HGNN-FSL

Hierarchical Graph Neural Networks for Few-Shot Learning

Requirements

• python 3.6.9
• pytorch 1.2.0
• torchvision 0.4.0
• tensorboardx
• numpy
• pandas
• tqdm

Dataset

Mini-Imagenet

You can download miniImagenet dataset from EGNN's author here

Copy them inside following directory:

.
├── ...
└── dataset
   └── compacted_datasets
   	├── mini_imagenet_train.pickle
   	├──	mini_imagenet_val.pickle
   	└── mini_imagenet_test.pickle 

Tiered-Imagenet

You can download tieredimagenet dataset from few-shot-ssl-public's author here

Copy them inside following directory:

.
├── ...
└── dataset
	└── tiered-imagenet
		├── train_images_png.pkl
		├── train_labels.pkl
		├── val_images_png.pkl
		├── val_labels.pkl
		├── test_images_png.pkl
		├── test_labels.pkl
		├── class_names.txt
		└── synsets.txt

Training

# ************************** miniImagenet, 5way 5shot *****************************
$ python train.py --device cuda:0 --dataset mini --num_ways 5 --num_shots 5 --transductive True --pool_mode kn --unet_mode addold
$ python train.py --device cuda:0 --dataset mini --num_ways 5 --num_shots 5 --transductive False --pool_mode kn --unet_mode addold

# ************************** miniImagenet, 5way 1shot *****************************
$ python train.py --device cuda:0 --dataset mini --num_ways 5 --num_shots 1 --transductive True --pool_mode kn --unet_mode addold
$ python train.py --device cuda:0 --dataset mini --num_ways 5 --num_shots 1 --transductive False --pool_mode kn --unet_mode addold

# ************************** tieredImagenet, 5way 5shot *****************************
$ python train.py --device cuda:0 --dataset tiered --num_ways 5 --num_shots 5 --transductive True --pool_mode kn --unet_mode addold
$ python train.py --device cuda:0 --dataset tiered --num_ways 5 --num_shots 5 --transductive False --pool_mode kn --unet_mode addold

Evaluation

The trained models are saved in the path './asset/checkpoints/', with the name of 'D-{dataset}_ N-{ways} _K-{shots} _Q-{num_queries} _B-{batch size} _T-{transductive} _P-{pooling mode} _Un-{unet mode}'. So, for example, if you want to test the trained model of 'miniImagenet, 5way 5shot, transductive, kngpooling, addold' setting, you can give --test_model argument as follow:

$ python3 eval.py --test_model D-mini_N-5_K-5_Q-5_B-40_T-True_P-kn_Un-addold

Result

You can download our experiment results and trained models from here

miniImagenet,non-tranductive

Model	5-way 5-shot acc(%)
GNN	66.41
EGNN	66.85
(ours)HGNN	69.05

miniImageNet, transductive

Model	5-way 5-shot acc(%)
GNN*	75.41
EGNN	76.37
(ours)HGNN	79.64

tieredImageNet, non-transductive

Model	5-way 5-shot acc(%)
GNN	69.45
EGNN	70.98
(ours)HGNN	73.01

tieredImageNet, transductive

Model	5-way 5-shot acc(%)
GNN*	81.89
EGNN	80.15
(ours)HGNN	83.34

GNN transductive mode was implemented in here by gaieepo.