Skip to content

Home

Jump to bottom
Ed Connell edited this page Apr 28, 2020 · 1 revision

Overview

SwiftRT is a computational framework research project written almost entirely in the Swift language with a small amount of C used to interface with system libraries and Cuda. The project goals are:

  • Simplify the model development process so engineers are able to successfully create and deploy models directly in their Swift applications
  • Develop a faster execution model to reduce training time and speed model design iteration
  • Leverage Google's Swift for TensorFlow Auto Differentiation support

Installation

  1. Install the latest Swift for TensorFlow toolchain for your platform.

  2. Clone the SwiftRT repository

git clone https://github.com/ewconnell/swiftrt.git

Setup for MacOS and Xcode

After downloading the toolchain, open the package and follow the installation instructions.

  • start Xcode and go to the menu Xcode/Preferences/Components and select the new toolchain
  • run the unit tests to verify that the installation is valid by pressing command + u

Setup for Linux and Cuda

  1. Install Cuda 10.2 for Ubuntu
  • First make sure your graphics card driver is up to date!
  • Then install using the Base Installer instructions
  1. Install cuDNN 7.4
  • First visit the NVIDIA cuDNN download site and register
  • Download "cuDNN v7.6.5 Library for Linux" for Cuda 10.2
  • Then install
sudo tar -xzf cudnn-10.0-linux-x64-v7.4.1.5.tgz -C /usr/local
rm cudnn-10.0-linux-x64-v7.4.1.5.tgz
sudo ldconfig
  1. Install SwiftRT dependencies (TODO update this)
sudo apt-get install
Clone this wiki locally