A collection of Libtorch Library based implementations with corresponding Pytorch (Python) based implementation.
Models can be prototyped in Python and exported to JIT and trained in C++ with CUDA or CPU - Multiprocessing training has been implemented with MPI and currently works only on CPU.
TrainerBase is a generic class in C++ (Trainer/TrainerBase.hpp) that can work with jit::Module and nn::Module in order to perform training and inference.
Python's utility in TorchScriptUtilities/ExportJIT/ can be used to export the JIT model. Once can prototype the model before the export. The exported model then can be trained with C++ Trainer. Additionally TrainerBase C++ can be inherited if one wishes to add
additional features as per their use-case. Trainer supports CUDA training and inference.
Toy Implementations included:
- CIFAR: Simple implementation for CIFAR dataset, includes files for dataset preparation and models.
libtorchonnxruntimeboost/logboost/mpiopencvopenmpiopenmp
To Build with CUDA
mkdir -p LibTorchPlayground-build
cd LibTorchPlayground-build
cmake -DCMAKE_PREFIX_PATH="/path/to/libtorch" -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_CXX_COMPILER=mpicxx -DCMAKE_CXX_STANDARD_REQUIRED=ON -Wl,--whole-archive on libc10_cuda.a ../LibTorchPlayground/
cmake --build . --config Release
make
To Build with CPU
mkdir -p LibTorchPlayground-build
cd LibTorchPlayground-build
cmake -DCMAKE_PREFIX_PATH="/path/to/libtorch" -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_CXX_COMPILER=mpicxx -DCMAKE_CXX_STANDARD_REQUIRED=ON ../LibTorchPlayground/
cmake --build . --config Release
make
You may need to pass path to onnxruntime build via -DCMAKE_PREFIX_PATH if it's not installed via package managers.
You can skip adding argument -DCMAKE_PREFIX_PATH for cmake if you have libtorch, onnxruntime and other packages installed via package manager (homebrew for example).
To run with mpi with 4 processes
cd LibTorchPlayground-build
mpirun -n 4 LibTorchPlayground