pgl's Introduction

Progressive Graph Learning for Open-Set Domain Adaptation

Requirements

Python 3.6
Pytorch 1.3

Datasets

The links of datasets will be released afterwards,

Syn2Real-O (VisDA-18) https://github.com/VisionLearningGroup/visda-2018-public/tree/master/openset
VisDA-17 https://github.com/VisionLearningGroup/taskcv-2017-public/tree/master/classification
Office-home https://www.hemanthdv.org/officeHomeDataset.html

Training

The general command for training is,

python3 train.py

Change arguments for different experiments:

dataset: "home" / "visda" / "visda18"
batch_size: mini_batch size
beta: The ratio of known target sample and Unk target sample in the pseudo label set
EF : Enlarging Factor α
num_layers: GNN's depth
adv_coeff: adversarial loss coefficient γ
node_loss: node classification loss μ For the detailed hyper-parameters setting for each dataset, please refer to Section 5.2 and Appendix 3.

Remember to change dataset_root to suit your own case

The training loss and validation accuracy will be automatically saved in './logs/', which can be visualized with tensorboard. The model weights will be saved in './checkpoints'

Graph Learning without Pseudo-labeling Results (ResNet-50)

VisDA-18 (alpha=1, beta=0.6)

Plane	Bike	Bus	Car	Horse	Knife	Motorcycle	Person	Plant	SkateB	Train	Truck	Unk	OS^*	OS
0.437	0.807	0.588	0.646	0.857	0.155	0.943	0.355	0.879	0.250	0.712	0.126	0.437	0.553	0.563

Office-Home

Src	R			A			C			P
Tar	A	C	P	C	P	R	R	P	A	A	C	R	Avg.
OS	0.722	0.499	0.763	0.505	0.523	0.826	0.727	0.622	0.599	0.589	0.446	0.752	0.639
OS*	0.733	0.506	0.777	0.511	0.632	0.840	0.739	0.631	0.607	0.567	0.449	0.765	0.649