Fewshot learning algorithm Comparsion

This repository is Experment Code for Matching Network, ProtoNet, MAML and Distribution Propagation Graph Network

Abstract

As an important branch of deep learning, few-shot Learning does not require a large amount of data but chooses a softer approach to solve problems, where it can be perfectly integrated with techniques such as meta learning and data augmentation. In this repository, I test four algorithms:Matching Network, ProtoNet, MAML and Distribution Propagation Graph NN on four FSL datasets: miniImageNet, Omniglot, CUB-200-2011 and CIFAR-FS.

Requirements

CUDA Version: 10.4

Python : 3.6.9

To install dependencies:

sudo pip3 install -r requirements.txt

Dataset

For your convenience, you can download the datasets directly from links on the left, or you can make them from scratch following the original splits on the right.

Dataset	Algorithms Paper
Mini-ImageNet	Matching Networks
Omniglot	ProtoNet
CIFAR-FS	MAML
CUB-200-2011	DP GNN

Experment obtained the following performance on mini-ImageNet, Omniglot, CUB-200-2011 and CIFAR-FS.

miniImageNet:

Method	Backbone	5way-1shot	5way-5shot
MatchingNet	ConvNet	43.56±0.84	55.31± 0.73
ProtoNet	ConvNet	49.42±0.78	68.20±0.66
MAML	ConvNet	48.70±1.84	55.31±0.73
DPGN	ConvNet	66.01±0.36	82.83±0.41

CUB-200-2011:

Method	backbone	5way-1shot	5way-5shot
MatchingNet	ConvNet	61.16±0.89%	72.86±0.7%
ProtoNet	ConvNet	51.31±0.9%	70.77±0.69%
MAML	ConvNet	55.92±0.95%	72.09±0.76%
DPGN	ConvNet	76.05±0.51%	89.08±0.38%

CIFAR-FS:

Method	backbone	5way-1shot	5way-5shot
MatchingNet	ConvNet	-	-
ProtoNet	ConvNet	55.5±0.7%	68.20±0.66%
MAML	ConvNet	58.9±1.9%	71.5±1.0%
DPGN	ConvNet	77.9±0.5%	90.2±0.4%

Ominiglot:

Method	backbone	5way-1shot	5way-5shot
MatchingNet	ConvNet	98.1%	98.9%
ProtoNet	ConvNet	97.4%	99.3%
MAML	ConvNet	98.7±0.4%	99.9±0.3%
DPGN	ConvNet	99.2%	99.7%