app.py

import torch

from PIL import Image
from transformers import VisionEncoderDecoderModel, VisionEncoderDecoderConfig  # , DonutProcessor


def demo_process(input_img):
    global pretrained_model, task_prompt, task_name
    # input_img = Image.fromarray(input_img)
    output = pretrained_model.inference(image=input_img, prompt=task_prompt)["predictions"][0]
    return output

task_prompt = f"<s>"

st.text('This model is trained with receipt images -> SROIE dataset.')
 
"""image = Image.open("./sample_image_1.png")
image.save("receipt1.png")
image = Image.open("./sample_image_2.png")
image.save("receipt2.png")

pretrained_model = VisionEncoderDecoderModel.from_pretrained("unstructured/donut-base-sroie")
pretrained_model.encoder.to(torch.bfloat16)
pretrained_model.eval()

# replace for streamlit widgets
demo = gr.Interface(
    fn=demo_process,
    inputs= gr.inputs.Image(type="pil"),
    outputs="json",
    title=f"Donut 🍩 demonstration for `cord-v2` task",
    description="""This model is trained with 800 Indonesian receipt images of CORD dataset. <br>
Demonstrations for other types of documents/tasks are available at https://github.com/clovaai/donut <br>
More CORD receipt images are available at https://huggingface.co/datasets/naver-clova-ix/cord-v2
More details are available at:
- Paper: https://arxiv.org/abs/2111.15664
- GitHub: https://github.com/clovaai/donut""",
    examples=[["receipt1.png"], ["receipt2.png"]],
    cache_examples=False,
)

demo.launch()"""