Upload 2 files
Browse files- README-75m.md +38 -0
- get_bigdocs_75m.py +357 -57
README-75m.md
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# BigDocs-7.5M
|
2 |
+
#### Training data for the paper: [BigDocs: An Open and Permissively-Licensed Dataset for Training Multimodal Models on Document and Code Tasks](https://huggingface.co/datasets/ServiceNow/BigDocs-Bench-Collections/)
|
3 |
+
|
4 |
+
🌐 [Homepage](https://bigdocs.github.io) | 📖 [arXiv](https://arxiv.org/pdf/2412.04626)
|
5 |
+
|
6 |
+
|
7 |
+
## Guide on Data Loading
|
8 |
+
Some parts of BigDocs-7.5M are distributed without their "image" column, and instead have an "img_id" column. The file `get_bigdocs_75m.py`, part of this repository, provides tooling to substitutes such images back in.
|
9 |
+
|
10 |
+
```python
|
11 |
+
from get_bigdocs_75m import get_bigdocs_75m
|
12 |
+
|
13 |
+
arxivocr = get_bigdocs_75m("ArxivOCR")
|
14 |
+
arxivtablecap = get_bigdocs_75m("ArxivTableCap")
|
15 |
+
cocotext = get_bigdocs_75m("COCOtext", user_local_path=".../train2014")
|
16 |
+
pubtables1m = get_bigdocs_75m("pubtables-1m", user_local_path=".../PubTables-1M-Detection/images")
|
17 |
+
textocr = get_bigdocs_75m("TextOCR", user_local_path=".../train")
|
18 |
+
tabfact = get_bigdocs_75m("TabFact", user_local_path=".../Table-Fact-Checking")
|
19 |
+
open4business = get_bigdocs_75m("Open4Business", user_local_path=".../Open4Business")
|
20 |
+
wikitq = get_bigdocs_75m("WikiTQ", user_local_path=".../WikiTableQuestions")
|
21 |
+
```
|
22 |
+
|
23 |
+
When specified, `user_local_path` must point to one of the third-party datasets listed below.
|
24 |
+
|
25 |
+
- COCOtext: http://images.cocodataset.org/zips/train2014.zip
|
26 |
+
- pubtables-1m: https://www.microsoft.com/en-us/research/publication/pubtables-1m
|
27 |
+
- TextOCR: https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip
|
28 |
+
- TabFact: https://github.com/wenhuchen/Table-Fact-Checking
|
29 |
+
- Open4Business: https://github.com/amanpreet692/Open4Business
|
30 |
+
- WikiTQ: https://github.com/ppasupat/WikiTableQuestions
|
31 |
+
|
32 |
+
You may specify `num_proc` as you would for `datasets.map`. See the docstring in `get_bigdocs_75m.py` for more details.
|
33 |
+
|
34 |
+
|
35 |
+
## Licensing
|
36 |
+
The part of this repository generated by us is Copyright ServiceNow 2024 and licensed under the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license.
|
37 |
+
|
38 |
+
Multiple datasets, documents, and tools were involved in the generation of BigDocs-Bench. We document these dependencies on a per-sample basis through the `query_info`, `annotation_info` and `image_info` fields, respectively documenting the `query`, `annotations` and `image` fields of our datasets.
|
get_bigdocs_75m.py
CHANGED
@@ -1,13 +1,34 @@
|
|
|
|
1 |
from typing import Optional
|
2 |
import datasets
|
3 |
import io
|
4 |
-
import PIL
|
5 |
-
import PIL.PngImagePlugin
|
6 |
import os
|
|
|
7 |
import hashlib
|
8 |
import warnings
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
|
|
|
11 |
ASSEMBLED_COLUMNS = (
|
12 |
'sample_id',
|
13 |
'dataset_name',
|
@@ -21,21 +42,319 @@ ASSEMBLED_COLUMNS = (
|
|
21 |
'image_sha256'
|
22 |
)
|
23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
m = hashlib.sha256()
|
27 |
m.update(b)
|
28 |
return m.hexdigest()
|
29 |
|
30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
def get_bigdocs_75m(
|
32 |
-
formal_name:
|
33 |
-
user_local_path: Optional[str],
|
34 |
-
|
35 |
-
|
|
|
|
|
36 |
raise_on_missing: Optional[bool]=None,
|
37 |
-
|
38 |
-
bigdocs_load_dataset_kwargs: Optional[dict]=None
|
|
|
39 |
) -> datasets.DatasetDict:
|
40 |
"""
|
41 |
Get a subset of BigDocs-7.5M
|
@@ -50,76 +369,57 @@ def get_bigdocs_75m(
|
|
50 |
- COCOtext: http://images.cocodataset.org/zips/train2014.zip
|
51 |
- pubtables-1m: https://www.microsoft.com/en-us/research/publication/pubtables-1m
|
52 |
- TextOCR: https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip
|
|
|
|
|
|
|
53 |
|
54 |
Args:
|
55 |
-
formal_name (`
|
56 |
-
user_local_path (`Optional[str]
|
57 |
-
load_from_cache_file (`Optional[bool]
|
58 |
-
num_proc (`Optional[int]
|
59 |
-
|
|
|
60 |
Determines what to do when there is an error loading an image.
