LiLT (Language-Independent Layout Transformer) is a Document Understanding model that uses both layout and text in order to detect labels of bounding boxes.
It relies on an external OCR engine to get words and bboxes from the document image. Thus, let's run in this APP an OCR engine ourselves (PyTesseract) as we'll need to do it in real life to get the bounding boxes, then run LiLT (already fine-tuned on the DocLayNet dataset at paragraph level: pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512) on the individual tokens and visualize the result at paragraph level!
LayoutXLM was proposed in LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
It is a Document Understanding model that uses both layout and text in order to detect labels of bounding boxes. More, it’s a multilingual extension of the LayoutLMv2 model trained on 53 languages.
It relies on an external OCR engine to get words and bboxes from the document image. Thus, let's run in this APP an OCR engine ourselves (PyTesseract) as we'll need to do it in real life to get the bounding boxes, then run LayoutXLM base (already fine-tuned on the DocLayNet dataset at paragraph level: pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512) on the individual tokens and visualize the result at paragraph level!
The idea here is to get new probabilities per label for each paragraph by summing the probabilities of the 2 models at the label level. Then we select as the label of each paragraph the one with the highest normalized probability.
%%capture
!pip install -q torch==1.10.0+cu111 torchvision==0.11+cu111 -f https://download.pytorch.org/whl/torch_stable.html
!python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
%%capture
# !sudo apt-get install poppler-utils
# !pip install pdf2image
# source: https://levelup.gitconnected.com/4-python-libraries-to-convert-pdf-to-images-7a09eba83a09
# source: https://pypi.org/project/pypdfium2/
!pip install -U pypdfium2
%%capture
!pip install -q langdetect
%%capture
!sudo apt install tesseract-ocr-all # english + osd (Orientation and script detection module)
# !sudo apt-get install tesseract-ocr-por # portuguese
# import os
# print(os.popen(f'cat /etc/debian_version').read())
# print(os.popen(f'cat /etc/issue').read())
# print(os.popen(f'apt search tesseract').read())
!pip install pytesseract
# In Colab, it is needed in order to update libraries with their new installed version (pillow).
# import os
# os.kill(os.getpid(), 9)
!pip install -q transformers sentencepiece datasets gradio pypdf
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.8/6.8 MB 33.5 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.3/1.3 MB 40.8 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 468.7/468.7 KB 23.0 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 15.7/15.7 MB 58.6 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 246.9/246.9 KB 20.9 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 7.6/7.6 MB 83.5 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 199.8/199.8 KB 17.6 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 212.2/212.2 KB 13.5 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 132.9/132.9 KB 10.0 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.0/1.0 MB 33.8 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 110.5/110.5 KB 9.5 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 57.8/57.8 KB 6.0 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 129.5/129.5 KB 10.7 MB/s eta 0:00:00 Preparing metadata (setup.py) ... done ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 144.1/144.1 KB 5.1 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 45.7/45.7 KB 4.7 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 71.5/71.5 KB 5.7 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 57.1/57.1 KB 4.2 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 50.5/50.5 KB 4.3 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 158.8/158.8 KB 11.6 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 264.6/264.6 KB 8.9 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 114.2/114.2 KB 8.9 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 66.9/66.9 KB 4.8 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 69.6/69.6 KB 6.1 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 58.3/58.3 KB 3.2 MB/s eta 0:00:00 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 80.6/80.6 KB 7.1 MB/s eta 0:00:00 Building wheel for ffmpy (setup.py) ... done
import os
from operator import itemgetter
import collections
import string
import re
import pypdf
from pypdf import PdfReader
from pypdf.errors import PdfReadError
# import pdf2image
# from pdf2image import convert_from_path
import pypdfium2 as pdfium
import langdetect
from langdetect import detect_langs
import pytesseract
import pandas as pd
import numpy as np
import random
from google.colab import files
import tempfile
from matplotlib import font_manager
from PIL import Image, ImageDraw, ImageFont
import cv2
# In Colab, use cv2_imshow instead of cv2.imshow
from google.colab.patches import cv2_imshow
from IPython.display import display
import itertools
import gradio as gr
import pathlib
from pathlib import Path
import shutil
from functools import partial
import transformers
import datasets
# categories colors
label2color = {
'Caption': 'brown',
'Footnote': 'orange',
'Formula': 'gray',
'List-item': 'yellow',
'Page-footer': 'red',
'Page-header': 'red',
'Picture': 'violet',
'Section-header': 'orange',
'Table': 'green',
'Text': 'blue',
'Title': 'pink'
}
# bounding boxes start and end of a sequence
cls_box = [0, 0, 0, 0]
cls_box1, cls_box2 = cls_box, cls_box
sep_box_lilt = cls_box
sep_box1 = sep_box_lilt
sep_box_layoutxlm = [1000, 1000, 1000, 1000]
sep_box2 = sep_box_layoutxlm
# models
model_id_lilt = "pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512"
model_id1 = model_id_lilt
model_id_layoutxlm = "pierreguillou/layout-xlm-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512"
model_id2 = model_id_layoutxlm
# tokenizer for LayoutXLM
tokenizer_id_layoutxlm = "xlm-roberta-base"
# (tokenization) The maximum length of a feature (sequence)
if (str(384) in model_id_lilt) and (str(384) in model_id_layoutxlm):
max_length = 384
elif (str(512) in model_id_lilt) and (str(512) in model_id_layoutxlm):
max_length = 512
else:
print("Error with max_length of chunks!")
# (tokenization) overlap
doc_stride = 128 # The authorized overlap between two part of the context when splitting it is needed.
