Questions tagged [libtorch]

The PyTorch C++ frontend provides a high level, pure C++ modeling interface for neural network and general machine learning research and production use cases, largely following the Python API in design and provided functionality. The C++ frontend includes the following:

• An interface for defining machine learning models through a hierarchical module system (like torch.nn.Module);

• A “standard library” of pre-existing modules for the most common modeling purposes (e.g. convolutions, RNNs, batch normalization etc.);

• An optimization API, including implementations of popular optimizers such as SGD, Adam, RMSprop, and others;

• A means of representing datasets and data pipelines, including functionality to load data in parallel over many CPU cores;

• A serialization format for storing and loading checkpoints of a training session (like torch.utils.data.DataLoader);

• Automatic parallelization of models onto multiple GPUs (like torch.nn.parallel.DataParallel);

• Support code to easily bind C++ models into Python using pybind11;

• Entry points to the TorchScript JIT compiler;

• Helpful utilities to facilitate interfacing with the Aten and Autograd APIs.