app.py

import os
os.system('pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu')
os.system('sudo apt-get install tesseract-ocr')
os.system('pip install -q pytesseract')
os.environ["TOKENIZERS_PARALLELISM"] = "false"
import gradio as gr
import torch
from datasets import load_dataset, ClassLabel
from transformers import LayoutLMv3ForTokenClassification, LayoutLMv3Processor,LayoutLMv3FeatureExtractor
import pytesseract
import numpy as np
from PIL import ImageDraw, ImageFont

examples = [['example1.png'],['example2.png'],['example3.png']]
dataset = load_dataset("nielsr/cord-layoutlmv3")['train']

def get_label_list(labels):
    unique_labels = set()
    for label in labels:
        unique_labels = unique_labels | set(label)
    label_list = list(unique_labels)
    label_list.sort()
    return label_list

def convert_l2n_n2l(dataset):
    features = dataset.features
    label_column_name = "ner_tags"

    label_list = features[label_column_name].feature.names
    if isinstance(features[label_column_name].feature, ClassLabel):
        id2label = {k:v for k,v in enumerate(label_list)}
        label2id = {v:k for k,v in enumerate(label_list)}
    else:
        label_list = get_label_list(dataset[label_column_name])
        id2label = {k:v for k,v in enumerate(label_list)}
        label2id = {v:k for k,v in enumerate(label_list)}

    return label_list, id2label, label2id, len(label_list)

def label_colour(label):
    label2color ={'MENU.PRICE':'blue','MENU.NM':'red','other':None,'MENU.NUM':'orange','TOTAL.TOTAL_PRICE':'green'}
    if label in label2color:
        colour = label2color.get(label)
    else:
        colour = None
    return colour

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

def convert_results(words,tags):
    ents = set()
    completeword = ""
    for word, tag in zip(words, tags):
        if tag != "O":
            ent_position, ent_type = tag.split("-")
            if ent_position == "S":
                ents.add((word,ent_type))
            else:
                if ent_position == "B":
                    completeword = completeword+ " "+ word
                elif ent_position == "I":
                    completeword= completeword+ " " + word
                elif ent_position == "E":
                    completeword =completeword+" " + word
                
                ents.add((completeword,ent_type))
                completeword= ""
    return ents

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

def predict(image):
    model = LayoutLMv3ForTokenClassification.from_pretrained("keldrenloy/LayoutLMv3FineTunedwithCORDandSGReceipts") #add your model directory here
    processor = LayoutLMv3Processor.from_pretrained("microsoft/layoutlmv3-base")
    label_list,id2label,label2id, num_labels = convert_l2n_n2l(dataset)
    width, height = image.size
    
    encoding_inputs = processor(image,return_offsets_mapping=True, return_tensors="pt",truncation = True)
    offset_mapping = encoding_inputs.pop('offset_mapping')

    with torch.no_grad():
        outputs = model(**encoding_inputs)
    
    predictions = outputs.logits.argmax(-1).squeeze().tolist()
    token_boxes = encoding_inputs.bbox.squeeze().tolist()
    
    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)
        draw.rectangle(box, outline=label_colour(predicted_label))
        draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label_colour(predicted_label), font=font)
        
    words = text_extraction(image)
    extracted_words = convert_results(words,true_predictions)
    
    menu_list = []
    price_list = []
    for idx,item in enumerate(extracted_words):
        if item[1] == 'MENU.NM':
            menu_list.append(f"item {idx}.{item[0]}")
        if item[1] == 'MENU.PRICE':
            price_list.append(f"item {idx}. ${item[0]}")
    
    return image,menu_list,price_list
    
def text_extraction(image):
    feature_extractor = LayoutLMv3FeatureExtractor()
    encoding = feature_extractor(image, return_tensors="pt")
    return encoding['words'][0]

css = """.output_image, .input_image {height: 600px !important}"""

demo = gr.Interface(fn = predict,
                    inputs = gr.inputs.Image(type="pil"),
                    outputs = [gr.outputs.Image(type="pil", label="annotated image"),'text','text'],
                    css = css,
                    examples = examples,
                    allow_flagging=True,
                    flagging_options=["incorrect", "correct"],
                    flagging_callback = gr.CSVLogger(),             
                    flagging_dir = "flagged",
                    analytics_enabled = True, enable_queue=True
                   )
demo.launch(debug=False)