yonigozlan HF staff commited on
Commit
5664e18
1 Parent(s): 1b00d9c

add script

Browse files
Files changed (2) hide show
  1. README.md +246 -0
  2. coco_detection_dataset_script.py +325 -0
README.md CHANGED
@@ -1,3 +1,249 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ task_categories:
4
+ - object-detection
5
+ tags:
6
+ - COCO
7
+ - Detection
8
+ - '2017'
9
+ pretty_name: COCO detection dataset script
10
+ size_categories:
11
+ - 100K<n<1M
12
+ dataset_info:
13
+ config_name: '2017'
14
+ features:
15
+ - name: id
16
+ dtype: int64
17
+ - name: objects
18
+ struct:
19
+ - name: bbox_id
20
+ sequence: int64
21
+ - name: category_id
22
+ sequence:
23
+ class_label:
24
+ names:
25
+ '0': N/A
26
+ '1': person
27
+ '2': bicycle
28
+ '3': car
29
+ '4': motorcycle
30
+ '5': airplane
31
+ '6': bus
32
+ '7': train
33
+ '8': truck
34
+ '9': boat
35
+ '10': traffic light
36
+ '11': fire hydrant
37
+ '12': street sign
38
+ '13': stop sign
39
+ '14': parking meter
40
+ '15': bench
41
+ '16': bird
42
+ '17': cat
43
+ '18': dog
44
+ '19': horse
45
+ '20': sheep
46
+ '21': cow
47
+ '22': elephant
48
+ '23': bear
49
+ '24': zebra
50
+ '25': giraffe
51
+ '26': hat
52
+ '27': backpack
53
+ '28': umbrella
54
+ '29': shoe
55
+ '30': eye glasses
56
+ '31': handbag
57
+ '32': tie
58
+ '33': suitcase
59
+ '34': frisbee
60
+ '35': skis
61
+ '36': snowboard
62
+ '37': sports ball
63
+ '38': kite
64
+ '39': baseball bat
65
+ '40': baseball glove
66
+ '41': skateboard
67
+ '42': surfboard
68
+ '43': tennis racket
69
+ '44': bottle
70
+ '45': plate
71
+ '46': wine glass
72
+ '47': cup
73
+ '48': fork
74
+ '49': knife
75
+ '50': spoon
76
+ '51': bowl
77
+ '52': banana
78
+ '53': apple
79
+ '54': sandwich
80
+ '55': orange
81
+ '56': broccoli
82
+ '57': carrot
83
+ '58': hot dog
84
+ '59': pizza
85
+ '60': donut
86
+ '61': cake
87
+ '62': chair
88
+ '63': couch
89
+ '64': potted plant
90
+ '65': bed
91
+ '66': mirror
92
+ '67': dining table
93
+ '68': window
94
+ '69': desk
95
+ '70': toilet
96
+ '71': door
97
+ '72': tv
98
+ '73': laptop
99
+ '74': mouse
100
+ '75': remote
101
+ '76': keyboard
102
+ '77': cell phone
103
+ '78': microwave
104
+ '79': oven
105
+ '80': toaster
106
+ '81': sink
107
+ '82': refrigerator
108
+ '83': blender
109
+ '84': book
110
+ '85': clock
111
+ '86': vase
112
+ '87': scissors
113
+ '88': teddy bear
114
+ '89': hair drier
115
+ '90': toothbrush
116
+ - name: bbox
117
+ sequence:
118
+ sequence: float64
119
+ length: 4
120
+ - name: iscrowd
121
+ sequence: int64
122
+ - name: area
123
+ sequence: float64
124
+ - name: height
125
+ dtype: int64
126
+ - name: width
127
+ dtype: int64
128
+ - name: file_name
129
+ dtype: string
130
+ - name: coco_url
131
+ dtype: string
132
+ - name: image_path
133
+ dtype: string
134
+ splits:
135
+ - name: train
136
+ num_bytes: 87231216
137
+ num_examples: 117266
138
+ - name: validation
139
+ num_bytes: 3692192
140
+ num_examples: 4952
141
+ download_size: 20405354669
142
+ dataset_size: 90923408
143
  ---
144
+ ## Usage
145
+ For using the COCO dataset (2017), you need to download it manually first:
146
+ ```bash
147
+ wget http://images.cocodataset.org/zips/train2017.zip
148
+ wget http://images.cocodataset.org/zips/val2017.zip
149
+ wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
150
+ ```
151
+
152
+ Then to load the dataset:
153
+ ```python
154
+ COCO_DIR = ...(path to the downloaded dataset directory)...
