Algorithm 1 Pseudocode of Attention-Based Contrastive Loss Computation.

Input 1: $Q_{sup}\in {\mathbb{R}}^{M\times 1}$                   ▹ M: number of $Q_{sup}$

Input 2: $Q\in {\mathbb{R}}^{N\times 1}$                       ▹ N: count of Q

Output: $los{s}^{cl}$

$AttnMatrix\in {\mathbb{R}}^{M\times N}\leftarrow SoftMax(Norm(Q_{sup}\odot Q^T)$   ▹⊙: Matrix Multiplication

$Q_{sup}^{dup}\in {\mathbb{R}}^{M\times N}\leftarrow Duplicate\left\{Q_{sup}\right\}$

$los{s}^{cl}\leftarrow 0$

for $i\in \{0,1,2,...,M-1\}$ do

$T_i\in {\mathbb{R}}^{N\times 1}$                         ▹ i-th row of $Q_{sup}^{dup}$

$att{n}_i\in {\mathbb{R}}^{N\times 1}$                     ▹ i-th row of $AttnMatrix$

$los{s}_i^{cl}\leftarrow C{L}_{(att{n}_i\times T_i)\leftrightarrow Q}$    ▹ CL: Contrastive Loss; ×: element-wise multiplication

$los{s}^{cl}\leftarrow los{s}^{cl}+los{s}_i^{cl}$

end for

return $los{s}^{cl}\leftarrow \frac{1}{M}·los{s}^{cl}$