# max PDF page images that will be displayed
max_imgboxes = 2
# get files
examples_dir = 'files/'
Path(examples_dir).mkdir(parents=True, exist_ok=True)
from huggingface_hub import hf_hub_download
files = ["example.pdf", "blank.pdf", "blank.png", "languages_iso.csv", "languages_tesseract.csv", "wo_content.png"]
for file_name in files:
path_to_file = hf_hub_download(
repo_id = "pierreguillou/Inference-APP-Document-Understanding-at-paragraphlevel-v3",
filename = "files/" + file_name,
repo_type = "space"
)
shutil.copy(path_to_file,examples_dir)
# path to files
image_wo_content = examples_dir + "wo_content.png" # image without content
pdf_blank = examples_dir + "blank.pdf" # blank PDF
image_blank = examples_dir + "blank.png" # blank image
## get langdetect2Tesseract dictionary
t = "files/languages_tesseract.csv"
l = "files/languages_iso.csv"
df_t = pd.read_csv(t)
df_l = pd.read_csv(l)
langs_t = df_t["Language"].to_list()
langs_t = [lang_t.lower().strip().translate(str.maketrans('', '', string.punctuation)) for lang_t in langs_t]
langs_l = df_l["Language"].to_list()
langs_l = [lang_l.lower().strip().translate(str.maketrans('', '', string.punctuation)) for lang_l in langs_l]
langscode_t = df_t["LangCode"].to_list()
langscode_l = df_l["LangCode"].to_list()
Tesseract2langdetect, langdetect2Tesseract = dict(), dict()
for lang_t, langcode_t in zip(langs_t,langscode_t):
try:
if lang_t == "Chinese - Simplified".lower().strip().translate(str.maketrans('', '', string.punctuation)): lang_t = "chinese"
index = langs_l.index(lang_t)
langcode_l = langscode_l[index]
Tesseract2langdetect[langcode_t] = langcode_l
except:
continue
langdetect2Tesseract = {v:k for k,v in Tesseract2langdetect.items()}
Downloading (…)in/files/example.pdf: 0%| | 0.00/343k [00:00<?, ?B/s]
Downloading (…)main/files/blank.pdf: 0%| | 0.00/1.15k [00:00<?, ?B/s]
Downloading (…)main/files/blank.png: 0%| | 0.00/3.99k [00:00<?, ?B/s]
Downloading (…)es/languages_iso.csv: 0%| | 0.00/2.52k [00:00<?, ?B/s]
Downloading (…)guages_tesseract.csv: 0%| | 0.00/2.13k [00:00<?, ?B/s]
Downloading (…)files/wo_content.png: 0%| | 0.00/27.6k [00:00<?, ?B/s]
# get text and bounding boxes from an image
# https://stackoverflow.com/questions/61347755/how-can-i-get-line-coordinates-that-readed-by-tesseract
# https://medium.com/geekculture/tesseract-ocr-understanding-the-contents-of-documents-beyond-their-text-a98704b7c655
def get_data_paragraph(results, factor, conf_min=0):
data = {}
for i in range(len(results['line_num'])):
level = results['level'][i]
block_num = results['block_num'][i]
par_num = results['par_num'][i]
line_num = results['line_num'][i]
top, left = results['top'][i], results['left'][i]
width, height = results['width'][i], results['height'][i]
conf = results['conf'][i]
text = results['text'][i]
if not (text == '' or text.isspace()):
if conf >= conf_min:
tup = (text, left, top, width, height)
if block_num in list(data.keys()):
if par_num in list(data[block_num].keys()):
if line_num in list(data[block_num][par_num].keys()):
data[block_num][par_num][line_num].append(tup)
else:
data[block_num][par_num][line_num] = [tup]
else:
data[block_num][par_num] = {}
data[block_num][par_num][line_num] = [tup]
else:
data[block_num] = {}
data[block_num][par_num] = {}
data[block_num][par_num][line_num] = [tup]
# get paragraphs dicionnary with list of lines
par_data = {}
par_idx = 1
for _, b in data.items():
for _, p in b.items():
line_data = {}
line_idx = 1
for _, l in p.items():
line_data[line_idx] = l
line_idx += 1
par_data[par_idx] = line_data
par_idx += 1
# get lines of texts, grouped by paragraph
texts_pars = list()
row_indexes = list()
texts_lines = list()
texts_lines_par = list()
row_index = 0
for _,par in par_data.items():
count_lines = 0
lines_par = list()
for _,line in par.items():
if count_lines == 0: row_indexes.append(row_index)
line_text = ' '.join([item[0] for item in line])
texts_lines.append(line_text)
lines_par.append(line_text)
count_lines += 1
row_index += 1
# lines.append("\n")
row_index += 1
texts_lines_par.append(lines_par)
texts_pars.append(' '.join(lines_par))
# lines = lines[:-1]
# get paragraphes boxes (par_boxes)
# get lines boxes (line_boxes)
par_boxes = list()
par_idx = 1
line_boxes, lines_par_boxes = list(), list()
line_idx = 1
for _, par in par_data.items():
xmins, ymins, xmaxs, ymaxs = list(), list(), list(), list()
line_boxes_par = list()
count_line_par = 0
for _, line in par.items():
xmin, ymin = line[0][1], line[0][2]
xmax, ymax = (line[-1][1] + line[-1][3]), (line[-1][2] + line[-1][4])
line_boxes.append([int(xmin/factor), int(ymin/factor), int(xmax/factor), int(ymax/factor)])
line_boxes_par.append([int(xmin/factor), int(ymin/factor), int(xmax/factor), int(ymax/factor)])
xmins.append(xmin)
ymins.append(ymin)
xmaxs.append(xmax)
ymaxs.append(ymax)
line_idx += 1
count_line_par += 1
xmin, ymin, xmax, ymax = min(xmins), min(ymins), max(xmaxs), max(ymaxs)
par_bbox = [int(xmin/factor), int(ymin/factor), int(xmax/factor), int(ymax/factor)]
par_boxes.append(par_bbox)
lines_par_boxes.append(line_boxes_par)
par_idx += 1
return texts_lines, texts_pars, texts_lines_par, row_indexes, par_boxes, line_boxes, lines_par_boxes
# rescale image to get 300dpi
# https://stackoverflow.com/questions/54001029/how-to-change-the-dpi-or-density-when-saving-images-using-pil
def set_image_dpi_resize(image):
"""
Rescaling image to 300dpi while resizing
:param image: An image
:return: A rescaled image
"""
length_x, width_y = image.size
factor = min(1, float(1024.0 / length_x))
size = int(factor * length_x), int(factor * width_y)
# image_resize = image.resize(size, Image.Resampling.LANCZOS)
image_resize = image.resize(size, Image.LANCZOS)
temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='1.png')
temp_filename = temp_file.name
image_resize.save(temp_filename, dpi=(300, 300))
