[Python] PaddleOCR keeps trying to download en_PP-OCRv3_rec_infer.tar even after manual...

I'm trying to set up PaddleOCR with GPU support on my local Ubuntu 22.04 machine for a custom OCR pipeline to extract voter data from scanned PDFs. My objective is to run OCR with GPU acceleration using manually downloaded models because auto-downloads often fail or are slow. System Details:

OS: Ubuntu 22.04 LTS

Python: 3.8 (Conda environment ocr38)

PaddleOCR installed via pip install paddleocr

GPU: NVIDIA RTX 3050

GPU is enabled (use_gpu=True)


What I’m Trying To Do: Use PP-

• OCRv3 English models to extract text from images.
• Manually download and use models to avoid repeated auto-download issues.

What I Have Tried:

My minimal test script (test_ocr.py):

from paddleocr import PaddleOCR

ocr = PaddleOCR(use_angle_cls=True, lang='en', use_gpu=True)
ocr.ocr('test.jpg', cls=True)


On first run, it tries to download: https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar

So I manually downloaded the model:

wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar


Extracted it to the expected folder:

mkdir -p ~/.paddleocr/whl/rec/en/en_PP-OCRv3_rec_infer/
tar -xvf en_PP-OCRv3_rec_infer.tar -C ~/.paddleocr/whl/rec/en/en_PP-OCRv3_rec_infer/


Moved the contents up one level (since the archive adds a nested folder):

mv ~/.paddleocr/whl/rec/en/en_PP-OCRv3_rec_infer/en_PP-OCRv3_rec_infer/* ~/.paddleocr/whl/rec/en/en_PP-OCRv3_rec_infer/
rmdir ~/.paddleocr/whl/rec/en/en_PP-OCRv3_rec_infer/en_PP-OCRv3_rec_infer


Still Facing This Issue:

Even after placing all model files correctly, I get this error:

[ppocr ERROR]: Something went wrong while downloading models


And it again tries to download the same en_PP-OCRv3_rec_infer.tar model file — which is already there. What I Need Help With:

• How to make PaddleOCR recognize already downloaded models instead of downloading again?
• Is there a way to manually set the model path in code or config?
• How can I confirm that all required models are present and in the correct structure?

Note:

My goal is to run PaddleOCR in a GPU-accelerated setup, and I’ve already configured CUDA and verified that PaddlePaddle is using the GPU.

Any help would be deeply appreciated Thanks in advance! Regards, Ravi Verma

Continue reading...