Датасет Open Images V7#
Open Images V7 — масштабный датасет для обнаружения объектов, созданный Google. В конфигурации Ultralytics он содержит 1 743 042 обучающих и 41 620 валидационных изображений с 601 классом объектов. Исходная версия датасета охватывает около 9 миллионов изображений с разметкой на уровне изображений, рамками объектов, масками сегментации, визуальными связями и локализованными описаниями. Это один из самых обширных ресурсов для разметки в области компьютерного зрения.
Смотри: Обнаружение объектов с предварительно обученной моделью Open Images V7
Предварительно обученные модели Open Images V7#
Ultralytics публикует пять моделей YOLOv8, предварительно обученных на Open Images V7, чтобы ты мог обнаруживать объекты 601 класса, не скачивая датасет:
| Модель | размер (пиксели) | mAPval50-95 | Скорость CPU ONNX (мс) | Скорость A100 TensorRT (мс) | параметры (M) | FLOPs (B) |
|---|---|---|---|---|---|---|
| YOLOv8n | 640 | 18.4 | 142.4 | 1.21 | 3.5 | 10.4 |
| YOLOv8s | 640 | 27.7 | 183.1 | 1.40 | 11.4 | 29.7 |
| YOLOv8m | 640 | 33.6 | 408.5 | 2.26 | 26.2 | 80.6 |
| YOLOv8l | 640 | 34.9 | 596.9 | 2.43 | 44.1 | 167.4 |
| YOLOv8x | 640 | 36.3 | 860.6 | 3.56 | 68.7 | 260.6 |
На визуализации ниже показаны классы объектов, которые могут обнаруживать эти модели:
Ты можешь запустить предсказание или начать дообучение с этих контрольных точек следующим образом.
from ultralytics import YOLO
# Загрузи предварительно обученную модель YOLOv8n для Open Images V7
model = YOLO("yolov8n-oiv7.pt")
# Выполни предсказание
results = model.predict(source="image.jpg")
# Начни обучение с предварительно обученной контрольной точки
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)Основные особенности#
- Open Images V7 содержит около 9 миллионов изображений с разметкой для нескольких задач компьютерного зрения; в среднем на изображении отмечено 8,3 объекта.
- Около 90% из 16 миллионов ограничивающих рамок на 1,9 миллиона изображений были размечены вручную профессиональными разметчиками, что обеспечивает единообразную и качественную разметку.
- Разметка не ограничивается рамками: 3,3 миллиона аннотаций визуальных связей для 1 466 триплетов, маски сегментации для 2,8 миллиона объектов 350 классов (добавлены в V5), 675 тысяч локализованных описаний (V6) и 66,4 миллиона точечных меток на 1,4 миллиона изображений 5 827 классов (V7).
- 61,4 миллиона меток на уровне изображений для 20 638 классов делают датасет подходящим для классификации изображений и мультимодальных исследований наряду с обнаружением объектов и сегментацией экземпляров.
Структура набора данных#
Конфигурация Ultralytics open-images-v7.yaml загружает подмножество Open Images V7 для обнаружения объектов:
| Подвыборка | Изображения | Описание |
|---|---|---|
| Обучение | 1,743,042 | Изображения с разметкой рамками для обучения модели |
| Валидация | 41,620 | Отложенные изображения для оценки и сравнительного тестирования |
В конфигурации определён 601 класс объектов с индексами от 0 до 600 — от Accordion до Zucchini. В документации Google это число округлено до 600, но и в официальном списке классов с разметкой рамками, и в YAML-файле Ultralytics содержится 601 запись. Значение ключа test: в конфигурации оставлено пустым.
Применение#
Open Images V7 подходит для обучения и оценки моделей в различных задачах компьютерного зрения:
- Обнаружение объектов с большим словарём классов: благодаря 601 классу детекторы, обученные на Open Images V7, распознают гораздо больше категорий, чем модели, обученные на COCO.
- Обнаружение визуальных связей: 3,3 миллиона аннотаций связей помогают обучать модели понимать взаимодействие между объектами.
