/SAA

Primary LanguagePython

Activating discriminative semantic for few-shot classification

General image classification

Environment

This code requires Pytorch 1.10.0 and torchvision 0.11.0 or higher with cuda support.

Result

mini-ImageNet

Model 5-way 1-shot 5-way 5-shot
ProtoNet 62.39±0.21 80.53±0.14
Poodle* 74.21 83.71
Our 65.39±0.61 81.28±0.43
Our* 76.82±0.41 82.93±0.38

tiered-ImageNet

Model 5-way 1-shot 5-way 5-shot
ProtoNet 68.23±0.23 84.03±0.16
Poodle* 78.72 86.57
Our 70.03±0.72 84.53±0.49
Our* 84.41±0.40 86.15±0.36

Dataset

Download mini-ImageNet, tiered-ImageNet and CUB dataset and put them into ./ .

Run

For example, to train the 1-shot/5-shot 5-way model with ProtoNet backbone on MiniImageNet: python ./Mini_imageNet/pre_train_improve_k.py

to train the 1-shot/5-shot 5-way model with ProtoNet backbone on tieredImageNet: python ./Tiered_imagenet/pre_train_improve_k.py