Datasets:
Tasks:
Visual Question Answering
Formats:
parquet
Sub-tasks:
visual-question-answering
Languages:
English
Size:
100K - 1M
ArXiv:
License:
Delete loading script
Browse files- compguesswhat.py +0 -254
compguesswhat.py
DELETED
|
@@ -1,254 +0,0 @@
|
|
| 1 |
-
import gzip
|
| 2 |
-
import json
|
| 3 |
-
import os
|
| 4 |
-
|
| 5 |
-
import datasets
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
class CompguesswhatConfig(datasets.BuilderConfig):
|
| 9 |
-
"""BuilderConfig for CompGuessWhat?!"""
|
| 10 |
-
|
| 11 |
-
def __init__(self, data_url, splits, gameplay_scenario, **kwargs):
|
| 12 |
-
"""
|
| 13 |
-
|
| 14 |
-
Args:
|
| 15 |
-
gameplay_scenario: to specify if we want to load original CompGuessWhat?! split ('original') or
|
| 16 |
-
the zero-shot reference games based on NOCAPS images ('zero_shot')
|
| 17 |
-
**kwargs: keyword arguments forwarded to super.
|
| 18 |
-
"""
|
| 19 |
-
super(CompguesswhatConfig, self).__init__(
|
| 20 |
-
version=datasets.Version("0.2.0", "Second CompGuessWhat?! release"), **kwargs
|
| 21 |
-
)
|
| 22 |
-
assert gameplay_scenario in (
|
| 23 |
-
"original",
|
| 24 |
-
"zero_shot",
|
| 25 |
-
), "Invalid choice for parameter 'gameplay_scenario': {gameplay_scenario}. Valid values are ('original', 'zero_shot')."
|
| 26 |
-
|
| 27 |
-
self.gameplay_scenario = gameplay_scenario
|
| 28 |
-
self.splits = splits
|
| 29 |
-
self.data_url = data_url
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
class Compguesswhat(datasets.GeneratorBasedBuilder):
|
| 33 |
-
_CITATION = """\
|
| 34 |
-
@inproceedings{suglia2020compguesswhat,
|
| 35 |
-
title={CompGuessWhat?!: a Multi-task Evaluation Framework for Grounded Language Learning},
|
| 36 |
-
author={Suglia, Alessandro, Konstas, Ioannis, Vanzo, Andrea, Bastianelli, Emanuele, Desmond Elliott, Stella Frank and Oliver Lemon},
|
| 37 |
-
booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
|
| 38 |
-
year={2020}
|
| 39 |
-
}
|
| 40 |
-
"""
|
| 41 |
-
|
| 42 |
-
_DESCRIPTION = """
|
| 43 |
-
CompGuessWhat?! is an instance of a multi-task framework for evaluating the quality of learned neural representations,
|
| 44 |
-
in particular concerning attribute grounding. Use this dataset if you want to use the set of games whose reference
|
| 45 |
-
scene is an image in VisualGenome. Visit the website for more details: https://compguesswhat.github.io
|
| 46 |
-
"""
|
| 47 |
-
|
| 48 |
-
BUILDER_CONFIGS = [
|
| 49 |
-
CompguesswhatConfig(
|
| 50 |
-
name="compguesswhat-original",
|
| 51 |
-
gameplay_scenario="original",
|
| 52 |
-
description="CompGuessWhat?! subset of games from the original GuessWhat?! dataset",
|
| 53 |
-
data_url="https://www.dropbox.com/s/qd9wlydpkpmq8rr/compguesswhat-original.zip?dl=1",
|
| 54 |
-
splits={
|
| 55 |
-
"train": "compguesswhat.train.jsonl.gz",
|
| 56 |
-
"valid": "compguesswhat.valid.jsonl.gz",
|
| 57 |
-
"test": "compguesswhat.test.jsonl.gz",
|
| 58 |
-
},
|
| 59 |
-
),
|
| 60 |
-
CompguesswhatConfig(
|
| 61 |
-
name="compguesswhat-zero_shot",
|
| 62 |
-
gameplay_scenario="zero_shot",
|
| 63 |
-
description="CompGuessWhat?! reference set of games for zero-shot evaluation using NOCAPS images",
|
| 64 |
-
data_url="https://www.dropbox.com/s/f5o2t7aiok7kpcm/compguesswhat-zero_shot.zip?dl=1",
|
| 65 |
-
splits={
|
| 66 |
-
"nd_valid": "compguesswhat.nd_valid.jsonl.gz",
|
| 67 |
-
"nd_test": "compguesswhat.nd_test.jsonl.gz",
|
| 68 |
-
"od_valid": "compguesswhat.od_valid.jsonl.gz",
|
| 69 |
-
"od_test": "compguesswhat.od_test.jsonl.gz",
|
| 70 |
-
},
|
| 71 |
-
),
|
| 72 |
-
]
|
| 73 |
-
|
| 74 |
-
VERSION = datasets.Version("0.2.0")
|
| 75 |
-
|
| 76 |
-
def _info(self):
|
| 77 |
-
if self.config.gameplay_scenario == "original":