return factor, temp_filename
# it is important that each bounding box should be in (upper left, lower right) format.
# source: https://github.com/NielsRogge/Transformers-Tutorials/issues/129
def upperleft_to_lowerright(bbox):
x0, y0, x1, y1 = tuple(bbox)
if bbox[2] < bbox[0]:
x0 = bbox[2]
x1 = bbox[0]
if bbox[3] < bbox[1]:
y0 = bbox[3]
y1 = bbox[1]
return [x0, y0, x1, y1]
# convert boundings boxes (left, top, width, height) format to (left, top, left+widght, top+height) format.
def convert_box(bbox):
x, y, w, h = tuple(bbox) # the row comes in (left, top, width, height) format
return [x, y, x+w, y+h] # we turn it into (left, top, left+widght, top+height) to get the actual box
# LiLT model gets 1000x10000 pixels images
def normalize_box(bbox, width, height):
return [
int(1000 * (bbox[0] / width)),
int(1000 * (bbox[1] / height)),
int(1000 * (bbox[2] / width)),
int(1000 * (bbox[3] / height)),
]
# LiLT model gets 1000x10000 pixels images
def denormalize_box(bbox, width, height):
return [
int(width * (bbox[0] / 1000)),
int(height * (bbox[1] / 1000)),
int(width* (bbox[2] / 1000)),
int(height * (bbox[3] / 1000)),
]
# get back original size
def original_box(box, original_width, original_height, coco_width, coco_height):
return [
int(original_width * (box[0] / coco_width)),
int(original_height * (box[1] / coco_height)),
int(original_width * (box[2] / coco_width)),
int(original_height* (box[3] / coco_height)),
]
def get_blocks(bboxes_block, categories, texts):
# get list of unique block boxes
bbox_block_dict, bboxes_block_list, bbox_block_prec = dict(), list(), list()
for count_block, bbox_block in enumerate(bboxes_block):
if bbox_block != bbox_block_prec:
bbox_block_indexes = [i for i, bbox in enumerate(bboxes_block) if bbox == bbox_block]
bbox_block_dict[count_block] = bbox_block_indexes
bboxes_block_list.append(bbox_block)
bbox_block_prec = bbox_block
# get list of categories and texts by unique block boxes
category_block_list, text_block_list = list(), list()
for bbox_block in bboxes_block_list:
count_block = bboxes_block.index(bbox_block)
bbox_block_indexes = bbox_block_dict[count_block]
category_block = np.array(categories, dtype=object)[bbox_block_indexes].tolist()[0]
category_block_list.append(category_block)
text_block = np.array(texts, dtype=object)[bbox_block_indexes].tolist()
text_block = [text.replace("\n","").strip() for text in text_block]
if id2label[category_block] == "Text" or id2label[category_block] == "Caption" or id2label[category_block] == "Footnote":
text_block = ' '.join(text_block)
else:
text_block = '\n'.join(text_block)
text_block_list.append(text_block)
return bboxes_block_list, category_block_list, text_block_list
# function to sort bounding boxes
def get_sorted_boxes(bboxes):
# sort by y from page top to bottom
sorted_bboxes = sorted(bboxes, key=itemgetter(1), reverse=False)
y_list = [bbox[1] for bbox in sorted_bboxes]
# sort by x from page left to right when boxes with same y
if len(list(set(y_list))) != len(y_list):
y_list_duplicates_indexes = dict()
y_list_duplicates = [item for item, count in collections.Counter(y_list).items() if count > 1]
for item in y_list_duplicates:
y_list_duplicates_indexes[item] = [i for i, e in enumerate(y_list) if e == item]
bbox_list_y_duplicates = sorted(np.array(sorted_bboxes, dtype=object)[y_list_duplicates_indexes[item]].tolist(), key=itemgetter(0), reverse=False)
np_array_bboxes = np.array(sorted_bboxes)
np_array_bboxes[y_list_duplicates_indexes[item]] = np.array(bbox_list_y_duplicates)
sorted_bboxes = np_array_bboxes.tolist()
return sorted_bboxes
# sort data from y = 0 to end of page (and after, x=0 to end of page when necessary)
def sort_data(bboxes, categories, texts):
sorted_bboxes = get_sorted_boxes(bboxes)
sorted_bboxes_indexes = [bboxes.index(bbox) for bbox in sorted_bboxes]
sorted_categories = np.array(categories, dtype=object)[sorted_bboxes_indexes].tolist()
sorted_texts = np.array(texts, dtype=object)[sorted_bboxes_indexes].tolist()
return sorted_bboxes, sorted_categories, sorted_texts
# sort data from y = 0 to end of page (and after, x=0 to end of page when necessary)
def sort_data_wo_labels(bboxes, texts):
sorted_bboxes = get_sorted_boxes(bboxes)
sorted_bboxes_indexes = [bboxes.index(bbox) for bbox in sorted_bboxes]
sorted_texts = np.array(texts, dtype=object)[sorted_bboxes_indexes].tolist()
return sorted_bboxes, sorted_texts
# get filename and images of PDF pages
def pdf_to_images(uploaded_pdf):
# Check if None object
if uploaded_pdf is None:
path_to_file = pdf_blank
filename = path_to_file.replace(examples_dir,"")
msg = "Invalid PDF file."
images = [Image.open(image_blank)]
else:
# path to the uploaded PDF
path_to_file = uploaded_pdf.name
filename = path_to_file# .replace("/tmp/","")
try:
PdfReader(path_to_file)
except PdfReadError:
path_to_file = pdf_blank
filename = path_to_file.replace(examples_dir,"")
msg = "Invalid PDF file."
images = [Image.open(image_blank)]
else:
try:
# images = convert_from_path(path_to_file, last_page=max_imgboxes)
pdf = pdfium.PdfDocument(str(filename))
version = pdf.get_version() # get the PDF standard version
n_pages = len(pdf) # get the number of pages in the document
last_page = max_imgboxes
page_indices = [i for i in range(last_page)] # pages until last_page
images = list(pdf.render(
pdfium.PdfBitmap.to_pil,
page_indices = page_indices,
scale = 300/72, # 300dpi resolution
))
num_imgs = len(images)
msg = f'The PDF "{filename}" was converted into {num_imgs} images.'
except:
msg = f'Error with the PDF "{filename}": it was not converted into images.'
