import os
import gradio as gr
import re
import string
from operator import itemgetter
import collections
import pypdf
from pypdf import PdfReader
from pypdf.errors import PdfReadError
import pdf2image
from pdf2image import convert_from_path
import langdetect
from langdetect import detect_langs
import pandas as pd
import numpy as np
import random
import tempfile
import itertools
from matplotlib import font_manager
from PIL import Image, ImageDraw, ImageFont
import cv2
## files
import sys
sys.path.insert(0, 'files/')
import functions
from functions import *
# update pip
os.system('python -m pip install --upgrade pip')
# model
from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-linelevel-ml384"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id);
model.to(device);
# APP outputs
def app_outputs(uploaded_pdf):
filename, msg, images = pdf_to_images(uploaded_pdf)
num_images = len(images)
if not msg.startswith("Error with the PDF"):
# Extraction of image data (text and bounding boxes)
dataset, lines, row_indexes, par_boxes, line_boxes = extraction_data_from_image(images)
# prepare our data in the format of the model
encoded_dataset = dataset.map(prepare_inference_features, batched=True, batch_size=64, remove_columns=dataset.column_names)
custom_encoded_dataset = CustomDataset(encoded_dataset, tokenizer)
# Get predictions (token level)
outputs, images_ids_list, chunk_ids, input_ids, bboxes = predictions_token_level(images, custom_encoded_dataset)
# Get predictions (line level)
probs_bbox, bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, df = predictions_line_level(dataset, outputs, images_ids_list, chunk_ids, input_ids, bboxes)
# Get labeled images with lines bounding boxes
images = get_labeled_images(dataset, images_ids_list, bboxes_list_dict, probs_dict_dict)
img_files = list()
# get image of PDF without bounding boxes
for i in range(num_images):
if filename != "files/blank.png": img_file = f"img_{i}_" + filename.replace(".pdf", ".png")
else: img_file = filename.replace(".pdf", ".png")
images[i].save(img_file)
img_files.append(img_file)
if num_images < max_imgboxes:
img_files += [image_blank]*(max_imgboxes - num_images)
images += [Image.open(image_blank)]*(max_imgboxes - num_images)
for count in range(max_imgboxes - num_images):
df[num_images + count] = pd.DataFrame()
else:
img_files = img_files[:max_imgboxes]
images = images[:max_imgboxes]
df = dict(itertools.islice(df.items(), max_imgboxes))
# save
csv_files = list()
for i in range(max_imgboxes):
csv_file = f"csv_{i}_" + filename.replace(".pdf", ".csv")
csv_files.append(gr.File.update(value=csv_file, visible=True))
df[i].to_csv(csv_file, encoding="utf-8", index=False)
else:
img_files, images, csv_files = [""]*max_imgboxes, [""]*max_imgboxes, [""]*max_imgboxes
img_files[0], img_files[1] = image_blank, image_blank
images[0], images[1] = Image.open(image_blank), Image.open(image_blank)
csv_file = "csv_wo_content.csv"
csv_files[0], csv_files[1] = gr.File.update(value=csv_file, visible=True), gr.File.update(value=csv_file, visible=True)
df, df_empty = dict(), pd.DataFrame()
df[0], df[1] = df_empty.to_csv(csv_file, encoding="utf-8", index=False), df_empty.to_csv(csv_file, encoding="utf-8", index=False)
return msg, img_files[0], img_files[1], images[0], images[1], csv_files[0], csv_files[1], df[0], df[1]
# gradio APP
with gr.Blocks(title="Inference APP for Document Understanding at line level (v1 - LiLT base)", css=".gradio-container") as demo:
gr.HTML("""
Inference APP for Document Understanding at line level (v1 - LiLT base)
LiLT (Language-Independent Layout Transformer) is a Document Understanding model that uses both layout and text in order to detect labels of bounding boxes. Combined with the model XML-RoBERTa base, this finetuned model has the capacity to understand any language. Finetuned on the dataset DocLayNet base, it can classifly any bounding box (and its OCR text) to 11 labels (Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, Title).
It relies on an external OCR engine to get words and bounding boxes from the document image. Thus, let's run in this APP an OCR engine (PyTesseract) to get the bounding boxes, then run LiLT (already fine-tuned on the dataset DocLayNet base at line level) on the individual tokens and then, visualize the result at line level!
It allows to get all pages of any PDF (of any language) with bounding boxes labeled at line level and the associated dataframes with labeled data (bounding boxes, texts, labels) :-)
However, the inference time per page can be high when running the model on CPU due to the number of line predictions to be made. Therefore, to avoid running this APP for too long, only the first 2 pages are processed by this APP. If you want to increase this limit, you can either clone this APP in Hugging Face Space (or run its notebook on your own plateform) and change the value of the parameter max_imgboxes, or run the inference notebook "Document AI | Inference at line level with a Document Understanding model (LiLT fine-tuned on DocLayNet dataset)" on your own platform as it does not have this limit.
More information about the DocLayNet datasets, the finetuning of the model and this APP in the following blog posts: