app.py

import os
os.system('git clone https://github.com/facebookresearch/detectron2.git')
os.system('pip install -e detectron2')
os.system("git clone https://github.com/microsoft/unilm.git")
os.system("sed -i 's/from collections import Iterable/from collections.abc import Iterable/' unilm/dit/object_detection/ditod/table_evaluation/data_structure.py")
os.system("curl -LJ -o publaynet_dit-b_cascade.pth 'https://layoutlm.blob.core.windows.net/dit/dit-fts/publaynet_dit-b_cascade.pth?sv=2022-11-02&ss=b&srt=o&sp=r&se=2033-06-08T16:48:15Z&st=2023-06-08T08:48:15Z&spr=https&sig=a9VXrihTzbWyVfaIDlIT1Z0FoR1073VB0RLQUMuudD4%3D'")

import sys
sys.path.append("unilm")
sys.path.append("detectron2")

## 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, ImageColor

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

dataset = load_dataset("ivan-wald/cord-layoutlmv3", split="test", trust_remote_code=True)
image = Image.open("./test0.jpeg")
labels = dataset.features['ner_tags'].feature.names
id2label = {v: k for v, k in enumerate(labels)}

#Need to get discrete colors for each labels
label_ints = np.random.randint(0, len(ImageColor.colormap.items()), 61)
label_color_pil = [k for k,_ in ImageColor.colormap.items()]
label_color = [label_color_pil[i] for i in label_ints]
label2color = {}
for k,v in id2label.items():
    label2color[v[2:]]=label_color[k]

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 = "LayoutLMv3 - CORD"
description = "description"
article = "article"
examples =[['test0.jpeg'],['test1.jpeg'],['test2.jpeg']]
css = ".output-image, .input-image, .image-preview {height: 600px !important}"

iface = gr.Interface(fn=process_image, 
                     inputs=gr.Image(type="pil"), 
                     outputs=gr.Image(type="pil", label="annotated image"),
                     title=title,
                     examples=examples,
                     css=css)

iface.launch(debug=True)