How smart_batching_collate works?

[InhyeokYoo]

Hi.

I've read SBERT paper but there's one thing I can not understand.
In section 7 Computational Efficiency in the paper, SBERT uses 'smart batching' for computation efficiency by reducing computational overhead from padding.

For improved computation of sentence embeddings, we implemented a smart batching strategy: Sentences with similar lengths are grouped together and are only padded to the longest element in a mini-batch. This drastically reduces computational overhead from padding tokens.

However, the code doesn't use this sentences with similar lengths information.
It seems to me that the collate function just extract text and label from example, and then tokenize the texts and return.

Could please somebody let me know what am I missing?

[nreimers]

Sentences are sorted by length in the encode method

[InhyeokYoo]

@nreimers

Does it mean smart_batching_collate actually not the smart batching of the paper?

And one more question: why did you use it in encode method not also fit method?

[nreimers]

Only in the encode. In the fit we what random shuffle

[InhyeokYoo]

Thank you. Totally understood.