|
61 |
- `True`: raise an error.
|
62 |
- `None`: print a warning and skip the sample (default).
|
63 |
- `False`: silently skip the sample.
|
64 |
-
use_bad_sha256 (`Optional[bool]
|
65 |
Determines what to do when the sha256 integrity test fails.
|
66 |
- `True`: ignore the sha256 integrity test.
|
67 |
- `None`: print a warning and skip samples with bad sha256 (default).
|
68 |
- `False`: silently skip entries with bad sha256.
|
69 |
-
|
|
|
70 |
"""
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
except Exception as e:
|
81 |
-
if raise_on_missing:
|
82 |
-
raise RuntimeError(f"Error loading image at {img_path}\n{e}")
|
83 |
-
if raise_on_missing is None:
|
84 |
-
warnings.warn(f"Skipping due to error loading image {img_path}", RuntimeWarning)
|
85 |
-
image = None # Sample will be filtered out
|
86 |
-
if image is not None:
|
87 |
-
# Place into `buffer` using PNG image format
|
88 |
-
buffer = io.BytesIO()
|
89 |
-
image.save(buffer, "png")
|
90 |
-
# Reload the image with guaranteed PNG format
|
91 |
-
image = PIL.Image.open(buffer)
|
92 |
-
# Check sha256
|
93 |
-
if not skip_bad_sha256:
|
94 |
-
sha256 = _hash_bytes(buffer.getvalue())
|
95 |
-
if sha256 != sample["image_sha256"]:
|
96 |
-
image = None # Sample will be filtered out
|
97 |
-
if skip_bad_sha256 is None:
|
98 |
-
warnings.warn(f"Skipping due to bad sha256 for {img_path}", RuntimeWarning)
|
99 |
-
return {"image": image}
|
100 |
-
|
101 |
# Get the correct processor
|
102 |
try:
|
103 |
-
|
104 |
-
"COCOtext": on_disk_processor,
|
105 |
-
"pubtables-1m": on_disk_processor,
|
106 |
-
"TextOCR": on_disk_processor,
|
107 |
-
}[formal_name]
|
108 |
except KeyError:
|
109 |
raise ValueError(f"Unknown formal_name: {formal_name}")
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
if processor is None:
|
114 |
processed = unprocessed
|
115 |
else:
|
|
|
116 |
processed = unprocessed.map(
|
117 |
processor,
|
118 |
remove_columns="img_id",
|
119 |
load_from_cache_file=load_from_cache_file,
|
120 |
-
num_proc=num_proc
|
|
|
121 |
)
