Repository files navigation

TorchSharp.PyBridge is an extension library forTorchSharp, providing seamless interoperability between .NET and Python for model serialization. It simplifies the process of saving and loading PyTorch models in a .NET environment, enabling developers to easily develop models in both .NET and Python and transfer models easily.

• module.load_py(...),optim.load_py(...): Extension method for modules and optimizers for easily loading PyTorch models saved in the standard Python format (usingtorch.save) directly into TorchSharp.

  This only works for when thestate_dict was saved and not the whole model, see example below.

• module.save_py(...),optim.save_py(...): Extension method for modules and optimizers for easily saving TorchSharp models in a format that can be directly loaded in PyTorch (usingtorch.load), offering cross-platform model compatibility.

• module.load_safetensors(...),module.save_safetensors(...): Extension methods for modules for easily saving and loading model weights using thesafetensors format.

• module.load_checkpoint(...): Extension method for loading in a checkpoint (both safetensors and regular pytorch, including sharded models) from a directory saved using HuggingFace'sPreTrainedModel.save_pretrained() method.

TorchSharp.PyBridge is available on NuGet. You can install it using the following command:

dotnet add package TorchSharp.PyBridge

Install-Package TorchSharp.PyBridge

• .NET SDK
• TorchSharp library

Saving the model in Python:

importtorchmodel= ...torch.save(model.state_dict(),'path_to_your_model.pth')

Loading it in C#:

usingTorchSharp.PyBridge;varmodel= ...;model.load_py("path_to_your_model.pth");

To save a model in a format compatible with PyTorch:

usingTorchSharp.PyBridge;varmodel= ...;model.save_py("path_to_save_model.pth");

And loading it in in Python:

importtorchmodel= ...model.load_state_dict(torch.load('path_to_save_model.pth'))

Contributions to TorchSharp.PyBridge are welcome.

This project makes use of thepickle library, a Java and .NET implementation of Python's pickle serialization protocol, developed by Irmen de Jong. Thepickle library plays a vital role in enabling the serialization features within TorchSharp.PyBridge. We extend our thanks to the developer for their significant contributions to the open-source community. For more details about thepickle library, please visit theirGitHub repository.

For support, questions, or feedback, please open an issue in theGitHub repository.