155
+ ds = datasets.load_dataset(
156
+ "yonigozlan/coco_2017_detection_script",
157
+ "2017",
158
+ data_dir=COCO_DIR,
159
+ trust_remote_code=True,
160
+ )
161
+ ```
162
+
163
+ ## Benchmarking
164
+ Here is an example of how to benchmark a 🤗 Transformers object detection model on the validation data of the COCO dataset:
165
+
166
+ ```python
167
+ import datasets
168
+ import torch
169
+ from PIL import Image
170
+ from torch.utils.data import DataLoader
171
+ from torchmetrics.detection.mean_ap import MeanAveragePrecision
172
+ from tqdm import tqdm
173
+
174
+ from transformers import AutoImageProcessor, AutoModelForObjectDetection
175
+
176
+ # prepare data
177
+ COCO_DIR = ...(path to the downloaded dataset directory)...
178
+ ds = datasets.load_dataset(
179
+ "yonigozlan/coco_2017_detection_script",
180
+ "2017",
181
+ data_dir=COCO_DIR,
182
+ trust_remote_code=True,
183
+ )
184
+ val_data = ds["validation"]
185
+ categories = val_data.features["objects"]["category_id"].feature.names
186
+ id2label = {index: x for index, x in enumerate(categories, start=0)}
187
+ label2id = {v: k for k, v in id2label.items()}
188
+ checkpoint = "facebook/detr-resnet-50"
189
+
190
+ # load model and processor
191
+ model = AutoModelForObjectDetection.from_pretrained(
192
+ checkpoint, torch_dtype=torch.float16
193
+ ).to("cuda")
194
+ id2label_model = model.config.id2label
195
+ processor = AutoImageProcessor.from_pretrained(checkpoint)
196
+
197
+
198
+ def collate_fn(batch):
199
+ data = {}
200
+ images = [Image.open(x["image_path"]).convert("RGB") for x in batch]
201
+ data["images"] = images
202
+ annotations = []
203
+ for x in batch:
204
+ boxes = x["objects"]["bbox"]
205
+ # convert to xyxy format
206
+ boxes = [[box[0], box[1], box[0] + box[2], box[1] + box[3]] for box in boxes]
207
+ labels = x["objects"]["category_id"]
208
+ boxes = torch.tensor(boxes)
209
+ labels = torch.tensor(labels)
210
+ annotations.append({"boxes": boxes, "labels": labels})
211
+ data["original_size"] = [(x["height"], x["width"]) for x in batch]
212
+ data["annotations"] = annotations
213
+ return data
214
+
215
+
216
+ # prepare dataloader
217
+ dataloader = DataLoader(val_data, batch_size=8, collate_fn=collate_fn)
218
+
219
+ # prepare metric
220
+ metric = MeanAveragePrecision(box_format="xyxy", class_metrics=True)
221
+
222
+ # evaluation loop
223
+ for i, batch in tqdm(enumerate(dataloader), total=len(dataloader)):
224
+ inputs = (
225
+ processor(batch["images"], return_tensors="pt").to("cuda").to(torch.float16)
226
+ )
227
+ with torch.no_grad():
228
+ outputs = model(**inputs)
229
