refactor dataset loading script

#1
by davanstrien HF staff - opened
Files changed (1) hide show
  1. DocLayNet.py +89 -85
DocLayNet.py CHANGED
@@ -6,10 +6,10 @@ https://huggingface.co/datasets/ydshieh/coco_dataset_script/blob/main/coco_datas
6
  import json
7
  import os
8
  import datasets
 
9
 
10
 
11
  class COCOBuilderConfig(datasets.BuilderConfig):
12
-
13
  def __init__(self, name, splits, **kwargs):
14
  super().__init__(name, **kwargs)
15
  self.splits = splits
@@ -43,12 +43,10 @@ _LICENSE = "CDLA-Permissive-1.0"
43
  # The HuggingFace dataset library don't host the datasets but only point to the original files
44
  # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
45
 
46
- # This script is supposed to work with local (downloaded) COCO dataset.
47
  _URLs = {
48
  "core": "https://codait-cos-dax.s3.us.cloud-object-storage.appdomain.cloud/dax-doclaynet/1.0.0/DocLayNet_core.zip",
49
  }
50
 
51
-
52
  # Name of the dataset usually match the script name with CamelCase instead of snake_case
53
  class COCODataset(datasets.GeneratorBasedBuilder):
54
  """An example dataset script to work with the local (downloaded) COCO dataset"""
@@ -57,28 +55,51 @@ class COCODataset(datasets.GeneratorBasedBuilder):
57
 
58
  BUILDER_CONFIG_CLASS = COCOBuilderConfig
59
  BUILDER_CONFIGS = [
60
- COCOBuilderConfig(name='2022.08', splits=['train', 'val', 'test']),
61
  ]
62
  DEFAULT_CONFIG_NAME = "2022.08"
63
 
64
  def _info(self):
65
- # This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
66
-
67
- feature_dict = {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
  "id": datasets.Value("int64"),
69
- "height": datasets.Value("int64"),
70
- "width": datasets.Value("int64"),
71
- "file_name": datasets.Value("string"),
72
-
73
- # Custom fields
74
- "doc_category": datasets.Value("string"), # high-level document category
75
- "collection": datasets.Value("string"), # sub-collection name
76
- "doc_name": datasets.Value("string"), # original document filename
77
- "page_no": datasets.Value("int64"), # page number in original document
78
- # "precedence": datasets.Value("int64"), # annotation order, non-zero in case of redundant double- or triple-annotation
79
  }
80
-
81
- features = datasets.Features(feature_dict)
82
 
83
  return datasets.DatasetInfo(
84
  # This is the description that will appear on the datasets page.
@@ -99,53 +120,41 @@ class COCODataset(datasets.GeneratorBasedBuilder):
99
 
100
  def _split_generators(self, dl_manager):
101
  """Returns SplitGenerators."""
102
- # This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
103
- # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
104
-
105
- # data_dir = self.config.data_dir
106
- # if not data_dir:
107
- # raise ValueError(
108
- # "This script is supposed to work with local (downloaded) COCO dataset. The argument `data_dir` in `load_dataset()` is required."
109
- # )
110
-
111
- # _DL_URLS = {
112
- # "train": os.path.join(data_dir, "train2017.zip"),
113
- # "val": os.path.join(data_dir, "val2017.zip"),
114
- # "test": os.path.join(data_dir, "test2017.zip"),
115
- # "annotations_trainval": os.path.join(data_dir, "annotations_trainval2017.zip"),
116
- # "image_info_test": os.path.join(data_dir, "image_info_test2017.zip"),
117
- # }
118
  archive_path = dl_manager.download_and_extract(_URLs)
119
- print("archive_path: ", archive_path)
120
-
121
  splits = []
122
  for split in self.config.splits:
123
- if split == 'train':
124
  dataset = datasets.SplitGenerator(
125
  name=datasets.Split.TRAIN,
126
  # These kwargs will be passed to _generate_examples
127
  gen_kwargs={
128
- "json_path": os.path.join(archive_path["core"], "COCO", "train.json"),
 
 
129
  "image_dir": os.path.join(archive_path["core"], "PNG"),
130
  "split": "train",
131
- }
132
  )
133
- elif split in ['val', 'valid', 'validation', 'dev']:
134
  dataset = datasets.SplitGenerator(
135
  name=datasets.Split.VALIDATION,
136
  # These kwargs will be passed to _generate_examples
137
  gen_kwargs={
138
- "json_path": os.path.join(archive_path["core"], "COCO", "val.json"),
 
