#!/usr/bin/env python
# coding: utf-8
# # Document AI | ACCURACY at paragraph level with a Document Understanding layout model (Layout XLM base) fine-tuned on DocLayNet dataset
# - Credit:
# - notebook created from the notebook [Fine_tuning_LayoutXLM_on_XFUND_for_token_classification_using_HuggingFace_Trainer.ipynb](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutXLM/Fine_tuning_LayoutXLM_on_XFUND_for_token_classification_using_HuggingFace_Trainer.ipynb)
# - dataset from IBM Research ([DocLayNet](https://github.com/DS4SD/DocLayNet))
# - Author of this notebook: [Pierre GUILLOU](https://www.linkedin.com/in/pierreguillou/)
# - Date: 22/04/2023
# - Blog posts:
# - Layout XLM base
# - (03/31/2023) [Document AI | Inference APP and fine-tuning notebook for Document Understanding at paragraph level with LayoutXLM base](https://medium.com/@pierre_guillou/document-ai-inference-app-and-fine-tuning-notebook-for-document-understanding-at-paragraph-level-3507af80573d)
# - (03/25/2023) [Document AI | APP to compare the Document Understanding LiLT and LayoutXLM (base) models at line level](https://medium.com/@pierre_guillou/document-ai-app-to-compare-the-document-understanding-lilt-and-layoutxlm-base-models-at-line-1c53eb481a15)
# - (03/05/2023) [Document AI | Inference APP and fine-tuning notebook for Document Understanding at line level with LayoutXLM base](https://medium.com/@pierre_guillou/document-ai-inference-app-and-fine-tuning-notebook-for-document-understanding-at-line-level-with-b08fdca5f4dc)
# - LiLT base
# - (02/16/2023) [Document AI | Inference APP and fine-tuning notebook for Document Understanding at paragraph level](https://medium.com/@pierre_guillou/document-ai-inference-app-and-fine-tuning-notebook-for-document-understanding-at-paragraph-level-c18d16e53cf8)
# - (02/14/2023) [Document AI | Inference APP for Document Understanding at line level](https://medium.com/@pierre_guillou/document-ai-inference-app-for-document-understanding-at-line-level-a35bbfa98893)
# - (02/10/2023) [Document AI | Document Understanding model at line level with LiLT, Tesseract and DocLayNet dataset](https://medium.com/@pierre_guillou/document-ai-document-understanding-model-at-line-level-with-lilt-tesseract-and-doclaynet-dataset-347107a643b8)
# - (01/31/2023) [Document AI | DocLayNet image viewer APP](https://medium.com/@pierre_guillou/document-ai-doclaynet-image-viewer-app-3ac54c19956)
# - (01/27/2023) [Document AI | Processing of DocLayNet dataset to be used by layout models of the Hugging Face hub (finetuning, inference)](https://medium.com/@pierre_guillou/document-ai-processing-of-doclaynet-dataset-to-be-used-by-layout-models-of-the-hugging-face-hub-308d8bd81cdb)
# - Notebooks (paragraph level)
# - Layout XLM base
# - [Document AI | Inference at paragraph level with a Document Understanding model (LayoutXLM base fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/inference_on_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb)
# - [Document AI | Inference APP at paragraph level with a Document Understanding model (LayoutXLM base fine-tuned on DocLayNet base dataset)](https://github.com/piegu/language-models/blob/master/Gradio_inference_on_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb)
# - [Document AI | Fine-tune LayoutXLM base on DocLayNet base in any language at paragraph level (chunk of 512 tokens with overlap)](https://github.com/piegu/language-models/blob/master/Fine_tune_LayoutXLM_base_on_DocLayNet_base_in_any_language_at_paragraphlevel_ml_512.ipynb)
# - LiLT base
# - [Document AI | Inference APP at paragraph level with a Document Understanding model (LiLT fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/Gradio_inference_on_LiLT_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb)
# - [Document AI | Inference at paragraph level with a Document Understanding model (LiLT fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/inference_on_LiLT_model_finetuned_on_DocLayNet_base_in_any_language_at_levelparagraphs_ml512.ipynb)
# - [Document AI | Fine-tune LiLT on DocLayNet base in any language at paragraph level (chunk of 512 tokens with overlap)](https://github.com/piegu/language-models/blob/master/Fine_tune_LiLT_on_DocLayNet_base_in_any_language_at_paragraphlevel_ml_512.ipynb)
# - Notebooks (line level)
# - Layout XLM base
# - [Document AI | Inference at line level with a Document Understanding model (LayoutXLM base fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/inference_on_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levellines_ml384.ipynb)
# - [Document AI | Inference APP at line level with a Document Understanding model (LayoutXLM base fine-tuned on DocLayNet base dataset)](https://github.com/piegu/language-models/blob/master/Gradio_inference_on_LayoutXLM_base_model_finetuned_on_DocLayNet_base_in_any_language_at_levellines_ml384.ipynb)
# - [Document AI | Fine-tune LayoutXLM base on DocLayNet base in any language at line level (chunk of 384 tokens with overlap)](https://github.com/piegu/language-models/blob/master/Fine_tune_LayoutXLM_base_on_DocLayNet_base_in_any_language_at_linelevel_ml_384.ipynb)
# - LiLT base
# - [Document AI | Inference at line level with a Document Understanding model (LiLT fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/inference_on_LiLT_model_finetuned_on_DocLayNet_base_in_any_language_at_levellines_ml384.ipynb)
# - [Document AI | Inference APP at line level with a Document Understanding model (LiLT fine-tuned on DocLayNet dataset)](https://github.com/piegu/language-models/blob/master/Gradio_inference_on_LiLT_model_finetuned_on_DocLayNet_base_in_any_language_at_levellines_ml384.ipynb)
# - [Document AI | Fine-tune LiLT on DocLayNet base in any language at line level (chunk of 384 tokens with overlap)](https://github.com/piegu/language-models/blob/master/Fine_tune_LiLT_on_DocLayNet_base_in_any_language_at_linelevel_ml_384.ipynb)
# - [DocLayNet image viewer APP](https://github.com/piegu/language-models/blob/master/DocLayNet_image_viewer_APP.ipynb)
# - [Processing of DocLayNet dataset to be used by layout models of the Hugging Face hub (finetuning, inference)](processing_DocLayNet_dataset_to_be_used_by_layout_models_of_HF_hub.ipynb)
# ## Inference at paragraph level
# ### layoutXLM
# LayoutXLM was proposed in [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) 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](https://github.com/madmaze/pytesseract#python-tesseract)) 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](https://huggingface.co/pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512)) on the individual tokens and visualize the result at paragraph level!