|
| 78 |
-
return datasets.DatasetInfo(
|
| 79 |
-
# This is the description that will appear on the datasets page.
|
| 80 |
-
description=self._DESCRIPTION,
|
| 81 |
-
# datasets.features.FeatureConnectors
|
| 82 |
-
features=datasets.Features(
|
| 83 |
-
{
|
| 84 |
-
"id": datasets.Value("int32"),
|
| 85 |
-
"target_id": datasets.Value("int32"),
|
| 86 |
-
"timestamp": datasets.Value("string"),
|
| 87 |
-
"status": datasets.Value("string"),
|
| 88 |
-
"image": {
|
| 89 |
-
# this is the image ID in GuessWhat?! which corresponds to the MSCOCO id
|
| 90 |
-
"id": datasets.Value("int32"),
|
| 91 |
-
"file_name": datasets.Value("string"),
|
| 92 |
-
"flickr_url": datasets.Value("string"),
|
| 93 |
-
"coco_url": datasets.Value("string"),
|
| 94 |
-
"height": datasets.Value("int32"),
|
| 95 |
-
"width": datasets.Value("int32"),
|
| 96 |
-
# this field represents the corresponding image metadata that can be found in VisualGenome
|
| 97 |
-
# in the file image_data.json
|
| 98 |
-
# We copy it over so that we avoid any confusion or possible wrong URL
|
| 99 |
-
# Please use the original image files to resolve photos
|
| 100 |
-
"visual_genome": {
|
| 101 |
-
"width": datasets.Value("int32"),
|
| 102 |
-
"height": datasets.Value("int32"),
|
| 103 |
-
"url": datasets.Value("string"),
|
| 104 |
-
"coco_id": datasets.Value("int32"),
|
| 105 |
-
# this is the actual VisualGenome image ID
|
| 106 |
-
# because we can't rely store it as an integer we same it as string
|
| 107 |
-
"flickr_id": datasets.Value("string"),
|
| 108 |
-
"image_id": datasets.Value("string"),
|
| 109 |
-
},
|
| 110 |
-
},
|
| 111 |
-
"qas": datasets.features.Sequence(
|
| 112 |
-
{
|
| 113 |
-
"question": datasets.Value("string"),
|
| 114 |
-
"answer": datasets.Value("string"),
|
| 115 |
-
"id": datasets.Value("int32"),
|
| 116 |
-
}
|
| 117 |
-
),
|
| 118 |
-
"objects": datasets.features.Sequence(
|
| 119 |
-
{
|
| 120 |
-
"id": datasets.Value("int32"),
|
| 121 |
-
"bbox": datasets.Sequence(datasets.Value("float32"), length=4),
|
| 122 |
-
"category": datasets.Value("string"),
|
| 123 |
-
"area": datasets.Value("float32"),
|
| 124 |
-
"category_id": datasets.Value("int32"),
|
| 125 |
-
"segment": datasets.features.Sequence(
|
| 126 |
-
datasets.features.Sequence(datasets.Value("float32"))
|
| 127 |
-
),
|
| 128 |
-
}
|
| 129 |
-
),
|
| 130 |
-
}
|
| 131 |
-
),
|
| 132 |
-
# If there's a common (input, target) tuple from the features,
|
| 133 |
-
# specify them here. They'll be used if as_supervised=True in
|
| 134 |
-
# builder.as_dataset.
|
| 135 |
-
supervised_keys=None,
|
| 136 |
-
# Homepage of the dataset for documentation
|
| 137 |
-
homepage="https://compguesswhat.github.io/",
|
| 138 |
-
citation=self._CITATION,
|
| 139 |
-
)
|
| 140 |
-
elif self.config.gameplay_scenario == "zero_shot":