- Сегментация экземпляров: маски для 2,8 миллиона объектов позволяют анализировать сцены на уровне пикселей.
- Мультимодальное обучение: локализованные описания, объединяющие голос, текст и следы перемещения мыши, связывают визуальные данные с подробными описаниями.
- Оценка в режиме zero-shot: широкий охват классов помогает оценить, как модели zero-shot справляются с объектами, которые не встречались при обучении.
Чтобы размечать собственные изображения, обучать модели и управлять крупными датасетами в браузере, выполни весь рабочий процесс на платформе Ultralytics.
Файл YAML набора данных#
Файл open-images-v7.yaml задаёт конфигурацию датасета: пути к данным, названия классов и другие метаданные. Файл поддерживается в репозитории Ultralytics по адресу https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/open-images-v7.yaml.
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
# Open Images v7 dataset https://storage.googleapis.com/openimages/web/index.html by Google
# Documentation: https://docs.ultralytics.com/datasets/detect/open-images-v7
# Example usage: yolo train data=open-images-v7.yaml
# parent
# ├── ultralytics
# └── datasets
# └── open-images-v7 ← downloads here (561 GB)
# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: open-images-v7 # dataset root dir
train: images/train # train images (relative to 'path') 1743042 images
val: images/val # val images (relative to 'path') 41620 images
test: # test images (optional)
# Classes
names:
0: Accordion
1: Adhesive tape
2: Aircraft
3: Airplane
4: Alarm clock
5: Alpaca
6: Ambulance
7: Animal
8: Ant
9: Antelope
10: Apple
11: Armadillo
12: Artichoke
13: Auto part
14: Axe
15: Backpack
16: Bagel
17: Baked goods
18: Balance beam
19: Ball
20: Balloon
21: Banana
22: Band-aid
23: Banjo
24: Barge
25: Barrel
26: Baseball bat
27: Baseball glove
28: Bat (Animal)
29: Bathroom accessory
30: Bathroom cabinet
31: Bathtub
32: Beaker
33: Bear
34: Bed
35: Bee
36: Beehive
37: Beer
38: Beetle
39: Bell pepper
40: Belt
41: Bench
42: Bicycle
43: Bicycle helmet
44: Bicycle wheel
45: Bidet
46: Billboard
47: Billiard table
48: Binoculars
49: Bird
50: Blender
51: Blue jay
52: Boat
53: Bomb
54: Book
55: Bookcase
56: Boot
57: Bottle
58: Bottle opener
59: Bow and arrow
60: Bowl
61: Bowling equipment
62: Box
63: Boy
64: Brassiere
65: Bread
66: Briefcase
67: Broccoli
68: Bronze sculpture
69: Brown bear
70: Building
71: Bull
72: Burrito
73: Bus
74: Bust
75: Butterfly
76: Cabbage
77: Cabinetry
78: Cake
79: Cake stand
80: Calculator
81: Camel
82: Camera
83: Can opener
84: Canary
85: Candle
86: Candy
87: Cannon
88: Canoe
89: Cantaloupe
90: Car
91: Carnivore
92: Carrot
93: Cart
94: Cassette deck
95: Castle
96: Cat
97: Cat furniture
98: Caterpillar
99: Cattle
100: Ceiling fan
101: Cello
102: Centipede
103: Chainsaw
104: Chair
105: Cheese
106: Cheetah
107: Chest of drawers
108: Chicken
109: Chime
110: Chisel
111: Chopsticks
112: Christmas tree
113: Clock
114: Closet
115: Clothing
116: Coat
117: Cocktail
118: Cocktail shaker
119: Coconut
120: Coffee
121: Coffee cup
122: Coffee table
123: Coffeemaker
124: Coin
125: Common fig
126: Common sunflower
