holylovenia commited on
Commit
21c9833
1 Parent(s): 069cd33

Upload total_defense_meme.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. total_defense_meme.py +277 -0
total_defense_meme.py ADDED
@@ -0,0 +1,277 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 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
+
16
+ import json
17
+ from pathlib import Path
18
+ from typing import Dict, List, Tuple
19
+
20
+ import datasets
21
+ import gdown
22
+
23
+ from seacrowd.utils import schemas
24
+ from seacrowd.utils.configs import SEACrowdConfig
25
+ from seacrowd.utils.constants import TASK_TO_SCHEMA, Licenses, Tasks
26
+
27
+ _CITATION = """\
28
+ @inproceedings{10.1145/3587819.3592545,
29
+ author = {Prakash, Nirmalendu and Hee, Ming Shan and Lee, Roy Ka-Wei},
30
+ title = {TotalDefMeme: A Multi-Attribute Meme dataset on Total Defence in Singapore},
31
+ year = {2023},
32
+ isbn = {9798400701481},
33
+ publisher = {Association for Computing Machinery},
34
+ address = {New York, NY, USA},
35
+ url = {https://doi.org/10.1145/3587819.3592545},
36
+ doi = {10.1145/3587819.3592545},
37
+ booktitle = {Proceedings of the 14th Conference on ACM Multimedia Systems},
38
+ pages = {369–375},
39
+ numpages = {7},
40
+ keywords = {multimodal, meme, dataset, topic clustering, stance classification},
41
+ location = {Vancouver, BC, Canada},
42
+ series = {MMSys '23}
43
+ }
44
+ """
45
+
46
+ _DATASETNAME = "total_defense_meme"
47
+
48
+ _DESCRIPTION = """\
49
+ This is a large-scale multimodal and multi-attribute dataset containing memes
50
+ about Singapore's Total Defence policy from different social media platforms.
51
+ The type (Singaporean or generic), pillars (military, civil, economic, social,
52
+ psychological, digital, others), topics and stances (against, neutral,
53
+ supportive) of each meme are manually identified by annotators.
54
+ """
55
+
56
+ _HOMEPAGE = "https://gitlab.com/bottle_shop/meme/TotalDefMemes"
57
+
58
+ _LANGUAGES = ["eng"]
59
+
60
+ _LICENSE = Licenses.UNKNOWN.value
61
+
62
+ _LOCAL = False
63
+
64
+ _URLS = {
65
+ "image": "https://drive.google.com/file/d/1oJIh4QQS3Idff2g6bZORstS5uBROjUUz/view?usp=share_link",
66
+ "annotations": "https://gitlab.com/bottle_shop/meme/TotalDefMemes/-/raw/main/report/annotation.json?ref_type=heads",
67
+ }
68
+
69
+ _SUPPORTED_TASKS = [Tasks.OPTICAL_CHARACTER_RECOGNITION, Tasks.IMAGE_CLASSIFICATION_MULTILABEL]
70
+ _SEACROWD_SCHEMA = {
71
+ task.value: f"seacrowd_{TASK_TO_SCHEMA[task].lower()}" for task in _SUPPORTED_TASKS
72
+ } # ocr: imtext, imc_multi: image_multi
73
+
74
+ _SOURCE_VERSION = "1.0.0"
75
+
76
+ _SEACROWD_VERSION = "2024.06.20"
77
+
78
+
79
+ class TotalDefenseMemeDataset(datasets.GeneratorBasedBuilder):
80
+ """Multimodal dataset containing memes about Singapore's Total Defence policy"""
81
+
82
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
83
+ SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
84
+
85
+ BUILDER_CONFIGS = [
86
+ SEACrowdConfig(
87
+ name=f"{_DATASETNAME}_source",
88
+ version=SOURCE_VERSION,
89
+ description=f"{_DATASETNAME} source schema",
90
+ schema="source",
91
+ subset_id=_DATASETNAME,
92
+ ),
93
+ SEACrowdConfig(
94
+ name=f"{_DATASETNAME}_{_SEACROWD_SCHEMA['OCR']}",
95
+ version=SEACROWD_VERSION,
96
+ description=f"{_DATASETNAME} SEACrowd schema",
97
+ schema=_SEACROWD_SCHEMA["OCR"],
98
+ subset_id=_DATASETNAME,
99
+ ),
100
+ SEACrowdConfig(
101
+ name=f"{_DATASETNAME}_{_SEACROWD_SCHEMA['IMC_MULTI']}",
102
+ version=SEACROWD_VERSION,
103
+ description=f"{_DATASETNAME} SEACrowd schema",
104
+ schema=_SEACROWD_SCHEMA["IMC_MULTI"],
105
+ subset_id=_DATASETNAME,
106
+ ),
107
+ ]
108
+
109
+ DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_source"
110
+
111
+ def _info(self) -> datasets.DatasetInfo:
112
+ # define labelling
113
+ meme_type = ["Non_Memes", "Non_SG_Memes", "SG_Memes"]
114
+ pillar_type = [
115
+ "Social",
116
+ "Economic",
117
+ "Psychological",
118
+ "Military",
119
+ "Civil",
120
+ "Digital",
121
+ "Others",
122
+ ]
123
+ stance_type = ["Against", "Neutral", "Supportive"]
124
+
125
+ if self.config.schema == "source":
126
+ features = datasets.Features(
127
+ {
128
+ "image_path": datasets.Value("string"),
129
+ "meme_type": datasets.Sequence(datasets.ClassLabel(names=meme_type)),
130
+ "text": datasets.Value("string"),
131
+ "tags": datasets.Sequence(datasets.Value("string")),
132
+ "pillar_stances": datasets.Sequence(
133
+ {
134
+ "category": datasets.ClassLabel(names=pillar_type),
135
+ "stance": datasets.Sequence(datasets.ClassLabel(names=stance_type)),
136
+ }
137
+ ),
138
+ }
139
+ )
140
+
141
+ elif self.config.schema == _SEACROWD_SCHEMA["OCR"]: # all images
142
+ features = schemas.image_text_features(label_names=meme_type)
143
+ features["metadata"] = {
144
+ "tags": datasets.Sequence(datasets.Value("string")),
145
+ "pillar_stances": datasets.Sequence(
146
+ {
147
+ "category": datasets.ClassLabel(names=pillar_type),
148
+ "stance": datasets.Sequence(datasets.ClassLabel(names=stance_type)),
149
+ }
150
+ ),
151
+ }
152
+ elif self.config.schema == _SEACROWD_SCHEMA["IMC_MULTI"]: # sg meme images only
153
+ features = schemas.image_multi_features(label_names=pillar_type)
154
+ features["metadata"] = {
155
+ "tags": datasets.Sequence(datasets.Value("string")),
156
+ "stances": datasets.Sequence(datasets.Sequence(datasets.ClassLabel(names=stance_type))),
157
+ }
158
+
159
+ return datasets.DatasetInfo(
160
+ description=_DESCRIPTION,
161
+ features=features,
162
+ homepage=_HOMEPAGE,
163
+ license=_LICENSE,
164
+ citation=_CITATION,
165
+ )
166
+
167
+ def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
168
+ """Returns SplitGenerators."""