images = [Image.open(image_wo_content)]
return filename, msg, images
# Extraction of image data (text and bounding boxes)
def extraction_data_from_image(images):
num_imgs = len(images)
if num_imgs > 0:
# https://pyimagesearch.com/2021/11/15/tesseract-page-segmentation-modes-psms-explained-how-to-improve-your-ocr-accuracy/
custom_config = r'--oem 3 --psm 3 -l eng' # default config PyTesseract: --oem 3 --psm 3 -l eng+deu+fra+jpn+por+spa+rus+hin+chi_sim
results, texts_lines, texts_pars, texts_lines_par, row_indexes, par_boxes, line_boxes, lines_par_boxes, images_pixels = dict(), dict(), dict(), dict(), dict(), dict(), dict(), dict(), dict()
images_ids_list, texts_lines_list, texts_pars_list, texts_lines_par_list, par_boxes_list, line_boxes_list, lines_par_boxes_list, images_list, images_pixels_list, page_no_list, num_pages_list = list(), list(), list(), list(), list(), list(), list(), list(), list(), list(), list()
try:
for i,image in enumerate(images):
# image preprocessing
# https://docs.opencv.org/3.0-beta/doc/py_tutorials/py_imgproc/py_thresholding/py_thresholding.html
img = image.copy()
factor, path_to_img = set_image_dpi_resize(img) # Rescaling to 300dpi
img = Image.open(path_to_img)
img = np.array(img, dtype='uint8') # convert PIL to cv2
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # gray scale image
ret,img = cv2.threshold(img,127,255,cv2.THRESH_BINARY)
# OCR PyTesseract | get langs of page
txt = pytesseract.image_to_string(img, config=custom_config)
txt = txt.strip().lower()
txt = re.sub(r" +", " ", txt) # multiple space
txt = re.sub(r"(\n\s*)+\n+", "\n", txt) # multiple line
# txt = os.popen(f'tesseract {img_filepath} - {custom_config}').read()
try:
langs = detect_langs(txt)
langs = [langdetect2Tesseract[langs[i].lang] for i in range(len(langs))]
langs_string = '+'.join(langs)
except:
langs_string = "eng"
langs_string += '+osd'
custom_config = f'--oem 3 --psm 3 -l {langs_string}' # default config PyTesseract: --oem 3 --psm 3
# OCR PyTesseract | get data
results[i] = pytesseract.image_to_data(img, config=custom_config, output_type=pytesseract.Output.DICT)
# results[i] = os.popen(f'tesseract {img_filepath} - {custom_config}').read()
# get image pixels
images_pixels[i] = feature_extractor(images[i], return_tensors="pt").pixel_values
texts_lines[i], texts_pars[i], texts_lines_par[i], row_indexes[i], par_boxes[i], line_boxes[i], lines_par_boxes[i] = get_data_paragraph(results[i], factor, conf_min=0)
texts_lines_list.append(texts_lines[i])
texts_pars_list.append(texts_pars[i])
texts_lines_par_list.append(texts_lines_par[i])
par_boxes_list.append(par_boxes[i])
line_boxes_list.append(line_boxes[i])
lines_par_boxes_list.append(lines_par_boxes[i])
images_ids_list.append(i)
images_pixels_list.append(images_pixels[i])
images_list.append(images[i])
page_no_list.append(i)
num_pages_list.append(num_imgs)
except:
print(f"There was an error within the extraction of PDF text by the OCR!")
else:
from datasets import Dataset
dataset = Dataset.from_dict({"images_ids": images_ids_list, "images": images_list, "images_pixels": images_pixels_list, "page_no": page_no_list, "num_pages": num_pages_list, "texts_line": texts_lines_list, "texts_par": texts_pars_list, "texts_lines_par": texts_lines_par_list, "bboxes_par": par_boxes_list, "bboxes_lines_par":lines_par_boxes_list})
# print(f"The text data was successfully extracted by the OCR!")
return dataset, texts_lines, texts_pars, texts_lines_par, row_indexes, par_boxes, line_boxes, lines_par_boxes
def prepare_inference_features_paragraph(example, tokenizer, max_length, cls_box, sep_box):
images_ids_list, chunks_ids_list, input_ids_list, attention_mask_list, bb_list, images_pixels_list = list(), list(), list(), list(), list(), list()
# get batch
# batch_page_hash = example["page_hash"]
batch_images_ids = example["images_ids"]
batch_images = example["images"]
batch_images_pixels = example["images_pixels"]
batch_bboxes_par = example["bboxes_par"]
batch_texts_par = example["texts_par"]
batch_images_size = [image.size for image in batch_images]
batch_width, batch_height = [image_size[0] for image_size in batch_images_size], [image_size[1] for image_size in batch_images_size]
# add a dimension if not a batch but only one image
if not isinstance(batch_images_ids, list):
batch_images_ids = [batch_images_ids]
batch_images = [batch_images]
batch_images_pixels = [batch_images_pixels]
batch_bboxes_par = [batch_bboxes_par]
batch_texts_par = [batch_texts_par]
batch_width, batch_height = [batch_width], [batch_height]
# process all images of the batch
for num_batch, (image_id, image_pixels, boxes, texts_par, width, height) in enumerate(zip(batch_images_ids, batch_images_pixels, batch_bboxes_par, batch_texts_par, batch_width, batch_height)):
tokens_list = []
bboxes_list = []
# add a dimension if only on image
if not isinstance(texts_par, list):
texts_par, boxes = [texts_par], [boxes]
# convert boxes to original
normalize_bboxes_par = [normalize_box(upperleft_to_lowerright(box), width, height) for box in boxes]
# sort boxes with texts
# we want sorted lists from top to bottom of the image
boxes, texts_par = sort_data_wo_labels(normalize_bboxes_par, texts_par)
count = 0
for box, text_par in zip(boxes, texts_par):
tokens_par = tokenizer.tokenize(text_par)
num_tokens_par = len(tokens_par) # get number of tokens
tokens_list.extend(tokens_par)
bboxes_list.extend([box] * num_tokens_par) # number of boxes must be the same as the number of tokens
# use of return_overflowing_tokens=True / stride=doc_stride
# to get parts of image with overlap
# source: https://huggingface.co/course/chapter6/3b?fw=tf#handling-long-contexts
encodings = tokenizer(" ".join(texts_par),
truncation=True,
padding="max_length",
max_length=max_length,
stride=doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True
)
otsm = encodings.pop("overflow_to_sample_mapping")
offset_mapping = encodings.pop("offset_mapping")
# Let's label those examples and get their boxes
sequence_length_prev = 0
for i, offsets in enumerate(offset_mapping):
# truncate tokens, boxes and labels based on length of chunk - 2 (special tokens <s> and </s>)
sequence_length = len(encodings.input_ids[i]) - 2
if i == 0: start = 0
else: start += sequence_length_prev - doc_stride
end = start + sequence_length
sequence_length_prev = sequence_length
# get tokens, boxes and labels of this image chunk
bb = [cls_box] + bboxes_list[start:end] + [sep_box]
# as the last chunk can have a length < max_length
# we must to add [tokenizer.pad_token] (tokens), [sep_box] (boxes) and [-100] (labels)
if len(bb) < max_length:
bb = bb + [sep_box] * (max_length - len(bb))
# append results
input_ids_list.append(encodings["input_ids"][i])
