first version of pdftoimage and ocr
Browse files- DUDE_imdb_loader.py +421 -0
DUDE_imdb_loader.py
ADDED
@@ -0,0 +1,421 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
"""DUDE dataset loader"""
|
16 |
+
|
17 |
+
import os
|
18 |
+
from pathlib import Path
|
19 |
+
import time
|
20 |
+
import copy
|
21 |
+
import json
|
22 |
+
import numpy as np
|
23 |
+
import pandas as pd
|
24 |
+
from tqdm import tqdm
|
25 |
+
from io import BytesIO
|
26 |
+
|
27 |
+
tqdm.pandas()
|
28 |
+
from joblib import Parallel, delayed
|
29 |
+
|
30 |
+
# import pdf2image
|
31 |
+
import PyPDF2
|
32 |
+
|
33 |
+
from datasets import load_dataset_builder, load_dataset
|
34 |
+
from PIL import Image as PIL_Image
|
35 |
+
|
36 |
+
|
37 |
+
MAX_PAGES = 50
|
38 |
+
MAX_PDF_SIZE = 100000000 # almost 100MB
|
39 |
+
MIN_WIDTH, MIN_HEIGHT = 150, 150
|
40 |
+
|
41 |
+
|
42 |
+
def load_json(json_path):
|
43 |
+
return json.load(open(json_path, "r"))
|
44 |
+
|
45 |
+
|
46 |
+
def save_json(json_path, data):
|
47 |
+
with open(json_path, "w") as f:
|
48 |
+
json.dump(data, f)
|
49 |
+
|
50 |
+
|
51 |
+
# 0. PDF to images
|
52 |
+
|
53 |
+
# 1. OCR metadata
|
54 |
+
|
55 |
+
|
56 |
+
def pdf_to_images(document_filepath):
|
57 |
+
def images_to_pagenames(images, document_filepath, page_image_dir):
|
58 |
+
"/home/jordy/Downloads/DUDE_train-val-test_binaries/PDF/val/001d6f557c342ef5a67cd38a29da9e83.pdf"
|
59 |
+
|
60 |
+
page_image_names = []
|
61 |
+
for page_idx, page_image in enumerate(images):
|
62 |
+
page_image_name = document_filepath.replace("PDF", "images").replace(
|
63 |
+
".pdf", f"_{page_idx}.jpg"
|
64 |
+
)
|
65 |
+
|
66 |
+
# page_image_names.append(page_image_name.replace(page_images_dir, "")) #without dir
|
67 |
+
|
68 |
+
if not os.path.exists(page_image_name):
|
69 |
+
page_image.save(page_image_name)
|
70 |
+
return page_image_names
|
71 |
+
|
72 |
+
example = {}
|
73 |
+
example["document"] = document_filepath
|
74 |
+
example["pages"] = 0
|
75 |
+
example["page_image_names"] = []
|
76 |
+
images = []
|
77 |
+
|
78 |
+
page_image_dir = "/".join(example["document"].split("/")[:-1]).replace("PDF", "images")
|
79 |
+
if not os.path.exists(page_image_dir):
|
80 |
+
os.makedirs(page_image_dir)
|
81 |
+
|
82 |
+
# if len(example["document"]) > MAX_PDF_SIZE:
|
83 |
+
# logger.warning(f"too large document {len(example['document'])}")
|
84 |
+
# return example
|
85 |
+
try:
|
86 |
+
reader = PyPDF2.PdfReader(example["document"])
|
87 |
+
except Exception as e:
|
88 |
+
logger.warning(f"read_pdf {e}")
|
89 |
+
return example
|
90 |
+
|
91 |
+
reached_page_limit = False
|
92 |
+
page_iterator = reader.pages
|
93 |
+
|
94 |
+
for p, page in enumerate(page_iterator):
|
95 |
+
if reached_page_limit:
|
96 |
+
break
|
97 |
+
for image in page.images:
|
98 |
+
# try:
|
99 |
+
# except Exception as e:
|
100 |
+
# logger.warning(f"get_images {e}")
|
101 |
+
if len(images) == MAX_PAGES:
|
102 |
+
reached_page_limit = True
|
103 |
+
break
|
104 |
+
|
105 |
+