# ## Install Detectron 2, Pypdfium2, LangDetect & PyTesseract OCR
# ### Detectron 2
# In[ ]:
get_ipython().run_cell_magic('capture', '', "!pip install -q torch==1.10.0+cu111 torchvision==0.11+cu111 -f https://download.pytorch.org/whl/torch_stable.html\n!python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'\n")
# ### Pypdfium2
# In[ ]:
get_ipython().run_cell_magic('capture', '', '# !sudo apt-get install poppler-utils\n# !pip install pdf2image\n\n# source: https://levelup.gitconnected.com/4-python-libraries-to-convert-pdf-to-images-7a09eba83a09\n# source: https://pypi.org/project/pypdfium2/\n!pip install -U pypdfium2\n')
# ### LangDetect
# In[ ]:
get_ipython().run_cell_magic('capture', '', '!pip install -q langdetect\n')
# ### Tesseract OCR
# In[ ]:
get_ipython().run_cell_magic('capture', '', "!sudo apt install tesseract-ocr-all # english + osd (Orientation and script detection module)\n# !sudo apt-get install tesseract-ocr-por # portuguese\n\n# import os\n# print(os.popen(f'cat /etc/debian_version').read())\n# print(os.popen(f'cat /etc/issue').read())\n# print(os.popen(f'apt search tesseract').read())\n\n!pip install pytesseract\n")
# ## Set-up environment
# In[ ]:
from google.colab import drive
drive.mount('/content/drive', force_remount=True)
# ### Libraries
# In[ ]:
get_ipython().system('pip install -q transformers sentencepiece datasets pypdf')
# In[ ]:
import os
from operator import itemgetter
import collections
import string
import re
import pypdf
from pypdf import PdfReader
from pypdf.errors import PdfReadError
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
import matplotlib.pyplot as plt
from matplotlib import font_manager
from PIL import Image, ImageDraw, ImageFont
font = ImageFont.load_default()
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 pathlib
from pathlib import Path
import shutil
from ipywidgets import widgets
from IPython.display import display, HTML
import transformers
import datasets
from datasets import concatenate_datasets
# ### Key parameters
# In[ ]:
# model
model_id = "pierreguillou/layout-xlm-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512" # layout
# model_id = "pierreguillou/lilt-xlm-roberta-base-finetuned-with-DocLayNet-base-at-paragraphlevel-ml512" # lilt
if "layout" in model_id:
# tokenizer
tokenizer_id = "xlm-roberta-base"
# In[ ]:
# 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'
}
domains = ["Financial Reports", "Manuals", "Scientific Articles", "Laws & Regulations", "Patents", "Government Tenders"]
domain_names = [domain_name.lower().replace(" ", "_").replace("&", "and") for domain_name in domains]
# bounding boxes start and end of a sequence
if "layout" in model_id:
cls_box = [0, 0, 0, 0]
sep_box = [1000, 1000, 1000, 1000]
elif "lilt" in model_id:
cls_box = [0, 0, 0, 0]
sep_box = cls_box
# DocLayNet dataset
# dataset_name = "pierreguillou/DocLayNet-small"
dataset_name = "pierreguillou/DocLayNet-base"
dataset_name_suffix = dataset_name.replace("pierreguillou/DocLayNet-", "")
# PAD token index
label_pad_token_id = -100
# parameters de TrainingArguments
batch_size=8 # WARNING: change this value according to your GPU RAM
# (tokenization) The maximum length of a feature (sequence)
if str(384) in model_id:
max_length = 384
elif str(512) in model_id:
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.
