Active Learning - Uncertainty Sampling

Pytorch implementation of "A New Active Labeling Method for Deep Learning. IJCNN 2014".

Environments

NVIDIA pytorch docker [ link ]

docker pull nvcr.io/nvidia/pytorch:22.12-py3

requirements.txt

accelerate
wandb
torchvision

Methods

./query_strategies

• Least Confidence
• Margin Sampling
• Entropy

Results

Experiment Setting

• Model: ResNet18
• Batch Size: 128
• Optimizer: SGD
• Learning Rate: 0.1
• Learning Rate Scheduler: Cosine Annealing with Warm-up

Active Learning

• The Number of Initial Labeled Images: 5,000
• The Number of Query Images: 500
• The Number of Iteration: 20

Reference

• cure-lab/deep-active-learning [ link ]
• Active Learning. Yi Zhang (CMU) [ link ]