Echo9Zulu/PP-DocLayoutV3-FP16-OpenVINO

FP16 conversion of PP-DocLayoutV3 for use with OpenVINO. Right now only CPU has been tested but other devices should work.

Inference code is provided as ov_infer.py.

I converted this model using deepseek-flash-v4 (in pi) by providing the safetensors weights and all four files from the tranformers implementation; deepseek was able to discover on its own several bugs in the openvino conversion process while it worked, and chose to export into ONNX first, to resolve some isses with incomplete dynamic shapes.

Usage

PP-DoclayoutV3 is not meant to be used on its own; you can, of course- but it's better as a first pass over complext documents to crop by discovered elements, determine natural reading order or to filter images for an element we are interested in