+ target_sizes = torch.tensor([x for x in batch["original_size"]]).to("cuda")
230
+ results = processor.post_process_object_detection(
231
+ outputs, threshold=0.0, target_sizes=target_sizes
232
+ )
233
+
234
+ # convert predicted label id to dataset label id
235
+ if len(id2label_model) != len(id2label):
236
+ for result in results:
237
+ result["labels"] = torch.tensor(
238
+ [label2id.get(id2label_model[x.item()], 0) for x in result["labels"]]
239
+ )
240
+ # put results back to cpu
241
+ for result in results:
242
+ for k, v in result.items():
243
+ if isinstance(v, torch.Tensor):
244
+ result[k] = v.to("cpu")
245
+ metric.update(results, batch["annotations"])
246
+
247
+ metrics = metric.compute()
248
+ print(metrics)
249
+ ```
coco_detection_dataset_script.py ADDED
@@ -0,0 +1,325 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ import os
3
+
4
+ import datasets
5
+
6
+
7
+ class COCOBuilderConfig(datasets.BuilderConfig):
8
+ def __init__(self, name, splits, **kwargs):
9
+ super().__init__(name, **kwargs)
10
+ self.splits = splits
11
+
12
+
13
+ # Add BibTeX citation
14
+ # Find for instance the citation on arxiv or on the dataset repo/website
15
+ _CITATION = """\
16
+ @article{DBLP:journals/corr/LinMBHPRDZ14,
17
+ author = {Tsung{-}Yi Lin and
18
+ Michael Maire and
19
+ Serge J. Belongie and
20
+ Lubomir D. Bourdev and
21
+ Ross B. Girshick and
22
+ James Hays and
23
+ Pietro Perona and
24
+ Deva Ramanan and
25
+ Piotr Doll{'{a} }r and
26
+ C. Lawrence Zitnick},
27
+ title = {Microsoft {COCO:} Common Objects in Context},
28
+ journal = {CoRR},
29
+ volume = {abs/1405.0312},
30
+ year = {2014},
31
+ url = {http://arxiv.org/abs/1405.0312},
32
+ archivePrefix = {arXiv},
33
+ eprint = {1405.0312},
34
+ timestamp = {Mon, 13 Aug 2018 16:48:13 +0200},
35
+ biburl = {https://dblp.org/rec/bib/journals/corr/LinMBHPRDZ14},
36
+ bibsource = {dblp computer science bibliography, https://dblp.org}
37
+ }
38
+ """
39
+
40
+ # Add description of the dataset here
41
+ # You can copy an official description
42
+ _DESCRIPTION = """\
43
+ COCO is a large-scale object detection, segmentation, and captioning dataset.
44
+ """
45
+
46
+ # Add a link to an official homepage for the dataset here
47
+ _HOMEPAGE = "http://cocodataset.org/#home"
48
+
49
+ # Add the licence for the dataset here if you can find it
50
+ _LICENSE = ""
51
+
52
+ # Add link to the official dataset URLs here
53
+ # The HuggingFace dataset library don't host the datasets but only point to the original files
54
+ # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
55
+
56
+ # This script is supposed to work with local (downloaded) COCO dataset.
