small patch for answer variants
Browse files- DUDE_imdb_loader.py +73 -24
- DUDE_loader.py +1 -2
- test_loader.py +1 -1
DUDE_imdb_loader.py
CHANGED
@@ -36,7 +36,7 @@ from datasets import load_dataset_builder, load_dataset, logging
|
|
36 |
|
37 |
logger = logging.get_logger(__name__)
|
38 |
|
39 |
-
PIL_Image.MAX_IMAGE_PIXELS = None
|
40 |
|
41 |
MAX_PAGES = 50
|
42 |
MAX_PDF_SIZE = 100000000 # almost 100MB
|
@@ -63,7 +63,9 @@ def get_images_pdf2image(document_filepath, chunksize=10):
|
|
63 |
try:
|
64 |
images.extend(
|
65 |
pdf2image.convert_from_path(
|
66 |
-
document_filepath,
|
|
|
|
|
67 |
)
|
68 |
)
|
69 |
except Exception as e:
|
@@ -78,7 +80,9 @@ def pdf_to_images(document_filepath, converter="PyPDF2"):
|
|
78 |
page_image_name = document_filepath.replace("PDF", "images").replace(
|
79 |
".pdf", f"_{page_idx}.jpg"
|
80 |
)
|
81 |
-
page_image_names.append(
|
|
|
|
|
82 |
if not os.path.exists(page_image_name):
|
83 |
page_image.convert("RGB").save(page_image_name)
|
84 |
return page_image_names
|
@@ -211,7 +215,7 @@ def get_document_info(documents_metadata, docId):
|
|
211 |
|
212 |
|
213 |
def format_answers(answers_list):
|
214 |
-
answers_list = list(set([answer
|
215 |
return answers_list
|
216 |
|
217 |
|
@@ -220,7 +224,6 @@ def create_imdb_record_from_json(
|
|
220 |
):
|
221 |
|
222 |
docId = record["docId"].split("_")[0]
|
223 |
-
# document_filepath = documents_dict[docId]
|
224 |
try:
|
225 |
num_pages, page_image_names = get_document_info(documents_metadata, docId)
|
226 |
document_ocr_info = documents_ocr_info[docId]
|
@@ -237,12 +240,13 @@ def create_imdb_record_from_json(
|
|
237 |
else:
|
238 |
answers = None
|
239 |
|
240 |
-
if include_variants:
|
241 |
-
|
242 |
-
|
243 |
|
244 |
-
#fix if wrongly saved pagenames
|
245 |
-
page_image_names = [
|
|
|
|
|
246 |
|
247 |
imdb_record = {
|
248 |
"question_id": record["questionId"],
|
@@ -253,7 +257,7 @@ def create_imdb_record_from_json(
|
|
253 |
"ocr_tokens": document_ocr_info["ocr_tokens"],
|
254 |
"ocr_normalized_boxes": document_ocr_info["ocr_boxes"],
|
255 |
"set_name": split,
|
256 |
-
"answers": answers,
|
257 |
"answer_page": None,
|
258 |
"extra": {
|
259 |
# 'question_type': record['qtype'],
|
@@ -267,14 +271,20 @@ def create_imdb_record_from_json(
|
|
267 |
|
268 |
|
269 |
def create_imdb_from_json(
|
270 |
-
data,
|
|
|
|
|
|
|
|
|
|
|
|
|
271 |
):
|
272 |
imdb_header = create_header(split, version, include_answers)
|
273 |
|
274 |
imdb_records = []
|
275 |
for record in tqdm(data):
|
276 |
imdb_record = create_imdb_record_from_json(
|
277 |
-
record, documents_metadata, documents_ocr_info, split, include_answers
|
278 |
)
|
279 |
if imdb_record:
|
280 |
imdb_records.append(imdb_record)
|
@@ -284,12 +294,53 @@ def create_imdb_from_json(
|
|
284 |
return imdb
|
285 |
|
286 |
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
292 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
splits = dataset.keys()
|
294 |
|
295 |
for split in splits:
|
@@ -297,7 +348,7 @@ if __name__ == "__main__":
|
|
297 |
OCR_paths = []
|
298 |
document_paths = []
|
299 |
for i, x in enumerate(dataset[split]):
|
300 |
-
if x["data_split"] != split:
|
301 |
continue
|
302 |
if x["document"] not in document_paths:
|
303 |
document_paths.append(x["document"])
|
@@ -371,12 +422,10 @@ if __name__ == "__main__":
|
|
371 |
save_json(documents_ocr_filename, documents_ocr_info)
|
372 |
|
373 |
imdb_filename = f"imdb_{split}.npy"
|
374 |
-
if os.path.exists(imdb_filename):
|
375 |
print(f"Loading from disk: {imdb_filename}")
|
376 |
imdb = np.load(imdb_filename, allow_pickle=True)
|
377 |
|
378 |
-
import pdb; pdb.set_trace() # breakpoint ff49174b //
|
379 |
-
|
380 |
else:
|
381 |
