Deep Active Learning with Augmentation-based Consistency Estimation

Official code for "Deep Active Learning with Augmentation-based Consistency Estimation".

Arxiv preprint: https://arxiv.org/abs/2011.02666

@misc{hong2020deep, title={Deep Active Learning with Augmentation-based Consistency Estimation}, author={SeulGi Hong and Heonjin Ha and Junmo Kim and Min-Kook Choi}, year={2020}, eprint={2011.02666}, archivePrefix={arXiv}, primaryClass={cs.CV} }

Deep Active Learning

Python implementations of the following active learning algorithms:

• Random Sampling
• Cutout Sampling
• CutMix Sampling
• Entropy Sampling
• Margin Sampling

Available Dataset

• CIFAR10
• FashionMNIST

Docker

• nvcr.io/nvidia/pytorch:19.07-py3
• pip install torchvision==0.2.1

Prerequisites

• pytorch, matplotlib, scikit-learn, pandas
• torchvision 0.2.1

Commandline Arguments

• round : # of learning cycle

• epoch : # of epoch per round

• initnum : initial data before applying AL strategy

• pick : # of data to be picked per round (notated as K)

• train_al : Cutout loss (Consistency-based Loss)

• train_cm : CutMix loss (Consistency-based Loss)

• SEED : torch.manual_seed(SEED)

• drop : epoch that apply learning rate drop (default 160)

• dataset : choose dataset among available ones

• pretrained_path : path of round0 network (for model initialization)

• ALtype : AL strategy. Random, CutoutSampling, ...

• alpha : cutout scaling