57
+ _URLs = {}
58
+
59
+
60
+ # Name of the dataset usually match the script name with CamelCase instead of snake_case
61
+ class COCODataset(datasets.GeneratorBasedBuilder):
62
+ """An example dataset script to work with the local (downloaded) COCO dataset"""
63
+
64
+ VERSION = datasets.Version("0.0.0")
65
+
66
+ BUILDER_CONFIG_CLASS = COCOBuilderConfig
67
+ BUILDER_CONFIGS = [
68
+ COCOBuilderConfig(name="2017", splits=["train", "val"]),
69
+ ]
70
+ DEFAULT_CONFIG_NAME = "2017"
71
+
72
+ def _info(self):
73
+ # This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
74
+
75
+ feature_dict = {
76
+ "id": datasets.Value("int64"),
77
+ "objects": {
78
+ "bbox_id": datasets.Sequence(datasets.Value("int64")),
79
+ "category_id": datasets.Sequence(
80
+ datasets.ClassLabel(
81
+ names=[
82
+ "N/A",
83
+ "person",
84
+ "bicycle",
85
+ "car",
86
+ "motorcycle",
87
+ "airplane",
88
+ "bus",
89
+ "train",
90
+ "truck",
91
+ "boat",
92
+ "traffic light",
93
+ "fire hydrant",
94
+ "street sign",
95
+ "stop sign",
96
+ "parking meter",
97
+ "bench",
98
+ "bird",
99
+ "cat",
100
+ "dog",
101
+ "horse",
102
+ "sheep",
103
+ "cow",
104
+ "elephant",
105
+ "bear",
106
+ "zebra",
107
+ "giraffe",
108
+ "hat",
109
+ "backpack",
110
+ "umbrella",
111
+ "shoe",
112
+ "eye glasses",
113
+ "handbag",
114
+ "tie",
115
+ "suitcase",
116
+ "frisbee",
117
+ "skis",
118
+ "snowboard",
119
+ "sports ball",
120
+ "kite",
121
+ "baseball bat",
122
+ "baseball glove",
123
+ "skateboard",
124
+ "surfboard",
125
+ "tennis racket",
126
+ "bottle",
127
+ "plate",
128
+ "wine glass",
129
+ "cup",
130
+ "fork",
131
+ "knife",
132
+ "spoon",
133
+ "bowl",
134
+ "banana",
135
+ "apple",
136
+ "sandwich",
137
+ "orange",
138
+ "broccoli",
139
+ "carrot",
140
+ "hot dog",
141
+ "pizza",
142
+ "donut",
143
+ "cake",
144
+ "chair",
145
+ "couch",
146
+ "potted plant",
147
+ "bed",
148
+ "mirror",
149
+ "dining table",
150
+ "window",
151
+ "desk",
152
+ "toilet",
153
+ "door",
154
+ "tv",
155
+ "laptop",
156
+ "mouse",
157
+ "remote",
158
+ "keyboard",
159
+ "cell phone",
160
+ "microwave",
161
+ "oven",
162
+ "toaster",
163
+ "sink",
164
+ "refrigerator",
165
+ "blender",
166
+ "book",
167
+ "clock",
168
+ "vase",
169
+ "scissors",
170
+ "teddy bear",
171
+ "hair drier",
172
+ "toothbrush",
173
+ ]
174
+ )
175
+ ),
176
+ "bbox": datasets.Sequence(
177
+ datasets.Sequence(datasets.Value("float64"), length=4)
178
+ ),
179
+ "iscrowd": datasets.Sequence(datasets.Value("int64")),
180
+ "area": datasets.Sequence(datasets.Value("float64")),
181
+ },
182
+ "height": datasets.Value("int64"),
183
+ "width": datasets.Value("int64"),
184
+ "file_name": datasets.Value("string"),
185
+ "coco_url": datasets.Value("string"),
186
+ "image_path": datasets.Value("string"),
187
+ }
188
+
189
+ features = datasets.Features(feature_dict)
190
+
191
+ return datasets.DatasetInfo(
192
+ # This is the description that will appear on the datasets page.
193
+ description=_DESCRIPTION,
194
+ # This defines the different columns of the dataset and their types
195
+ features=features, # Here we define them above because they are different between the two configurations
196
+ # If there's a common (input, target) tuple from the features,
197
+ # specify them here. They'll be used if as_supervised=True in
198
+ # builder.as_dataset.
199
+ supervised_keys=None,
200
+ # Homepage of the dataset for documentation
201
+ homepage=_HOMEPAGE,
202
+ # License for the dataset if available
203
+ license=_LICENSE,
204
+ # Citation for the dataset
205
+ citation=_CITATION,
206
+ )
207
+
208
+ def _split_generators(self, dl_manager):
209
+ """Returns SplitGenerators."""
