Can't Run Demo Without Streamlit

[csimlinger]

Hi,

I was running the official demo (PyTorch) successfully and did some changes to the code - it still run successfully.
Once I remove the graphical user interface and try to run it via python demo/app.js, I get an error:

WARNING:root:Invalid model URL, using default initialization.

When I run the adjusted same script like that USE_TORCH=1 streamlit run demo/app.py, it works perfectly fine.

I can't figure out why - what am I missing?

Here is my script:

import os
import cv2
import numpy as np
import matplotlib.pyplot as plt

from doctr.io import DocumentFile
from doctr.utils.visualization import visualize_page
import torch
from backend.pytorch import DET_ARCHS, RECO_ARCHS, forward_image, load_predictor

# Set the environment variable to ensure PyTorch is used
os.environ['USE_TORCH'] = '1'

forward_device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

def save_image(image, path, cmap=None):
    """Helper function to save an image"""
    plt.figure(figsize=(10, 10))
    plt.imshow(image, cmap=cmap)
    plt.axis('off')
    plt.savefig(path)
    plt.close()

def main(det_archs, reco_archs):
    # Model configuration
    det_arch = "fast_tiny"
    reco_arch = "master"
    assume_straight_pages = True
    straighten_pages = False
    bin_thresh = 0.3
    box_thresh = 0.1

    # Load document
    image_path = os.path.join(os.path.dirname(__file__), "frame_5.jpg")
    doc = DocumentFile.from_images(image_path)
    page = doc[0]

    # Load model
    predictor = load_predictor(
        det_arch, reco_arch, assume_straight_pages, straighten_pages, bin_thresh, box_thresh, forward_device
    )

    # Process image
    seg_map = forward_image(predictor, page, forward_device)
    seg_map = np.squeeze(seg_map)
    seg_map = cv2.resize(seg_map, (page.shape[1], page.shape[0]), interpolation=cv2.INTER_LINEAR)

    # Save outputs
    save_image(seg_map, "output_segmentation_heatmap.jpg", cmap='viridis')  # Save segmentation heatmap

    # Process OCR and save results
    out = predictor([page])
    ocr_output = out.pages[0].export()
    if assume_straight_pages or (not assume_straight_pages and straighten_pages):
        img = out.pages[0].synthesize()
        cv2.imwrite("output_page_reconstitution.jpg", img)  # Save synthesized page image

    # Save OCR results as JSON
    import json
    with open("ocr_output.json", "w") as json_file:
        json.dump(ocr_output, json_file)

if __name__ == "__main__":
    main(DET_ARCHS, RECO_ARCHS)

Thanks a lot!

[felixdittrich92]

Hi @csimlinger
See: https://github.com/mindee/doctr/releases/tag/v0.8.1

To use the pretrained fast models you need to clone the main branch and install with pip install -e .[torch]
Afterwards it should run :)

[csimlinger]

@felixdittrich92 - thanks a lot, now I finally made it run :)