Added classes for multi-label classification
Browse files- coco_dataset_multi_label_script.py +139 -15
coco_dataset_multi_label_script.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import json
|
2 |
import os
|
3 |
import datasets
|
|
|
4 |
|
5 |
|
6 |
class COCOBuilderConfig(datasets.BuilderConfig):
|
@@ -81,6 +82,8 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
81 |
"file_name": datasets.Value("string"),
|
82 |
"coco_url": datasets.Value("string"),
|
83 |
"image_path": datasets.Value("string"),
|
|
|
|
|
84 |
}
|
85 |
|
86 |
features = datasets.Features(feature_dict)
|
@@ -120,7 +123,6 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
120 |
"annotations_trainval": os.path.join(data_dir, "annotations_trainval2017.zip"),
|
121 |
"image_info_test": os.path.join(data_dir, "image_info_test2017.zip"),
|
122 |
}
|
123 |
-
archive_path = dl_manager.download_and_extract(_DL_URLS)
|
124 |
|
125 |
splits = []
|
126 |
for split in self.config.splits:
|
@@ -129,8 +131,9 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
129 |
name=datasets.Split.TRAIN,
|
130 |
# These kwargs will be passed to _generate_examples
|
131 |
gen_kwargs={
|
132 |
-
"
|
133 |
-
"
|
|
|
134 |
"split": "train",
|
135 |
}
|
136 |
)
|
@@ -139,8 +142,9 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
139 |
name=datasets.Split.VALIDATION,
|
140 |
# These kwargs will be passed to _generate_examples
|
141 |
gen_kwargs={
|
142 |
-
"
|
143 |
-
"
|
|
|
144 |
"split": "valid",
|
145 |
},
|
146 |
)
|
@@ -149,8 +153,9 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
149 |
name=datasets.Split.TEST,
|
150 |
# These kwargs will be passed to _generate_examples
|
151 |
gen_kwargs={
|
152 |
-
"
|
153 |
-
"
|
|
|
154 |
"split": "test",
|
155 |
},
|
156 |
)
|
@@ -160,42 +165,150 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
160 |
splits.append(dataset)
|
161 |
|
162 |
return splits
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
163 |
|
164 |
def _generate_examples(
|
165 |
# method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
|
166 |
-
self,
|
167 |
):
|
168 |
-
""" Yields examples as (key, example) tuples. """
|
169 |
# This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
|
170 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
171 |
|
172 |
-
_features = ["image_id", "caption_id", "caption", "height", "width", "file_name", "coco_url", "image_path", "id"]
|
173 |
features = list(_features)
|
174 |
|
175 |
if split in "valid":
|
176 |
split = "val"
|
177 |
|
178 |
-
with open(
|
179 |
-
|
|
|
|
|
|
|
180 |
|
181 |
# list of dict
|
182 |
-
images =
|
|
|
183 |
entries = images
|
184 |
|
|
|
|
|
|
|
185 |
# build a dict of image_id -> image info dict
|
186 |
d = {image["id"]: image for image in images}
|
187 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
# list of dict
|
189 |
if split in ["train", "val"]:
|
190 |
-
annotations =
|
191 |
|
192 |
# build a dict of image_id ->
|
193 |
for annotation in annotations:
|
194 |
_id = annotation["id"]
|
195 |
-
|
|
|
196 |
annotation.update(image_info)
|
197 |
annotation["id"] = _id
|
198 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
entries = annotations
|
200 |
|
201 |
for id_, entry in enumerate(entries):
|
@@ -213,3 +326,14 @@ class COCODataset(datasets.GeneratorBasedBuilder):
|
|
213 |
entry = {k: entry[k] for k in _features if k in entry}
|
214 |
|
215 |
yield str((entry["image_id"], entry["caption_id"])), entry
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import json
|
2 |
import os
|
3 |
import datasets
|
4 |
+
import torch
|
5 |
|
6 |
|
7 |
class COCOBuilderConfig(datasets.BuilderConfig):
|
|
|
82 |
"file_name": datasets.Value("string"),
|
83 |
"coco_url": datasets.Value("string"),
|
84 |
"image_path": datasets.Value("string"),
|
85 |
+
"category_ids": datasets.Sequence(datasets.Value("int64")),
|
86 |
+
"category_one_hot": datasets.Sequence(datasets.Value("int64")),
|
87 |