# ### Functions
# #### General
# In[ ]:
# 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
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
# In[ ]:
# 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
def get_blocks_evaluation(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, texts_lines_list, text_block_list = 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]
texts_lines_list.append(text_block) # list of text lines by 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)
lines_par_boxes_list = [[bboxes_block[group[j]] for j in range(len(group))]for group in list(bbox_block_dict.values())]
return bboxes_block_list, category_block_list, texts_lines_list, text_block_list, lines_par_boxes_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
# #### Dataset
# In[ ]:
# get PDF image and its data
def generate_annotated_image(index_image=None, split="all"):
# get dataset
example = dataset
# get split
if split == "all":
example = concatenate_datasets([example["train"], example["validation"], example["test"]])
else:
example = example[split]
# get random image & PDF data
if index_image == None: index_image = random.randint(0, len(example)-1)
example = example[index_image]
image = example["image"] # original image
coco_width, coco_height = example["coco_width"], example["coco_height"]
original_width, original_height = example["original_width"], example["original_height"]
original_filename = example["original_filename"]
page_no = example["page_no"]
num_pages = example["num_pages"]
# resize image to original
image = image.resize((original_width, original_height))
# get corresponding annotations
texts = example["texts"]
bboxes_block = example["bboxes_block"]
bboxes_line = example["bboxes_line"]
categories = example["categories"]
domain = example["doc_category"]
# get domain name
index_domain = domain_names.index(domain)
domain = domains[index_domain]
# convert boxes to original
original_bboxes_block = [original_box(convert_box(box), original_width, original_height, coco_width, coco_height) for box in bboxes_block]
original_bboxes_line = [original_box(convert_box(box), original_width, original_height, coco_width, coco_height) for box in bboxes_line]
##### block boxes #####
# get unique blocks and its data
bboxes_blocks_list, category_block_list, text_block_list = get_blocks(original_bboxes_block, categories, texts)
# sort data from y = 0 to end of page (and after, x=0 to end of page when necessary)
sorted_original_bboxes_block_list, sorted_category_block_list, sorted_text_block_list = sort_data(bboxes_blocks_list, category_block_list, text_block_list)
##### line boxes ####
# sort data from y = 0 to end of page (and after, x=0 to end of page when necessary)
sorted_original_bboxes_line_list, sorted_category_line_list, sorted_text_line_list = sort_data(original_bboxes_line, categories, texts)
# group paragraphs and lines outputs
sorted_original_bboxes = [sorted_original_bboxes_block_list, sorted_original_bboxes_line_list]
sorted_categories = [sorted_category_block_list, sorted_category_line_list]
sorted_texts = [sorted_text_block_list, sorted_text_line_list]
# get annotated boudings boxes on images
images = [image.copy(), image.copy()]
imgs, df_paragraphs, df_lines = dict(), pd.DataFrame(), pd.DataFrame()
for i, img in enumerate(images):
img = img.convert('RGB') # Convert to RGB
draw = ImageDraw.Draw(img)
for box, label_idx, text in zip(sorted_original_bboxes[i], sorted_categories[i], sorted_texts[i]):
label = id2label[label_idx]
color = label2color[label]
draw.rectangle(box, outline=color)
text = text.encode('latin-1', 'replace').decode('latin-1') # https://stackoverflow.com/questions/56761449/unicodeencodeerror-latin-1-codec-cant-encode-character-u2013-writing-to
draw.text((box[0] + 10, box[1] - 10), text=label, fill=color, font=font)
if i == 0:
imgs["paragraphs"] = img
df_paragraphs["paragraphs"] = list(range(len(sorted_original_bboxes_block_list)))
df_paragraphs["categories"] = [id2label[label_idx] for label_idx in sorted_category_block_list]
df_paragraphs["texts"] = sorted_text_block_list
df_paragraphs["bounding boxes"] = [str(bbox) for bbox in sorted_original_bboxes_block_list]
else:
imgs["lines"] = img
df_lines["lines"] = list(range(len(sorted_original_bboxes_line_list)))
df_lines["categories"] = [id2label[label_idx] for label_idx in sorted_category_line_list]
df_lines["texts"] = sorted_text_line_list
df_lines["bounding boxes"] = [str(bbox) for bbox in sorted_original_bboxes_line_list]
return imgs, original_filename, page_no, num_pages, domain, df_paragraphs, df_lines
# In[ ]:
# display PDF image and its data
def display_pdf_blocks_lines(index_image=None, split="all"):
# get image and image data
images, original_filename, page_no, num_pages, domain, df_paragraphs, df_lines = generate_annotated_image(index_image=index_image, split=split)
print(f"PDF: {original_filename} (page: {page_no+1} / {num_pages}; domain: {domain})\n")
# left widget
style1 = {'overflow': 'scroll' ,'white-space': 'nowrap', 'width':'50%'}
output1 = widgets.Output(description = "PDF image with bounding boxes of paragraphs", style=style1)
with output1:
# display image
print(">> PDF image with bounding boxes of paragraphs\n")
open_cv_image = np.array(images["paragraphs"]) # PIL to cv2
# Convert RGB to BGR
open_cv_image = open_cv_image[:, :, ::-1].copy()
# cv2.imshow('',open_cv_image) # lambda
cv2_imshow(open_cv_image) # Colab
cv2.waitKey(0)
# display DataFrame
print("\n>> Paragraphs dataframe\n")
display(df_paragraphs)
# right widget
style2 = style1
output2 = widgets.Output(description = "PDF image with bounding boxes of lines", style=style2)
with output2:
# display image
print(">> PDF image with bounding boxes of lines\n")
open_cv_image = np.array(images["lines"]) # PIL to cv2
# Convert RGB to BGR
open_cv_image = open_cv_image[:, :, ::-1].copy()
# cv2.imshow('',open_cv_image) # lambda
cv2_imshow(open_cv_image) # Colab
cv2.waitKey(0)
# display DataFrame
print("\n>> Lines dataframe\n")
display(df_lines)
## Side by side thanks to HBox widgets
sidebyside = widgets.HBox([output1,output2])
## Finally, show.
display(sidebyside)
# #### PDF processing
# In[ ]:
# get filename and images of PDF pages
def pdf_to_images(uploaded_pdf):
# file name of the uploaded PDF
filename = next(iter(uploaded_pdf))
try:
PdfReader(filename)
except PdfReadError:
print("Invalid PDF file.")
