob(data, new long[]{1, data.length});
  IValue output = module.forward(IValue.tensor(input));
  float[] scores = output.getTensor().getDataAsFloatArray();

Mobile (Experimental)

PyTorch supports an end-to-end workflow from Python to deployment on iOS and Android. It extends the PyTorch API to cover common preprocessing and integration tasks needed for incorporating ML in mobile applications.

Tools & Libraries

An active community of researchers and developers have built a rich ecosystem of tools and libraries for extending PyTorch and supporting development in areas from computer vision to reinforcement learning.

  import torchvision.models as models
  resnet18 = models.resnet18(pretrained=True)
  alexnet = models.alexnet(pretrained=True)
  squeezenet = models.squeezenet1_0(pretrained=True)
  vgg16 = models.vgg16(pretrained=True)
  densenet = models.densenet161(pretrained=True)
  inception = models.inception_v3(pretrained=True)

  import torch.onnx
  import torchvision

  dummy_input = torch.randn(1, 3, 224, 224)
  model = torchvision.models.alexnet(pretrained=True)
  torch.onnx.export(model, dummy_input, "alexnet.onnx")

Native ONNX Support

Export models in the standard ONNX (Open Neural Network Exchange) format for direct access to ONNX-compatible platforms, runtimes, visualizers, and more.

C++ Front-End

The C++ frontend is a pure C++ interface to PyTorch that follows the design and architecture of the established Python frontend. It is intended to enable research in high performance, low latency and bare metal C++ applications.

  #include <torch/torch.h>

  torch::nn::Linear model(num_features, 1);
  torch::optim::SGD optimizer(model->parameters());
  auto data_loader = torch::data::data_loader(dataset);

  for (size_t epoch = 0; epoch < 10; ++epoch) {
    for (auto batch : data_loader) {
      auto prediction = model->forward(batch.data);
      auto loss = loss_function(prediction, batch.target);
      loss.backward();
      optimizer.step();
    }
  }

  export IMAGE_FAMILY="pytorch-latest-cpu"
  export ZONE="us-west1-b"
  export INSTANCE_NAME="my-instance"
  
  gcloud compute instances create $INSTANCE_NAME \
    --zone=$ZONE \
    --image-family=$IMAGE_FAMILY \
    --image-project=deeplearning-platform-release

Cloud Partners

PyTorch is well supported on major cloud platforms, providing frictionless development and easy scaling through prebuilt images, large scale training on GPUs, ability to run models in a production scale environment, and more.