imdb = create_imdb_from_json(
|
382 |
dataset[split], # .select(split_indices),
|
@@ -385,6 +434,6 @@ if __name__ == "__main__":
|
|
385 |
split=split,
|
386 |
version="0.1",
|
387 |
include_answers=True,
|
|
|
388 |
)
|
389 |
-
np.save(imdb_filename, imdb)
|
390 |
-
#Missing: 'a7b500f3e0244a50571769aaef4fabc7' a7b500f3e0244a50571769aaef4fabc7
|
|
|
36 |
|
37 |
logger = logging.get_logger(__name__)
|
38 |
|
39 |
+
PIL_Image.MAX_IMAGE_PIXELS = None # 933120000
|
40 |
|
41 |
MAX_PAGES = 50
|
42 |
MAX_PDF_SIZE = 100000000 # almost 100MB
|
|
|
63 |
try:
|
64 |
images.extend(
|
65 |
pdf2image.convert_from_path(
|
66 |
+
document_filepath,
|
67 |
+
first_page=page,
|
68 |
+
last_page=min(page + chunksize - 1, maxPages),
|
69 |
)
|
70 |
)
|
71 |
except Exception as e:
|
|
|
80 |
page_image_name = document_filepath.replace("PDF", "images").replace(
|
81 |
".pdf", f"_{page_idx}.jpg"
|
82 |
)
|
83 |
+
page_image_names.append(
|
84 |
+
page_image_name.replace(page_image_dir, page_image_dir.split("/")[-1])
|
85 |
+
) # without dir
|
86 |
if not os.path.exists(page_image_name):
|
87 |
page_image.convert("RGB").save(page_image_name)
|
88 |
return page_image_names
|
|
|
215 |
|
216 |
|
217 |
def format_answers(answers_list):
|
218 |
+
answers_list = list(set([answer for answer in answers_list])) # standardization is done
|
219 |
return answers_list
|
220 |
|
221 |
|
|
|
224 |
):
|
225 |
|
226 |
docId = record["docId"].split("_")[0]
|
|
|
227 |
try:
|
228 |
num_pages, page_image_names = get_document_info(documents_metadata, docId)
|
229 |
document_ocr_info = documents_ocr_info[docId]
|
|
|
240 |
else:
|
241 |
answers = None
|
242 |
|
243 |
+
if include_variants and record['answers_variants']:
|
244 |
+
answers += record["answers_variants"]
|
|
|
245 |
|
246 |
+
# fix if wrongly saved pagenames
|
247 |
+
# page_image_names = [
|
248 |
+
# split + image_name for image_name in page_image_names if image_name.startswith("/")
|
249 |
+
# ]
|
250 |
|
251 |
imdb_record = {
|
252 |
"question_id": record["questionId"],
|
|
|
257 |
"ocr_tokens": document_ocr_info["ocr_tokens"],
|
258 |
"ocr_normalized_boxes": document_ocr_info["ocr_boxes"],
|
259 |
"set_name": split,
|
260 |
+
"answers": answers, # list
|
261 |
"answer_page": None,
|
262 |
"extra": {
|
263 |
# 'question_type': record['qtype'],
|
|
|
271 |
|
272 |
|
273 |
def create_imdb_from_json(
|
274 |
+
data,
|
275 |
+
documents_metadata,
|
276 |
+
documents_ocr_info,
|
277 |
+
split,
|
278 |
+
version,
|
279 |
+
include_answers=True,
|
280 |
+
include_variants=False,
|
281 |
):
|
282 |
imdb_header = create_header(split, version, include_answers)
|
283 |
|
284 |
imdb_records = []
|
285 |
for record in tqdm(data):
|
286 |
imdb_record = create_imdb_record_from_json(
|
287 |
+
record, documents_metadata, documents_ocr_info, split, include_answers, include_variants
|
288 |
)
|
289 |
if imdb_record:
|
290 |
imdb_records.append(imdb_record)
|
|
|
294 |
return imdb
|
295 |
|
296 |
|
297 |
+
def parse_arguments():
|
298 |
+
import argparse
|
299 |
+
|
300 |
+
parser = argparse.ArgumentParser(
|
301 |
+
description="Instantiate HuggingFace dataloader and convert to ImDB format"
|
302 |
+
)
|
303 |
+
|
304 |
+
parser.add_argument(
|
305 |
+
"--redo-imdb-build",
|
306 |
+
action="store_true",
|
307 |
+
default=False,
|
308 |
+
help="Whether to rebuild the imdb record and overwrite",
|
309 |
+
)
|
310 |
+
|
311 |
+
# Optional
|
312 |
+
parser.add_argument(
|
313 |
+
"--no-include-variants",
|
314 |
+
action="store_false",
|
315 |
+
default=True,
|
316 |
+
help="Whether to include variants as full ground truths",
|
317 |
+
)
|
318 |
+
|
319 |
+
# Overwrite/Pass HF parameters
|
320 |
+
parser.add_argument(
|
321 |
+