169
+ # download image from gdrive
170
+ output_dir = Path.cwd() / "data" / _DATASETNAME
171
+ output_dir.mkdir(parents=True, exist_ok=True)
172
+ output_file = output_dir / f"{_DATASETNAME}.zip"
173
+ if not output_file.exists():
174
+ gdown.download(_URLS["image"], str(output_file), fuzzy=True)
175
+ else:
176
+ print(f"File already downloaded: {str(output_file)}")
177
+ # extract image data
178
+ image_dir = Path(dl_manager.extract(output_file)) / "TD_Memes"
179
+
180
+ # download annotations
181
+ annotation_path = Path(dl_manager.download(_URLS["annotations"]))
182
+ return [
183
+ datasets.SplitGenerator(
184
+ name=datasets.Split.TRAIN,
185
+ gen_kwargs={
186
+ "image_dir": image_dir,
187
+ "annotation_file": annotation_path,
188
+ },
189
+ ),
190
+ ]
191
+
192
+ def _generate_examples(self, image_dir: Path, annotation_file: Path) -> Tuple[int, Dict]:
193
+ """Yields examples as (key, example) tuples."""
194
+ # load annotation
195
+ with open(annotation_file, "r", encoding="utf-8") as file:
196
+ annotation = json.load(file)
197
+
198
+ # get unique image names
199
+ image_names = sorted(
200
+ list(
201
+ set(annotation["Non_Memes"])
202
+ | set(annotation["Non_SG_Memes"])
203
+ | set(annotation["SG_Memes"])
204
+ )
205
+ )
206
+
207
+ # annotation data is a list of dict, instead of dict of image names
208
+ def get_value(image_name, list_of_dicts):
209
+ for dictionary in list_of_dicts:
210
+ if image_name in dictionary:
211
+ return dictionary[image_name]
212
+ return None
213
+
214
+ key = 0
215
+ for image_name in image_names:
216
+ # assert image exist in directory
217
+ assert (image_dir / image_name).exists(), f"Image {image_name} not found"
218
+ image_path = str(image_dir / image_name)
219
+
220
+ # get categories, can be multiple
221
+ categories = []
222
+ if image_name in annotation["Non_Memes"]:
223
+ categories.append("Non_Memes")
224
+ if image_name in annotation["Non_SG_Memes"]:
225
+ categories.append("Non_SG_Memes")
226
+ if image_name in annotation["SG_Memes"]:
227
+ categories.append("SG_Memes")
228
+
229
+ # get attributes
230
+ text = get_value(image_name, annotation["Text"])
231
+ tags = get_value(image_name, annotation["Tags"])
232
+ raw_pillar_stances = get_value(image_name, annotation["Pillar_Stances"])
233
+
234
+ # process pillar stances
235
+ pillar_stances = []
236
+ if raw_pillar_stances:
237
+ for pillar, stances in raw_pillar_stances:
238
+ category = pillar.split(" ")[0]
239
+ pillar_stances.append({"category": category, "stance": stances})
240
+
241
+ # source schema
242
+ if self.config.schema == "source":
243
+ yield key, {
244
+ "image_path": image_path,
245
+ "meme_type": categories,
246
+ "text": text,
247
+ "tags": tags,
248
+ "pillar_stances": pillar_stances,
249
+ }
250
+ key += 1
251
+
252
+ # ocr seacrowd schema
253
+ elif self.config.schema == _SEACROWD_SCHEMA["OCR"]:
254
+ yield key, {
255
+ "id": str(key),
256
+ "image_paths": [image_path],
257
+ "texts": text,
258
+ "metadata": {
259
+ "tags": tags,
260
+ "pillar_stances": pillar_stances,
261
+ },
262
+ }
263
+ key += 1
264
+
265
+ # pillar/topic classification seacrowd schema
266
+ elif self.config.schema == _SEACROWD_SCHEMA["IMC_MULTI"]:
267
+ if pillar_stances: # only those with pillar stances
268
+ yield key, {
269
+ "id": str(key),
270
+ "labels": [pillar["category"] for pillar in pillar_stances],
271
+ "image_path": image_path,
272
+ "metadata": {
273
+ "tags": tags,
274
+ "stances": [pillar["stance"] for pillar in pillar_stances],
275
+ },
276
+ }
277
+ key += 1