Layout Order random

#48
by Chgayot - opened

Hi all,

Super nice model when I run it through the demo, but when running it vllm/Docker, I'm having a completely random layout order on clean documents, and tables always have a "null" order. I couldn't find any option that would help, but I noticed that processing directly PPStructureV3 does not show the same issue. What am I missing?

I have tried
pipeline = PaddleOCRVL(vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8118/v1", use_layout_detection=True)
and
pipeline = PaddleOCRVL(vl_rec_backend="vllm-server", vl_rec_server_url="http://127.0.0.1:8118/v1")
with the same results.

ad8232_0_layout_order_res

PaddlePaddle org

Hello, the order of tables, charts, and images does not affect the parsing of markdown, nor is it the focus of our optimization, so we have not visualized this part.

I'm not sure I understand. Having the wrong layout order affects directly the markdown parsing. See result below.

My problem is:

  • The online demo gives the right layout order (see below), but the self deployed docker version does not.
    Also, I tried adding the preprocessing step with the same results.

When running the Online Demo:
ad8

On Docker, resulting parsed markdown with wrong order:

APPLICATIONS

GENERAL DESCRIPTION

Fully integrated single-lead ECG front end

Low supply current: 170 $ \mu $ A (typical)

Common-mode rejection ratio: 80 dB (dc to 60 Hz)

Two or three electrode configurations

High signal gain (G = 100) with dc blocking capabilities

2-pole adjustable high-pass filter

Accepts up to $ \pm $ 300 mV of half cell potential

Fast restore feature improves filter settling

Uncommitted op amp

3-pole adjustable low-pass filter with adjustable gain

Leads off detection: ac or dc options

Integrated right leg drive (RLD) amplifier

Single-supply operation: 2.0 V to 3.5 V

Integrated reference buffer generates virtual ground

Rail-to-rail output

Internal RFI filter

8 kV HBM ESD rating

Shutdown pin

20-lead, 4 mm $ \times $ 4 mm LFCSP and LFCSP_SS package

Qualified for automotive applications

FEATURES

Fitness and activity heart rate monitors

Portable ECG

Remote health monitors

Gaming peripherals

Biopotential signal acquisition

Data Sheet

An uncommitted operational amplifier enables the AD8232 to create a three-pole low-pass filter to remove additional noise. The user can select the frequency cutoff of all filters to suit different types of applications.

...

Hi, i got the same issue when i do it throught windows and paddlex server

but when i try this on WSL ubuntu, somehow it fixed with perfect order.

Interesting @saikanov !
But I just tried and still get the same bad results.
What exactly are the differences between your Win and WSL setups?

Interesting @saikanov !
But I just tried and still get the same bad results.
What exactly are the differences between your Win and WSL setups?

i literally copy the Windows notebook to WSL so i dont think it has any differences

Sign up or log in to comment