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
page_indices = [i for i in range(n_pages)] # 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)
print(f'The PDF "{filename}"" was converted into {num_imgs} images.')
print("Now, you can extract data from theses images (text, bounding boxes...).")
except:
print(f"Error with the PDF {filename}: it was not converted into images.")
print()
else:
# display images
if num_imgs > 0:
import matplotlib.pyplot as plt
get_ipython().run_line_magic('matplotlib', 'inline')
plt.figure(figsize=(20,10))
columns = 5
for i, image in enumerate(images):
plt.subplot(int(num_imgs / columns + 1), columns, i + 1)
plt.xticks(color="white")
plt.yticks(color="white")
plt.tick_params(bottom = False)
plt.tick_params(left = False)
plt.imshow(image)
return filename, images
# In[ ]:
# Extraction of image data (text and bounding boxes)
def extraction_data_from_dataset(example, model_id):
ids = example['id']
texts = example['texts']
bboxes_blocks = example['bboxes_block']
bboxes_lines = example['bboxes_line']
categories = example['categories']
images = example['image']
page_hashs = example['page_hash']
original_filenames = example['original_filename']
page_nos = example['page_no']
num_pages = example['num_pages']
original_widths = example['original_width']
original_heights = example['original_height']
coco_widths = example['coco_width']
coco_heights = example['coco_height']
collections = example['collection']
doc_categorys = example['doc_category']
num_imgs = len(images)
if num_imgs > 0:
# 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()
if "layout" in model_id:
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, category_block_list = list(), list(), list(), list(), list(), list(), list(), list(), list(), list(), list(), list()
elif "lilt" in model_id:
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, page_no_list, num_pages_list, category_block_list = list(), list(), list(), list(), list(), list(), list(), list(), list(), list(), list()
for i in range(num_imgs):
images_ids_list.append(i)
image = images[i].resize((original_widths[i], original_heights[i]))
images_list.append(image)
if "layout" in model_id:
images_pixels = feature_extractor(image, return_tensors="pt").pixel_values
images_pixels_list.append(images_pixels)
page_no_list.append(1)
num_pages_list.append(1)
# convert boxes to original
original_bboxes_block = [upperleft_to_lowerright(original_box(convert_box(box), original_widths[i], original_heights[i], coco_widths[i], coco_heights[i])) for box in bboxes_blocks[i]]
# original_bboxes_line = [upperleft_to_lowerright(original_box(convert_box(box), original_widths[i], original_heights[i], coco_widths[i], coco_heights[i]) for box in bboxes_lines[i]]
##### block boxes #####
# get unique blocks and its data
bboxes_blocks_list_img, category_block_list_img, texts_lines_par_list_img, text_block_list_img, lines_par_boxes_list_img = get_blocks_evaluation(original_bboxes_block, categories[i], texts[i])
# bboxes_block_list, category_block_list, texts_lines_list, text_block_list, lines_par_boxes_list
# sort data from y = 0 to end of page (and after, x=0 to end of page when necessary)
sorted_original_bboxes_block_list_img, sorted_category_block_list_img, sorted_text_block_list_img = sort_data(bboxes_blocks_list_img, category_block_list_img, text_block_list_img)
# texts_lines_par_list.append(texts_lines_par_list_img)
texts_pars_list.append(sorted_text_block_list_img)
par_boxes_list.append(sorted_original_bboxes_block_list_img) # bboxes_block_list
# lines_par_boxes_list.append(lines_par_boxes_list_img)
category_block_list.append(sorted_category_block_list_img)
# except:
# print(f"There was an error within the extraction of PDF text by the OCR!")
from datasets import Dataset
if "layout" in model_id:
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_par": texts_pars_list, "bboxes_par": par_boxes_list, "category_block_list": category_block_list})
elif "lilt" in model_id:
dataset = Dataset.from_dict({"images_ids": images_ids_list, "images": images_list, "page_no": page_no_list, "num_pages": num_pages_list, "texts_par": texts_pars_list, "bboxes_par": par_boxes_list, "category_block_list": category_block_list})
# print(f"The text data was successfully extracted by the OCR!")
return dataset, texts_pars_list, par_boxes_list, category_block_list
# #### Inference
# In[ ]:
def prepare_inference_features_paragraph(example, model_id=model_id, cls_box = cls_box, sep_box = sep_box):
if "layout" in model_id:
images_ids_list, chunks_ids_list, input_ids_list, attention_mask_list, bb_list, cc_list, images_pixels_list = list(), list(), list(), list(), list(), list(), list()
elif "lilt" in model_id:
images_ids_list, chunks_ids_list, input_ids_list, attention_mask_list, bb_list, cc_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"]
if "layout" in model_id:
batch_images_pixels = example["images_pixels"]
batch_bboxes_par = example["bboxes_par"]
batch_texts_par = example["texts_par"]
batch_categories_blocks = example["category_block_list"]
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]
if "layout" in model_id:
batch_images_pixels = [batch_images_pixels]
batch_bboxes_par = [batch_bboxes_par]
batch_texts_par = [batch_texts_par]
batch_categories_blocks = [batch_categories_blocks]
batch_width, batch_height = [batch_width], [batch_height]
# process all images of the batch
if "layout" in model_id:
for num_batch, (image_id, image_pixels, boxes, texts_par, width, height, categories_blocks) in enumerate(zip(batch_images_ids, batch_images_pixels, batch_bboxes_par, batch_texts_par, batch_width, batch_height, batch_categories_blocks)):
tokens_list = []
bboxes_list = []
categories_blocks_list = []
# add a dimension if only on image
if not isinstance(texts_par, list):
texts_par, boxes, categories_blocks = [texts_par], [boxes], [categories_blocks]
# 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)
boxes, categories_blocks, texts_par = sort_data(normalize_bboxes_par, categories_blocks, texts_par)
count = 0
for box, category_block, text_par in zip(boxes, categories_blocks, 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
categories_blocks_list.extend([category_block] * num_tokens_par)
# 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 and )
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]
cc = [-100] + categories_blocks_list[start:end] + [-100]