|
122 |
-
# Drop missing images.
|
123 |
if not raise_on_missing:
|
124 |
processed = processed.filter((lambda image: image is not None), input_columns="image", num_proc=num_proc)
|
125 |
# Column order
|
|
|
1 |
+
import abc
|
2 |
from typing import Optional
|
3 |
import datasets
|
4 |
import io
|
|
|
|
|
5 |
import os
|
6 |
+
import json
|
7 |
import hashlib
|
8 |
import warnings
|
9 |
+
import csv
|
10 |
+
import textwrap
|
11 |
+
|
12 |
+
# Tested with pandas=2.2.2
|
13 |
+
import pandas as pd
|
14 |
+
|
15 |
+
# Tested with pillow==10.4.0
|
16 |
+
import PIL
|
17 |
+
import PIL.Image
|
18 |
+
import PIL.PngImagePlugin
|
19 |
+
|
20 |
+
# Tested with PyMuPDF==1.24.7 PyMuPDFb==1.24.6
|
21 |
+
import pymupdf
|
22 |
+
|
23 |
+
# Tested with reportlab==4.2.2
|
24 |
+
import reportlab
|
25 |
+
import reportlab.lib.colors
|
26 |
+
import reportlab.lib.pagesizes
|
27 |
+
import reportlab.platypus
|
28 |
+
import reportlab.pdfgen
|
29 |
|
30 |
|
31 |
+
# Once properly assembled, the datasets should have these columns
|
32 |
ASSEMBLED_COLUMNS = (
|
33 |
'sample_id',
|
34 |
'dataset_name',
|
|
|
42 |
'image_sha256'
|
43 |
)
|
44 |
|
45 |
+
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
|
46 |
+
# Now scroll down to the very bottom for `get_bigdocs_75m`! #
|
47 |
+
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
|
48 |
+
|
49 |
+
|
50 |
+
def pdf_to_image(pdf_stream, crop: bool=False) -> PIL.PngImagePlugin.PngImageFile:
|
51 |
+
doc = pymupdf.open(stream=pdf_stream)
|
52 |
+
# Render each page to an image in a list of images
|
53 |
+
images = [
|
54 |
+
PIL.Image.open(io.BytesIO(page.get_pixmap(dpi=144).tobytes("png")))
|
55 |
+
for page in doc
|
56 |
+
]
|
57 |
+
# Crop the footer if crop is True
|
58 |
+
if crop:
|
59 |
+
for i in range(len(images)):
|
60 |
+
images[i] = images[i].crop(
|
61 |
+
(120, 120, images[i].width - 120, images[i].height - 140)
|
62 |
+
)
|
63 |
+
# Determine the total width and height of the combined image
|
64 |
+
total_width = max(im.width for im in images)
|
65 |
+
total_height = sum(im.height for im in images)
|
66 |
+
# Create a new image with the combined size
|
67 |
+
combined_image = PIL.Image.new("RGB", (total_width, total_height), "white")
|
68 |
+
# Paste each page image into the combined image
|
69 |
+
y_offset = 0
|
70 |
+
for im in images:
|
71 |
+
combined_image.paste(im, (0, y_offset))
|
72 |
+
y_offset += im.height
|
73 |
+
# At this point combined_image is what we want, but under the wrong Python type
|
74 |
+
# Place into buffer using PNG image format
|
75 |
+
buffer = io.BytesIO()
|
76 |
+
combined_image.save(buffer, "png")
|
77 |
+
# Reload as PIL.PngImagePlugin.PngImageFile
|
78 |
+
return PIL.Image.open(buffer)
|
79 |
+
|
80 |
+
|
81 |
+
WHITE = reportlab.lib.colors.white
|
82 |
+
BLACK = reportlab.lib.colors.black
|
83 |
+
TABLE_STYLE = reportlab.platypus.TableStyle([
|
84 |
+
("BACKGROUND", (0, 0), (-1, 0), WHITE),
|
85 |
+
("TEXTCOLOR", (0, 0), (-1, 0), BLACK),
|
86 |
+
("ALIGN", (0, 0), (-1, -1), "CENTER"),
|
87 |
+
("FONTNAME", (0, 0), (-1, 0), "Helvetica-Bold"),
|
88 |
+
("BOTTOMPADDING", (0, 0), (-1, 0), 12),
|
89 |
+
("BACKGROUND", (0, 1), (-1, -1), WHITE),
|
90 |
+
("GRID", (0, 0), (-1, -1), 1, BLACK),
|
91 |
+
])
|
92 |
+
|
93 |
+
|
94 |
+
def csv_to_image(csv_path: str) -> PIL.PngImagePlugin.PngImageFile:
|
95 |
+
# Load the data
|
96 |
+
with open(csv_path, newline="") as csv_stream:
|
97 |
+
reader = csv.reader(csv_stream, delimiter="#")
|
98 |
+
data = [row for row in reader]
|
99 |
+
# Create a table with CSV data
|
100 |
+
table = reportlab.platypus.Table(data)
|
101 |
+
table.setStyle(TABLE_STYLE)
|
102 |
+
# Virtual PDF file
|
103 |
+
pdf_stream = io.BytesIO()
|
104 |
+
