paddle ocr text detection, fp16

[FahNos]

Hi,
I want to covert model PP-OCRv5_server_det from fp32 to fp16,
I already use the method:
convert model to onnx, and use from onnxconverter_common import float16 to convert model to fp16
but the inference result very bad,

In pytorch very simple to convert model to fp16 by use model.half()
but with paddle I do not how to do it?
In paddle have any method similar with model.half() in pytorch ?

Plese instruct me.

[liuhongen1234567]

Hello, if you want to use an fp16 model, you can simply set the value of precision to fp16 to achieve this. For details, please refer to the Method and parameter descriptions in documentation： https://paddlepaddle.github.io/PaddleOCR/main/en/version3.x/module_usage/text_detection.html#3-quick-start

from paddleocr import TextDetection
model = TextDetection(model_name="PP-OCRv5_server_det", precision="fp16")
output = model.predict("general_ocr_001.png", batch_size=1)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/res.json")

[FahNos]

Thank you so much,
It very helpful
But Do you know how to export model pdiparams with fp16 ?
I need it to reduce the volume of pdiparams file.

[leo-q8]

Paddle currently does not support exporting model pdiparams in FP16.