Algorithm 2: Pixel class level adaptation

Input: iris data $X$, pixel label $Y, \mathrm{segmentation} \mathrm{network} G$, training epochs $\mathsf{\phi }$, ratio coefficient $\gamma$

Initialize: extended discriminator $D_{pixel}$

For t = 1,…, $\mathsf{\phi }$ do

Unfreeze the $D_{pixel}$ and freeze the $G$

$\mathrm{Compute} \mathrm{the} \mathrm{logit} \mathrm{maps}: P=softmax\left(G\left(X\right)\right)$

Extract class constraint knowledge: ${\mu }_i=\frac{e^{a_i/T}}{{{\displaystyle \sum }}_{j=1}^K{e}^{a_j/T}}$

Compute the $D_{pixel}$’s loss: $L^{D_{pixel}}\left(X_{easy},X_{hard}\right)=-{\displaystyle \sum }_{i=1}^K{\mu }_i^{\left(easy\right)}\log P_{D_{pixel}}^{\left(i\right)}\left(E_{easy}\right)+{\displaystyle \sum }_{j=1}^K{\mu }_j^{\left(hard\right)}\log P_{D_{pixel}}^{\left(j+K\right)}\left(E_{hard}\right)$

$\underset{D_{pixel}}{\min }{L}^{D_{pixel}}$ to train $D_{pixel}$

Unfreeze the $G$ and freeze the $D_{pixel}$

$\mathrm{Compute} \mathrm{the} \mathrm{cross}-\mathrm{entropy}: L_{cross}^{seg}\left(X\right)=-{\displaystyle \sum }_h^H{\displaystyle \sum }_w^W{\displaystyle \sum }_c^K{Y}^{\left(h,w,c\right)}\cdot \log \left(P^{\left(h,w,c\right)}\right)$

Compute the class level adversarial loss: $L_{class}^{adv}\left(X_{easy},X_{hard}\right)=-{\displaystyle \sum }_{j=1}^K{\mu }_j^{\left(hard\right)}\log P_{D_{pixel}}^{\left(j+K\right)}\left(E_{hard}\right)$

$\underset{G}{\min }{L}_{cross}^{seg}+L_{class}^{adv}$ to train $G$