Installation

Getting Code

If retrieving the most up-to-date version of the code, clone the repository with

git clone --recursive https://github.com/darchr/AutoTM
export AUTOTM_HOME=$(pwd)/AutoTM

If retrieving code via DOI, run

wget https://zenodo.org/record/3612698/files/autotm.tar.gz
tar -xvf autotm.tar.gz
export AUTOTM_HOME=$(pwd)/AutoTM

Setup

A simple setup needs to be performed to indicate how the project will be used. To enter the setup, run

cd $AUTOTM_HOME
julia --color=yes setup.jl

The following selections can be made - choose which are appropriate for your system:

Use NVDIMMs in 1LM (requires a Cascade Lake system with Optane DC NVDIMMs)
Use of a GPU (requires CUDA 10.1 or CUDA 10.2)
Use Gurobi as the ILP solver (requires a Gurobi license (see below)). If Gurobi is not selected, the open source Cbc solver will be used. Please note that the original experiments were run with Gurobi.

Building

Launch Julia from the AutoTM project

cd $AUTOTM_HOME/AutoTM
julia --project

In the Julia REPL, press ] to switch to package (pkg) mode and run following commands:

julia> ]
(AutoTM) pkg> instantiate
(AutoTM) pkg> build -v

This will trigger the build process for our custom version of ngraph. Passing the -v command to build will helpfully display any errors that occur during the build process.

Using the Gurobi ILP solver (optional)

The results in the AutoTM paper use Gurobi for the ILP solver. However, Gurobi requires a license to run. Free trial and academic licenses are available from the Gurobi website: https://www.gurobi.com

If using Gurobi, please obtain a license and install the software according the instructions on the website.

Then, when building the project, make sure to run

julia> ENV["GUROBI_HOME"] = "path/to/gurobi"

in Julia before executing the build step above.

Note

Using the Gurobi ILP solver is optional. If not selected during the setup step, an open-source solver Cbc will be used. However, since Cbc is considerably less powerful than Gurobi, larger ILP models will likely not be solvable in a reasonable period of time.

Acceptable Build Warnings

CMake warnings regarding variable TBB_ROOT can be ignored. The version of ngraph used for this project does not use any of the TBB based code but does not build with TBB disabled.