cc

RECom C++ Examples

This directory contains examples of RECom using TensorFlow C++. You can use these scripts to reproduce the evaluation results in our paper.

You can follow the subsequent instructions to run the C++ examples.

How to build

To run TensorFlow C++ examples, you should first check out the TensorFlow submodule:

git submodule update --init --recursive

Second, apply the RECom example patch to the TensorFlow submodule:

./apply_patch.sh

Third, install the necessary Python packages:

pip install numpy==1.19
pip install keras_preprocessing --no-deps

Then, you can run the configure script in the TensorFlow repository to launch a configuration session. You should enter 'y' for the CUDA support only and specify the Python and CUDA paths according to your systems.

cd tensorflow-v2.6.2 
./configure

You can set the environment variables before running the configuration script to skip the interactive Q&A.

export TF_CONFIGURE_IOS=False
export TF_SET_ANDROID_WORKSPACE=False
export TF_NEED_ROCM=0
export TF_CUDA_CLANG=0
export TF_NEED_TENSORRT=0
export GCC_HOST_COMPILER_PATH=/usr/bin/gcc
export CC_OPT_FLAGS=-Wno-sign-compare
export PYTHON_BIN_PATH=/usr/bin/python
export USE_DEFAULT_PYTHON_LIB_PATH=1
export TF_CUDA_COMPUTE_CAPABILITIES=7.5,8.6

Finally, build the targets:

bazel build --config=cuda --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" //tensorflow/recom:benchmark_multi_thread

More details about how to run TensorFlow C++ examples can be found in TensorFlow - Build from source.

Run the examples

After building, you can find the executable binaries in bazel-bin/tensorflow/recom under the TensorFlow directory. Use -h or --help to show the usage (we use cxxopts to parse the arguments).

You can run the Python examples in this repository to create TensorFlow saved models:

python ../../python/microbenchmark.py -N 1000 --save_path microbenchmark_1000

Then, execute the binaries of C++ examples:

# TF-GPU
./bazel-bin/tensorflow/recom/benchmark_multi_thread --model_path microbenchmark_1000 --batch_size 256 --serve_workers 1

# TF-CPU
./bazel-bin/tensorflow/recom/benchmark_multi_thread --model_path microbenchmark_1000 --batch_size 256 --serve_workers 1 --disable_gpu

# RECom
./bazel-bin/tensorflow/recom/benchmark_multi_thread --model_path microbenchmark_1000 --batch_size 256 --serve_workers 1 --lib_path ../../../bazel-bin/tensorflow_addons/librecom.so