attention_mask_list.append(encodings["attention_mask"][i])
bb_list.append(bb)
images_ids_list.append(image_id)
chunks_ids_list.append(i)
images_pixels_list.append(image_pixels)
return {
"images_ids": images_ids_list,
"chunk_ids": chunks_ids_list,
"input_ids": input_ids_list,
"attention_mask": attention_mask_list,
"normalized_bboxes": bb_list,
"images_pixels": images_pixels_list
}
from torch.utils.data import Dataset
class CustomDataset(Dataset):
def __init__(self, dataset, tokenizer):
self.dataset = dataset
self.tokenizer = tokenizer
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
# get item
example = self.dataset[idx]
encoding = dict()
encoding["images_ids"] = example["images_ids"]
encoding["chunk_ids"] = example["chunk_ids"]
encoding["input_ids"] = example["input_ids"]
encoding["attention_mask"] = example["attention_mask"]
encoding["bbox"] = example["normalized_bboxes"]
encoding["images_pixels"] = example["images_pixels"]
return encoding
import torch.nn.functional as F
# get predictions at token level
def predictions_token_level(images, custom_encoded_dataset, model_id, model):
num_imgs = len(images)
if num_imgs > 0:
chunk_ids, input_ids, bboxes, pixels_values, outputs, token_predictions = dict(), dict(), dict(), dict(), dict(), dict()
images_ids_list = list()
for i,encoding in enumerate(custom_encoded_dataset):
# get custom encoded data
image_id = encoding['images_ids']
chunk_id = encoding['chunk_ids']
input_id = torch.tensor(encoding['input_ids'])[None]
attention_mask = torch.tensor(encoding['attention_mask'])[None]
bbox = torch.tensor(encoding['bbox'])[None]
pixel_values = torch.tensor(encoding["images_pixels"])
# save data in dictionnaries
if image_id not in images_ids_list: images_ids_list.append(image_id)
if image_id in chunk_ids: chunk_ids[image_id].append(chunk_id)
else: chunk_ids[image_id] = [chunk_id]
if image_id in input_ids: input_ids[image_id].append(input_id)
else: input_ids[image_id] = [input_id]
if image_id in bboxes: bboxes[image_id].append(bbox)
else: bboxes[image_id] = [bbox]
if image_id in pixels_values: pixels_values[image_id].append(pixel_values)
else: pixels_values[image_id] = [pixel_values]
# get prediction with forward pass
with torch.no_grad():
if model_id == model_id_lilt:
output = model(
input_ids=input_id.to(device),
attention_mask=attention_mask.to(device),
bbox=bbox.to(device),
)
elif model_id == model_id_layoutxlm:
output = model(
input_ids=input_id.to(device),
attention_mask=attention_mask.to(device),
bbox=bbox.to(device),
image=pixel_values.to(device)
)
# save probabilities of predictions in dictionnary
if image_id in outputs: outputs[image_id].append(F.softmax(output.logits.squeeze(), dim=-1))
else: outputs[image_id] = [F.softmax(output.logits.squeeze(), dim=-1)]
return outputs, images_ids_list, chunk_ids, input_ids, bboxes
else:
print("An error occurred while getting predictions!")
from functools import reduce
# Get predictions (paragraph level)
def predictions_probs_paragraph_level(max_length, tokenizer, id2label, dataset, outputs, images_ids_list, chunk_ids, input_ids, bboxes, cls_box, sep_box):
ten_probs_dict, ten_input_ids_dict, ten_bboxes_dict = dict(), dict(), dict()
bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, df = dict(), dict(), dict(), dict()
if len(images_ids_list) > 0:
for i, image_id in enumerate(images_ids_list):
# get image information
images_list = dataset.filter(lambda example: example["images_ids"] == image_id)["images"]
image = images_list[0]
width, height = image.size
# get data
chunk_ids_list = chunk_ids[image_id]
outputs_list = outputs[image_id]
input_ids_list = input_ids[image_id]
bboxes_list = bboxes[image_id]
# create zeros tensors
ten_probs = torch.zeros((outputs_list[0].shape[0] - 2)*len(outputs_list), outputs_list[0].shape[1])
ten_input_ids = torch.ones(size=(1, (outputs_list[0].shape[0] - 2)*len(outputs_list)), dtype =int)
ten_bboxes = torch.zeros(size=(1, (outputs_list[0].shape[0] - 2)*len(outputs_list), 4), dtype =int)
if len(outputs_list) > 1:
for num_output, (output, input_id, bbox) in enumerate(zip(outputs_list, input_ids_list, bboxes_list)):
start = num_output*(max_length - 2) - max(0,num_output)*doc_stride
end = start + (max_length - 2)
if num_output == 0:
ten_probs[start:end,:] += output[1:-1]
ten_input_ids[:,start:end] = input_id[:,1:-1]
ten_bboxes[:,start:end,:] = bbox[:,1:-1,:]
else:
ten_probs[start:start + doc_stride,:] += output[1:1 + doc_stride]
ten_probs[start:start + doc_stride,:] = ten_probs[start:start + doc_stride,:] * 0.5
ten_probs[start + doc_stride:end,:] += output[1 + doc_stride:-1]
ten_input_ids[:,start:start + doc_stride] = input_id[:,1:1 + doc_stride]
ten_input_ids[:,start + doc_stride:end] = input_id[:,1 + doc_stride:-1]
ten_bboxes[:,start:start + doc_stride,:] = bbox[:,1:1 + doc_stride,:]
ten_bboxes[:,start + doc_stride:end,:] = bbox[:,1 + doc_stride:-1,:]
else:
ten_probs += outputs_list[0][1:-1]
ten_input_ids = input_ids_list[0][:,1:-1]
ten_bboxes = bboxes_list[0][:,1:-1]
ten_probs_list, ten_input_ids_list, ten_bboxes_list = ten_probs.tolist(), ten_input_ids.tolist()[0], ten_bboxes.tolist()[0]
bboxes_list = list()
input_ids_dict, probs_dict = dict(), dict()
bbox_prev = [-100, -100, -100, -100]
for probs, input_id, bbox in zip(ten_probs_list, ten_input_ids_list, ten_bboxes_list):
bbox = denormalize_box(bbox, width, height)
if bbox != bbox_prev and bbox != cls_box and bbox != sep_box and bbox[0] != bbox[2] and bbox[1] != bbox[3]:
bboxes_list.append(bbox)
input_ids_dict[str(bbox)] = [input_id]
probs_dict[str(bbox)] = [probs]
elif bbox != cls_box and bbox != sep_box and bbox[0] != bbox[2] and bbox[1] != bbox[3]:
input_ids_dict[str(bbox)].append(input_id)
probs_dict[str(bbox)].append(probs)
bbox_prev = bbox
probs_bbox = dict()
for i,bbox in enumerate(bboxes_list):
probs = probs_dict[str(bbox)]
probs = np.array(probs).T.tolist()
probs_label = list()
for probs_list in probs:
prob_label = reduce(lambda x, y: x*y, probs_list)
prob_label = prob_label**(1./(len(probs_list))) # normalization
probs_label.append(prob_label)
# max_value = max(probs_label)
# max_index = probs_label.index(max_value)
# probs_bbox[str(bbox)] = max_index
probs_bbox[str(bbox)] = probs_label
bboxes_list_dict[image_id] = bboxes_list
input_ids_dict_dict[image_id] = input_ids_dict
probs_dict_dict[image_id] = probs_bbox
# df[image_id] = pd.DataFrame()
# df[image_id]["bboxes"] = bboxes_list
# df[image_id]["texts"] = [tokenizer.decode(input_ids_dict[str(bbox)]) for bbox in bboxes_list]
# df[image_id]["labels"] = [id2label[probs_bbox[str(bbox)]] for bbox in bboxes_list]
return probs_bbox, bboxes_list_dict, input_ids_dict_dict, probs_dict_dict #, df
else:
print("An error occurred while getting predictions!")
