app.py

import torch
from sconf import Config
from PIL import Image, ImageOps
from donut import DonutConfig, DonutModel
import warnings

warnings.filterwarnings("ignore")

from transformers import logging

logging.set_verbosity_warning()

config = Config(default="./config.yaml")

has_cuda = torch.cuda.is_available()
print(f"Is CUDA available: {has_cuda}")
if has_cuda:
    print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")

model = DonutModel.from_pretrained(
                config.pretrained_model_name_or_path,
                input_size=config.input_size,
                max_length=config.max_position_embeddings, #self.config.max_length,
                align_long_axis=config.align_long_axis,
                ignore_mismatched_sizes=True,
            )

#device = "cuda" if torch.cuda.is_available() else "cpu"

if has_cuda:
    model.half()
    model = model.to("cuda")


task_name = "matricula"
task_prompt = f"<s_{task_name}>"
    
def predict_matricula(model, task_name, image):
    image = ImageOps.exif_transpose(image)
    image = image.resize(size=(1280, 960), 
                         resample=Image.Resampling.NEAREST)

    model.eval()
    
    result = model.inference(image=image, prompt=f"<s_{task_name}>")["predictions"][0]
    return result


import gradio as gr

demo = gr.Interface(
    fn=lambda x:predict_matricula(model, task_name="matricula", image=x),
    title="Demo: Donut 🍩 for DR Matriculas",
    description="Dominican Vehicle **Matriculas OCR** Infering",
    inputs=gr.Image(label="Matricula", sources="upload", type="pil", show_label=True), 
    outputs=[gr.JSON(label="Matricula JSON", show_label=True, value={})]
)

demo.launch(share=True)