127: Computer keyboard
128: Computer monitor
129: Computer mouse
130: Container
131: Convenience store
132: Cookie
133: Cooking spray
134: Corded phone
135: Cosmetics
136: Couch
137: Countertop
138: Cowboy hat
139: Crab
140: Cream
141: Cricket ball
142: Crocodile
143: Croissant
144: Crown
145: Crutch
146: Cucumber
147: Cupboard
148: Curtain
149: Cutting board
150: Dagger
151: Dairy Product
152: Deer
153: Desk
154: Dessert
155: Diaper
156: Dice
157: Digital clock
158: Dinosaur
159: Dishwasher
160: Dog
161: Dog bed
162: Doll
163: Dolphin
164: Door
165: Door handle
166: Donut
167: Dragonfly
168: Drawer
169: Dress
170: Drill (Tool)
171: Drink
172: Drinking straw
173: Drum
174: Duck
175: Dumbbell
176: Eagle
177: Earrings
178: Egg (Food)
179: Elephant
180: Envelope
181: Eraser
182: Face powder
183: Facial tissue holder
184: Falcon
185: Fashion accessory
186: Fast food
187: Fax
188: Fedora
189: Filing cabinet
190: Fire hydrant
191: Fireplace
192: Fish
193: Flag
194: Flashlight
195: Flower
196: Flowerpot
197: Flute
198: Flying disc
199: Food
200: Food processor
201: Football
202: Football helmet
203: Footwear
204: Fork
205: Fountain
206: Fox
207: French fries
208: French horn
209: Frog
210: Fruit
211: Frying pan
212: Furniture
213: Garden Asparagus
214: Gas stove
215: Giraffe
216: Girl
217: Glasses
218: Glove
219: Goat
220: Goggles
221: Goldfish
222: Golf ball
223: Golf cart
224: Gondola
225: Goose
226: Grape
227: Grapefruit
228: Grinder
229: Guacamole
230: Guitar
231: Hair dryer
232: Hair spray
233: Hamburger
234: Hammer
235: Hamster
236: Hand dryer
237: Handbag
238: Handgun
239: Harbor seal
240: Harmonica
241: Harp
242: Harpsichord
243: Hat
244: Headphones
245: Heater
246: Hedgehog
247: Helicopter
248: Helmet
249: High heels
250: Hiking equipment
251: Hippopotamus
252: Home appliance
253: Honeycomb
254: Horizontal bar
255: Horse
256: Hot dog
257: House
258: Houseplant
259: Human arm
260: Human beard
261: Human body
262: Human ear
263: Human eye
264: Human face
265: Human foot
266: Human hair
267: Human hand
268: Human head
269: Human leg
270: Human mouth
271: Human nose
272: Humidifier
273: Ice cream
274: Indoor rower
275: Infant bed
276: Insect
277: Invertebrate
278: Ipod
279: Isopod
280: Jacket
281: Jacuzzi
282: Jaguar (Animal)
283: Jeans
284: Jellyfish
285: Jet ski
286: Jug
287: Juice
288: Kangaroo
289: Kettle
290: Kitchen & dining room table
291: Kitchen appliance
292: Kitchen knife
293: Kitchen utensil
294: Kitchenware
295: Kite
296: Knife
297: Koala
298: Ladder
299: Ladle
300: Ladybug
301: Lamp
302: Land vehicle
303: Lantern
304: Laptop
305: Lavender (Plant)
306: Lemon
307: Leopard
308: Light bulb
309: Light switch
310: Lighthouse
311: Lily
312: Limousine
313: Lion
314: Lipstick
315: Lizard
316: Lobster
317: Loveseat
318: Luggage and bags
319: Lynx
320: Magpie
321: Mammal
322: Man
323: Mango
324: Maple
325: Maracas
326: Marine invertebrates
327: Marine mammal
328: Measuring cup
329: Mechanical fan
330: Medical equipment
331: Microphone
332: Microwave oven
333: Milk
334: Miniskirt
335: Mirror
336: Missile
337: Mixer
338: Mixing bowl
339: Mobile phone
340: Monkey
