Custom PyTorch C++ Kernel

• Compiled for Python 3.6.x
• Custom functions written in C++ v14, using torch v1.x C++ API
• Functions bound to a Python module using pybind11
• Uses setuptools for building, and pytest for testing
• Implements in C++ a custom LSTM-like layer (forward and backward passes), and plugs it in a regression model, based on these tutorials:
  • https://pytorch.org/tutorials/advanced/cpp_extension.html (PyTorch C++ kernel)
  • https://machinelearningmastery.com/building-a-regression-model-in-pytorch/ (PyTorch regression model)

To compile, run

pip install -r requirements.txt
python setup.py build

This will also install the module torch_kernel in the Python system path.

To run the unit tests, run

pytest

To import the module, import first torch

import torch
import torch_kernel

To train a sample regressor neural network (one LSTM-like layer followed by a linear layer) run:

python applications/custom_nn.py

This script contains PyTorch classes wrapping up the custom layer. It also shows how to run stochastic gradient descent to update the network weights, and solve a sample Scikit-learn regression problem.