Table 2 Training hyperparameters

Dim of embedded \(\varvec{B}_{g}\)

128

Dim of \(\varvec{B}_{t}\)

128

Multi-heads

6

Number of attention modules

2

Dim of MLP

256

Length of beam codeword

64

Option of latent beam codebook \(N_{CB}\)

64, 128, 256

Batch size N

128

Number of positive sample

1

Number of negative samples

127

Anchor t

16

Upper bound of k

8

Optimizer

Adam

Learning rate

\(10^{-3}\)

Number of epochs

100

Dropout percentage

\(20\%\)

Dataset size (\(100\%\))

\(10\times 10^{4}\)

Dataset split

\(70\%;30\%\)