/ML-Model-CI

MLModelCI is a complete platform for managing, converting, profiling, and deploying models as cloud services (MLaaS)

Primary LanguagePythonApache License 2.0Apache-2.0

...

Machine Learning Model CI

codebeat badge

中文简介FeaturesInstallationQuick StartNotebookTutorialContributingCitationLicense

Introduction

Machine Learning Model CI is a one-stop machine learning MLOps platform on clouds, aiming to solve the "last mile" problem between model training and model serving. We implement a highly automated pipeline between the trained models and the online machine learning applications.

We offer the following features and users 1) can register models to our system and enjoy the automated pipeline, 2) or use them individually.

  • Housekeeper provides a refined management for model (service) registration, deletion, update and selection.
  • Converter is designed to convert models to serialized and optimized formats so that the models can be deployed to cloud. Support Tensorflow SavedModel, ONNX, TorchScript, TensorRT
  • Profiler simulates the real service behavior by invoking a gRPC client and a model service, and provides a detailed report about model runtime performance (e.g. P99-latency and throughput) in production environment.
  • Dispatcher launches a serving system to load a model in a containerized manner and dispatches the MLaaS to a device. Support Tensorflow Serving, Trion Inference Serving, ONNX runtime, Web Framework (e.g., FastAPI)
  • Controller receives data from the monitor and node exporter, and controls the whole workflow of our system.

Several features are in beta testing and will be available in the next release soon. You are welcome to discuss them with us in the issues.

  • Automatic model quantization and pruning.
  • Model visulization and fine-tune.

The system is currently under rapid iterative development. Some APIs or CLIs may be broken. Please go to Wiki for more details

If your want to join in our development team, please contact huaizhen001 @ e.ntu.edu.sg

Demo

The below figures illusrates the web interface of our system and overall workflow.

Web frontend Workflow
drawing drawing

Installation Guide

Prerequisites

  • A GNU/Linux environment(Ubuntu preferred)
  • Docker
  • Docker Compose (Optional, for Installation via Docker)
  • TVM and tvm Python module(Optional)
  • TensorRT and Python API(Optional)
  • Python >= 3.7

Installation using pip

# install modelci from GitHub
pip install git+https://github.com/cap-ntu/ML-Model-CI.git@master

Once you have installed, make sure the docker daemon is running, then you can start MLModelCI service on a leader server by:

modelci service init

CLI start service

Or stop the service by:

modelci service stop

CLI stop service

Installation using Docker

For CPU-only Machines

docker pull mlmodelci/mlmodelci:cpu

Start basic services by Docker Compose:

docker-compose -f ML-Model-CI/docker/docker-compose-cpu-modelhub.yml up -d

Stop the services by:

docker-compose -f ML-Model-CI/docker/docker-compose-cpu-modelhub.yml down

For CUDA10.2 Machine

docker pull mlmodelci/mlmodelci:cuda10.2-cudnn8

Start basic services by Docker Compose:

docker-compose -f ML-Model-CI/docker/docker-compose-gpu-modelhub.yml up -d

docker-compose start service

Stop the services by:

docker-compose -f ML-Model-CI/docker/docker-compose-gpu-modelhub.yml down

docker-compose stop service

Usage

We provide three options for users to use MLModelCI: CLI, Running Programmatically and Web interface

1. CLI

# publish a model to the system
modelci@modelci-PC:~$ modelci modelhub publish -f example/resnet50.yml
{'data': {'id': ['60746e4bc3d5598e0e7a786d']}, 'status': True}

Please refer to WIKI for more CLI options.

2. Running Programmatically

# utilize the convert function
from modelci.hub.converter import convert
from modelci.types.bo import IOShape

# the system can trigger the function automaticlly
# users can call the function individually 
convert(
    '<torch model>',
    src_framework='pytorch', 
    dst_framework='onnx',
    save_path='<path to export onnx model>', 
    inputs=[IOShape([-1, 3, 224, 224], dtype=float)], 
    outputs=[IOShape([-1, 1000], dtype=float)], 
    opset=11)

3. Web Interface

If you have installed MLModelCI via pip, you should start the frontend service manually.

# Navigate to the frontend folder
cd frontend
# Install dependencies
yarn install
# Start the frontend
yarn start

The frontend will start on http://localhost:3333

Quickstart with Notebook

Tutorial

After the Quick Start, we provide detailed tutorials for users to understand our system.

Contributing

MLModelCI welcomes your contributions! Please refer to here to get start.

Citation

If you use MLModelCI in your work or use any functions published in MLModelCI, we would appreciate if you could cite:

@inproceedings{10.1145/3394171.3414535,
  author = {Zhang, Huaizheng and Li, Yuanming and Huang, Yizheng and Wen, Yonggang and Yin, Jianxiong and Guan, Kyle},
  title = {MLModelCI: An Automatic Cloud Platform for Efficient MLaaS},
  year = {2020},
  url = {https://doi.org/10.1145/3394171.3414535},
  doi = {10.1145/3394171.3414535},
  booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},
  pages = {4453–4456},
  numpages = {4},
  location = {Seattle, WA, USA},
  series = {MM '20}
}

Contact

Please feel free to contact our team if you meet any problem when using this source code. We are glad to upgrade the code meet to your requirements if it is reasonable.

We also open to collaboration based on this elementary system and research idea.

huaizhen001 AT e.ntu.edu.sg

License

   Copyright 2021 Nanyang Technological University, Singapore

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.