Awesome

Codebase Readme

This is the codebase for paper: "Revisiting Unsupervised Domain Adaptation Models: a Smoothness Perspective" (ACCV 2022)

Environment

conda env create -f leco.yaml
conda activate leco

Prepare the datasets

Office-31 can be found here.
Office-Home can be found here.
Visda-C can be found here.
DomainNet can be found here.

Training guides

Visda-C

For MCC:

python da_visda.py --dset visda --lr 0.001 --net resnet101 --gpu_id 0 --batch_size 36 --base MCC --method Blank --interval 2 --s 0 --t 1

For MCC + LECO:

python da_visda.py --dset visda --lr 0.001 --net resnet101 --gpu_id 0 --batch_size 36 --base MCC --method LECO --interval 2 --s 0 --t 1 --warm_up 3000 --lamda 3

We set seed=[2020, 2021, 2022], showing the stable improvements to MCC. Logs can refer to TV.

Methods	plane	bcycl	bus	car	horse	knife	mcycl	person	plant	sktbrd	train	truck	Per-class
MCC(2020)	94.3	80.35	75.93	64.03	92.45	97.16	85.23	83.12	89.23	86.01	82.11	53.26	81.93
MCC(2021)	93.69	84.06	76.35	65.71	91.39	94.94	86.04	77.62	92.44	89.57	81.52	54.29	82.30
MCC(2022)	93.25	81.18	73.73	57.23	90.94	71.08	83.09	77.05	82.63	86.94	81.89	55.73	77.90
MCC+LeCo(2020)	97.12	85.96	83.86	89.66	96.55	97.45	89.06	84.05	95.91	90.79	85.08	43.82	86.61
MCC+LeCo(2021)	95.72	86.33	86.46	91.55	96.18	96.82	92.53	74.18	96.07	92.85	84.07	38.09	85.90
MCC+LeCo(2022)	96.49	87.02	79.17	90.46	95.86	96.43	91.24	82.55	94.55	92.42	88.36	40.57	86.26

For CDAN:

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base CDAN --method Blank --interval 2 --s 0 --t 1 --warm_up 3000 --lamda 3 --lr_decay2 0.1

For CDAN + LECO:

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base CDAN --method LECO --interval 2 --s 0 --t 1 --warm_up 3000 --lamda 0.5 --lr_decay2 0.1

For BNM:

python da_visda.py --dset visda --lr 0.001 --net resnet101 --gpu_id 0 --batch_size 36 --base BNM --method Blank --interval 2 --s 0 --t 1

FOr BNM + LeCo:

python da_visda.py --dset visda --lr 0.001 --net resnet101 --gpu_id 0 --batch_size 36 --base BNM --method LECO --interval 2 --s 0 --t 1 --warm_up 3000 --lamda 2

Office-home

For MCC:

python da_home.py --dset office-home --lr 0.01 --net resnet50 --gpu_id 0 --batch_size 36 --base MCC --method Blank --interval 2

For MCC + LECO:

python da_home.py --dset office-home --lr 0.01 --net resnet50 --gpu_id 0 --batch_size 36 --base MCC --method LECO --interval 2 --warm_up 3000 --lamda 2

For CDAN:

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base CDAN --method Blank --interval 2 --lr_decay2 0.1

For CDAN + LECO:

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base CDAN --method LECO --interval 2 --warm_up 3000 --lamda 2 --lr_decay2 0.1

For BNM

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base BNM --method Blank --interval 2

For BNM + LECO

python da_visda.py --dset visda --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base BNM --method LECO --interval 2 --lambda 3

DomainNet

For MCC + LECO

python da_domainNet.py --dset com-dn --lr 0.01 --net resnet101 --gpu_id 0 --batch_size 36 --base MCC --method LECO --interval 5 --warm_up 3000 --lamda 2

Office-31

This code file is borrowed from BNM. And you need to specify the source and target domain like follows:
For baseline: MCC, and method: LECO

python da_office.py --gpu_id 0 --base MCC --method LECO --num_iterations 8004  --dset office --s dslr --t amazon --test_interval 2000  --lambda_method 3

Visualization

Intra-class variance and inter-class variance visualization can refer to files (cal_cluster_intra.py, cal_cluster_inter.py).

Validation

Choosing the best hyper-parameter can refer to file: dev_loss.py.

BibTeX

@inproceedings{wang2022revisiting,
  title={Revisiting Unsupervised Domain Adaptation Models: a Smoothness Perspective},
  author={Wang, Xiaodong and Zhuo, Junbao and Zhang, Mengru and Wang, Shuhui and Fang, Yuejian},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  pages={1504--1521},
  year={2022}
}