# Build a document from the table
|
105 |
+
reportlab.platypus.SimpleDocTemplate(pdf_stream, pagesize=reportlab.lib.pagesizes.letter).build([table])
|
106 |
+
|
107 |
+
return pdf_to_image(pdf_stream)
|
108 |
+
|
109 |
+
|
110 |
+
def text_to_image(text: str, font_size: int = 10) -> PIL.PngImagePlugin.PngImageFile:
|
111 |
+
pdf_stream = io.BytesIO()
|
112 |
+
c = reportlab.pdfgen.canvas.Canvas(
|
113 |
+
pdf_stream,
|
114 |
+
pagesize=reportlab.lib.pagesizes.letter
|
115 |
+
)
|
116 |
+
c.setFont("Helvetica", font_size)
|
117 |
+
|
118 |
+
# Wrap the text for better handling in PDF
|
119 |
+
wrapped_text = textwrap.wrap(text, width=100)
|
120 |
+
|
121 |
+
# Starting position on the page
|
122 |
+
x_position = 72 # 1 inch from the left margin
|
123 |
+
y_position = 11 * 72 - 72 # Start 1 inch from the top of an 11-inch page
|
124 |
+
|
125 |
+
text_height = font_size * 1.2 # Approximate line height
|
126 |
|
127 |
+
for line in wrapped_text:
|
128 |
+
if y_position < 72: # Check if we're near the bottom of the page
|
129 |
+
c.showPage()
|
130 |
+
c.setFont("Helvetica", font_size)
|
131 |
+
y_position = 11 * 72 - 72 # Reset position to top of new page
|
132 |
+
|
133 |
+
c.drawString(x_position, y_position, line)
|
134 |
+
y_position -= text_height # Move down for next line
|
135 |
+
|
136 |
+
c.save()
|
137 |
+
|
138 |
+
return pdf_to_image(pdf_stream, crop=True)
|
139 |
+
|
140 |
+
|
141 |
+
def csv_to_markdown(csv_path: str, sep: str) -> str:
|
142 |
+
# Format with pandas, but ensure there are no consecutive spaces
|
143 |
+
df = pd.read_csv(csv_path, sep=sep)
|
144 |
+
return " ".join(df.to_markdown(index=False).split())
|
145 |
+
|
146 |
+
|
147 |
+
def get_sha256(
|
148 |
+
image: PIL.PngImagePlugin.PngImageFile,
|
149 |
+
b: Optional[bytes]=None
|
150 |
+
) -> str:
|
151 |
+
# Ignore image if bytes representation b is already provided
|
152 |
+
if b is None:
|
153 |
+
buffer = io.BytesIO()
|
154 |
+
image.save(buffer, "png")
|
155 |
+
b = buffer.getvalue()
|
156 |
m = hashlib.sha256()
|
157 |
m.update(b)
|
158 |
return m.hexdigest()
|
159 |
|
160 |
|
161 |
+
class Assembler(abc.ABC):
|
162 |
+
def __init__(
|
163 |
+
self,
|
164 |
+
user_local_path: Optional[str],
|
165 |
+
raise_on_missing: Optional[bool],
|
166 |
+
use_bad_sha256: Optional[bool]
|
167 |
+
):
|
168 |
+
self.user_local_path = user_local_path
|
169 |
+
self.raise_on_missing = raise_on_missing
|
170 |
+
self.use_bad_sha256 = use_bad_sha256
|
171 |
+
|
172 |
+
@abc.abstractmethod
|
173 |
+
def __call__(self, sample) -> dict:
|
174 |
+
"""Processor called by `map` on each sample"""
|
175 |
+
|
176 |
+
def keep_image(
|
177 |
+
self,
|
178 |
+
image: Optional[PIL.PngImagePlugin.PngImageFile],
|
179 |
+
expected_sha256: str,
|
180 |
+
b: Optional[bytes]=None
|
181 |
+
) -> bool:
|
182 |
+
if image is None:
|
183 |
+
# No image to check
|
184 |
+
return False
|
185 |
+
if self.use_bad_sha256:
|
186 |
+
# We're going to use the image whether or not the sha256 is good
|
187 |
+
return True
|
188 |
+
sha256 = get_sha256(image, b)
|
189 |
+
if sha256 != expected_sha256:
|
190 |
+
# Give warning if didn't explicitly set use_bad_sha256 to False
|
191 |
+
if self.use_bad_sha256 is None:
|
192 |
+
warnings.warn(f"Skipping due to bad sha256", RuntimeWarning)
|
193 |
+
# Sample will be filtered out
|
194 |
+
return False
|
195 |
+
else:
|
196 |
+
# Sample will be used
|
197 |
+
return True
|
198 |
+
|
199 |
+
|
200 |
+
class AssembleFromDisk(Assembler):
|
201 |
+
def __init__(self, *args, **kwargs):
|
202 |
+
super().__init__(*args, **kwargs)
|
203 |
+
assert self.user_local_path is not None, f"user_local_path is mandatory for this dataset"
|
204 |
+
|
205 |
+
def __call__(self, sample) -> dict:
|
206 |
+
img_path = os.path.join(self.user_local_path, sample['img_id'])
|
207 |
+
# Load the image
|
208 |
+
try:
|
209 |
+
image = PIL.Image.open(img_path)
|
210 |
+
except Exception as e:
|
211 |
+
if self.raise_on_missing:
|
212 |
+
raise RuntimeError(f"Error loading image at {img_path}\n{e}")
|
213 |
+
if self.raise_on_missing is None:
|
214 |
+
warnings.warn(f"Skipping due to error loading image {img_path}\n{e}", RuntimeWarning)
|
215 |
+
image = None # Sample will be filtered out
|
216 |
+
if image is not None:
|
217 |
+
# Place into `buffer` using PNG image format
|
218 |
+
buffer = io.BytesIO()
|
219 |
+
image.save(buffer, "png")
|
220 |
+
# Reload the image with guaranteed PNG format
|
221 |
+
image = PIL.Image.open(buffer)
|
222 |
+
# Check sha256
|
223 |
+
if not self.keep_image(image, sample["image_sha256"], b=buffer.getvalue()):
|
224 |
+
image = None
|
225 |
+
return {"image": image}
|
226 |
+
|
227 |
+
|
228 |
+
class AssembleTabFact(Assembler):
|
229 |
+
def __init__(self, *args, **kwargs):
|
230 |
+
super().__init__(*args, **kwargs)
|
231 |
+
# Get annotations
|
232 |
+
json_path = os.path.join(self.user_local_path, "tokenized_data/total_examples.json")
|
233 |
+
with open(json_path, "rt") as fp:
|
234 |
+
self.tables_data = json.load(fp)
|
235 |
+
|
236 |
+
def __call__(self, sample) -> dict:
|
237 |
+
csv_path = os.path.join(self.user_local_path, "data/all_csv", sample["img_id"])
|
238 |
+
image = csv_to_image(csv_path)
|
239 |
+
# Check sha256
|
240 |
+
if not self.keep_image(image, sample["image_sha256"]):
|
241 |
+
# Skip both image and annotations (will be filtered out)
|
242 |
+
return {"image": None, "annotations": [""]}
|
243 |
+
# Annotations
|
244 |
+
if sample["task_name"] == "table_parsing2md":
|
245 |
+
annotations = [csv_to_markdown(csv_path, sep="#")]
|
246 |
+
else:
|
247 |
+
facts, entails, title = self.tables_data[sample["img_id"]]
|
248 |
+
if sample["task_name"] == "caption":
|
249 |
+
# The "caption" is the table's title
|
250 |
+
annotations = [title]
|
251 |
+
else:
|
252 |
+
assert sample["task_name"] == "summary"
|
253 |
+
# select only entries in facts with entailment label as 1
|
254 |
+
facts = [fact for fact, entailment in zip(facts, entails) if entailment == 1]
|
255 |
+
# concat facts with numbered bullets and new lines
|
256 |
+
annotations = ["\n".join([f"{i+1}. {fact}" for i, fact in enumerate(facts)])]
|
257 |
+
return {"image": image, "annotations": annotations}
|
258 |
+
|
259 |
+
|
260 |
+
class AssembleOpen4Business(Assembler):
|
261 |
+
def __init__(self, *args, **kwargs):
|
262 |
+
super().__init__(*args, **kwargs)
|
263 |
+
# Get annotations
|
264 |
+
self.texts, self.summaries = {}, {}
|
265 |
+
for split in ["train", "val", "test"]:
|
266 |
+
with open(os.path.join(self.user_local_path, f"{split}.source"), "rt") as fp:
|
267 |
+
# Don't strip here because that may alter the image sha256
|
268 |
+
self.texts[split] = list(line for line in fp)
|
269 |
+
with open(os.path.join(self.user_local_path, f"{split}.target"), "rt") as fp:
|
270 |
+
self.summaries[split] = list(line.strip() for line in fp)
|
271 |
+
|
272 |
+
def __call__(self, sample) -> dict:
|
273 |
+
split, line_number = sample["img_id"].split("_")
|
274 |
+
line_number = int(line_number)
|
275 |
+
print(split, line_number, sample["task_name"])
|
276 |
+
try:
|
277 |
+
text = self.texts[split][line_number]
|
278 |
+
image = text_to_image(text)
|
279 |
+
# Check sha256
|
280 |
+
if not self.keep_image(image, sample["image_sha256"]):
|
281 |
+
# Skip both image and annotations (will be filtered out)
|
282 |
+
return {"image": None, "annotations": [""]}
|
283 |
+
# Annotations
|
284 |
+
if sample["task_name"] == "extraction":