 
139
  "image_dir": os.path.join(archive_path["core"], "PNG"),
140
  "split": "val",
141
  },
142
  )
143
- elif split == 'test':
144
  dataset = datasets.SplitGenerator(
145
  name=datasets.Split.TEST,
146
  # These kwargs will be passed to _generate_examples
147
  gen_kwargs={
148
- "json_path": os.path.join(archive_path["core"], "COCO", "test.json"),
 
 
149
  "image_dir": os.path.join(archive_path["core"], "PNG"),
150
  "split": "test",
151
  },
@@ -154,53 +163,48 @@ class COCODataset(datasets.GeneratorBasedBuilder):
154
  continue
155
 
156
  splits.append(dataset)
157
-
158
  return splits
159
 
160
  def _generate_examples(
161
  # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
162
- self, json_path, image_dir, split
 
 
 
163
  ):
164
- """ Yields examples as (key, example) tuples. """
165
  # This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
166
  # The `key` is here for legacy reason (tfds) and is not important in itself.
167
-
168
- _features = ["image_id", "image_path", "doc_category", "collection", "height", "width", "file_name", "doc_name", "page_no", "id"]
169
- features = list(_features)
170
-
171
- with open(json_path, 'r', encoding='UTF-8') as fp:
172
- data = json.load(fp)
173
-
174
- # list of dict
175
- images = data["images"]
176
- entries = images
177
-
178
- # build a dict of image_id -> image info dict
179
- d = {image["id"]: image for image in images}
180
-
181
- # list of dict
182
- if split in ["train", "val"]:
183
- annotations = data["annotations"]
184
-
185
- # build a dict of image_id ->
186
  for annotation in annotations:
187
- _id = annotation["id"]
188
- image_info = d[annotation["image_id"]]
189
- annotation.update(image_info)
190
- annotation["id"] = _id
191
-
192
- entries = annotations
193
-
194
- for id_, entry in enumerate(entries):
195
 
196
- entry = {k: v for k, v in entry.items() if k in features}
197
-
198
- if split == "test":
199
- entry["image_id"] = entry["id"]
200
- entry["id"] = -1
201
-
202
- entry["image_path"] = os.path.join(image_dir, entry["file_name"])
203
-
204
- entry = {k: entry[k] for k in _features if k in entry}
205
-
206
- yield str((entry["image_id"], entry["id"])), entry
 
 
6
  import json
7
  import os
8
  import datasets
9
+ import collections
10
 
11
 
12
  class COCOBuilderConfig(datasets.BuilderConfig):
 
13
  def __init__(self, name, splits, **kwargs):
14
  super().__init__(name, **kwargs)
15
  self.splits = splits
 
43
  # The HuggingFace dataset library don't host the datasets but only point to the original files
44
  # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
45
 
 
46
  _URLs = {
47
  "core": "https://codait-cos-dax.s3.us.cloud-object-storage.appdomain.cloud/dax-doclaynet/1.0.0/DocLayNet_core.zip",
48
  }
49
 
 
50
  # Name of the dataset usually match the script name with CamelCase instead of snake_case
51
  class COCODataset(datasets.GeneratorBasedBuilder):
52
  """An example dataset script to work with the local (downloaded) COCO dataset"""
 
55
 
56
  BUILDER_CONFIG_CLASS = COCOBuilderConfig
57
  BUILDER_CONFIGS = [
58
+ COCOBuilderConfig(name="2022.08", splits=["train", "val", "test"]),
59
  ]
60
  DEFAULT_CONFIG_NAME = "2022.08"
61
 
