Open_set_domain_adaptation

Tensorflow Implementation of 《Open Set Domain Adaptation by Backpropagation》

On SVHN->MNIST and MNIST->USPS, USPS->MNIST

Usage:

python osda_train.py

Major Difference: Flip the sign of L_adv(xt)

(After confirmation with the authors, my thoughts are correct!)

paper:

Ladv(xt) = tlog(p(y = K + 1|xt)) + (1 − t) log(1 − p(y = K + 1|xt)

C: Ls(xs, ys) + Ladv(xt)

G: Ls(xs, ys) - Ladv(xt)

my implementation:

Ladv(xt) = -(tlog(p(y = K + 1|xt)) + (1 − t) log(1 − p(y = K + 1|xt))

C: Ls(xs, ys) + Ladv(xt)

G: Ls(xs, ys) - Ladv(xt)

Results:

my implementation

OS 85.7 OS* 85.6 ALL 85.8 UNK 85.9

paper

OS 92.3 OS* 91.2 ALL 94.4 UNK 97.6

I'm trying to fix this gap at present. Any advice is welcome.