from functools import reduce
# Get predictions (paragraph level)
def predictions_paragraph_level(max_length, tokenizer1, id2label, dataset, outputs1, images_ids_list1, chunk_ids1, input_ids1, bboxes1, cls_box1, sep_box1, tokenizer2, outputs2, images_ids_list2, chunk_ids2, input_ids2, bboxes2, cls_box2, sep_box2):
bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, df = dict(), dict(), dict(), dict()
probs_bbox1, bboxes_list_dict1, input_ids_dict_dict1, probs_dict_dict1 = predictions_probs_paragraph_level(max_length, tokenizer1, id2label, dataset, outputs1, images_ids_list1, chunk_ids1, input_ids1, bboxes1, cls_box1, sep_box1)
probs_bbox2, bboxes_list_dict2, input_ids_dict_dict2, probs_dict_dict2 = predictions_probs_paragraph_level(max_length, tokenizer2, id2label, dataset, outputs2, images_ids_list2, chunk_ids2, input_ids2, bboxes2, cls_box2, sep_box2)
if len(images_ids_list1) > 0:
for i, image_id in enumerate(images_ids_list1):
bboxes_list1 = bboxes_list_dict1[image_id]
input_ids_dict1 = input_ids_dict_dict1[image_id]
probs_bbox1 = probs_dict_dict1[image_id]
bboxes_list2 = bboxes_list_dict2[image_id]
input_ids_dict2 = input_ids_dict_dict2[image_id]
probs_bbox2 = probs_dict_dict2[image_id]
probs_bbox = dict()
for bbox in bboxes_list1:
prob_bbox = [(p1+p2)/2 for p1,p2 in zip(probs_bbox1[str(bbox)], probs_bbox2[str(bbox)])]
max_value = max(prob_bbox)
max_index = prob_bbox.index(max_value)
probs_bbox[str(bbox)] = max_index
bboxes_list_dict[image_id] = bboxes_list1
input_ids_dict_dict[image_id] = input_ids_dict1
probs_dict_dict[image_id] = probs_bbox
df[image_id] = pd.DataFrame()
df[image_id]["bboxes"] = bboxes_list1
df[image_id]["texts"] = [tokenizer1.decode(input_ids_dict1[str(bbox)]) for bbox in bboxes_list1]
df[image_id]["labels"] = [id2label[probs_bbox[str(bbox)]] for bbox in bboxes_list1]
return bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, df
else:
print("An error occurred while getting predictions!")
# Get labeled images with lines bounding boxes
def get_labeled_images(id2label, dataset, images_ids_list, bboxes_list_dict, probs_dict_dict):
labeled_images = list()
for i, image_id in enumerate(images_ids_list):
# get image
images_list = dataset.filter(lambda example: example["images_ids"] == image_id)["images"]
image = images_list[0]
width, height = image.size
# get predicted boxes and labels
bboxes_list = bboxes_list_dict[image_id]
probs_bbox = probs_dict_dict[image_id]
draw = ImageDraw.Draw(image)
# https://stackoverflow.com/questions/66274858/choosing-a-pil-imagefont-by-font-name-rather-than-filename-and-cross-platform-f
font = font_manager.FontProperties(family='sans-serif', weight='bold')
font_file = font_manager.findfont(font)
font_size = 30
font = ImageFont.truetype(font_file, font_size)
for bbox in bboxes_list:
predicted_label = id2label[probs_bbox[str(bbox)]]
draw.rectangle(bbox, outline=label2color[predicted_label])
draw.text((bbox[0] + 10, bbox[1] - font_size), text=predicted_label, fill=label2color[predicted_label], font=font)
labeled_images.append(image)
return labeled_images
# get data of encoded chunk
def get_encoded_chunk_inference(tokenizer, dataset, encoded_dataset, index_chunk=None):
# get datasets
example = dataset
encoded_example = encoded_dataset
# get randomly a document in dataset
if index_chunk == None: index_chunk = random.randint(0, len(encoded_example)-1)
encoded_example = encoded_example[index_chunk]
encoded_image_ids = encoded_example["images_ids"]
# get the image
example = example.filter(lambda example: example["images_ids"] == encoded_image_ids)[0]
image = example["images"] # original image
width, height = image.size
page_no = example["page_no"]
num_pages = example["num_pages"]
# get boxes, texts, categories
bboxes, input_ids = encoded_example["normalized_bboxes"][1:-1], encoded_example["input_ids"][1:-1]
bboxes = [denormalize_box(bbox, width, height) for bbox in bboxes]
num_tokens = len(input_ids) + 2
# get unique bboxes and corresponding labels
bboxes_list, input_ids_list = list(), list()
input_ids_dict = dict()
bbox_prev = [-100, -100, -100, -100]
for i, (bbox, input_id) in enumerate(zip(bboxes, input_ids)):
if bbox != bbox_prev:
bboxes_list.append(bbox)
input_ids_dict[str(bbox)] = [input_id]
else:
input_ids_dict[str(bbox)].append(input_id)