im = PIL_Image.open(BytesIO(image.data))
|
106 |
+
if im.width < MIN_WIDTH and im.height < MIN_HEIGHT:
|
107 |
+
continue
|
108 |
+
images.append(im)
|
109 |
+
example["pages"] = len(images)
|
110 |
+
if len(images) == 0:
|
111 |
+
return example
|
112 |
+
|
113 |
+
example["page_image_names"] = images_to_pagenames(images, example["document"], page_image_dir)
|
114 |
+
|
115 |
+
return example
|
116 |
+
|
117 |
+
|
118 |
+
def pdf_to_images_block(document_paths_blocks):
|
119 |
+
new_doc_metadata = {}
|
120 |
+
for document_filepath in document_paths_blocks:
|
121 |
+
docId = document_filepath.split("/")[-1].replace(".pdf", "")
|
122 |
+
new_doc_metadata[docId] = pdf_to_images(document_filepath)
|
123 |
+
return new_doc_metadata
|
124 |
+
|
125 |
+
|
126 |
+
"""
|
127 |
+
def get_document_metadata(docs_metadata, doc_id, document_filepath):
|
128 |
+
|
129 |
+
if doc_id in docs_metadata and docs_metadata[doc_id]["num_pages"] != -1:
|
130 |
+
num_pages = docs_metadata[doc_id]["num_pages"]
|
131 |
+
page_image_names = docs_metadata[doc_id]["page_image_names"]
|
132 |
+
|
133 |
+
else:
|
134 |
+
try:
|
135 |
+
images = pdf2image.convert_from_path(document_filepath)
|
136 |
+
except:
|
137 |
+
print(doc_id)
|
138 |
+
return -1, -1
|
139 |
+
num_pages = len(images)
|
140 |
+
page_image_dir = ("/").join(
|
141 |
+
document_filepath.replace(documents_dir, page_images_dir).split("/")[:-1]
|
142 |
+
)
|
143 |
+
if not os.path.exists(page_image_dir):
|
144 |
+
os.makedirs(page_image_dir)
|
145 |
+
|
146 |
+
page_image_names = []
|
147 |
+
for page_idx, page_image in enumerate(images):
|
148 |
+
page_image_name = document_filepath.replace(documents_dir, page_images_dir).replace(
|
149 |
+
".pdf", f"_{page_idx}.jpg"
|
150 |
+
)
|
151 |
+
page_image_names.append(page_image_name.replace(page_images_dir, ""))
|
152 |
+
|
153 |
+
if not os.path.exists(page_image_name):
|
154 |
+
page_image.save(page_image_name)
|
155 |
+
|
156 |
+
return num_pages, page_image_names
|
157 |
+
|
158 |
+
|
159 |
+
def get_document_metadata_block(docs_metadata, documents_path_dict, documents_blocks):
|
160 |
+
new_doc_metadata = {}
|
161 |
+
for doc_id in documents_blocks:
|
162 |
+
document_filepath = documents_path_dict[doc_id]
|
163 |
+
num_pages, page_image_names = get_document_metadata(
|
164 |
+
docs_metadata, doc_id, document_filepath
|
165 |
+
)
|
166 |
+
new_doc_metadata[doc_id] = {"num_pages": num_pages, "page_image_names": page_image_names}
|
167 |
+
|
168 |
+
return new_doc_metadata
|
169 |
+
"""
|
170 |
+
|
171 |
+
|
172 |
+
def parse_textract_bbox(box):
|
173 |
+
# 0.47840896, 0.12897822, 0.5341576 , 0.14347914 # x,w,y,h
|
174 |
+
return np.array([box["Left"], box["Width"], box["Top"], box["Height"]])
|
175 |
+
|
176 |
+
|
177 |
+
def parse_azure_box(box, page_width, page_height):
|
178 |
+
# Box in Azure are in format X top left, Y top left, X top right, Y top right, X bottom right, Y bottom right, X bottom left, Y bottom left
|
179 |
+
