cord_inference.py

import torch
import numpy as np
from transformers import LayoutLMv3TokenizerFast, LayoutLMv3Processor, LayoutLMv3ForTokenClassification
from PIL import Image, ImageDraw, ImageFont
from utils import OCR, unnormalize_box

label_list = [
    "O",
    "B-MENU.CNT",
    "B-MENU.DISCOUNTPRICE",
    "B-MENU.NM",
    "B-MENU.NUM",
    "B-MENU.PRICE",
    "B-MENU.SUB.CNT",
    "B-MENU.SUB.NM",
    "B-MENU.SUB.PRICE",
    "B-MENU.UNITPRICE",
    "B-SUB_TOTAL.DISCOUNT_PRICE",
    "B-SUB_TOTAL.ETC",
    "B-SUB_TOTAL.SERVICE_PRICE",
    "B-SUB_TOTAL.SUBTOTAL_PRICE",
    "B-SUB_TOTAL.TAX_PRICE",
    "B-TOTAL.CASHPRICE",
    "B-TOTAL.CHANGEPRICE",
    "B-TOTAL.CREDITCARDPRICE",
    "B-TOTAL.MENUQTY_CNT",
    "B-TOTAL.TOTAL_PRICE",
    "I-MENU.CNT",
    "I-MENU.DISCOUNTPRICE",
    "I-MENU.NM",
    "I-MENU.NUM",
    "I-MENU.PRICE",
    "I-MENU.SUB.CNT",
    "I-MENU.SUB.NM",
    "I-MENU.SUB.PRICE",
    "I-MENU.UNITPRICE",
    "I-SUB_TOTAL.DISCOUNT_PRICE",
    "I-SUB_TOTAL.ETC",
    "I-SUB_TOTAL.SERVICE_PRICE",
    "I-SUB_TOTAL.SUBTOTAL_PRICE",
    "I-SUB_TOTAL.TAX_PRICE",
    "I-TOTAL.CASHPRICE",
    "I-TOTAL.CHANGEPRICE",
    "I-TOTAL.CREDITCARDPRICE",
    "I-TOTAL.MENUQTY_CNT",
    "I-TOTAL.TOTAL_PRICE"
]

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

tokenizer = LayoutLMv3TokenizerFast.from_pretrained("mp-02/layoutlmv3-finetuned-cord", apply_ocr=False)
processor = LayoutLMv3Processor.from_pretrained("mp-02/layoutlmv3-finetuned-cord", apply_ocr=False)
model = LayoutLMv3ForTokenClassification.from_pretrained("mp-02/layoutlmv3-finetuned-cord")

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)


def prediction(image):
    boxes, words = OCR(image)
    encoding = processor(image, words, boxes=boxes, return_offsets_mapping=True, return_tensors="pt", truncation=True)
    offset_mapping = encoding.pop('offset_mapping')

    for k, v in encoding.items():
        encoding[k] = v.to(device)

    outputs = model(**encoding)

    predictions = outputs.logits.argmax(-1).squeeze().tolist()
    token_boxes = encoding.bbox.squeeze().tolist()
    probabilities = torch.softmax(outputs.logits, dim=-1)
    confidence_scores = probabilities.max(-1).values.squeeze().tolist()

    inp_ids = encoding.input_ids.squeeze().tolist()
    inp_words = [tokenizer.decode(i) for i in inp_ids]

    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]]
    true_confidence_scores = [confidence_scores[idx] for idx, conf in enumerate(confidence_scores) if not is_subword[idx]]
    true_words = []

    for id, i in enumerate(inp_words):
        if not is_subword[id]:
            true_words.append(i)
        else:
            true_words[-1] = true_words[-1]+i

    true_predictions = true_predictions[1:-1]
    true_boxes = true_boxes[1:-1]
    true_words = true_words[1:-1]
    true_confidence_scores = true_confidence_scores[1:-1]

    """
    for i, conf in enumerate(true_confidence_scores):
        if conf < 0.5 :
            true_predictions[i] = "O"
    """            
    
    d = {}
    for id, i in enumerate(true_predictions):
        if i not in d.keys():
            d[i] = true_words[id]
        else:
            d[i] = d[i] + ", " + true_words[id]
    d = {k: v.strip() for (k, v) in d.items()}

    # TODO:process the json

    draw = ImageDraw.Draw(image, "RGBA")
    font = ImageFont.load_default()

    for prediction, box, confidence in zip(true_predictions, true_boxes, true_confidence_scores):
        draw.rectangle(box)
        draw.text((box[0]+10, box[1]-10), text=prediction+ ", "+ str(confidence), font=font, fill="black", font_size="15")

    return d, image