Overview

This is the artifact for ChamleonAPI: Automatic and Efficient Customization of Neural Networks for ML Applications (to appear at OSDI 24).

This artifact contains the code and trained model weights required to reproduce the key results in the paper (you may find it at the HotCRP website).

The installation of required packages can be found at install.md, and for reproducing the results, you may find it at scripts/eval.md.

Structure of the repo

scripts/: This contains the scripts to plot Table 3.

image_classification/ : This contains the scripts to run evaluation for the image classification apps.

object_cc/ : This contains the scripts to run evaluation for the object detection apps.

nlp/ : This contains the scripts to run evaluation for the text topic classification apps.

Installation

Go to install.md

Kick-the-tire experiments

Please check if you are able to download ChameleonAPI (and baseline) models in the links below:

wget https://s3-us-east-2.amazonaws.com/chameleonapi/image_classification_models.zip

wget https://s3-us-east-2.amazonaws.com/chameleonapi/object_detection_models.zip

wget https://s3-us-east-2.amazonaws.com/chameleonapi/nlp_models.zip

To run a "hello-world" experiment, first, make sure the environment is installed: We provide two ways to install the packages,

Docker image

See install_with_docker.md.

Install with Miniconda

See install.md.

Then, run the following code:

bash scripts/run.sh

Reproducing the main results

Details in here

Contact

Yuhan Liu ([email protected])