Skip to content

Latest commit

 

History

History
64 lines (43 loc) · 1.71 KB

package_description.rst

File metadata and controls

64 lines (43 loc) · 1.71 KB

primitiv: A Neural Network Toolkit. (Python frontend)

Features

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

Install

Prerequisites:

  • Python 3 (3.5 or later)
  • NumPy (1.11.0 or later)
  • Cython (0.27 or later)
  • CMake (3.1.0 or later)
  • scikit-build (0.6.1 or later)
  • (optional) CUDA (7.5 or later)
  • (optional) OpenCL (1.2 or later) and OpenCL C++ binding v2

Install required packages:

pip3 install numpy cython cmake scikit-build

Build and install primitiv without CUDA and OpenCL:

pip3 install primitiv

Build and install primitiv with CUDA and/or OpenCL support:

# Enable only CUDA
pip3 install primitiv --global-option --enable-cuda

# Enable both CUDA and OpenCL
pip3 install primitiv --global-option --enable-cuda --global-option --enable-opencl

--enable-eigen flag that enables Eigen backend is added by default in the package contained in PyPI. To disable the Eigen backend, use --disable-eigen flag. Note that Eigen is bundled with the package contained in PyPI.

Notes

We are providing only a source pacakge for now, and pip command downloads the source package and builds it before installing. This is mainly because of keeping compatibility with the manylinux1 standard described in PEP 513 while maintaining supports of non-standard backends such as CUDA/OpenCL.

Resources