Algorithm 1: Training process of ZIC-LDM
Input: Training process iteration rounds $epochs$, batch size $m$, learning rate $lr$, semantic features $v$, FC parameter $W_h$ initialized for semantic features mapping, visual feature $f_{\phi }\left(x\right)$, FC parameter $W_g$ for visual features mapping and relation module $r_{\omega }$.
Output: Optimized FC parameter $W_h^{\ast }$ for semantic features mapping, FC parameter $W_g^{\ast }$ for visual features mapping and relation module $r_{\omega }^{*}$.
1	for $epoch=0,1,2,\dots ,$ $epochs-1$do
2	for $i=0,1,2,\dots ,$ $ntrain/m-1$ do
3	Sampling $m$ training samples $x$ and corresponding label $y$ from seen classes;
4	Mapping $f_{\phi }\left(x\right)$ into common space: $g_{\phi }\left(x\right)\leftarrow W_g\times f_{\phi }\left(x\right)$;
5	Mapping $v$ into common space: $h_{\phi }\left(v\right)\leftarrow W_h\times v$;
6	Concatenate $g_{\phi }\left(x\right)$ and $h_{\phi }\left(v\right)$;
7	Calculate similarity score: $s_{p,m}\leftarrow r_{\omega }\left(g_{\phi }\left(x\right),h_{\phi }\left(v\right)\right)$;
8	Calculate MSE loss: $L=MSE\left(s_{p,m},y_v\right)$;
9	Update FC parameters for semantic features mapping, FC parameters for visual features mapping and relation module:    $W_g^{*},W_h^{*},r_{\omega }^{*}\leftarrow Adan\left(\nabla W_g,W_h,r_{\omega }[L],W_g,W_h,r_{\omega },lr\right)$;
10	end for
11	end for