yejg2017 / fewshot-egnn

fewshot-egnn

Introduction

The current project page provides pytorch code that implements the following CVPR2019 paper:
Title: "Edge-labeling Graph Neural Network for Few-shot Learning"
Authors: Jongmin Kim, Taesup Kim, Sungwoong Kim, Chang D.Yoo

Institution: KAIST, KaKaoBrain
Code: https://github.com/khy0809/fewshot-egnn
Arxiv: https://arxiv.org/abs/1905.01436

Abstract: In this paper, we propose a novel edge-labeling graph neural network (EGNN), which adapts a deep neural network on the edge-labeling graph, for few-shot learning. The previous graph neural network (GNN) approaches in few-shot learning have been based on the node-labeling framework, which implicitly models the intra-cluster similarity and the inter-cluster dissimilarity. In contrast, the proposed EGNN learns to predict the edge-labels rather than the node-labels on the graph that enables the evolution of an explicit clustering by iteratively updating the edgelabels with direct exploitation of both intra-cluster similarity and the inter-cluster dissimilarity. It is also well suited for performing on various numbers of classes without retraining, and can be easily extended to perform a transductive inference. The parameters of the EGNN are learned by episodic training with an edge-labeling loss to obtain a well-generalizable model for unseen low-data problem. On both of the supervised and semi-supervised few-shot image classification tasks with two benchmark datasets, the proposed EGNN significantly improves the performances over the existing GNNs.

Citation

If you find this code useful you can cite us using the following bibTex:

@article{kim2019egnn,
  title={Edge-labeling Graph Neural Network for Few-shot Learning},
  author={Jongmin Kim, Taesup Kim, Sungwoong Kim, Chang D. Yoo},
  journal={arXiv preprint arXiv:1905.01436},
  year={2019}
}

Platform

This code was developed and tested with pytorch version 1.0.1

Setting

You can download miniImagenet dataset from here.

Download 'mini_imagenet_train/val/test.pickle', and put them in the path 'tt.arg.dataset_root/mini-imagenet/compacted_dataset/'

In train.py, replace the dataset root directory with your own: tt.arg.dataset_root = '/data/private/dataset'

Training

# ************************** miniImagenet, 5way 1shot *****************************
$ python3 train.py --dataset mini --num_ways 5 --num_shots 1 --transductive False
$ python3 train.py --dataset mini --num_ways 5 --num_shots 1 --transductive True

# ************************** miniImagenet, 5way 5shot *****************************
$ python3 trainer.py --dataset mini --num_ways 5 --num_shots 5 --trainsductive False
$ python3 trainer.py --dataset mini --num_ways 5 --num_shots 5 --trainsductive True

# ************************** miniImagenet, 10way 5shot *****************************
$ python3 trainer.py --dataset mini --num_ways 10 --num_shots 5 --meta_batch_size 20 --trainsductive True

# ************************** tieredImagenet, 5way 5shot *****************************
$ python3 trainer.py --dataset tiered --num_ways 5 --num_shots 5 --trainsductive False
$ python3 trainer.py --dataset tiered --num_ways 5 --num_shots 5 --trainsductive True

# **************** miniImagenet, 5way 5shot, 20% labeled (semi) *********************
$ python3 train.py --dataset mini --num_ways 5 --num_shots 5 --num_unlabeled 4 --transductive False
$ python3 train.py --dataset mini --num_ways 5 --num_shots 5 --num_unlabeled 4 --transductive True

Evaluation

The trained models are saved in the path './asset/checkpoints/', with the name of 'D-{dataset}-N-{ways}-K-{shots}-U-{num_unlabeld}-L-{num_layers}-B-{batch size}-T-{transductive}'. So, for example, if you want to test the trained model of 'miniImagenet, 5way 1shot, transductive' setting, you can give --test_model argument as follow:

$ python3 eval.py --test_model D-mini_N-5_K-1_U-0_L-3_B-40_T-True

Result

Here are some experimental results presented in the paper. You should be able to reproduce all the results by using the trained models which can be downloaded from here.

miniImageNet, non-transductive

Model	5-way 5-shot acc (%)
Matching Networks [1]	55.30
Reptile [2]	62.74
Prototypical Net [3]	65.77
GNN [4]	66.41
(ours) EGNN	66.85

miniImageNet, transductive

Model	5-way 5-shot acc (%)
MAML [5]	63.11
Reptile + BN [2]	65.99
Relation Net [6]	67.07
MAML + Transduction [5]	66.19
TPN [7]	69.43
TPN (Higher K) [7]	69.86
(ours) EGNN	76.37

tieredImageNet, non-transductive

Model	5-way 5-shot acc (%)
Reptile [2]	66.47
Prototypical Net [3]	69.57
(ours) EGNN	70.98

tieredImageNet, transductive

Model	5-way 5-shot acc (%)
MAML [5]	70.30
Reptile + BN [2]	71.03
Relation Net [6]	71.31
MAML + Transduction [5]	70.83
TPN [7]	72.58
(ours) EGNN	80.15

miniImageNet, semi-supervised, 5-way 5-shot

Model	20%	40%	60%	100%
GNN-LabeledOnly [4]	50.33	56.91	-	66.41
GNN-Semi [4]	52.45	58.76	-	66.41
EGNN-LabeledOnly	52.86	-	-	66.85
EGNN-Semi	61.88	62.52	63.53	66.85
EGNN-LabeledOnly (Transductive)	59.18	-	-	76.37
EGNN-Semi (Transductive)	63.62	64.32	66.37	76.37

miniImageNet, cross-way experiment

Model	train way	test way	Accuracy
GNN	5	5	66.41
GNN	5	10	N/A
GNN	10	10	51.75
GNN	10	5	N/A
EGNN	5	5	76.37
EGNN	5	10	56.35
EGNN	10	10	57.61
EGNN	10	5	76.27

References

[1] O. Vinyals et al. Matching networks for one shot learning.
[2] A Nichol, J Achiam, J Schulman, On first-order meta-learning algorithms.
[3] J. Snell, K. Swersky, and R. S. Zemel. Prototypical networks for few-shot learning.
[4] V Garcia, J Bruna, Few-shot learning with graph neural network.
[5] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks.
[6] F. Sung et al, Learning to Compare: Relation Network for Few-Shot Learning.
[7] Y Liu, J Lee, M Park, S Kim, Y Yang, Transductive propagation network for few-shot learning.