|
285 |
+
# Don't forget the strip!
|
286 |
+
annotations = [text.strip()]
|
287 |
+
else:
|
288 |
+
assert sample["task_name"] == "summary"
|
289 |
+
annotations = [self.summaries[split][line_number]]
|
290 |
+
except Exception as e:
|
291 |
+
print(f"EXCEPTION on {split}, {line_number}, {sample['task_name']}. Original error: {e.__str__}")
|
292 |
+
raise e
|
293 |
+
return {"image": image, "annotations": annotations}
|
294 |
+
|
295 |
+
|
296 |
+
class AssembleWikiTQ(Assembler):
|
297 |
+
def __init__(self, *args, **kwargs):
|
298 |
+
super().__init__(*args, **kwargs)
|
299 |
+
# Get annotations
|
300 |
+
self.annotations = pd.concat([
|
301 |
+
pd.read_csv(
|
302 |
+
os.path.join(self.user_local_path, "data", tsv_file),
|
303 |
+
sep="\t", on_bad_lines="skip", index_col="id"
|
304 |
+
) for tsv_file in {
|
305 |
+
"training.tsv",
|
306 |
+
"pristine-seen-tables.tsv",
|
307 |
+
"pristine-unseen-tables.tsv"
|
308 |
+
}
|
309 |
+
])
|
310 |
+
|
311 |
+
def __call__(self, sample) -> dict:
|
312 |
+
question, context, answer = self.annotations.loc[sample["img_id"]]
|
313 |
+
csv_path = os.path.join(self.user_local_path, context)
|
314 |
+
image = csv_to_image(csv_path)
|
315 |
+
# Check sha256
|
316 |
+
if not self.keep_image(image, sample["image_sha256"]):
|
317 |
+
# Skip both image and annotations (will be filtered out)
|
318 |
+
return {"image": None, "annotations": [""]}
|
319 |
+
# Annotations
|
320 |
+
if sample["task_name"] == "table_parsing2md":
|
321 |
+
annotations = [csv_to_markdown(csv_path, sep=",")]
|
322 |
+
return {"image": image, "annotations": annotations}
|
323 |
+
else:
|
324 |
+
assert sample["task_name"] == "qa"
|
325 |
+
query = [sample["query"][0].format(question=question)]
|
326 |
+
answers = str(answer).split("|")
|
327 |
+
answer = " or ".join(answers) if len(answers) > 1 else answers[0]
|
328 |
+
annotations = [sample["annotations"][0].format(answer=answer)]
|
329 |
+
return {"image": image, "query": query, "annotations": annotations}
|
330 |
+
|
331 |
+
|
332 |
+
KNOWN_ASSEMBLERS = {
|
333 |
+
"ArxivOCR": None,
|
334 |
+
"ArxivTableCap": None,
|
335 |
+
"COCOtext": AssembleFromDisk,
|
336 |
+
"pubtables-1m": AssembleFromDisk,
|
337 |
+
"TextOCR": AssembleFromDisk,
|
338 |
+
"TabFact": AssembleTabFact,
|
339 |
+
"Open4Business": AssembleOpen4Business,
|
340 |
+
"WikiTQ": AssembleWikiTQ,
|
341 |
+
}
|
342 |
+
|
343 |
+
|
344 |
+
# # # # # # # # # # # # # # # # # # # # # # #
|
345 |
+
# This is the function you are looking for! #
|
346 |
+
# # # # # # # # # # # # # # # # # # # # # # #
|
347 |
def get_bigdocs_75m(
|
348 |
+
formal_name: str,
|
349 |
+
user_local_path: Optional[str]=None,
|
350 |
+
*,
|
351 |
+
load_from_cache_file: Optional[bool]=None,
|
352 |
+
num_proc: Optional[int]=4,
|
353 |
+
writer_batch_size: int=100,
|
354 |
raise_on_missing: Optional[bool]=None,
|
355 |
+
use_bad_sha256: Optional[bool]=None,
|
356 |
+
bigdocs_load_dataset_kwargs: Optional[dict]=None,
|
357 |
+
unprocessed: Optional[datasets.DatasetDict]=None
|
358 |
) -> datasets.DatasetDict:
|
359 |
"""
|
360 |
Get a subset of BigDocs-7.5M
|
|
|
369 |
- COCOtext: http://images.cocodataset.org/zips/train2014.zip
|
370 |
- pubtables-1m: https://www.microsoft.com/en-us/research/publication/pubtables-1m
|
371 |
- TextOCR: https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip
|
372 |
+
- TabFact: https://github.com/wenhuchen/Table-Fact-Checking
|
373 |
+
- Open4Business: https://github.com/amanpreet692/Open4Business
|
374 |
+
- WikiTQ: https://github.com/ppasupat/WikiTableQuestions
|
375 |
|
376 |
Args:
|
377 |
+
formal_name (`str`): The desired subset of BigDocs-7.5M .