# [14.1592, 3.8494, 28.668, 3.8494, 28.668, 8.0487, 13.9844, 7.8738]
|
180 |
+
left = min(box[0], box[6])
|
181 |
+
right = max(box[2], box[4])
|
182 |
+
top = min(box[1], box[3])
|
183 |
+
bottom = max(box[5], box[7])
|
184 |
+
width = right - left
|
185 |
+
height = bottom - top
|
186 |
+
|
187 |
+
# Normalize
|
188 |
+
left = left / page_width
|
189 |
+
top = top / page_height
|
190 |
+
width = width / page_width
|
191 |
+
height = height / page_height
|
192 |
+
|
193 |
+
return [left, width, top, height]
|
194 |
+
|
195 |
+
|
196 |
+
def get_ocr_information(doc_id):
|
197 |
+
ocr_info = load_json(os.path.join(ocr_dir, doc_id + "_original.json"))
|
198 |
+
num_pages, _ = get_document_metadata(documents_metadata, doc_id, documents[doc_id])
|
199 |
+
|
200 |
+
page_ocr_tokens = [[] for page_ix in range(num_pages)]
|
201 |
+
page_ocr_boxes = [[] for page_ix in range(num_pages)]
|
202 |
+
for ocr_block in ocr_info:
|
203 |
+
for ocr_extraction in ocr_block["Blocks"]:
|
204 |
+
if ocr_extraction["BlockType"] == "WORD":
|
205 |
+
text = ocr_extraction["Text"].lower()
|
206 |
+
bounding_box = parse_textract_bbox(ocr_extraction["Geometry"]["BoundingBox"])
|
207 |
+
page = ocr_extraction["Page"] - 1
|
208 |
+
|
209 |
+
page_ocr_tokens[page].append(text)
|
210 |
+
page_ocr_boxes[page].append(bounding_box)
|
211 |
+
|
212 |
+
for page in range(num_pages):
|
213 |
+
page_ocr_boxes[page] = np.array(page_ocr_boxes[page])
|
214 |
+
|
215 |
+
page_ocr_boxes = page_ocr_boxes
|
216 |
+
return page_ocr_tokens, page_ocr_boxes
|
217 |
+
|
218 |
+
|
219 |
+
def create_header(split, version, has_answer):
|
220 |
+
header = {
|
221 |
+
"creation_time": time.time(),
|
222 |
+
"version": version,
|
223 |
+
"dataset_type": split,
|
224 |
+
"has_answer": has_answer,
|
225 |
+
}
|
226 |
+
|
227 |
+
return header
|
228 |
+
|
229 |
+
|
230 |
+
def get_document_info(documents_metadata, doc_id):
|
231 |
+
doc_metadata = documents_metadata[doc_id]
|
232 |
+
num_pages = doc_metadata["num_pages"]
|
233 |
+
page_image_names = doc_metadata["page_image_names"]
|
234 |
+
return num_pages, page_image_names
|
235 |
+
|
236 |
+
|
237 |
+
def format_answers(answers_list):
|
238 |
+
answers_list = list(set([answer.lower() for answer in answers_list]))
|
239 |
+
return answers_list
|
240 |
+
|
241 |
+
|
242 |
+
def create_imdb_record_from_json(
|
243 |
+
record, documents_metadata, documents_ocr_information, split, include_answers
|
244 |
+
):
|
245 |
+
|
246 |
+
doc_id = record["docId"]
|
247 |
+
# document_filepath = documents_dict[doc_id]
|
248 |
+
num_pages, page_image_names = get_document_info(documents_metadata, doc_id)
|
249 |
+
document_ocr_info = documents_ocr_information[doc_id]
|
250 |
+
|
251 |
+
if include_answers:
|
252 |
+
answers = format_answers(record["answers"])
|
253 |
+
else:
|
254 |
+
answers = None
|
255 |
+
|
256 |
+
imdb_record = {
|
257 |
+
"question_id": record["questionId"],
|
258 |
+
"question": record["question"],
|
259 |
+
"docId": doc_id,
|
260 |
+
