Creating lite, small or medium versions of inference models

[danielvfung]

I managed to create the inference versions of the SAST, DB++ and even DB det_algorithms as per the instructions on the Paddle Docs.
They all seem to segment words better than the default en_PP-OCRv3 lite weight model.
The inference models i created using the export script are all 100 MB and larger. my question is how can we generate small sized models say medium, or smaller sized to be used for inference but still get the detection levels that these algorithms provide over the default line block level detection.

[danielvfung]

Justification is the need for word level bounding boxes, for NER and NLP type labelling and inference tasks.