# 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))
cc = cc + [-100] * (max_length - len(cc))
# append results
input_ids_list.append(encodings["input_ids"][i])
attention_mask_list.append(encodings["attention_mask"][i])
bb_list.append(bb)
cc_list.append(cc)
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,
"category_blocks": cc_list
}
elif "lilt" in model_id:
for num_batch, (image_id, boxes, texts_par, width, height, categories_blocks) in enumerate(zip(batch_images_ids, batch_bboxes_par, batch_texts_par, batch_width, batch_height, batch_categories_blocks)):
tokens_list = []
bboxes_list = []
categories_blocks_list = []
# add a dimension if only on image
if not isinstance(texts_par, list):
texts_par, boxes, categories_blocks = [texts_par], [boxes], [categories_blocks]
# 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)
boxes, categories_blocks, texts_par = sort_data(normalize_bboxes_par, categories_blocks, texts_par)
count = 0
for box, category_block, text_par in zip(boxes, categories_blocks, 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
categories_blocks_list.extend([category_block] * num_tokens_par)
# 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 and )
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]
cc = [-100] + categories_blocks_list[start:end] + [-100]
# 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))
cc = cc + [-100] * (max_length - len(cc))
# append results
input_ids_list.append(encodings["input_ids"][i])
attention_mask_list.append(encodings["attention_mask"][i])
bb_list.append(bb)
cc_list.append(cc)
images_ids_list.append(image_id)
chunks_ids_list.append(i)
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,
"category_blocks": cc_list
}
# In[ ]:
if "layout" in model_id:
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"]
encoding["category_blocks"] = example["category_blocks"]
return encoding
elif "lilt" in model_id:
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["category_blocks"] = example["category_blocks"]
return encoding
# In[ ]:
import torch.nn.functional as F
# get predictions at token level
def predictions_token_level(images, custom_encoded_dataset, model_id):
num_imgs = len(images)
if num_imgs > 0:
if "layout" in model_id:
chunk_ids, input_ids, bboxes, pixels_values, outputs, token_predictions, categories_blocks = dict(), dict(), dict(), dict(), dict(), dict(), dict()
elif "lilt" in model_id:
chunk_ids, input_ids, bboxes, outputs, token_predictions, categories_blocks = 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]
if "layout" in model_id:
pixel_values = torch.tensor(encoding["images_pixels"])
category_blocks = torch.tensor(encoding['category_blocks'])[None]
# 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 "layout" in model_id:
if image_id in pixels_values: pixels_values[image_id].append(pixel_values)
else: pixels_values[image_id] = [pixel_values]
if image_id in categories_blocks: categories_blocks[image_id].append(category_blocks)
else: categories_blocks[image_id] = [category_blocks]
# get prediction with forward pass
with torch.no_grad():
if "layout" in model_id:
output = model(
input_ids=input_id.to(device),
attention_mask=attention_mask.to(device),
bbox=bbox.to(device),
image=pixel_values.to(device)
)
elif "lilt" in model_id:
output = model(
input_ids=input_id,
attention_mask=attention_mask,
bbox=bbox
)
# 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, categories_blocks
else:
print("An error occurred while getting predictions!")
# In[ ]:
from functools import reduce
# Get predictions (line level)
def predictions_paragraph_level(dataset, outputs, images_ids_list, chunk_ids, input_ids, bboxes, categories_blocks):
ten_probs_dict, ten_input_ids_dict, ten_bboxes_dict = dict(), dict(), dict()
bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, categories_blocks_dict_dict, df = dict(), 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]
categories_blocks_list = categories_blocks[image_id]
# create zeros tensors
ten_probs = torch.zeros((outputs_list[0].shape[0] - 2)*len(outputs_list), outputs_list[0].shape[1]).to(device)
ten_input_ids = torch.ones(size=(1, (outputs_list[0].shape[0] - 2)*len(outputs_list)), dtype =int).to(device)
ten_bboxes = torch.zeros(size=(1, (outputs_list[0].shape[0] - 2)*len(outputs_list), 4), dtype =int).to(device)
ten_categories_blocks = torch.ones(size=(1, (outputs_list[0].shape[0] - 2)*len(outputs_list)), dtype =int).to(device)
if len(outputs_list) > 1:
for num_output, (output, input_id, bbox, category_blocks) in enumerate(zip(outputs_list, input_ids_list, bboxes_list, categories_blocks_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,:]
ten_categories_blocks[:,start:end] = category_blocks[:,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,:]
ten_categories_blocks[:,start:start + doc_stride] = category_blocks[:,1:1 + doc_stride]
ten_categories_blocks[:,start + doc_stride:end] = category_blocks[:,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_categories_blocks = categories_blocks_list[0][:,1:-1]
ten_probs_list, ten_input_ids_list, ten_bboxes_list, ten_categories_blocks_list = ten_probs.tolist(), ten_input_ids.tolist()[0], ten_bboxes.tolist()[0], ten_categories_blocks.tolist()[0]
bboxes_list = list()
input_ids_dict, probs_dict, categories_blocks_dict = dict(), dict(), dict()
bbox_prev = [-100, -100, -100, -100]
for probs, input_id, bbox, category_block in zip(ten_probs_list, ten_input_ids_list, ten_bboxes_list, ten_categories_blocks_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]
categories_blocks_dict[str(bbox)] = category_block
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)
# categories_blocks_dict[str(bbox)].append(category_block)
bbox_prev = bbox
probs_bbox = dict()
for num_box,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
bboxes_list_dict[image_id] = bboxes_list
input_ids_dict_dict[image_id] = input_ids_dict
probs_dict_dict[image_id] = probs_bbox
categories_blocks_dict_dict[image_id] = categories_blocks_dict
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]["pred_labels"] = [id2label[probs_bbox[str(bbox)]] for bbox in bboxes_list]
df[image_id]["true_labels"] = [id2label[categories_blocks_dict[str(bbox)]] for bbox in bboxes_list]
if (i % 10 == 0 and i != 0) or (i == len(images_ids_list) - 1): print(f"{i}/{len(images_ids_list)}")
return probs_bbox, bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, categories_blocks_dict_dict, df
else:
print("An error occurred while getting predictions!")
