Learning Loss for Active Learning

An unofficial PyTorch implementation of the paper Learning Loss for Active Learning.

Requirements

torch
torchvision
matplotlib
tqdm
cv2
imageio
imutils

Usage

python main.py --task {clf OR detection OR hpe}

• task
  • clf: image classification
  • detection: object detection
  • hpe: human pose estimation

image classification

• dataset: CIFAR10 & CIFAR100
• model: ResNet
• metric: Accuracy

python main.py --task clf --dataset CIFAR10 --subset 10000 --num_epoch 200 --batch_size 128 --lr 0.1 --epoch_loss 120 --weights 1.0 --milestone 160

object detection

• dataset: PASCAL VOC2007 & 2012
• model: SSD (Single Shot Multibox Detector)
• metric: mAP

python main.py --task detection --dataset VOC0712 --num_epoch 300 --batch_size 32 --lr 0.001 --epoch_loss 240 --weights 1.0 --milestone 240 

human pose estimation

• dataset: MPII
• model: SHN (Stacked Hourglass Networks)
• metric: [email protected]

python main.py --task hpe --dataset mpii --subset 5000 --num_epoch 125 --batch_size 6 --wdecay 0 --lr 0.00025 --epoch_loss 75 --weights 0.0001 --milestone 100 

References

• https://github.com/Mephisto405/Learning-Loss-for-Active-Learning
• https://github.com/amdegroot/ssd.pytorch
• https://github.com/bearpaw/pytorch-pose
• https://github.com/kuangliu/pytorch-cifar