# start_indexes_list.append(i)
bbox_prev = bbox
# do not keep "</s><pad><pad>..."
if input_ids_dict[str(bboxes_list[-1])][0] == (tokenizer.convert_tokens_to_ids('</s>')):
del input_ids_dict[str(bboxes_list[-1])]
bboxes_list = bboxes_list[:-1]
# get texts by line
input_ids_list = input_ids_dict.values()
texts_list = [tokenizer.decode(input_ids) for input_ids in input_ids_list]
# display DataFrame
df = pd.DataFrame({"texts": texts_list, "input_ids": input_ids_list, "bboxes": bboxes_list})
return image, df, num_tokens, page_no, num_pages
# display chunk of PDF image and its data
def display_chunk_lines_inference(dataset, encoded_dataset, index_chunk=None):
# get image and image data
image, df, num_tokens, page_no, num_pages = get_encoded_chunk_inference(dataset, encoded_dataset, index_chunk=index_chunk)
# get data from dataframe
input_ids = df["input_ids"]
texts = df["texts"]
bboxes = df["bboxes"]
print(f'Chunk ({num_tokens} tokens) of the PDF (page: {page_no+1} / {num_pages})\n')
# display image with bounding boxes
print(">> PDF image with bounding boxes of lines\n")
draw = ImageDraw.Draw(image)
labels = list()
for box, text in zip(bboxes, texts):
color = "red"
draw.rectangle(box, outline=color)
# resize image to original
width, height = image.size
image = image.resize((int(0.5*width), int(0.5*height)))
# convert to cv and display
img = np.array(image, dtype='uint8') # PIL to cv2
cv2_imshow(img)
cv2.waitKey(0)
# display image dataframe
print("\n>> Dataframe of annotated lines\n")
cols = ["texts", "bboxes"]
df = df[cols]
display(df)
# APP outputs by model
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, texts_lines, texts_pars, texts_lines_par, row_indexes, par_boxes, line_boxes, lines_par_boxes = extraction_data_from_image(images)
# prepare our data in the format of the model
# model1
prepare_inference_features_partial1 = partial(prepare_inference_features_paragraph, tokenizer=tokenizer1, max_length=max_length, cls_box=cls_box1, sep_box=sep_box1)
encoded_dataset1 = dataset.map(prepare_inference_features_partial1, batched=True, batch_size=64, remove_columns=dataset.column_names)
custom_encoded_dataset1 = CustomDataset(encoded_dataset1, tokenizer1)
# model2
prepare_inference_features_partial2 = partial(prepare_inference_features_paragraph, tokenizer=tokenizer2, max_length=max_length, cls_box=cls_box2, sep_box=sep_box2)
encoded_dataset2 = dataset.map(prepare_inference_features_partial2, batched=True, batch_size=64, remove_columns=dataset.column_names)
custom_encoded_dataset2 = CustomDataset(encoded_dataset2, tokenizer2)
# Get predictions (token level)
# model1
outputs1, images_ids_list1, chunk_ids1, input_ids1, bboxes1 = predictions_token_level(images, custom_encoded_dataset1, model_id1, model1)
# model2
outputs2, images_ids_list2, chunk_ids2, input_ids2, bboxes2 = predictions_token_level(images, custom_encoded_dataset2, model_id2, model2)
# Get predictions (paragraph level)
bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, df = predictions_paragraph_level(max_length, tokenizer1, id2label, dataset, outputs1, images_ids_list1, chunk_ids1, input_ids1, bboxes1, cls_box1, sep_box1, tokenizer2, outputs2, images_ids_list2, chunk_ids2, input_ids2, bboxes2, cls_box2, sep_box2)
# Get labeled images with lines bounding boxes
images = get_labeled_images(id2label, dataset, images_ids_list1, 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")
img_file = img_file.replace("/", "_")
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_file = csv_file.replace("/", "_")
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]
# get device
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
## model LiLT
import transformers
from transformers import AutoTokenizer, AutoModelForTokenClassification
tokenizer_lilt = AutoTokenizer.from_pretrained(model_id_lilt)
model_lilt = AutoModelForTokenClassification.from_pretrained(model_id_lilt);
model_lilt.to(device);
tokenizer1 = tokenizer_lilt
model1 = model_lilt
## model LayoutXLM
from transformers import LayoutLMv2ForTokenClassification # LayoutXLMTokenizerFast,
model_layoutxlm = LayoutLMv2ForTokenClassification.from_pretrained(model_id_layoutxlm);
model_layoutxlm.to(device);
# feature extractor
from transformers import LayoutLMv2FeatureExtractor
feature_extractor = LayoutLMv2FeatureExtractor(apply_ocr=False)
# tokenizer
from transformers import AutoTokenizer
tokenizer_layoutxlm = AutoTokenizer.from_pretrained(tokenizer_id_layoutxlm)
tokenizer2 = tokenizer_layoutxlm
model2 = model_layoutxlm
Downloading (…)okenizer_config.json: 0%| | 0.00/463 [00:00<?, ?B/s]
Downloading tokenizer.json: 0%| | 0.00/17.1M [00:00<?, ?B/s]
Downloading (…)cial_tokens_map.json: 0%| | 0.00/280 [00:00<?, ?B/s]
Downloading (…)lve/main/config.json: 0%| | 0.00/1.27k [00:00<?, ?B/s]
Downloading pytorch_model.bin: 0%| | 0.00/1.13G [00:00<?, ?B/s]
Downloading (…)lve/main/config.json: 0%| | 0.00/3.32k [00:00<?, ?B/s]
Downloading pytorch_model.bin: 0%| | 0.00/1.48G [00:00<?, ?B/s]
/usr/local/lib/python3.9/dist-packages/transformers/models/layoutlmv2/feature_extraction_layoutlmv2.py:30: FutureWarning: The class LayoutLMv2FeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please use LayoutLMv2ImageProcessor instead. warnings.warn(
Downloading (…)lve/main/config.json: 0%| | 0.00/615 [00:00<?, ?B/s]
Downloading (…)tencepiece.bpe.model: 0%| | 0.00/5.07M [00:00<?, ?B/s]
Downloading (…)/main/tokenizer.json: 0%| | 0.00/9.10M [00:00<?, ?B/s]
# get labels
id2label = model_lilt.config.id2label
label2id = model_lilt.config.label2id
num_labels = len(id2label)
# Gradio APP
with gr.Blocks(title='Inference APP for Document Understanding at paragraph level (v3 - Ensemble "LiLT + LayoutXLM" base)', css=".gradio-container") as demo:
gr.HTML("""
<div style="font-family:'Times New Roman', 'Serif'; font-size:26pt; font-weight:bold; text-align:center;"><h1>Inference APP for Document Understanding at paragraph level (v3 - Ensemble "LiLT + LayoutXLM" base)</h1></div>
<div style="margin-top: 40px"><p>(04/04/2023) This Inference APP uses an ensemble of 2 Document Understanding models finetuned on the dataset DocLayNet base at paragraph level (chunk size of 512 tokens) and combined with XLM-RoBERTa base: <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/pierreguillou/lilt-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512" target="_blank">LiLT base</a> and <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/pierreguillou/layout-xlm-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512" target="_blank">LayoutXLM base</a>.</p><p>This ensemble calculates the probabilities of each block from the outputs of the models for each label before selecting the label with the highest sum of the normalized probabilities.</p>
<p>Note: <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://arxiv.org/abs/2202.13669" target="_blank">LiLT (Language-Independent Layout Transformer)</a> and <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://arxiv.org/abs/2104.08836" target="_blank">LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding</a> are Document Understanding models that use both layout and text in order to detect labels of bounding boxes. Combined with the model <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/xlm-roberta-base" target="_blank">XML-RoBERTa base</a>, this finetuned model has the capacity to <b>understand any language</b>. Finetuned on the dataset <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/datasets/pierreguillou/DocLayNet-base" target="_blank">DocLayNet base</a>, they can <b>classifly any bounding box (and its OCR text) to 11 labels</b> (Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, Title).</p>