# In[ ]:
# Get labeled images with lines bounding boxes
def get_labeled_images(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
# In[ ]:
# get data of encoded chunk
def get_encoded_chunk_inference(dataset, index_chunk=None):
# get datasets
example = dataset
encoded_example = encoded_dataset
print(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"]
print(encoded_image_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 "..."
if input_ids_dict[str(bboxes_list[-1])][0] == (tokenizer.convert_tokens_to_ids('')):
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
# In[ ]:
# display chunk of PDF image and its data
def display_chunk_paragraphs_inference(dataset, index_chunk=None):
# get image and image data
image, df, num_tokens, page_no, num_pages = get_encoded_chunk_inference(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 paragraphs\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 paragraphs\n")
cols = ["texts", "bboxes"]
df = df[cols]
display(df)
# #### Accuracy
# In[ ]:
def blocks_acc_by_image(df):
image_ids, num_blocks, exact_matches, acc_imgs, pct_accs = list(), list(), list(), list(), list()
mat_res = dict()
for label in labels:
mat_res[label] = [0]* len(labels)
for image_id in df.keys():
df_img = df[image_id]
num_blocks_img = len(df_img)
if num_blocks_img > 0:
# global exact matches by image
exact_match_img = sum(df_img["pred_labels"] == df_img["true_labels"])
# exact match by label and by image
for label in labels:
df_img_label = df_img[df_img["true_labels"] == label].copy()
num_blocks_img_label = len(df_img_label)
# exact_match_img_label = sum(df_img_label["pred_labels"] == df_img_label["true_labels"])
for index, row in df_img_label.iterrows():
pred_label = row["pred_labels"]
mat_res[label][labels.index(pred_label)] += 1
# acc by img (number of blocks with correct predicted label for each image)
acc_img = exact_match_img / num_blocks_img
pct_acc_img = round(acc_img*100, 2)
# append
image_ids.append(image_id)
num_blocks.append(num_blocks_img)
exact_matches.append(exact_match_img)
acc_imgs.append(acc_img)
pct_accs.append(pct_acc_img)
# acc for all images (number of blocks with correct predicted label for all images)
acc_dataset = sum(exact_matches) / sum(num_blocks)
pct_acc_dataset = round(acc_dataset*100, 2)
return image_ids, num_blocks, exact_matches, acc_imgs, pct_accs, acc_dataset, pct_acc_dataset, mat_res
# ## HF login
# In[ ]:
get_ipython().system('huggingface-cli login')
# ## Model & tokenizer
# In[ ]:
if "layout" in model_id:
from transformers import LayoutLMv2ForTokenClassification # LayoutXLMTokenizerFast,
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model
# tokenizer = LayoutXLMTokenizerFast.from_pretrained(model_id)
model = LayoutLMv2ForTokenClassification.from_pretrained(model_id);
model.to(device);
# feature extractor
from transformers import LayoutLMv2FeatureExtractor
feature_extractor = LayoutLMv2FeatureExtractor(apply_ocr=False)
# tokenizer
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
elif "lilt" in model_id:
from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id);
model.to(device);
# In[ ]:
# get labels
id2label = model.config.id2label
label2id = model.config.label2id
num_labels = len(id2label)
# ## Download DocLayNet
# ### Download
# In[ ]:
local_dataset_name = "/content/drive/MyDrive/DocLayNet/datasets/" + dataset_name.replace("pierreguillou/DocLayNet-", "")
# In[ ]:
# from datasets import load_dataset
# dataset = load_dataset(dataset_name)
# # save locally
# dataset.save_to_disk(local_dataset_name)
# In[ ]:
# load
from datasets import load_from_disk
dataset = load_from_disk(local_dataset_name)
# In[ ]:
dataset
# In[ ]:
dataset["test"].features
# In[ ]:
labels = dataset["test"].features["categories"].feature.names
id2label = {id:label for id, label in enumerate(labels)}
label2id = {label:id for id, label in enumerate(labels)}
num_labels = len(labels)
print(id2label)
# In[ ]:
example = dataset["test"][1]
example["image"]
# In[ ]:
example["texts"]
# In[ ]:
example["bboxes_block"]
# In[ ]:
example["bboxes_line"]
# In[ ]:
example
# ### Checking of the dataset
# Select a dataset split and display a random annotated image from it and its dataframe.
#
# In[ ]:
# choose your dataset
splits = ["all", "train", "validation", "test"]
index_split = 3
split = splits[index_split]
# display random PDF image and its data
display_pdf_blocks_lines(split=split)
# ## Inference
# As Colab can stop the evaluation (lack of time, etc.), we chunck our test dataset into 3 and save results.
# In[ ]:
dset = dataset["test"].select(list(range(170)))
# dset = dataset["test"].select(list(range(340))[170:])
# dset = dataset["test"].select(list(range(499))[340:])
dataset_inference, texts_pars_list, par_boxes_list, category_block_list = extraction_data_from_dataset(dset, model_id)
dataset_inference
# ### Data encoding
# Now, we need to prepare our data in the format of the model.
# In[ ]:
encoded_dataset = dataset_inference.map(prepare_inference_features_paragraph, batched=True, batch_size=64, remove_columns=dataset_inference.column_names)
encoded_dataset
# #### Checking of the encoded dataset
# (Checking) Display a random annotated chunk image and its dataframe.
#
# **Note**: the image is squared because of its normalization to 1000px vs 1000px in the encoded dataset (necessary for training the model).