}
|
88 |
|
89 |
features = datasets.Features(feature_dict)
|
|
|
123 |
"annotations_trainval": os.path.join(data_dir, "annotations_trainval2017.zip"),
|
124 |
"image_info_test": os.path.join(data_dir, "image_info_test2017.zip"),
|
125 |
}
|
|
|
126 |
|
127 |
splits = []
|
128 |
for split in self.config.splits:
|
|
|
131 |
name=datasets.Split.TRAIN,
|
132 |
# These kwargs will be passed to _generate_examples
|
133 |
gen_kwargs={
|
134 |
+
"captions_json_path": os.path.join(data_dir, "annotations", "captions_train2017.json"),
|
135 |
+
"instances_json_path": os.path.join(data_dir, "annotations", "instances_train2017.json"),
|
136 |
+
"image_dir": os.path.join(data_dir, "train2017"),
|
137 |
"split": "train",
|
138 |
}
|
139 |
)
|
|
|
142 |
name=datasets.Split.VALIDATION,
|
143 |
# These kwargs will be passed to _generate_examples
|
144 |
gen_kwargs={
|
145 |
+
"captions_json_path": os.path.join(data_dir, "annotations", "captions_val2017.json"),
|
146 |
+
"instances_json_path": os.path.join(data_dir, "annotations", "instances_val2017.json"),
|
147 |
+
"image_dir": os.path.join(data_dir, "val2017"),
|
148 |
"split": "valid",
|
149 |
},
|
150 |
)
|
|
|
153 |
name=datasets.Split.TEST,
|
154 |
# These kwargs will be passed to _generate_examples
|
155 |
gen_kwargs={
|
156 |
+
"captions_json_path": os.path.join(data_dir, "annotations", "image_info_test2017.json"),
|
157 |
+
"instances_json_path": os.path.join(data_dir, "annotations", "image_info_test2017.json"), # "instances_test2017.json
|
158 |
+
"image_dir": os.path.join(data_dir, "test2017"),
|
159 |
"split": "test",
|
160 |
},
|
161 |
)
|
|
|
165 |
splits.append(dataset)
|
166 |
|
167 |
return splits
|
168 |
+
|
169 |
+
# instances.json
|
170 |
+
# {
|
171 |
+
# "info": {
|
172 |
+
# "year": "2020",
|
173 |
+
# "version": "1",
|
174 |
+
# "description": "Exported from roboflow.ai",
|
175 |
+
# "contributor": "Roboflow",
|
176 |
+
# "url": "https://app.roboflow.ai/datasets/hard-hat-sample/1",
|
177 |
+
# "date_created": "2000-01-01T00:00:00+00:00"
|
178 |
+
# },
|
179 |
+
# "licenses": [
|
180 |
+
# {
|
181 |
+
# "id": 1,
|
182 |
+
# "url": "https://creativecommons.org/publicdomain/zero/1.0/",
|
183 |
+
# "name": "Public Domain"
|
184 |
+
# }
|
185 |
+
# ],
|
186 |
+
# "categories": [
|
187 |
+
# {
|
188 |
+
# "id": 0,
|
189 |
+
# "name": "Workers",
|
190 |
+
# "supercategory": "none"
|
191 |
+
# },
|
192 |
+
# {
|
193 |
+
# "id": 1,
|
194 |
+
# "name": "head",
|
195 |
+
# "supercategory": "Workers"
|
196 |
+
# },
|
197 |
+
# {
|
198 |
+
# "id": 2,
|
199 |
+
# "name": "helmet",
|
200 |
+
# "supercategory": "Workers"
|
201 |
+
# },
|
202 |
+
# {
|
203 |
+
# "id": 3,
|
204 |
+
# "name": "person",
|
205 |
+
# "supercategory": "Workers"
|
206 |
+
# }
|
207 |
+
# ],
|
208 |
+
# "images": [
|
209 |
+
# {
|
210 |
+
# "id": 0,
|
211 |
+
# "license": 1,
|
212 |
+
# "file_name": "0001.jpg",
|
213 |
+
# "height": 275,
|
214 |
+
# "width": 490,
|
215 |
+
# "date_captured": "2020-07-20T19:39:26+00:00"
|
216 |
+
# }
|
217 |
+
# ],
|
218 |
+
# "annotations": [
|
219 |
+
# {
|
220 |
+
# "id": 0,
|
221 |
+
# "image_id": 0,
|
222 |
+
# "category_id": 2,
|
223 |
+
# "bbox": [
|
224 |
+
# 45,
|
225 |
+
# 2,
|
226 |
+
# 85,
|
227 |
+
# 85
|
228 |
+
# ],
|
229 |
+
# "area": 7225,
|
230 |
+
# "segmentation": [],
|
231 |
+
# "iscrowd": 0
|
232 |
+
# },
|
233 |
+
# {
|
234 |
+
# "id": 1,
|
235 |
+
# "image_id": 0,
|
236 |
+
# "category_id": 2,
|
237 |
+
# "bbox": [
|
238 |
+
# 324,
|
239 |
+
# 29,
|
240 |
+
# 72,
|
241 |
+
# 81
|
242 |
+
# ],
|
243 |
+
# "area": 5832,
|
244 |
+
# "segmentation": [],
|
245 |
+
# "iscrowd": 0
|
246 |
+
# }
|
247 |
+
# ]
|
248 |
+
# }
|
249 |
|
250 |
def _generate_examples(
|
251 |
# method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
|
252 |
+
self, captions_json_path, instances_json_path, image_dir, split
|
253 |
):
|
254 |
+
""" Yields examples as (key, example, categories) tuples. """
