active-learning

This repository contains code developed for Kili Technology, to investigate the use of active learning to accelerate the training pipeline. Active learning is a used to select which samples out of an unlabeled dataset should be added to the training data to maximize your model accuracy.

Summary

The repository is divided as follows :

• In active-learning, you can find a library containing models, dataset, and algorithms used for active learning.
  • In algorithms, different classes of algorithms are reproduced.
  • In dataset, academic dataset wrappers adapted to the active learning framework are defined.
  • In experiments, there are useful functions for setting up experiments.
  • In helpers, you can find things like a logger, a timer, etc...
  • In model, you have backbones of models used to produce the experiments in model_zoo/ and a wrapper around those models to support active learning at the root of the folder.
  • In train, you have active_train.py which contains a class used to train a model in an active-learning fashion.
• In experiments, you can find the code for different experiments ran.

Get started

git clone https://github.com/kili-technology/active-learning
cd active-learning
pip install .

As an example on how to use the library, check out /experiments/siim-isic-melanoma-classification/ : this presents a use case on how to create a training pipeline.

• In data_processing.py, an ActiveDataset dataset object is created, MelanomaDataset.
• In model.py, an ActiveModel learner object is created, SEResnext50_32x4dLearner.
• In main.py, those objects are combined in an ActiveTrain trainer object, together with an active learning algorithm.