62
  def _info(self):
63
+ features = datasets.Features(
64
+ {
65
+ "image_id": datasets.Value("int64"),
66
+ "image": datasets.Image(),
67
+ "width": datasets.Value("int32"),
68
+ "height": datasets.Value("int32"),
69
+ # Custom fields
70
+ "doc_category": datasets.Value(
71
+ "string"
72
+ ), # high-level document category
73
+ "collection": datasets.Value("string"), # sub-collection name
74
+ "doc_name": datasets.Value("string"), # original document filename
75
+ "page_no": datasets.Value("int64"), # page number in original document
76
+ }
77
+ )
78
+ object_dict = {
79
+ "category_id": datasets.ClassLabel(
80
+ names=[
81
+ "Caption",
82
+ "Footnote",
83
+ "Formula",
84
+ "List-item",
85
+ "Page-footer",
86
+ "Page-header",
87
+ "Picture",
88
+ "Section-header",
89
+ "Table",
90
+ "Text",
91
+ "Title",
92
+ ]
93
+ ),
94
+ "image_id": datasets.Value("string"),
95
  "id": datasets.Value("int64"),
96
+ "area": datasets.Value("int64"),
97
+ "bbox": datasets.Sequence(datasets.Value("float32"), length=4),
98
+ "segmentation": [[datasets.Value("float32")]],
99
+ "iscrowd": datasets.Value("bool"),
100
+ "precedence": datasets.Value("int32"),
 
 
 
 
 
101
  }
102
+ features["objects"] = [object_dict]
 
103
 
104
  return datasets.DatasetInfo(
105
  # This is the description that will appear on the datasets page.
 
120
 
121
  def _split_generators(self, dl_manager):
122
  """Returns SplitGenerators."""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
123
  archive_path = dl_manager.download_and_extract(_URLs)
 
 
124
  splits = []
125
  for split in self.config.splits:
126
+ if split == "train":
127
  dataset = datasets.SplitGenerator(
128
  name=datasets.Split.TRAIN,
129
  # These kwargs will be passed to _generate_examples
130
  gen_kwargs={
131
+ "json_path": os.path.join(
132
+ archive_path["core"], "COCO", "train.json"
133
+ ),
134
  "image_dir": os.path.join(archive_path["core"], "PNG"),
135
  "split": "train",
136
+ },
137
  )
138
+ elif split in ["val", "valid", "validation", "dev"]:
139
  dataset = datasets.SplitGenerator(
140
  name=datasets.Split.VALIDATION,
141
  # These kwargs will be passed to _generate_examples
142
  gen_kwargs={
143
+ "json_path": os.path.join(
144
+ archive_path["core"], "COCO", "val.json"
145
+ ),
146
  "image_dir": os.path.join(archive_path["core"], "PNG"),
147
  "split": "val",
148
  },
149
  )
150
+ elif split == "test":
151
  dataset = datasets.SplitGenerator(
152
  name=datasets.Split.TEST,
153
  # These kwargs will be passed to _generate_examples
154
  gen_kwargs={
155
+ "json_path": os.path.join(
156
+ archive_path["core"], "COCO", "test.json"
157
+ ),
158
  "image_dir": os.path.join(archive_path["core"], "PNG"),
159
  "split": "test",
160
  },
 
163
  continue
164
 
165
  splits.append(dataset)
 
166
  return splits
167
 
168
  def _generate_examples(
169
  # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
170
+ self,
171
+ json_path,
172
+ image_dir,
173
+ split,
174
  ):
175
+ """Yields examples as (key, example) tuples."""
176
  # This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
177
  # The `key` is here for legacy reason (tfds) and is not important in itself.
178
+ def _image_info_to_example(image_info, image_dir):
179
+ image = image_info["file_name"]
180
+ return {
181
+ "image_id": image_info["id"],
182
+ "image": os.path.join(image_dir, image),
183
+ "width": image_info["width"],
184
+ "height": image_info["height"],
185
+ "doc_category": image_info["doc_category"],
186
+ "collection": image_info["collection"],
187
+ "doc_name": image_info["doc_name"],
188
+ "page_no": image_info["page_no"],
189
+ }
190
+
191
+ with open(json_path, encoding="utf8") as f:
192
+ annotation_data = json.load(f)
193
+ images = annotation_data["images"]
194
+ annotations = annotation_data["annotations"]
195
+ image_id_to_annotations = collections.defaultdict(list)
 
196
  for annotation in annotations:
197
+ image_id_to_annotations[annotation["image_id"]].append(annotation)
 
 
 
 
 
 
 
198
 
199
+ for idx, image_info in enumerate(images):
200
+ example = _image_info_to_example(image_info, image_dir)
201
+ annotations = image_id_to_annotations[image_info["id"]]
202
+ objects = []
203
+ for annotation in annotations:
204
+ category_id = annotation["category_id"] # Zero based counting
205
+ if category_id != -1:
206
+ category_id = category_id - 1
207
+ annotation["category_id"] = category_id
208
+ objects.append(annotation)
209
+ example["objects"] = objects
210
+ yield idx, example