|
255 |
# This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
|
256 |
# The `key` is here for legacy reason (tfds) and is not important in itself.
|
257 |
|
258 |
+
_features = ["image_id", "caption_id", "caption", "height", "width", "file_name", "coco_url", "image_path", "id", "category_ids", "category_one_hot"]
|
259 |
features = list(_features)
|
260 |
|
261 |
if split in "valid":
|
262 |
split = "val"
|
263 |
|
264 |
+
with open(captions_json_path, 'r', encoding='UTF-8') as fp:
|
265 |
+
captions_data = json.load(fp)
|
266 |
+
|
267 |
+
with open(instances_json_path, 'r', encoding='UTF-8') as fp:
|
268 |
+
instances_data = json.load(fp)
|
269 |
|
270 |
# list of dict
|
271 |
+
images = captions_data["images"]
|
272 |
+
instances_annotations = instances_data["annotations"]
|
273 |
entries = images
|
274 |
|
275 |
+
self.classes = list(map(lambda x: {'id': x['id'], 'name': x['name']}, instances_data['categories']))
|
276 |
+
self.num_classes = len(self.classes)
|
277 |
+
|
278 |
# build a dict of image_id -> image info dict
|
279 |
d = {image["id"]: image for image in images}
|
280 |
|
281 |
+
# build a dict of image_id -> list of category_ids
|
282 |
+
cat_ids_dict = {}
|
283 |
+
|
284 |
+
for annotation in instances_annotations:
|
285 |
+
image_id = annotation["image_id"]
|
286 |
+
category_id = annotation["category_id"]
|
287 |
+
if image_id not in cat_ids_dict:
|
288 |
+
cat_ids_dict[image_id] = set([])
|
289 |
+
cat_ids_dict[image_id].add(category_id)
|
290 |
+
|
291 |
# list of dict
|
292 |
if split in ["train", "val"]:
|
293 |
+
annotations = captions_data["annotations"]
|
294 |
|
295 |
# build a dict of image_id ->
|
296 |
for annotation in annotations:
|
297 |
_id = annotation["id"]
|
298 |
+
image_id = annotation["image_id"]
|
299 |
+
image_info = d[image_id]
|
300 |
annotation.update(image_info)
|
301 |
annotation["id"] = _id
|
302 |
|
303 |
+
# Add the category_ids to the annotation
|
304 |
+
annotation["category_ids"] = cat_ids_dict[annotation["image_id"]] if annotation["image_id"] in cat_ids_dict else []
|
305 |
+
|
306 |
+
annotation['category_one_hot'] = torch.zeros(len(self.classes))
|
307 |
+
for category_id in annotation["category_ids"]:
|
308 |
+
# Get index of category_id in self.classes
|
309 |
+
index = next((index for (index, d) in enumerate(self.classes) if d["id"] == category_id), None)
|
310 |
+
annotation['category_one_hot'][index] = 1
|
311 |
+
|
312 |
entries = annotations
|
313 |
|
314 |
for id_, entry in enumerate(entries):
|
|
|
326 |
entry = {k: entry[k] for k in _features if k in entry}
|
327 |
|
328 |
yield str((entry["image_id"], entry["caption_id"])), entry
|
329 |
+
|
330 |
+
from datasets import load_dataset
|
331 |
+
if __name__ == "__main__":
|
332 |
+
dataset = load_dataset(
|
333 |
+
"coco_dataset_multi_label_script/coco_dataset_multi_label_script.py",
|
334 |
+
"2017",
|
335 |
+
keep_in_memory=False,
|
336 |
+
splits=["valid"],
|
337 |
+
data_dir="/workspace/pixt/clip-training/data/mscoco",
|
338 |
+
)
|
339 |
+
print(dataset["validation"][0])
|