QimingChen / cpp-torch

cpp-torch is a C++ library, implemented as a wrapper around Torch libraries.

Introduction

cpp-torch is a C++ library, implemented as a wrapper around torch C libraries (not lua libraries).

Using this library, you can:

• load torch data tensor from .t7 file
• load torch network model from .t7 file
• feed data into model, perform forward pass and get output

All in C++, without touching lua.

Pretty handy when you want to deploy an off-the-shelf torch model.

Install

Check our install script for Linux, Windows(TODO) and MacOS(TODO).

Install cpp-torch on Linux

The following commands install our C++ wrapper: cpp-torch in default torch install dir:

git clone https://github.com/yytdfc/cpp-torch
cd cpp-torch
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=~/torch/install -DCMAKE_PREFIX_PATH=~/torch/install ..
# For GPU version, set -DBUILD_CUDA=ON:
cmake  -DBUILD_CUDA=ON -DCMAKE_INSTALL_PREFIX=~/torch/install -DCMAKE_PREFIX_PATH=~/torch/install ..
make
make install
cd ..

Get started

The following code loads a float tensor and a float network from file, and forwards the tensor into the network:

// read input tensor
std::ifstream fs_input("input_tensor.t7", std::ios::binary);
auto obj_input = cpptorch::load(fs_input);
auto input = cpptorch::read_tensor<float>(obj_input.get());     // load float tensor

// read network
std::ifstream fs_net("net.t7", std::ios::binary);
auto obj_net = cpptorch::load(fs_net);
auto net = cpptorch::read_net<float>(obj_net.get());          // load float network

// forward
auto output = net->forward(input);

// display
std::cout << input << std::endl;
std::cout << *net << std::endl;
std::cout << output << std::endl;

If tensor and network type is double, change the template type accordingly:

auto input = cpptorch::read_tensor<double>(obj_input.get());     // load double tensor
auto net = cpptorch::read_tensor<double>(obj_net.get());     // load double network

To use GPU, use read_cuda_tensor() function:

auto input = cpptorch::read_cuda_tensor(obj_input.get());     // load cuda tensor
auto net = cpptorch::read_cuda_net(obj_net.get());          // load cuda network

We also provides an example script to test the famous CMU OpenFace model. This example transfers a 3 * 96 * 96 face image(OpenCV Mat) into cpp-torch tensor, forwards it through the network, receives a 128 * 1 identity feature tensor, and print the result.

Progress

Currently, this library supports forward pass of

• some modules in nn package
• related functions in torch7 package
• a few modules in dpnn package.

Check this list to see supported modules.

You are more than welcome to add new modules to cpp-torch. Please check our developer guide.

FAQ

-- How can I train my own model with this wrapper?

-- We don't support backward functions, so training is impossible. Use the original torch.