Home

Build ngraph-he

1. Build HE Transformer without TF

• Create build directory

cd /home/yixing/repo/he-transformer/
mkdir build
cd build

• To build without ngraph-tf, we can build nGraph automatically

cmake ..
make -j

• Run C++ Native example: in build

./test/unit-test --gtest_filter="HE_SEAL.ab"
./test/unit-test --gtest_filter="HE_HEAAN.ab"

2. Build HE Transformer with TF

• Crate python virtualenv

virtualenv /home/yixing/repo/venvs/he3 -p python3
source ~/repo/venvs/he3/bin/activate

• Install tensorflow-1.9 pip inside virtualenv

cd ~/repo/ngraph-tf
pip install tensorflow-1.9.0-cp35-cp35m-linux_x86_64.whl

• Build ngraph-tf

cd ~/repo/ngraph-tf
mkdir build
cd build
cmake ..
make -j

#  Install ngraph-tf
make install

# Inside the virtualenv
pip install python/dist/ngraph-0.5.0-py2.py3-none-linux_x86_64.whl

# Try importing -- this shouldn't crash
python -c "import ngraph"

• To config with ngraph-tf First we need to clean up our original he-transformer build

cd /home/yixing/repo/he-transformer
rm -rf build
mkdir build
cd build

Next we’ll configure he-transformer with the ngrpah-tf

cmake .. -DCMAKE_BUILD_TYPE=Debug -DPREBUILD_NGRAPH_INCLUDE_DIR=/home/yixing/repo/ngraph-tf/build/ngraph/ngraph_dist/include -DPREBUILD_NGRAPH_LIB_DIR=/home/yixing/repo/venvs/he3/lib/python3.5/site-packages/ngraph # <- missing "h" here

make -j

# this make install will install libhe_backend.so into the python site-packages
make install

• Run TF python model

# source activate virtualenv first
source ~/repo/venvs/he3/bin/activate
cd /home/yixing/repo/ngraph-tf/examples

# run on ngraph-CPU
python axpy.py

# run on SEAL
LD_LIBRARY_PATH=$HOME/repo/venvs/he3/lib/python3.5/site-packages/ngraph NGRAPH_TF_BACKEND=HE_SEAL NGRAPH_ENABLE_SERIALIZE=1 python axpy.py

# run on HEAAN
LD_LIBRARY_PATH=$HOME/repo/venvs/he3/lib/python3.5/site-packages/ngraph NGRAPH_TF_BACKEND=HE_HEAAN NGRAPH_ENABLE_SERIALIZE=1 python axpy.py