<p>They rely 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 (<a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://github.com/madmaze/pytesseract#python-tesseract" target="_blank">PyTesseract</a>) to get the bounding boxes, then run the 2 models (already fine-tuned on the dataset DocLayNet base at paragraph level) on the individual tokens and then, normalized the sum of block probabilities as explained, and visualize the result at paragraph level!</p>
<p><b>It allows to get all pages of any PDF (of any language) with bounding boxes labeled at paragraph level and the associated dataframes with labeled data (bounding boxes, texts, labels) :-)</b></p></div>
<div><p>However, the inference time per page can be high when running the model on CPU due to the number of paragraph predictions to be made. Therefore, to avoid running this APP for too long, <b>only the first 2 pages are processed by this APP</b>. If you want to increase this limit, you can either clone this APP in Hugging Face Space (or run its <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://github.com/piegu/language-models/blob/master/Gradio_inference_on_Ensemble_LiLT_&_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb" target="_blank">notebook</a> on your own plateform) and change the value of the parameter <code>max_imgboxes</code>, or run the inference notebook "<a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://github.com/piegu/language-models/blob/master/inference_on_Ensemble_LiLT_&_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb" target="_blank">Document AI | Inference at paragraph level by using the association of 2 Document Understanding models (LiLT and LayoutXLM base fine-tuned on DocLayNet base dataset)</a>" on your own platform as it does not have this limit.</p></div>
<div style="margin-top: 20px"><p>Links to Document Understanding APPs:</p><ul><li>Line level: <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-linelevel-v1" target="_blank">v1 (LiLT base)</a> | <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-linelevel-v2" target="_blank">v2 (LayoutXLM base)</a> | <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-linelevel-v3" target="_blank">v3 (LilT base vs LayoutXLM base)</a></li><li>Paragraph level: <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-paragraphlevel-v1" target="_blank">v1 (LiLT base)</a> | <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-paragraphlevel-v2" target="_blank">v2 (LayoutXLM base)</a> | <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://huggingface.co/spaces/pierreguillou/Inference-APP-Document-Understanding-at-paragraphlevel-v3" target="_blank">v3 (LilT base vs LayoutXLM base)</a></li></ul></div>
<div style="margin-top: 20px"><p>More information about the DocLayNet datasets, the finetuning of the model and this APP in the following blog posts:</p>
<ul><li>(03/31/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-inference-app-and-fine-tuning-notebook-for-document-understanding-at-paragraph-level-3507af80573d" target="_blank">Document AI | Inference APP and fine-tuning notebook for Document Understanding at paragraph level with LayoutXLM base</a></li><li>(03/25/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-app-to-compare-the-document-understanding-lilt-and-layoutxlm-base-models-at-line-1c53eb481a15" target="_blank">Document AI | APP to compare the Document Understanding LiLT and LayoutXLM (base) models at line level</a></li><li>(03/05/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-inference-app-and-fine-tuning-notebook-for-document-understanding-at-line-level-with-b08fdca5f4dc" target="_blank">Document AI | Inference APP and fine-tuning notebook for Document Understanding at line level with LayoutXLM base</a></li><li>(02/14/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-inference-app-for-document-understanding-at-line-level-a35bbfa98893" target="_blank">Document AI | Inference APP for Document Understanding at line level</a></li><li>(02/10/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-document-understanding-model-at-line-level-with-lilt-tesseract-and-doclaynet-dataset-347107a643b8" target="_blank">Document AI | Document Understanding model at line level with LiLT, Tesseract and DocLayNet dataset</a></li><li>(01/31/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-doclaynet-image-viewer-app-3ac54c19956" target="_blank">Document AI | DocLayNet image viewer APP</a></li><li>(01/27/2023) <a style="text-decoration: none; border-bottom: #64b5f6 0.125em solid; color: #64b5f6" href="https://medium.com/@pierre_guillou/document-ai-processing-of-doclaynet-dataset-to-be-used-by-layout-models-of-the-hugging-face-hub-308d8bd81cdb" target="_blank">Document AI | Processing of DocLayNet dataset to be used by layout models of the Hugging Face hub (finetuning, inference)</a></li></ul></div>
""")
with gr.Row():
pdf_file = gr.File(label="PDF")
with gr.Row():
submit_btn = gr.Button(f"Display first {max_imgboxes} labeled PDF pages")
reset_btn = gr.Button(value="Clear")
with gr.Row():
output_msg = gr.Textbox(label="Output message")
with gr.Row():
fileboxes = []
for num_page in range(max_imgboxes):
file_path = gr.File(visible=True, label=f"Image file of the PDF page n°{num_page}")
fileboxes.append(file_path)
with gr.Row():
imgboxes = []
for num_page in range(max_imgboxes):
img = gr.Image(type="pil", label=f"Image of the PDF page n°{num_page}")
imgboxes.append(img)
with gr.Row():
csvboxes = []
for num_page in range(max_imgboxes):
csv = gr.File(visible=True, label=f"CSV file at paragraph level (page {num_page})")
csvboxes.append(csv)
with gr.Row():
dfboxes = []
for num_page in range(max_imgboxes):
df = gr.Dataframe(
headers=["bounding boxes", "texts", "labels"],
datatype=["str", "str", "str"],
col_count=(3, "fixed"),
visible=True,
label=f"Data of page {num_page}",
type="pandas",
wrap=True
)
dfboxes.append(df)
outputboxes = [output_msg] + fileboxes + imgboxes + csvboxes + dfboxes
submit_btn.click(app_outputs, inputs=[pdf_file], outputs=outputboxes)
# https://github.com/gradio-app/gradio/pull/2044/files#diff-a91dd2749f68bb7d0099a0f4079a4fd2d10281e299e7b451cb1bb876a7c21975R91
reset_btn.click(
lambda: [pdf_file.update(value=None), output_msg.update(value=None)] + [filebox.update(value=None) for filebox in fileboxes] + [imgbox.update(value=None) for imgbox in imgboxes] + [csvbox.update(value=None) for csvbox in csvboxes] + [dfbox.update(value=None) for dfbox in dfboxes],
inputs=[],
outputs=[pdf_file, output_msg] + fileboxes + imgboxes + csvboxes + dfboxes
)
gr.Examples(
[["files/example.pdf"]],
[pdf_file],
outputboxes,
fn=app_outputs,
cache_examples=True,
)
demo.launch()
Caching examples at: '/content/gradio_cached_examples/22'
Map: 0%| | 0/2 [00:00<?, ? examples/s]
Map: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Filter: 0%| | 0/2 [00:00<?, ? examples/s]
Colab notebook detected. This cell will run indefinitely so that you can see errors and logs. To turn off, set debug=False in launch(). Note: opening Chrome Inspector may crash demo inside Colab notebooks. To create a public link, set `share=True` in `launch()`.
Keyboard interruption in main thread... closing server.