|
| 141 |
-
return datasets.DatasetInfo(
|
| 142 |
-
# This is the description that will appear on the datasets page.
|
| 143 |
-
description=self._DESCRIPTION,
|
| 144 |
-
# datasets.features.FeatureConnectors
|
| 145 |
-
features=datasets.Features(
|
| 146 |
-
{
|
| 147 |
-
"id": datasets.Value("int32"),
|
| 148 |
-
"target_id": datasets.Value("string"),
|
| 149 |
-
"status": datasets.Value("string"),
|
| 150 |
-
"image": {
|
| 151 |
-
"id": datasets.Value("int32"),
|
| 152 |
-
"file_name": datasets.Value("string"),
|
| 153 |
-
"coco_url": datasets.Value("string"),
|
| 154 |
-
"height": datasets.Value("int32"),
|
| 155 |
-
"width": datasets.Value("int32"),
|
| 156 |
-
"license": datasets.Value("int32"),
|
| 157 |
-
"open_images_id": datasets.Value("string"),
|
| 158 |
-
"date_captured": datasets.Value("string"),
|
| 159 |
-
},
|
| 160 |
-
"objects": datasets.features.Sequence(
|
| 161 |
-
{
|
| 162 |
-
"id": datasets.Value("string"),
|
| 163 |
-
"bbox": datasets.Sequence(datasets.Value("float32"), length=4),
|
| 164 |
-
"category": datasets.Value("string"),
|
| 165 |
-
"area": datasets.Value("float32"),
|
| 166 |
-
"category_id": datasets.Value("int32"),
|
| 167 |
-
"IsOccluded": datasets.Value("int32"),
|
| 168 |
-
"IsTruncated": datasets.Value("int32"),
|
| 169 |
-
"segment": datasets.features.Sequence(
|
| 170 |
-
{
|
| 171 |
-
"MaskPath": datasets.Value("string"),
|
| 172 |
-
"LabelName": datasets.Value("string"),
|
| 173 |
-
"BoxID": datasets.Value("string"),
|
| 174 |
-
"BoxXMin": datasets.Value("string"),
|
| 175 |
-
"BoxXMax": datasets.Value("string"),
|
| 176 |
-
"BoxYMin": datasets.Value("string"),
|
| 177 |
-
"BoxYMax": datasets.Value("string"),
|
| 178 |
-
"PredictedIoU": datasets.Value("string"),
|
| 179 |
-
"Clicks": datasets.Value("string"),
|
| 180 |
-
}
|
| 181 |
-
),
|
| 182 |
-
}
|
| 183 |
-
),
|
| 184 |
-
}
|
| 185 |
-
),
|
| 186 |
-
# If there's a common (input, target) tuple from the features,
|
| 187 |
-
# specify them here. They'll be used if as_supervised=True in
|
| 188 |
-
# builder.as_dataset.
|
| 189 |
-
supervised_keys=None,
|
| 190 |
-
# Homepage of the dataset for documentation
|
| 191 |
-
homepage="https://compguesswhat.github.io/",
|
| 192 |
-
citation=self._CITATION,
|
| 193 |
-
)
|
| 194 |
-
|
| 195 |
-
def _split_generators(self, dl_manager):
|
| 196 |
-
"""Returns SplitGenerators."""
|
| 197 |
-
dl_dir = dl_manager.download_and_extract(self.config.data_url)
|
| 198 |
-
|
| 199 |
-
splits_gen = []
|
| 200 |
-
|
| 201 |
-
for split_id, split_filename in self.config.splits.items():
|
| 202 |
-
if self.config.gameplay_scenario == "original":
|
| 203 |
-
if "train" in split_id:
|
| 204 |
-
split_name = datasets.Split.TRAIN
|
| 205 |
-
elif "valid" in split_id:
|
| 206 |
-
split_name = datasets.Split.VALIDATION
|
| 207 |
-
elif "test" in split_id:
|
| 208 |
-
split_name = datasets.Split.TEST
|
| 209 |
-
else:
|
| 210 |
-
split_name = datasets.Split(split_id)
|
| 211 |
-
|
| 212 |
-
full_split_name = "-".join(["compguesswhat", self.config.gameplay_scenario])
|
| 213 |
-
splits_gen.append(
|
| 214 |
-
datasets.SplitGenerator(
|
| 215 |
-
name=split_name,
|
| 216 |
-
gen_kwargs={
|
| 217 |
-
"filepath": os.path.join(
|
| 218 |
-
dl_dir,
|
| 219 |
-
full_split_name,
|
| 220 |
-
self.VERSION.version_str,
|
| 221 |
-
split_filename,
|
| 222 |
-
)
|
| 223 |
-
},
|
| 224 |
-
)
|
| 225 |
-
)
|
| 226 |
-
|
| 227 |
-
return splits_gen
|
| 228 |
-
|
| 229 |
-
def _generate_examples(self, filepath):
|
| 230 |
-
def _extract_game_tuple(data):
|
| 231 |
-
data = data.decode("utf-8")
|
| 232 |
-
game = json.loads(data.strip("\n"))
|
| 233 |
-
|
| 234 |
-
# we refactor the data structure a bit to fit with the new version
|
| 235 |
-
game["target_id"] = game["object_id"]
|
| 236 |
-
if "object_id" in game:
|
| 237 |
-
del game["object_id"]
|
| 238 |
-
|
| 239 |
-
if "questioner_id" in game:
|
| 240 |
-
del game["questioner_id"]
|
| 241 |
-
###
|
| 242 |
-
|
| 243 |
-
if "visual_genome" in game["image"]:
|
| 244 |
-
# We need to cast it to string so that we avoid issues with int size
|
| 245 |
-
game["image"]["visual_genome"]["image_id"] = str(game["image"]["visual_genome"]["image_id"])
|
| 246 |
-
game["image"]["visual_genome"]["flickr_id"] = str(game["image"]["visual_genome"]["flickr_id"])
|
| 247 |
-
|
| 248 |
-
return game["id"], game
|
| 249 |
-
|
| 250 |
-
"""Yields examples."""
|
| 251 |
-
with open(filepath, "rb") as gzip_file:
|
| 252 |
-
with gzip.open(gzip_file) as in_file:
|
| 253 |
-
for data in in_file:
|
| 254 |
-
yield _extract_game_tuple(data)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|