Alleviating Imbalanced Pseudo-label Distribution: Self-Supervised Multi-Source Domain Adaptation with Label-specific Confidence

Official implementation for S3DA-LC (Based on SImpAl)

Parameters:

• --tau : refer to $\tau$ in paper
• --w_k : whether to use weights $w_k$
• --UTF : refer to $\lambda$ in paper

Please refer main.py for the detailed parameters setting.

Training:

python main.py --dataset office-31 --task DW_A --tau 0.9 --UTF 1.5 --w_k 1

t-SNE visualization

after warm up
after converge

Results on Office-31:

Method	$\rightarrow$ A	$\rightarrow$ W	$\rightarrow$ D	Avg
CAiDA	75.8	98.9	99.8	91.6
DECISION	75.4	98.4	99.6	91.1
SPS	73.8	99.3	100.0	91.10
S3DA-lc	78.1	99.0	100.0	92.4

Results on Office-Home:

Method	$\rightarrow$ Ar	$\rightarrow$ Cl	$\rightarrow$ Pr	$\rightarrow$ Rw	Avg
CAiDA	75.2	60.5	84.7	84.2	76.2
DECISION	74.5	59.4	84.4	83.6	75.5
SPS	75.1	66.0	84.4	84.2	77.4
S3DA-lc	78.1	70.0	87.4	87.2	80.7

Results on DomainNet:

Method	$\rightarrow$ Clp	$\rightarrow$ Inf	$\rightarrow$ Pnt	$\rightarrow$ Qdr	$\rightarrow$ Rel	$\rightarrow$ Skt	Avg
MSCAN	69.3	28.0	58.6	30.3	73.3	59.5	53.2
KD3A	72.5	23.4	60.9	16.4	72.7	60.6	51.1
STEM	72.0	28.2	61.5	25.7	72.6	60.2	53.4
S3DA-lc	71.9	31.3	61.3	27.1	75.7	61.2	54.8