You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Exploring the MMT code, I discovered it is doing a re-segmentation of sentences to process the segmented result as a batch.
First of all, it would be nice to explain somewhere that, when no adaptation is needed, very higher performances can be obtained by sending multiple sentences together. They will be then really processed one by one (not as a whole), but sent as a mini-batch to the GPU. The constraint is that the multi-sentence text should be re-segmentable (for example, with a point at the end of each sentence).
Secondly, this may cause problems: MMT can split a single sentence in several parts, depending on its content, then not translating it as a single sentence. It would be nice to add an option somewhere to prevent from re-segmentation, knowing the input sentence is to be taken as a whole (not to be split).
The text was updated successfully, but these errors were encountered:
Exploring the MMT code, I discovered it is doing a re-segmentation of sentences to process the segmented result as a batch.
First of all, it would be nice to explain somewhere that, when no adaptation is needed, very higher performances can be obtained by sending multiple sentences together. They will be then really processed one by one (not as a whole), but sent as a mini-batch to the GPU. The constraint is that the multi-sentence text should be re-segmentable (for example, with a point at the end of each sentence).
Secondly, this may cause problems: MMT can split a single sentence in several parts, depending on its content, then not translating it as a single sentence. It would be nice to add an option somewhere to prevent from re-segmentation, knowing the input sentence is to be taken as a whole (not to be split).
The text was updated successfully, but these errors were encountered: