AE

RECom Artifact Evaluation

The artifact contains the necessary software components to validate the main results in the RECom paper. We provide a Dockfile for users to build the docker image, which contains the basic environment used to build and run the RECom examples.

Run with a single command

We provide a single script to run RECom examples and reproduce Figures 10 and 11 in our paper. Note: if the compute capability of your GPU is not 7.5 or 8.6, you should modify the Dockerfile correspondingly.

./run_all.sh

Run step by step

You can also run RECom examples step by step.

Firstly, use the Dockerfile to build the docker image.

docker build -f ./Dockerfile -t recom:latest .

Then you can launch the container:

docker run -d --gpus all --net=host --name recom_ae -it recom:latest
docker exec -it recom_ae bash

After launching the container, clone this repo:

git clone --recurse-submodules https://github.com/AlibabaResearch/recom.git recom

Then you can run a single script to perform the following steps:

1. Create the synthetic models E and F used in the paper.
2. Build the RECom addon.
3. Build the TensorFlow C++ examples.
4. Measure the inference latency of RECom and the TensorFlow baselines.
5. Draw the most important figures in the paper (Figures 10 and 11).

python recom/AE/build_and_run.py

Finally, you can open the generated figures to check the results.