Does not run without internet

[rh920]

I want to use surya-ocr offline. I have downloaded models DETECTOR_MODEL_CHECKPOINT and RECOGNITION_MODEL_CHECKPOINT and updated OCR (text recognition) script as shown bellow. However, the script still attemps to download vikp/surya_rec2

from PIL import Image
from surya.ocr import run_ocr
from surya.model.detection.model import load_model as load_det_model, load_processor as load_det_processor
from surya.model.recognition.model import load_model as load_rec_model
from surya.model.recognition.processor import load_processor as load_rec_processor

DETECTOR_MODEL_CHECKPOINT = './models/surya_det3'
RECOGNITION_MODEL_CHECKPOINT = './models/surya_rec2'

image = Image.open(IMAGE_PATH)
langs = ["en"] # Replace with your languages - optional but recommended
det_processor, det_model = load_det_processor(DETECTOR_MODEL_CHECKPOINT), load_det_model(DETECTOR_MODEL_CHECKPOINT)
rec_model, rec_processor = load_rec_model(RECOGNITION_MODEL_CHECKPOINT), load_rec_processor()

predictions = run_ocr([image], [langs], det_model, det_processor, rec_model, rec_processor)

I found that in surya.model.recognition.processor.load_processor there is a class named SuryaProcessor, that uses RECOGNITION_MODEL_CHECKPOINT. I suggest modifying load_processor to accept RECOGNITION_MODEL_CHECKPOINT as an input argument to specify the model path.

Alternatively, if there is a better solution to ensure the script runs without internet access, I would appreciate your guidance.

[xiaoyao9184]

This project uses Hugging Face's cached models and can operate in offline mode by setting HF_HUB_OFFLINE environment variable, which will prevent models from being downloaded.

To use offline mode, you need to pre-download the models. You can do this by running the project to download the models or by using the huggingface-cli download command (recommended).

The downloaded cache follows a dedicated folder structure; see manage-cache for details. You can also change the default cache location by setting the XDG_CACHE_HOME environment variable.

[xiaoyao9184]

After further try, it not only downloads models online but also fetches fonts from GitHub. If you want to run it completely offline, you can use the Docker image I created, which includes a Compose example.

[rh920]

After further try, it not only downloads models online but also fetches fonts from GitHub. If you want to run it completely offline, you can use the Docker image I created, which includes a Compose example.

Thanks a lot. There was range of docker compose file in your repository. Do you know anything about finetuning recognition model?

[xiaoyao9184]

After further try, it not only downloads models online but also fetches fonts from GitHub. If you want to run it completely offline, you can use the Docker image I created, which includes a Compose example.

Thanks a lot. There was range of docker compose file in your repository. Do you know anything about finetuning recognition model?

No, I haven't actually done it.