"--DUDE_config",
|
322 |
+
type=str,
|
323 |
+
default="Amazon_original",
|
324 |
+
help="HF Config to load to control OCR version",
|
325 |
+
)
|
326 |
+
parser.add_argument(
|
327 |
+
"--data_dir",
|
328 |
+
type=str,
|
329 |
+
default="/home/jordy/Downloads/DUDE_train-val-test_binaries",
|
330 |
+
help="Load PDFs and store images",
|
331 |
)
|
332 |
+
|
333 |
+
return parser.parse_args()
|
334 |
+
|
335 |
+
|
336 |
+
if __name__ == "__main__":
|
337 |
+
"""
|
338 |
+
Parse and redo
|
339 |
+
"""
|
340 |
+
args = parse_arguments()
|
341 |
+
|
342 |
+
dataset = load_dataset("jordyvl/DUDE_loader", args.DUDE_config, data_dir=args.data_dir)
|
343 |
+
|
344 |
splits = dataset.keys()
|
345 |
|
346 |
for split in splits:
|
|
|
348 |
OCR_paths = []
|
349 |
document_paths = []
|
350 |
for i, x in enumerate(dataset[split]):
|
351 |
+
if x["data_split"] != split: # redundant check
|
352 |
continue
|
353 |
if x["document"] not in document_paths:
|
354 |
document_paths.append(x["document"])
|
|
|
422 |
save_json(documents_ocr_filename, documents_ocr_info)
|
423 |
|
424 |
imdb_filename = f"imdb_{split}.npy"
|
425 |
+
if os.path.exists(imdb_filename) and not args.redo_imdb_build:
|
426 |
print(f"Loading from disk: {imdb_filename}")
|
427 |
imdb = np.load(imdb_filename, allow_pickle=True)
|
428 |
|
|
|
|
|
429 |
else:
|
430 |
imdb = create_imdb_from_json(
|
431 |
dataset[split], # .select(split_indices),
|
|
|
434 |
split=split,
|
435 |
version="0.1",
|
436 |
include_answers=True,
|
437 |
+
include_variants=(not args.no_include_variants),
|
438 |
)
|
439 |
+
np.save(imdb_filename, imdb)
|
|
DUDE_loader.py
CHANGED
@@ -49,9 +49,8 @@ _URLS = {
|
|
49 |
# "binaries": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_binaries.tar.gz", #
|
50 |
# "annotations": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_dataset-sample_gt.json" #"
|
51 |
"binaries": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_train-val-test_binaries.tar.gz", # DUDE_train-val-test_binaries.tar.gz
|
52 |
-
"annotations": "https://zenodo.org/record/
|
53 |
}
|
54 |
-
|
55 |
SKIP_DOC_IDS = ["ef03364aa27a0987c9870472e312aceb", "5c5a5880e6a73b4be2315d506ab0b15b"]
|
56 |
|
57 |
|
|
|
49 |
# "binaries": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_binaries.tar.gz", #
|
50 |
# "annotations": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_dataset-sample_gt.json" #"
|
51 |
"binaries": "https://huggingface.co/datasets/jordyvl/DUDE_loader/resolve/main/data/DUDE_train-val-test_binaries.tar.gz", # DUDE_train-val-test_binaries.tar.gz
|
52 |
+
"annotations": "https://zenodo.org/record/7662385/files/DUDE_gt_release-candidate_trainval.json?download=1",
|
53 |
}
|
|
|
54 |
SKIP_DOC_IDS = ["ef03364aa27a0987c9870472e312aceb", "5c5a5880e6a73b4be2315d506ab0b15b"]
|
55 |
|
56 |
|
test_loader.py
CHANGED
@@ -22,7 +22,7 @@ from codetiming import Timer
|
|
22 |
for binding in ["dict_annotations (new)"]: #"dict_PDF",
|
23 |
with Timer(name=f"{binding}", text=binding + " Elapsed time: {:.4f} seconds"):
|
24 |
if binding == "dict_annotations (new)":
|
25 |
-
ds = load_dataset("
|
26 |
else:
|
27 |
ds = load_dataset("jordyvl/DUDE_loader", revision='db20bbf751b14e14e8143170bc201948ef5ac83c')
|
28 |
|
|
|
22 |
for binding in ["dict_annotations (new)"]: #"dict_PDF",
|
23 |
with Timer(name=f"{binding}", text=binding + " Elapsed time: {:.4f} seconds"):
|
24 |
if binding == "dict_annotations (new)":
|
25 |
+
ds = load_dataset("jordyvl/DUDE_loader", 'Amazon_original', data_dir="/home/jordy/Downloads/DUDE_train-val-test_binaries") #ignore_verifications=True, , writer_batch_size=10
|
26 |
else:
|
27 |
ds = load_dataset("jordyvl/DUDE_loader", revision='db20bbf751b14e14e8143170bc201948ef5ac83c')
|
28 |
|