Help with PaddleOCR KIE

[hoangziet]

I am currently working on a project that requires the use of PaddleOCR KIE. But I failed on the run test and inference of how it works from this readme

This is my version of the packages I installed (on Windows 11):

paddlepaddle==2.5.2
paddlenlp==2.5.2
paddlex==3.0.1

I try to run inference with a single image like the instruction, but can't get the result because of the following bug:

python tools/infer_kie_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg

[2025/06/01 17:00:55] ppocr INFO: resume from ./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
Traceback (most recent call last):
  File "D:\Document\learn_paddle_ocr\PaddleOCR\tools\infer_kie_token_ser.py", line 129, in <module>
    ser_engine = SerPredictor(config)
  File "D:\Document\learn_paddle_ocr\PaddleOCR\tools\infer_kie_token_ser.py", line 76, in __init__
    self.ocr_engine = PaddleOCR(
  File "D:\Document\learn_paddle_ocr\PaddleOCR\paddleocr\_pipelines\ocr.py", line 144, in __init__
    super().__init__(**base_params)
  File "D:\Document\learn_paddle_ocr\PaddleOCR\paddleocr\_pipelines\base.py", line 60, in __init__
    self._common_args = parse_common_args(
  File "D:\Document\learn_paddle_ocr\PaddleOCR\paddleocr\_common_args.py", line 43, in parse_common_args
    raise ValueError(f"Unknown argument: {name}")
ValueError: Unknown argument: use_gpu

I have been stuck with this bug for hours. Is there anyone who can help me with it? I appreciate any help.
Anyway, are there any more clearly documented/tutorials for using PaddleOCR KIE?

[liuhongen1234567]

Hello, PaddleOCR recently underwent an upgrade, resulting in significant changes to the inference code. Unfortunately, due to time constraints, the code related to KIE models has not been updated accordingly. You can switch to the PaddleOCR 2.8 branch and use the KIE-related features by running git checkout remotes/origin/release/2.8 and then python -m pip install paddleocr==2.8.

[aymennasri]

Anything else we should keep in mind? Because there is a note that says use versions < 2.6

[liuhongen1234567]

Hello, my Paddle environment is version 3.0.0, but older versions should also work. The key is to download and install the PaddleOCR 2.8 code.

[aymennasri]

The docs in the 2.8 branch are all good?

[liuhongen1234567]

Currently, it seems that the discussion on the issue is okay. Since mainstream methods, such as PP-ChatOCRv4, are now using large models for information extraction, KIE tutorials are rarely updated. If you have any other questions, feel free to ask.

[aymennasri]

Well i'm using ppstructure to extract key/value pairs from a document that has (A, B, C, etc) field codes with their values close to them, this task is called KIE which can be done using SER and RE if i'm not mistaken.

The documents are also in french but setting the models to english works fine, my issue is with these version conflicts; I just want to use what's better and up to date and not old versions that might have issues with them especially 2.x

Currently i'm creating my custom dataset using PPOCRLabel and I don't know whether i can even finetune the models to work on my case.

[liuhongen1234567]

Hello, please refer to the sample data in the document.

wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar

The directory structure of this dataset is as follows:

The file class_list_xfun.txt defines the entity labels, which consist of 4 categories:

QUESTION: Represents the question in information extraction.
ANSWER: Represents the answer in information extraction.
OTHER: Indicates OCR text that does not require information extraction.
HEADER: Refers to key OCR information that still needs to be extracted, even though it is not presented in a Q&A format.

The file train.json contains the labels for KIE (Key Information Extraction).

"transcription": The recognized text content from OCR.
"label": The category label as defined above (e.g., QUESTION, ANSWER, OTHER, HEADER).
"id": The unique identifier for the OCR text.
"linking": A list of associated entities:
For OTHER and HEADER, it is an empty list ([]).
For ANSWER and QUESTION, it contains the ids of the linked entities (e.g., the corresponding QUESTION and ANSWER pairs).