relationNet

This repo contains the code for Learning to Compare: Relation Network for Few-Shot Learning in Pytorch

---

Usage

datasets

Please contact the authors for dataset (either miniImageNet from Vinyals or new miniImageNet from Ravi)

training

python -m visdom.server -port 5000
python demo.py --gpuid 0 --logport 5000

feel free to modify config.py to change arguments

weights

pretrained weights for miniImageNet on 5-way 1-shot is here

results

settings	5w1s acc
84x84 conv w padding	45.8%
80x80 centralCrop	48.2%
80x80 randCrop	49.0%

todos

• add ResNet version.
• merge training and testing