alaas 0.2.1

A Distributed MLOps System for Efficient Active Learning

ALaaS: Active Learning as a Service.

Active Learning as a Service (ALaaS) is a fast and scalable framework for automatically selecting a subset to be labeled from a full dataset so to reduce labeling cost. It provides a out-of-the-box and standalone experience for users to quickly utilize active learning.

ALaaS is featured for

• :hatching_chick: Easy-to-use With <10 lines of code to start the system to employ active learning.
• :rocket: Fast Use the stage-level parallellism to achieve over 10x speedup than under-optimized active learning process.
• :collision: Elastic Scale up and down multiple active workers, depending on the number of GPU devices.

The project is still under the active development. Welcome to join us!

• Demo on AWS
• Installation
• Quick Start
• ALaaS Server Customization (for Advance users)
• Strategy Zoo
• Citation

Demo on AWS :coffee:

Free ALaaS demo on AWS (Support HTTP & gRPC)

Use least confidence sampling with ResNet-18 to select images to be labeled for your tasks!

We have deployed ALaaS on AWS for demonstration. Try it by yourself!

Call ALaaS with HTTP 🌐

Call ALaaS with gRPC 🔐

curl \ -X POST http://13.213.29.8:8081/post \ -H 'Content-Type: application/json' \ -d '{"data":[{"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane1.png"}, {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane2.png"}, {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane3.png"}, {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane4.png"}, {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane5.png"}], "parameters": {"budget": 3}, "execEndpoint":"/query"}'

# pip install alaas from alaas.client import Client url_list = [ 'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane1.png', 'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane2.png', 'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane3.png', 'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane4.png', 'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane5.png' ] client = Client('grpc://13.213.29.8:60035') print(client.query_by_uri(url_list, budget=3))

Then you will see 3 data samples (the most informative) has been selected from all the 5 data points by ALaaS.

Installation :construction:

You can easily install the ALaaS by PyPI,

pip install alaas

The package of ALaaS contains both client and server parts. You can build an active data selection service on your own servers or just apply the client to perform data selection.

:warning: For deep learning frameworks like TensorFlow and Pytorch, you may need to install manually since the version to meet your deployment can be different (as well as transformers if you are running models from it).

You can also use Docker to run ALaaS:

docker pull huangyz0918/alaas

and start a service by the following command:

docker run -it --rm -p 8081:8081 \
        --mount type=bind,source=<config path>,target=/server/config.yml,readonly huangyz0918/alaas:latest

Quick Start :truck:

After the installation of ALaaS, you can easily start a local server, here is the simplest example that can be executed with only 2 lines of code.

from alaas.server import Server

Server.start()

The example code (by default) will start an image data selection (PyTorch ResNet-18 for image classification task) HTTP server in port 8081 for you. After this, you can try to get the selection results on your own image dataset, a client-side example is like

curl \
-X POST http://0.0.0.0:8081/post \
-H 'Content-Type: application/json' \
-d '{"data":[{"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane1.png"},
            {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane2.png"},
            {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane3.png"},
            {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane4.png"},
            {"uri": "https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane5.png"}], 
    "parameters": {"budget": 3},
    "execEndpoint":"/query"}'

You can also use alaas.Client to build the query request (for both http and grpc protos) like this,

from alaas.client import Client

url_list = [
    'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane1.png',
    'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane2.png',
    'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane3.png',
    'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane4.png',
    'https://www.cs.toronto.edu/~kriz/cifar-10-sample/airplane5.png'
]
client = Client('http://0.0.0.0:8081')
print(client.query_by_uri(url_list, budget=3))

The output data is a subset uris/data in your input dataset, which indicates selected results for further data labeling.

ALaaS Server Customization :wrench:

We support two different methods to start your server, 1. by input parameters 2. by YAML configuration

Input Parameters

You can modify your server by setting different input parameters,

from alaas.server import Server

Server.start(proto='http',                      # the server proto, can be 'grpc', 'http' and 'https'.
    port=8081,                                  # the access port of your server.
    host='0.0.0.0',                             # the access IP address of your server.
    job_name='default_app',                     # the server name.
    model_hub='pytorch/vision:v0.10.0',         # the active learning model hub, the server will automatically download it for data selection.
    model_name='resnet18',                      # the active learning model name (should be available in your model hub).
    device='cpu',                               # the deploy location/device (can be something like 'cpu', 'cuda' or 'cuda:0'). 
    strategy='LeastConfidence',                 # the selection strategy (read the document to see what ALaaS supports).
    batch_size=1,                               # the batch size of data processing.
    replica=1,                                  # the number of workers to select/query data.
    tokenizer=None,                             # the tokenizer name (should be available in your model hub), only for NLP tasks.
    transformers_task=None                      # the NLP task name (for Hugging Face [Pipelines](https://huggingface.co/docs/transformers/main_classes/pipelines)), only for NLP tasks.
)

YAML Configuration

You can also start the server by setting an input YAML configuration like this,

from alaas import Server

# start the server by an input configuration file.
Server.start_by_config('path_to_your_configuration.yml')

Details about building a configuration for your deployment scenarios can be found here.

Strategy Zoo :art:

Currently we supported several active learning strategies shown in the following table,

Type	Setting	Abbr	Strategy	Year	Reference
Random	Pool-base	RS	Random Sampling	-	-
Uncertainty	Pool	LC	Least Confidence Sampling	1994	DD Lew et al.
Uncertainty	Pool	MC	Margin Confidence Sampling	2001	T Scheffer et al.
Uncertainty	Pool	RC	Ratio Confidence Sampling	2009	B Settles et al.
Uncertainty	Pool	VRC	Variation Ratios Sampling	1965	EH Johnson et al.
Uncertainty	Pool	ES	Entropy Sampling	2009	B Settles et al.
Uncertainty	Pool	MSTD	Mean Standard Deviation	2016	M Kampffmeyer et al.
Uncertainty	Pool	BALD	Bayesian Active Learning Disagreement	2017	Y Gal et al.
Clustering	Pool	KCG	K-Center Greedy Sampling	2017	Ozan Sener et al.
Clustering	Pool	KM	K-Means Sampling	2011	Z Bodó et al.
Clustering	Pool	CS	Core-Set Selection Approach	2018	Ozan Sener et al.
Diversity	Pool	DBAL	Diverse Mini-batch Sampling	2019	Fedor Zhdanov
Adversarial	Pool	DFAL	DeepFool Active Learning	2018	M Ducoffe et al.

Citation

Our tech report of ALaaS is available on arxiv and NeurIPS 2022. Please cite as:

@article{huang2022active,
  title={Active-Learning-as-a-Service: An Efficient MLOps System for Data-Centric AI},
  author={Huang, Yizheng and Zhang, Huaizheng and Li, Yuanming and Lau, Chiew Tong and You, Yang},
  journal={arXiv preprint arXiv:2207.09109},
  year={2022}
}

Contributors ✨

Thanks goes to these wonderful people (emoji key):

Yizheng Huang 🚇 ⚠️ 💻

Huaizheng 🖋 ⚠️ 📖

Yuanming Li ⚠️ 💻

This project follows the all-contributors specification. Contributions of any kind welcome!

Acknowledgement

• Jina - Build cross-modal and multimodal applications on the cloud.
• Transformers - State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX.

License

The theme is available as open source under the terms of the Apache 2.0 License.