210
+ # This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
211
+ # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
212
+
213
+ data_dir = self.config.data_dir
214
+ if not data_dir:
215
+ raise ValueError(
216
+ "This script is supposed to work with local (downloaded) COCO dataset. The argument `data_dir` in `load_dataset()` is required."
217
+ )
218
+
219
+ _DL_URLS = {
220
+ "train": os.path.join(data_dir, "train2017.zip"),
221
+ "val": os.path.join(data_dir, "val2017.zip"),
222
+ "annotations_trainval": os.path.join(
223
+ data_dir, "annotations_trainval2017.zip"
224
+ ),
225
+ }
226
+ archive_path = dl_manager.download_and_extract(_DL_URLS)
227
+
228
+ splits = []
229
+ for split in self.config.splits:
230
+ if split == "train":
231
+ dataset = datasets.SplitGenerator(
232
+ name=datasets.Split.TRAIN,
233
+ # These kwargs will be passed to _generate_examples
234
+ gen_kwargs={
235
+ "json_path": os.path.join(
236
+ archive_path["annotations_trainval"],
237
+ "annotations",
238
+ "instances_train2017.json",
239
+ ),
240
+ "image_dir": os.path.join(archive_path["train"], "train2017"),
241
+ "split": "train",
242
+ },
243
+ )
244
+ elif split in ["val", "valid", "validation", "dev"]:
245
+ dataset = datasets.SplitGenerator(
246
+ name=datasets.Split.VALIDATION,
247
+ # These kwargs will be passed to _generate_examples
248
+ gen_kwargs={
249
+ "json_path": os.path.join(
250
+ archive_path["annotations_trainval"],
251
+ "annotations",
252
+ "instances_val2017.json",
253
+ ),
254
+ "image_dir": os.path.join(archive_path["val"], "val2017"),
255
+ "split": "valid",
256
+ },
257
+ )
258
+ else:
259
+ continue
260
+
261
+ splits.append(dataset)
262
+
263
+ return splits
264
+
265
+ def _generate_examples(
266
+ # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
267
+ self,
268
+ json_path,
269
+ image_dir,
270
+ split,
271
+ ):
272
+ """Yields examples as (key, example) tuples."""
273
+ # This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
274
+ # The `key` is here for legacy reason (tfds) and is not important in itself.
275
+
276
+ features = [
277
+ "id",
278
+ "objects",
279
+ "height",
280
+ "width",
281
+ "file_name",
282
+ "coco_url",
283
+ "image_path",
284
+ ]
285
+ object_features = [
286
+ "bbox_id",
287
+ "category_id",
288
+ "bbox",
289
+ "iscrowd",
290
+ "area",
291
+ ]
292
+
293
+ with open(json_path, "r", encoding="UTF-8") as fp:
294
+ data = json.load(fp)
295
+
296
+ images = data["images"]
297
+ images_entry = {image["id"]: image for image in images}
298
+ for image_id, image_entry in images_entry.items():
299
+ image_entry["image_path"] = os.path.join(
300
+ image_dir, image_entry["file_name"]
301
+ )
302
+ image_entry["objects"] = []
303
+
304
+ objects = data["annotations"]
305
+ for id_, object_entry in enumerate(objects):
306
+ image_id = object_entry["image_id"]
307
+
308
+ entry = {k: v for k, v in object_entry.items() if k in object_features}
309
+ entry["bbox_id"] = object_entry["id"]
310
+ if entry["iscrowd"]:
311
+ continue
312
+ images_entry[image_id]["objects"].append(entry)
313
+
314
+ for id_, entry in images_entry.items():
315
+ entry = {k: v for k, v in entry.items() if k in features}
316
+ # collate objects
317
+ objects = entry.pop("objects")
318
+ if not objects:
319
+ continue
320
+ entry["objects"] = {
321
+ object_feature: [obj[object_feature] for obj in objects]
322
+ for object_feature in object_features
323
+ }
324
+
325
+ yield str(entry["id"]), entry