app.py

import os

os.system("pip install pyyaml==5.1")
# 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==1.8.0+cu101 torchvision==0.9.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html"
)

# install detectron2 that matches pytorch 1.8
# See https://detectron2.readthedocs.io/tutorials/install.html for instructions
os.system(
    "pip install -q detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.8/index.html"
)

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

import gradio as gr
import numpy as np
from transformers import LayoutLMv3Processor, LayoutLMv3ForTokenClassification
from datasets import load_dataset
from PIL import Image, ImageDraw, ImageFont

processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base")
model = LayoutLMv3ForTokenClassification.from_pretrained(
    "nielsr/layoutlmv3-finetuned-funsd"
)

# load image example
dataset = load_dataset("nielsr/funsd", split="test")
image = Image.open(dataset[0]["image_path"]).convert("RGB")
image = Image.open("./invoice.png")
image.save("document.png")

labels = dataset.features["ner_tags"].feature.names
id2label = {v: k for v, k in enumerate(labels)}
label2color = {
    "question": "blue",
    "answer": "green",
    "header": "orange",
    "other": "violet",
}


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


def process_image(image):
    width, height = image.size

    # encode
    encoding = processor(
        image, truncation=True, return_offsets_mapping=True, return_tensors="pt"
    )
    offset_mapping = encoding.pop("offset_mapping")

    # forward pass
    outputs = model(**encoding)

    # get predictions
    predictions = outputs.logits.argmax(-1).squeeze().tolist()
    token_boxes = encoding.bbox.squeeze().tolist()

    # only keep non-subword predictions
    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 predictions over the image
    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,
        )

    return image


title = "Interactive demo: LayoutLMv3"
description = "Demo for Microsoft's LayoutLMv3, a Transformer for state-of-the-art document image understanding tasks. This particular model is fine-tuned on FUNSD, 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/2204.08387' target='_blank'>LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking</a> | <a href='https://github.com/microsoft/unilm' target='_blank'>Github Repo</a></p>"
examples = [["document.png"]]

css = ".output-image, .input-image {height: 40rem !important; width: 100% !important;}"
# css = "@media screen and (max-width: 600px) { .output_image, .input_image {height:20rem !important; width: 100% !important;} }"
# css = ".output_image, .input_image {height: 600px !important}"

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

iface = gr.Interface(
    fn=process_image,
    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)