zeef 0.1.0

A Python Framework for Deep Active Learning

Zeef: Active Learning for Data-Centric AI

Zeef is an active learning framework that can be applied to deep learning scenarios leak of labeled data. It contains many built-in data selection algorithms to reduce the labor of data annotation.

Installation

pip install zeef

For the local development, you can install from the Anaconda environment by

conda env create -f environment.yml

A quick MNIST CNN example can be found in here. Run

conda activate zeef
python main.py

to start the quick demonstration.

Quick Start

We can start from the most easy example: random select data points from an unlabeled data pool.

from zeef.data import Pool
from zeef.strategy import RandomSampling

# define the pool and active learning strategy. 
pool = Pool(torch_dataset_class, unlabeled_data)
strategy = RandomSampling(pool, network)

# start the active learning.
data_ids = strategy.query(1000)
# label those 1k sampled data points.
pool.label_by_ids(data_ids, data_labels) 
# retrain the model
strategy.learn()
# test the model
predictions = strategy.predict(test_x, test_y)

License

Apache License 2.0