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 an out-of-the-box and standalone experience for users to quickly utilize active learning.
ALaaS is featured for
- 🐣 Easy-to-use With <10 lines of code to start the system to employ active learning.
- 🚀 Fast Use the stage-level parallellism to achieve over 10x speedup than under-optimized active learning process.
- 💥 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!
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.
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
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.
We support two different methods to start your server, 1. by input parameters 2. by YAML configuration
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.
)
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.
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. |
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}
}
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!
- Jina - Build cross-modal and multimodal applications on the cloud.
- Transformers - State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX.
The theme is available as open source under the terms of the Apache 2.0 License.