"image_name": page_image_names,
|
261 |
+
"num_pages": num_pages,
|
262 |
+
"ocr_tokens": document_ocr_info["ocr_tokens"],
|
263 |
+
"ocr_normalized_boxes": document_ocr_info["ocr_boxes"],
|
264 |
+
"set_name": split,
|
265 |
+
"answers": answers,
|
266 |
+
"answer_page": None,
|
267 |
+
"extra": {
|
268 |
+
# 'question_type': record['qtype'],
|
269 |
+
# 'industry': record['industry'],
|
270 |
+
# 'category': record['category'],
|
271 |
+
"answer_type": record["answer_type"],
|
272 |
+
},
|
273 |
+
}
|
274 |
+
|
275 |
+
return imdb_record
|
276 |
+
|
277 |
+
|
278 |
+
def create_imdb_from_json(
|
279 |
+
data, documents_metadata, documents_ocr_information, split, version, include_answers
|
280 |
+
):
|
281 |
+
imdb_header = create_header(split, version, include_answers)
|
282 |
+
|
283 |
+
imdb_records = []
|
284 |
+
for record in tqdm(data):
|
285 |
+
imdb_records.append(
|
286 |
+
create_imdb_record_from_json(
|
287 |
+
record, documents_metadata, documents_ocr_information, split, include_answers
|
288 |
+
)
|
289 |
+
)
|
290 |
+
|
291 |
+
imdb = [imdb_header] + imdb_records
|
292 |
+
|
293 |
+
return imdb
|
294 |
+
|
295 |
+
|
296 |
+
if __name__ == "__main__":
|
297 |
+
dataset = load_dataset(
|
298 |
+
"../DUDE_loader/DUDE_loader.py",
|
299 |
+
"DUDE",
|
300 |
+
data_dir="/home/jordy/Downloads/DUDE_train-val-test_binaries",
|
301 |
+
)
|
302 |
+
splits = datasets.keys()
|
303 |
+
|
304 |
+
split = "val"
|
305 |
+
|
306 |
+
|
307 |
+
# 1. PDF to image dir and collect document metadata (num_pages, page_image_names)
|
308 |
+
|
309 |
+
document_paths = sorted(
|
310 |
+
set([x["document"] for x in dataset[split] if x["data_split"] == split])
|
311 |
+
)
|
312 |
+
|
313 |
+
document_paths = document_paths[:30]
|
314 |
+
|
315 |
+
# document_ids = [document_filepath.split("/")[-1].replace(".pdf", "") for document_filepath in document_paths]
|
316 |
+
documents_metadata_filename = "documents_metadata.json"
|
317 |
+
if os.path.exists(documents_metadata_filename):
|
318 |
+
documents_metadata = load_json(documents_metadata_filename)
|
319 |
+
else:
|
320 |
+
documents_metadata = {}
|
321 |
+
# for document_filepath in tqdm(document_paths):
|
322 |
+
# documents_metadata[docId] = pdf_to_images(document_filepath)
|
323 |
+
num_jobs = 6
|
324 |
+
block_size = int(len(document_paths) / num_jobs) + 1
|
325 |
+
print(f"{block_size} * {num_jobs} = {block_size*num_jobs} ({len(document_paths)})")
|
326 |
+
documents_blocks = [
|
327 |
+
document_paths[block_size * i : block_size * i + block_size] for i in range(num_jobs)
|
328 |
+
]
|
329 |
+
print(
|
330 |
+
"chunksize",
|
331 |
+
len(set([doc_id for doc_block in documents_blocks for doc_id in doc_block])),
|
332 |
+
)
|
333 |
+
parallel_results = Parallel(n_jobs=num_jobs)(
|
334 |
+
delayed(get_document_metadata_block)(documents_metadata, documents, document_paths[i])
|
335 |
+
for i in range(num_jobs)
|
336 |
+
)
|
337 |
+
|
338 |
+
for block_result in parallel_results:
|
339 |
+
