DPGN: Distribution Propagation Graph Network for Few-shot Learning

This repository is the official implementation of DPGN: Distribution Propagation Graph Network for Few-shot Learning.

Abstract

Most graph-network-based meta-learning approaches model instance-level relation of examples. We extend this idea further to explicitly model the distribution-level relation of one example to all other examples in a 1-vs-N manner. We propose a novel approach named distribution propagation graph network (DPGN) for few-shot learning. It conveys both the distribution-level relations and instance-level relations in each few-shot learning task. To combine the distribution-level relations and instance-level relations for all examples, we construct a dual complete graph network which consists of a point graph and a distribution graph with each node standing for an example. Equipped with dual graph architecture, DPGN propagates label information from labeled examples to unlabeled examples within several update generations. In extensive experiments on few-shot learning benchmarks, DPGN outperforms state-of-the-art results by a large margin in 5% ∼ 12% under supervised settings and 7% ∼ 13% under semi-supervised settings.

Requirements

CUDA Version: 10.1

Python : 3.5.2

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	Original Split
Mini-ImageNet	Matching Networks
Tiered-ImageNet	SSL
CIFAR-FS	R2D2
CUB-200-2011	Closer Look

The dataset directory should look like this:

├── dataset
    ├── mini-imagenet
        ├── mini_imagenet_test.pickle   
        ├── mini_imagenet_train.pickle  
        ├── mini_imagenet_val.pickle
    ├── tiered-imagenet
        ├── class_names.txt   
        ├── synsets.txt  
        ├── test_images.npz
        ├── test_labels.pkl   
        ├── train_images.npz  
        ├── train_labels.pkl
        ├── val_images.npz
        ├── val_labels.pkl
    ├── cifar-fs
        ├── cifar_fs_test.pickle   
        ├── cifar_fs_train.pickle  
        ├── cifar_fs_val.pickle
    ├── cub-200-2011
        ├── attributes   
        ├── bounding_boxes.txt 
        ├── classes.txt
        ├── image   
        ├── image_class_labels.txt 
        ├── images
        ├── images.txt   
        ├── parts
        ├── README
        ├── split
        ├── train_test_split.txt

Training

To train the model(s) in the paper, run:

python3 main.py --dataset_root dataset --config config/5way_1shot_resnet12_mini-imagenet.py --num_gpu 1 --mode train

Evaluation

To evaluate the model(s) in the paper, run:

python3 main.py --dataset_root dataset --config config/5way_1shot_resnet12_mini-imagenet.py --num_gpu 1 --mode eval

Pre-trained Models

This Google Drive contains pre-trained model under settings of 5way-1shot and 5way-5shots for mini-ImageNet dataset with ResNet12 backbone.

Results

# Default checkpoints directory is:
./checkpoints

# Default logs directory is:
./logs

Our model achieves the following performance on mini-ImageNet, tiered-ImageNet, CUB-200-2011 and CIFAR-FS (more detailed experimental results are in the paper).

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
RelationNet	ConvNet	50.44±0.82	65.32±0.70
MAML	ConvNet	48.70±1.84	55.31±0.73
GNN	ConvNet	50.33±0.36	66.41±0.63
TPN	ConvNet	55.51±0.86	69.86±0.65
Edge-label	ConvNet	59.63±0.52	76.34±0.48
DPGN	ConvNet	66.01±0.36	82.83±0.41
LEO	WRN	61.76±0.08	77.59±0.12
wDAE	WRN	61.07±0.15	76.75±0.11
DPGN	WRN	67.24±0.51	83.72±0.44
CloserLook	ResNet18	51.75±0.80	74.27±0.63
CTM	ResNet18	62.05±0.55	78.63±0.06
DPGN	ResNet18	66.63±0.51	84.07±0.42
MetaGAN	ResNet12	52.71±0.64	68.63±0.67
SNAIL	ResNet12	55.71±0.99	68.88±0.92
TADAM	ResNet12	58.50±0.30	76.70±0.30
Shot-Free	ResNet12	59.04±0.43	77.64±0.39
Meta-Transfer	ResNet12	61.20±1.80	75.53±0.80
FEAT	ResNet12	62.96±0.02	78.49±0.02
MetaOptNet	ResNet12	62.64±0.61	78.63±0.46
DPGN	ResNet12	67.77±0.32	84.60±0.43

tieredImageNet:

Method	backbone	5way-1shot	5way-5shot
MAML	ConvNet	51.67±1.81	70.30±1.75
ProtoNet	ConvNet	53.34±0.89	72.69±0.74
RelationNet	ConvNet	54.48±0.93	71.32±0.78
TPN	ConvNet	59.91±0.94	73.30±0.75
Edge-label	ConvNet	63.52±0.52	80.24±0.49
DPGN	ConvNet	69.43±0.49	85.92±0.42
CTM	ResNet18	64.78±0.11	81.05±0.52
DPGN	ResNet18	70.46±0.52	86.44±0.41
TapNet	ResNet12	63.08±0.15	80.26±0.12
Meta-Transfer	ResNet12	65.62±1.80	80.61±0.90
MetaOptNet	ResNet12	65.81±0.74	81.75±0.53
Shot-Free	ResNet12	66.87±0.43	82.64±0.39
DPGN	ResNet12	72.45±0.51	87.24±0.39

CUB-200-2011:

Method	backbone	5way-1shot	5way-5shot
ProtoNet	ConvNet	51.31±0.91	70.77±0.69
MAML	ConvNet	55.92±0.95	72.09±0.76
MatchingNet	ConvNet	61.16±0.89	72.86±0.70
RelationNet	ConvNet	62.45±0.98	76.11±0.69
CloserLook	ConvNet	60.53±0.83	79.34±0.61
DN4	ConvNet	53.15±0.84	81.90±0.60
DPGN	ConvNet	76.05±0.51	89.08±0.38
FEAT	ResNet12	68.87±0.22	82.90±0.15
DPGN	ResNet12	75.71±0.47	91.48±0.33

CIFAR-FS:

Method	backbone	5way-1shot	5way-5shot
ProtoNet	ConvNet	55.5±0.7	72.0±0.6
MAML	ConvNet	58.9±1.9	71.5±1.0
RelationNet	ConvNet	55.0±1.0	69.3±0.8
R2D2	ConvNet	65.3±0.2	79.4±0.1
DPGN	ConvNet	76.4±0.5	88.4±0.4
Shot-Free	ResNet12	69.2±0.4	84.7±0.4
MetaOptNet	ResNet12	72.0±0.7	84.2±0.5
DPGN	ResNet12	77.9±0.5	90.2±0.4