# In[ ]:
# get and image from random chunk
display_chunk_paragraphs_inference(dataset = dataset_inference)
# #### Create a custom dataset
# We end our data preparation with a new class that keeps only the information needed for inference.
# In[ ]:
custom_encoded_dataset = CustomDataset(encoded_dataset, tokenizer)
# Now, we can get the predictions!
# ### Get predictions
# LayoutXLM outputs labels at the token level, but we are interested in the predicted labels at the line level.
# In[ ]:
#@title Get predictions (token level)
outputs, images_ids_list, chunk_ids, input_ids, bboxes, categories_blocks = predictions_token_level(dataset_inference["images"], custom_encoded_dataset, model_id)
# In[ ]:
#@title Get predictions (paragraph level)
probs_bbox, bboxes_list_dict, input_ids_dict_dict, probs_dict_dict, categories_blocks_dict_dict, df = predictions_paragraph_level(dataset_inference, outputs, images_ids_list, chunk_ids, input_ids, bboxes, categories_blocks)
# In[ ]:
# save
import pickle
num_data = 170
# num_data = 340
# num_data = 499
# path to main folder
if "layout" in model_id:
path_to_main = "/content/drive/MyDrive/DocLayNet/results_" + "LayoutXLM_paragraph512_" + "DocLayNet_base_test/"
elif "lilt" in model_id:
path_to_main = "/content/drive/MyDrive/DocLayNet/results_" + "LiLT_paragraph512_" + "DocLayNet_base_test/"
Path(path_to_main).mkdir(parents=True, exist_ok=True)
# path to results folder
path_to_results_folder = path_to_main + str(num_data) + "/"
Path(path_to_results_folder).mkdir(parents=True, exist_ok=True)
# bboxes_list_dict
path_to_bboxes_list_dict = path_to_results_folder + "bboxes_list_dict" + "_" + str(num_data) + ".pkl"
with open(path_to_bboxes_list_dict, 'wb') as f:
pickle.dump(bboxes_list_dict, f)
# input_ids_dict_dict
path_to_input_ids_dict_dict = path_to_results_folder + "input_ids_dict_dict" + "_" + str(num_data) + ".pkl"
with open(path_to_input_ids_dict_dict, 'wb') as f:
pickle.dump(input_ids_dict_dict, f)
# probs_dict_dict
path_to_probs_dict_dict = path_to_results_folder + "probs_dict_dict" + "_" + str(num_data) + ".pkl"
with open(path_to_probs_dict_dict, 'wb') as f:
pickle.dump(probs_dict_dict, f)
# categories_blocks_dict_dict
path_to_categories_blocks_dict_dict = path_to_results_folder + "categories_blocks_dict_dict" + "_" + str(num_data) + ".pkl"
with open(path_to_categories_blocks_dict_dict, 'wb') as f:
pickle.dump(categories_blocks_dict_dict, f)
# categories_blocks_dict_dict
path_to_df = path_to_results_folder + "df" + "_" + str(num_data) + ".pkl"
with open(path_to_df, 'wb') as f:
pickle.dump(df, f)
# In[ ]:
# Load
# import pickle
# num_data = 170
# num_data = 340
# num_data = 499
# path_to_bboxes_list_dict = path_to_main + str(num_data) + "/" + "bboxes_list_dict" + "_" + str(num_data) + ".pkl"
# path_to_input_ids_dict_dict = path_to_main + str(num_data) + "/" + "input_ids_dict_dict" + "_" + str(num_data) + ".pkl"
# path_to_probs_dict_dict = path_to_main + str(num_data) + "/" + "probs_dict_dict" + "_" + str(num_data) + ".pkl"
# path_to_categories_blocks_dict_dict = path_to_main + str(num_data) + "/" + "categories_blocks_dict_dict" + "_" + str(num_data) + ".pkl"
# path_to_df = path_to_main + str(num_data) + "/" + "df" + "_" + str(num_data) + ".pkl"
# with open(path_to_bboxes_list_dict, 'rb') as f:
# bboxes_list_dict_170 = pickle.load(f)
# with open(path_to_input_ids_dict_dict, 'rb') as f:
# input_ids_dict_dict_170 = pickle.load(f)
# with open(path_to_probs_dict_dict, 'rb') as f:
# probs_dict_dict_170 = pickle.load(f)
# with open(path_to_categories_blocks_dict_dict, 'rb') as f:
# categories_blocks_dict_dict_170 = pickle.load(f)
# with open(path_to_df, 'rb') as f:
# df_170 = pickle.load(f)
# In[ ]:
#@title Get labeled images with paragraphs bounding boxes
labeled_images = get_labeled_images(dataset_inference, images_ids_list, bboxes_list_dict, probs_dict_dict)
# ### Display labeled images and get their dataframes
# In[ ]:
print(f"Number of PDF page images: {len(labeled_images)}")
# In[ ]:
# #@title Labeled images
# import matplotlib.pyplot as plt
# %matplotlib inline
# plt.figure(figsize=(15,10))
# columns = 3
# for i, image in enumerate(labeled_images):
# plt.subplot(int(len(labeled_images) / columns + 1), columns, i + 1)
# plt.xticks(color="white")
# plt.yticks(color="white")
# plt.tick_params(bottom = False)
# plt.tick_params(left = False)
# plt.imshow(image)
# In[ ]:
#@title Display one labeled image (full size)
num_page = 0
print(f"Image of the labeled page {num_page} (at paragraph level).")
labeled_images[num_page]
# ## Accuracy (dataset DocLayNet base test)
# As we got the test predictions in 3 dataframes, we concatenate them befor to get accuracy of the whole DocLayNet base test dataset.
