CorrProxies

Declaration

This repo is for paper: Optimizing Machine Learning Inference Queries with Correlative Proxy Models.

Setup ENV

Quick Start

  1. We provide a fully ready Docker Image ready to use out-of-box.
  2. Optionally, you can also follow the steps to build your own testing environment.

The Provided Docker Environment

Steps to run the Docker Environment

  • Get the docker image from this link.
  • Load the docker image.
    docker load -i corrproxies-image.tar
  • Run the docker image in a container.
    docker run --name=CorrProxies -i -t -d corrproxies-image

    • it will return you the docker container ID, for example d979af9a17f23345cb2894b22dc8527680acdfd7a7e1aaed6a7a28ea134e66e6.
  • Use CLI to control the container with the specific ID generated.
    docker exec -it d979af9a17f23345cb2894b22dc8527680acdfd7a7e1aaed6a7a28ea134e66e6 /bin/zsh

ENV Spec

File structure:

  • The home directory for CorrProxies locates at /home/CorrProxies.
  • The Python executable locates at /home/anaconda3/envs/condaenv/bin/python3.
  • The models locate at /home/CorrProxies/model.
  • The datasets locate at /home/CorrProxies/data.
  • The starting scripts locate at /home/CorrProxies/scripts.

Build Your Own Environment

This instruction is based on a clean distribution of [email protected]

  1. Install pre-requisites.

    apt-get update && apt-get install -y build-essential

  2. Install Anaconda.

    • wget https://repo.anaconda.com/archive/Anaconda3-5.3.1-Linux-x86_64.sh && bash Anaconda3-5.3.1-Linux-x86_64.sh -b -p <AnacondaHome>
    • export PATH="<AnacondaHome>/bin/:$PATH"

  3. Install [email protected] with Anaconda3.

    conda create -n condaenv python=3.6.6

  4. Activate the newly installed Python ENV.

    conda activate condaenv

  5. Install dependencies with pip.

    pip3 install -r requirements.txt

  6. Install Java (openjdk-8) (for standford-nlp usage).

    apt-get install -y openjdk-8-jdk

Queries & Datasets

  • We use Twitter text dataset, COCO image dataset and UCF101 video dataset as our benchmark datasets.
    Please see this page for examples of detailed Queries and Datasets examples we use in our experiments.

  • After you setup the environment, either manually or using the docker image provided by us, the next step is to download the datasets.

    • To get the COCO dataset:
      cd /home/CorrProxies/data/image/coco && ./get_coco_dataset.sh
    • To get the UCF101 dataset:
      cd /home/CorrProxies/data/video/ucf101 && wget -c https://www.crcv.ucf.edu/data/UCF101/UCF101.rar && unrar x UCF101.rar.

Execution

Please pull the latest code before executing the code. Command cd /home/CorrProxies && git pull

Run Operators Individually

To run and see each operator we used in our experiment, simply execute python3 <operator_name.py>. For example: python3 operators/ml_operators/image_video_operators/video_activity_recognition.py.

Run Experiments

We use scripts/run.sh to start experiments. The script will take in command line arguments.

  • Text(Twitter)

    • Since we do not provide text dataset, we will skip the experiment.

  • Image(COCO)

    Example: ./scripts/run.sh -w 2 -t 1 -i '1' -a 0.9 -s 3 -o 2 -e 1

  • Video(UCF101)

    Example: ./scripts/run.sh -w 2 -t 2 -i '1' -a 0.9 -s 3 -o 2 -e 1

  • arguments detail.

    • w int: experiment type in [1, 2, 3, 4] referring to /home/CorrProxies/ml_workflow/exps/WorkflowExp*.py;
    • t int: query type in [0, 1, 2]. Int 0, 1, 2 means queries on the Twitter, COCO, and UCF101 datasets, respectively;
    • i int: query index in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10];
    • a float: query accuracy;
    • s int: scheme in [0, 1, 2, 3, 4, 5, 6]. Int 0, 1, 2, 3, 4, 5, 6 means 'ORIG', 'NS', 'PP', 'CORE', 'COREa', 'COREh' and 'REORDER' schemes, respectively;
    • o int: number of threads used in optimization phase;
    • e int: number of threads used in execution phase after generating an optimized plan.

GitHub

View Github