A Neural Network Toolkit.

• Dynamic and incremental graph construction
• On-demand memory allocation
• Automatic minibatch broadcasting
• Mostly device-independent
• Simple usage

• CMake 3.1.0 or later
• GCC 4.8 or later
• Protocol Buffers
• (optional) Google Test
  • Required only when -DPRIMITIV_BUILD_TESTS=ON.
• (optional) CUDA 7.5 or later
  • Required only when -DPRIMITIV_USE_CUDA=ON

The default protobuf repository on Ubuntu 16.04 does not support proto3 format and users need to install newer library from source. Typical step to build/install protobuf is below:

sudo apt install autoconf automake build-essential cmake libtool unzip
git clone https://github.com/google/protobuf
cd protobuf
./autogen.sh
./configure
make [-j <threads>]
make check
sudo make install
sudo ldconfig

git clone <this repository>
cd primitiv
mkdir build
cd build
cmake .. [-DPRIMITIV_USE_CUDA=ON] [(Other options listed below if necessary)]
make [-j <threads>]
[make test]
[make install]

• PRIMITIV_BUILD_STATIC_LIBRARY (default=OFF)
  • Builds a static library instead of a shared object.
• PRIMITIV_BUILD_TESTS (default=OFF)
  • Builds test binaries and generates make test command.
• PRIMITIV_BUILD_TESTS_PROBABILISTIC (default=OFF)
  • Builds test cases that probabilistically fails.
• PRIMITIV_GTEST_SOURCE_DIR (default="")
  • Specifies the source directory of Google Test. If you want to use googletest module provided from Debian/Ubuntu repository, add -DPRIMITIV_GTEST_SOURCE_DIR=/usr/src/googletest/googletest together with -PRIMITIV_BUILD_TESTS=ON option.
• PRIMITIV_USE_CACHE (default=OFF)
  • Whether or not to use cached values to prevent increasing computation amount.
  • Libraries built with this flag will tend to consume more memory.
• PRIMITIV_USE_CUDA (default=OFF)
  • Enables CUDA backend (devices::CUDA class).
• Other available options:
  • CMake standard options.
  • FindCUDA options.
  • FindProtobuf options.
  • FindGTest options.

• Short Python tutorial with solving XOR problem.
• Other examples.

• yus.takara at gmail.com
• @odashi_t on Twitter

This project is supported by ASTREC in NICT.