|
378 |
+
user_local_path (`Optional[str]`): The local path containing the images to be linked.
|
379 |
+
load_from_cache_file (`Optional[bool]): Passed to `map`, `filter` and the likes.
|
380 |
+
num_proc (`Optional[int]): Passed to `map`, `filter` and the likes.
|
381 |
+
writer_batch_size (`int`, defaults to 100): Passed to `map`. Too large values may cause OOM.
|
382 |
+
raise_on_missing (`Optional[bool]`):
|
383 |
Determines what to do when there is an error loading an image.
|
384 |
- `True`: raise an error.
|
385 |
- `None`: print a warning and skip the sample (default).
|
386 |
- `False`: silently skip the sample.
|
387 |
+
use_bad_sha256 (`Optional[bool]):
|
388 |
Determines what to do when the sha256 integrity test fails.
|
389 |
- `True`: ignore the sha256 integrity test.
|
390 |
- `None`: print a warning and skip samples with bad sha256 (default).
|
391 |
- `False`: silently skip entries with bad sha256.
|
392 |
+
bigdocs_load_dataset_kwargs (`Optional[dict]`): Arguments passed to datasets.load_dataset when retrieving ServiceNow/BigDocs-7.5M .
|
393 |
+
unprocessed (Optional[datasets.DatasetDict]): If provided, will be used in stead of ServiceNow/BigDocs-7.5M .
|
394 |
"""
|
395 |
+
# Get the unprocessed ServiceNow/BigDocs-7.5M
|
396 |
+
if unprocessed is None:
|
397 |
+
if bigdocs_load_dataset_kwargs is None:
|
398 |
+
bigdocs_load_dataset_kwargs = {}
|
399 |
+
unprocessed = datasets.load_dataset(
|
400 |
+
"ServiceNow/BigDocs-7.5M",
|
401 |
+
formal_name,
|
402 |
+
**bigdocs_load_dataset_kwargs
|
403 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
404 |
# Get the correct processor
|
405 |
try:
|
406 |
+
assembler = KNOWN_ASSEMBLERS[formal_name]
|
|
|
|
|
|
|
|
|
407 |
except KeyError:
|
408 |
raise ValueError(f"Unknown formal_name: {formal_name}")
|
409 |
+
# Do the processing
|
410 |
+
if assembler is None:
|
411 |
+
assert user_local_path is None
|
|
|
412 |
processed = unprocessed
|
413 |
else:
|
414 |
+
processor = assembler(user_local_path, raise_on_missing, use_bad_sha256)
|
415 |
processed = unprocessed.map(
|
416 |
processor,
|
417 |
remove_columns="img_id",
|
418 |
load_from_cache_file=load_from_cache_file,
|
419 |
+
num_proc=num_proc,
|
420 |
+
writer_batch_size=writer_batch_size
|
421 |
)
|
422 |
+
# Drop missing images (we can skip if we raised on missing images).
|
423 |
if not raise_on_missing:
|
424 |
processed = processed.filter((lambda image: image is not None), input_columns="image", num_proc=num_proc)
|
425 |
# Column order
|