app.py

import os
os.system('pip install pyyaml==5.1')
os.system('pip install transformers==4.25.1')
os.system('pip install sentencepiece')
# workaround: install old version of pytorch since detectron2 hasn't released packages for pytorch 1.9 (issue: https://github.com/facebookresearch/detectron2/issues/3158)
os.system('pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html')
# install detectron2 that matches pytorch 1.8
os.system('pip install --upgrade detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cpu/torch1.8/index.html')

## install PyTesseract
os.system('pip install -q pytesseract')

import gradio as gr
import numpy as np
from transformers import LayoutXLMProcessor, LayoutLMv2ForTokenClassification
from datasets import load_dataset
import torch
from PIL import Image, ImageDraw, ImageFont
from itertools import chain

processor = LayoutXLMProcessor.from_pretrained("amir22010/layoutxlm-xfund-ja")
model = LayoutLMv2ForTokenClassification.from_pretrained("amir22010/layoutxlm-xfund-ja",num_labels = 7)

# load image example
#dataset = load_dataset("ranpox/xfund", 'xfund.ja', split="validation")
#image = Image.open(dataset[0]["image"][0]).convert("RGB")
image1 = Image.open("./ja_val_49.jpg").convert("RGB")
image1.save("document.jpg")

# define id2label, label2color
labels = [
    'O',
    'B-QUESTION',
    'B-ANSWER',
    'B-HEADER',
    'I-ANSWER',
    'I-QUESTION',
    'I-HEADER'
]

id2label = {v: k for v, k in enumerate(labels)}
label2id = {k: v for v, k in enumerate(labels)}

def unnormalize_box(bbox, width, height):
    return [
         width * (bbox[0] / 1000),
         height * (bbox[1] / 1000),
         width * (bbox[2] / 1000),
         height * (bbox[3] / 1000),
    ]

def iob_to_label(label):
    label = label[2:]
    if not label:
        return 'other'
    return label

label2color = {'question':'blue', 'answer':'green', 'header':'orange', 'other':'violet'}

def infer(image):
    # Use this if you're loading images
    #image = Image.open(img_path).convert("RGB")
    #image = image.convert("RGB") # loading PDFs
    try:
        encoding = processor(image, return_offsets_mapping=True, return_tensors="pt", truncation=True, max_length=514)#max_positional_embeddings
        offset_mapping = encoding.pop('offset_mapping')
        outputs = model(**encoding)
        predictions = outputs.logits.argmax(-1).squeeze().tolist()
        token_boxes = encoding.bbox.squeeze().tolist()
        width, height = image.size
        is_subword = np.array(offset_mapping.squeeze().tolist())[:,0] != 0
    
        true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]]
        true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]]
        draw = ImageDraw.Draw(image)
    
        font = ImageFont.load_default()
    
        for prediction, box in zip(true_predictions, true_boxes):
            predicted_label = iob_to_label(prediction).lower()
            draw.rectangle(box, outline=label2color[predicted_label])
            draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label2color[predicted_label], font=font)
    except Exception as e:
        print(e)

    return image

title = "Interactive demo: layoutxlm-ja"
description = "Demo for Microsoft's layoutxlm-ja, a Transformer for state-of-the-art document image understanding tasks. For More Information - https://huggingface.co/microsoft/layoutxlm-base. This particular model is fine-tuned on XFUND japanese, a dataset of manually annotated forms. It annotates the words appearing in the image as QUESTION/ANSWER/HEADER/OTHER. To use it, simply upload an image or use the example image below and click 'Submit'. Results will show up in a few seconds. If you want to make the output bigger, right-click on it and select 'Open image in new tab'."
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2104.08836' target='_blank'>LayoutXLM: LayoutXLM is a multilingual variant of LayoutLMv2. Pre-training for Visually-Rich Document Understanding</a> | <a href='https://github.com/microsoft/unilm' target='_blank'>Github Repo</a></p>"
examples =[['document.jpg']]

css = ".output-image, .input-image {height: 40rem !important; width: 100% !important;}"

css = ".image-preview {height: auto !important;}"

iface = gr.Interface(fn=infer, 
                     inputs=gr.inputs.Image(type="pil"), 
                     outputs=gr.outputs.Image(type="pil", label="annotated image"),
                     title=title,
                     description=description,
                     article=article,
                     examples=examples,
                     css=css,
                     enable_queue=True)
iface.launch(debug=True)