Table 1.

Algorithm details of the proposed method.

Algorithm: TLADA
Require: source data $X^s$; target data $X^t$; minibatch size m; critic training step n; learning rate for domain critic a1; learning rate for classification and feature learning a2;
Initialize feature extractor, domain critic, classifier with random weights ${\theta }_g,{\theta }_w,{\theta }_c$ repeat Sample minibatch ${\left\{x_i^s,y_i^t\right\}}_{i=1}^m,{\left\{x_i^t\right\}}_{i=1}^m$ from $X^s$ and $X^t$ for $t=1,\dots ,n$ do $h^s\leftarrow f_g\left(x^s\right),h^t\leftarrow f_g\left(x^t\right)$ sample h as the random points along straight lines between $h^s$ and $h^t$ $\widehat{h}\leftarrow \left\{h^s,h^t,h\right\}$ ${\theta }_w\leftarrow {\theta }_w+{\alpha }_1{\nabla }_{{\theta }_w}\left[{\mathcal{L}}_{wd}\left(x^s,x^t\right)-\rho {\mathcal{L}}_{grad}\left(\widehat{h}\right)\right]$ end for ${\theta }_c\leftarrow {\theta }_c-{\alpha }_2{\nabla }_{{\theta }_c}{\mathcal{L}}_c\left(x^s,y^s\right)$ ${\theta }_g\leftarrow {\theta }_g-{\alpha }_2{\nabla }_{{\theta }_g}\left[\mathcal{L}\left(x^s,y^s\right)+{\lambda }_1{\mathcal{L}}_{wd}\left(x^s,x^t\right)+{\lambda }_2{\mathcal{L}}_{tri}\left(x^s,x^t\right)\right]$ until ${\theta }_g,{\theta }_w,{\theta }_c$ converge