for doc_id, metadata in tqdm(block_result.items()):
|
340 |
+
if doc_id not in documents_metadata:
|
341 |
+
documents_metadata[doc_id] = metadata
|
342 |
+
|
343 |
+
save_json(documents_metadata_filename, documents_metadata)
|
344 |
+
|
345 |
+
#2. Process OCR to obtain doc_ocr_info
|
346 |
+
documents_ocr_filename = "documents_ocr.json"
|
347 |
+
|
348 |
+
if os.path.exists(documents_ocr_filename):
|
349 |
+
documents_ocr_info = load_json(documents_ocr_filename)
|
350 |
+
else:
|
351 |
+
documents_ocr_info = {}
|
352 |
+
no_ocr = []
|
353 |
+
error_ocr = []
|
354 |
+
|
355 |
+
for document_filepath in document_paths:
|
356 |
+
docId = document_filepath.split("/")[-1].replace(".pdf", "")
|
357 |
+
try:
|
358 |
+
ocr_tokens, ocr_boxes = get_ocr_information(docId)
|
359 |
+
doc_ocr_info[docId] = {'ocr_tokens': ocr_tokens, 'ocr_boxes': ocr_boxes}
|
360 |
+
except FileNotFoundError:
|
361 |
+
no_ocr.append(docId)
|
362 |
+
except KeyError:
|
363 |
+
error_ocr.append(docId)
|
364 |
+
|
365 |
+
save_json(documents_ocr_filename, documents_ocr_info)
|
366 |
+
|
367 |
+
# page_image_dir = '/'.join(dataset['val']['document'][0].split("/")[:-1]).replace('PDF', 'images')
|
368 |
+
# if not os.path.exists(page_image_dir):
|
369 |
+
# os.makedirs(page_image_dir)
|
370 |
+
# dataset.info.features
|
371 |
+
|
372 |
+
"""
|
373 |
+
Describe all steps that need to happen after loading HF DUDE dataset
|
374 |
+
Change functions
|
375 |
+
|
376 |
+
|
377 |
+
page_images_dir
|
378 |
+
|
379 |
+
|
380 |
+
|
381 |
+
2. Process OCR to obtain doc_ocr_info
|
382 |
+
"""
|
383 |
+
# update dataset with
|
384 |
+
# for split in SPLITS
|
385 |
+
|
386 |
+
# documents_metadata
|
387 |
+
# doc_ocr_info
|
388 |
+
|
389 |
+
# dict to unique docs
|
390 |
+
# documents_metadata[doc_id] = {"num_pages": num_pages, "page_image_names": image_names}
|
391 |
+
# doc_ocr_info[doc_id] = {"ocr_tokens": ocr_tokens, "ocr_boxes": ocr_boxes}
|
392 |
+
|
393 |
+
"""
|
394 |
+
train_imdb = create_imdb_from_json(
|
395 |
+
train_data,
|
396 |
+
documents_metadata=documents_metadata,
|
397 |
+
documents_ocr_information=doc_ocr_info,
|
398 |
+
split="train",
|
399 |
+
version="0.1",
|
400 |
+
include_answers=True,
|
401 |
+
)
|
402 |
+
val_imdb = create_imdb_from_json(
|
403 |
+
val_data,
|
404 |
+
documents_metadata=documents_metadata,
|
405 |
+
documents_ocr_information=doc_ocr_info,
|
406 |
+
split="train",
|
407 |
+
version="0.1",
|
408 |
+
include_answers=True,
|
409 |
+
)
|
410 |
+
np.save("Imdb/train_imdb.npy", train_imdb)
|
411 |
+
np.save("Imdb/val_imdb.npy", val_imdb)
|
412 |
+
|
413 |
+
document_paths = []
|
414 |
+
num_jobs = 6
|
415 |
+
block_size = int(len(document_ids) / num_jobs) + 1
|
416 |
+
print(f"{block_size} * {num_jobs} = {block_size*num_jobs} ({len(document_ids)})")
|
417 |
+
parallel_results = Parallel(n_jobs=num_jobs)(
|
418 |
+
delayed(get_document_metadata_block)(documents_metadata, documents, documents_blocks[i])
|
419 |
+
for i in range(num_jobs)
|
420 |
+
)
|
421 |
+
"""
|