# In[ ]:
import pickle
num_data = 170
path_to_df = path_to_main + str(num_data) + "/" + "df" + "_" + str(num_data) + ".pkl"
with open(path_to_df, 'rb') as f:
df_170 = pickle.load(f)
num_data = 340
path_to_df = path_to_main + str(num_data) + "/" + "df" + "_" + str(num_data) + ".pkl"
with open(path_to_df, 'rb') as f:
df_340 = pickle.load(f)
num_data = 499
path_to_df = path_to_main + str(num_data) + "/" + "df" + "_" + str(num_data) + ".pkl"
with open(path_to_df, 'rb') as f:
df_499 = pickle.load(f)
print(len(df_170),len(df_340),len(df_499),len(df_170)+len(df_340)+len(df_499))
df1 = df_170.copy()
df2 = {k+170:v for k,v in df_340.items()}
df3 = {k+2*170:v for k,v in df_499.items()}
df = df1
df.update(df2)
df.update(df3)
len(df)
# In[ ]:
image_ids, num_blocks, exact_matches, acc_imgs, pct_accs, acc_dataset, pct_acc_dataset, mat_res = blocks_acc_by_image(df)
if "layout" in model_id:
print("\nAccuracy of the finetuned LayoutXLM base model\n(number of blocks with correct predicted label for all images of the test DocLayNet base dataset)")
elif "lilt" in model_id:
print("\nAccuracy of the finetuned LiLT base model\n(number of blocks with correct predicted label for all images of the test DocLayNet base dataset)")
print(f"{pct_acc_dataset}%")
# In[ ]:
import numpy as np
accs = list()
for label in labels:
index = label2id[label]
res_label = mat_res[label]
num_true_blocks_label = sum(res_label)
num_pred_blocks_label = res_label[index]
acc = num_pred_blocks_label / num_true_blocks_label
acc = round(100*acc, 2)
accs.append(acc)
if "layout" in model_id:
print("Accuracy by label of the finetuned LayoutXLM base model")
elif "lilt" in model_id:
print("Accuracy by label of the finetuned LiLT base model")
for acc,label in zip (accs, labels):
print(f"- {label}: {acc}%")
# In[ ]:
import matplotlib.pyplot as plt
import pandas as pd
data = {'labels': labels,
'accuracy (%)': accs
}
df = pd.DataFrame(data)
df = df.sort_values(by = ['accuracy (%)'], ascending = False)
colors_green, colors_red = list(), list()
for acc in accs:
if acc >= 80:
color = "green"
colors_green.append(color)
else:
color = "red"
colors_red.append(color)
plt.bar(df['labels'][df['accuracy (%)'] >= 80], df['accuracy (%)'][df['accuracy (%)'] >= 80], color=colors_green)
plt.bar(df['labels'][df['accuracy (%)'] < 80], df['accuracy (%)'][df['accuracy (%)'] < 80], color=colors_red)
if "layout" in model_id:
title = 'Accuracy by label\nof the labeled paragraphs\n\ndataset: test DocLayNet base\nmodel: LayoutXLM base finetuned on DocLayNet base\n'
elif "lilt" in model_id:
title = 'Accuracy by label\nof the labeled paragraphs\n\ndataset: test DocLayNet base\nmodel: LiLT base finetuned on DocLayNet base\n'
plt.title(title, fontsize=14)
plt.xlabel('Labels', fontsize=14)
plt.xticks(rotation=70)
plt.ylabel('accuracy (%)', fontsize=14)
plt.grid(True)
plt.legend([">= 80%", "< 80%"], loc=1)
plt.show()
# In[ ]:
# print("Accuracy by image (number of blocks with correct predicted label by image)")
# for image_id, pct_acc in zip(image_ids, pct_accs):
# print(f"Image n°{image_id}: {pct_acc}%")
# In[ ]:
# # Confusion matrix
# # https://scikit-learn.org/stable/modules/generated/sklearn.metrics.ConfusionMatrixDisplay.html
# from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
# import matplotlib.pyplot as plt
# import numpy as np
# cm = list()
# for label in labels:
# cm.append(mat_res[label])
# cm = np.array(cm)
# disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=labels)
# disp.plot(xticks_rotation=70)
# if "layout" in model_id:
# plt.title('Confusion matrix of the labeled paragraphs\nof the test DocLayNet base dataset\n(model: LayoutXLM base finetuned on DocLayNet base)\n')
# elif "lilt" in model_id:
# plt.title('Confusion matrix of the labeled paragraphs\nof the test DocLayNet base dataset\n(model: LiLT base finetuned on DocLayNet base)\n')
# plt.show()
# In[ ]:
# # Confusion matrix
# # https://scikit-learn.org/stable/modules/generated/sklearn.metrics.ConfusionMatrixDisplay.html
# # https://stackoverflow.com/questions/64559225/normalizing-a-color-map-for-plotting-a-confusion-matrix-with-confusionmatrixdisp
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
import matplotlib.pyplot as plt
import numpy as np
cm = list()
for label in labels:
cm.append(mat_res[label])
cm = np.array(cm)
accuracies = cm/cm.sum(1)
fig, ax = plt.subplots(figsize=(10,8))
cb = ax.imshow(accuracies, cmap='Blues')
plt.xticks(range(len(labels)), labels, rotation=70)
plt.yticks(range(len(labels)), labels)
for i in range(len(labels)):
for j in range(len(labels)):
color='blue' if accuracies[i,j] < 0.5 else 'white'
ax.annotate(f'{cm[i,j]}', (j,i),
color=color, va='center', ha='center')
ax.set(
ylabel="True label",
xlabel="Predicted label",
)
plt.colorbar(cb, ax=ax)
if "layout" in model_id:
plt.title('Confusion matrix of the labeled paragraphs\nof the test DocLayNet base dataset\n(model: LayoutXLM base finetuned on DocLayNet base)\n')
elif "lilt" in model_id:
plt.title('Confusion matrix of the labeled paragraphs\nof the test DocLayNet base dataset\n(model: LiLT base finetuned on DocLayNet base)\n')
plt.show()
# # END