README.md

GearNet

AAAI 2022: GearNet: Stepwise Dual Learning for Weakly Supervised Domain Adaptation. (Pytorch implementation)

Requirements

• Python 3.8.3
• PyTorch 1.6.0

Taining

All the commands are shown in scripts files.

Results

UNIF-20%	A -> W	A -> D	W -> A	W -> D	D -> A	D -> W	Average
Standard	0.6197	0.6312	0.4861	0.9229	0.3164	0.8085	0.6308
Co-teaching	0.6158	0.6791	0.5614	0.9437	0.5482	0.845	0.6988
Jocor	0.6653	0.7416	0.5237	0.9770	0.5028	0.9036	0.7190
DAN	0.7031	0.7187	0.5809	0.9562	0.6299	0.9127	0.7502
DANN	0.7122	0.7500	0.6093	0.9687	0.6157	0.9270	0.7638
TCL	0.7773	0.8166	0.6075	0.9812	0.6100	0.9361	0.7881
G+Co-teaching	0.7343	0.7791	0.6058	0.9500	0.6051	0.9036	0.7629
G+DANN	0.7552	0.7916	0.6175	0.9729	0.6139	0.9361	0.7812
G+TCL	0.8177	0.8437	0.6168	0.9875	0.6225	0.9518	0.8066

UNIF-40%	A -> W	A -> D	W -> A	W -> D	D -> A	D -> W	Average
Standard	0.5585	0.5833	0.4098	0.8145	0.3870	0.7330	0.5810
Co-teaching	0.5937	0.6583	0.4971	0.8500	0.3338	0.5598	0.5821
Jocor	0.6302	0.6625	0.4723	0.9562	0.4524	0.8307	0.6673
DAN	0.6380	0.6708	0.5071	0.8916	0.5951	0.8971	0.6999
DANN	0.6731	0.7062	0.5223	0.9250	0.5696	0.8893	0.7142
TCL	0.7656	0.7562	0.5106	0.9437	0.5600	0.8880	0.7373
G+Co-teaching	0.6979	0.7250	0.5507	0.8583	0.3465	0.5833	0.6269
G+DANN	0.7434	0.7229	0.5326	0.9500	0.5742	0.9049	0.7380
G+TCL	0.7968	0.7916	0.4907	0.9458	0.5312	0.9322	0.7480

Flip-20%	A -> W	A -> D	W -> A	W -> D	D -> A	D -> W	Average
Standard	0.6015	0.6187	0.4421	0.9062	0.4005	0.7929	0.6269
Co-teaching	0.6054	0.6187	0.5191	0.8625	0.4609	0.7070	0.6289
Jocor	0.6484	0.6937	0.4847	0.8541	0.3469	0.7421	0.6283
DAN	0.6614	0.677	0.5500	0.8812	0.5802	0.8710	0.7034
DANN	0.6601	0.6541	0.5312	0.877	0.5717	0.8645	0.6931
TCL	0.7526	0.7833	0.5767	0.9229	0.6029	0.9322	0.76176
G+Co-teaching	0.7473	0.6791	0.5671	0.8770	0.4989	0.763	0.6887
G+DANN	0.7265	0.7312	0.5539	0.8854	0.5951	0.8945	0.7311
G+TCL	0.8776	0.8291	0.5873	0.9395	0.6072	0.9479	0.7981

Flip-40%	A -> W	A -> D	W -> A	W -> D	D -> A	D -> W	Average
Standard	0.4687	0.4812	0.3547	0.7208	0.3504	0.6328	0.5014
Co-teaching	0.5039	0.525	0.3615	0.5729	0.2929	0.4257	0.4469
Jocor	0.5429	0.602	0.4595	0.7125	0.3892	0.6914	0.5662
DAN	0.5312	0.5458	0.4218	0.7187	0.4353	0.6679	0.5534
DANN	0.5143	0.5062	0.4232	0.6937	0.4222	0.6757	0.5392
TCL	0.6601	0.6166	0.4595	0.7625	0.4335	0.6497	0.5969
G+Co-teaching	0.5742	0.5104	0.3806	0.5666	0.3164	0.4283	0.4627
G+DANN	0.5494	0.5208	0.4261	0.7354	0.4375	0.6731	0.5570
G+TCL	0.6992	0.6270	0.4630	0.7541	0.4225	0.6601	0.6043

Citation

If you find this useful in your research, please consider citing:

@article{xie2022gearnet,
  title={GearNet: Stepwise Dual Learning for Weakly Supervised Domain Adaptation},
  author={Xie, Renchunzi and Wei, Hongxin and Feng, Lei and An, Bo},
  journal={AAAI},
  year={2022}
}

Contact

If you have any problems about our code, feel free to contact

• XIER0002@e.ntu.edu.sg

or describe your problem in Issues.