sherlock

Sherlock is an output range analysis tool for deep neural networks. The only library that is needed is the MILP solver Gurobi. It's free for academic purposes and can be downloaded from here : http://www.gurobi.com/resources/getting-started/mip-basics

Instructions to Compile

Please modify the file Makefile.locale to help us find Gurobi.

For a Mac with the latest gurobi10.0 installed your likely settings will be:

HOST_ARCH=macos_universal2

GUROBI_PATH=/Library/gurobi1000

For a linux box, your settings will be: (Not tested yet)

ARCH=linux64 # if you are using a linux box

GUROBI_PATH=/opt/gurobi1000

You should feel free to modify these two variables. The Makefile will look for Gurobi headers under

$(GUROBI\_PATH)/$(HOST_ARCH)/include

and libraries under

$(GUROBI\_PATH)/$(HOST_ARCH)/include

Instructions to Install Protobuf Compiler

Install Protocol buffer compiler, using the following commands :

Once these are set, you should type

make

to compile. It should work out of the box.

Instructions to run

Please use the command line run_file to run all the benchmarks.

./run_file

Enjoy the output as it scrolls on your screen :-)

Instructions to test against randomly generated pytorch networks

Sherlock can read networks in the ONNX format. Right now the test script generates fixed architecture networks with random weights.

bash test_script.sh