341: Moths and butterflies
342: Motorcycle
343: Mouse
344: Muffin
345: Mug
346: Mule
347: Mushroom
348: Musical instrument
349: Musical keyboard
350: Nail (Construction)
351: Necklace
352: Nightstand
353: Oboe
354: Office building
355: Office supplies
356: Orange
357: Organ (Musical Instrument)
358: Ostrich
359: Otter
360: Oven
361: Owl
362: Oyster
363: Paddle
364: Palm tree
365: Pancake
366: Panda
367: Paper cutter
368: Paper towel
369: Parachute
370: Parking meter
371: Parrot
372: Pasta
373: Pastry
374: Peach
375: Pear
376: Pen
377: Pencil case
378: Pencil sharpener
379: Penguin
380: Perfume
381: Person
382: Personal care
383: Personal flotation device
384: Piano
385: Picnic basket
386: Picture frame
387: Pig
388: Pillow
389: Pineapple
390: Pitcher (Container)
391: Pizza
392: Pizza cutter
393: Plant
394: Plastic bag
395: Plate
396: Platter
397: Plumbing fixture
398: Polar bear
399: Pomegranate
400: Popcorn
401: Porch
402: Porcupine
403: Poster
404: Potato
405: Power plugs and sockets
406: Pressure cooker
407: Pretzel
408: Printer
409: Pumpkin
410: Punching bag
411: Rabbit
412: Raccoon
413: Racket
414: Radish
415: Ratchet (Device)
416: Raven
417: Rays and skates
418: Red panda
419: Refrigerator
420: Remote control
421: Reptile
422: Rhinoceros
423: Rifle
424: Ring binder
425: Rocket
426: Roller skates
427: Rose
428: Rugby ball
429: Ruler
430: Salad
431: Salt and pepper shakers
432: Sandal
433: Sandwich
434: Saucer
435: Saxophone
436: Scale
437: Scarf
438: Scissors
439: Scoreboard
440: Scorpion
441: Screwdriver
442: Sculpture
443: Sea lion
444: Sea turtle
445: Seafood
446: Seahorse
447: Seat belt
448: Segway
449: Serving tray
450: Sewing machine
451: Shark
452: Sheep
453: Shelf
454: Shellfish
455: Shirt
456: Shorts
457: Shotgun
458: Shower
459: Shrimp
460: Sink
461: Skateboard
462: Ski
463: Skirt
464: Skull
465: Skunk
466: Skyscraper
467: Slow cooker
468: Snack
469: Snail
470: Snake
471: Snowboard
472: Snowman
473: Snowmobile
474: Snowplow
475: Soap dispenser
476: Sock
477: Sofa bed
478: Sombrero
479: Sparrow
480: Spatula
481: Spice rack
482: Spider
483: Spoon
484: Sports equipment
485: Sports uniform
486: Squash (Plant)
487: Squid
488: Squirrel
489: Stairs
490: Stapler
491: Starfish
492: Stationary bicycle
493: Stethoscope
494: Stool
495: Stop sign
496: Strawberry
497: Street light
498: Stretcher
499: Studio couch
500: Submarine
501: Submarine sandwich
502: Suit
503: Suitcase
504: Sun hat
505: Sunglasses
506: Surfboard
507: Sushi
508: Swan
509: Swim cap
510: Swimming pool
511: Swimwear
512: Sword
513: Syringe
514: Table
515: Table tennis racket
516: Tablet computer
517: Tableware
518: Taco
519: Tank
520: Tap
521: Tart
522: Taxi
523: Tea
524: Teapot
525: Teddy bear
526: Telephone
527: Television
528: Tennis ball
529: Tennis racket
530: Tent
531: Tiara
532: Tick
533: Tie
534: Tiger
535: Tin can
536: Tire
537: Toaster
538: Toilet
539: Toilet paper
540: Tomato
541: Tool
542: Toothbrush
543: Torch
544: Tortoise
545: Towel
546: Tower
547: Toy
548: Traffic light
549: Traffic sign
550: Train
551: Training bench
552: Treadmill
553: Tree
554: Tree house
555: Tripod
556: Trombone
557: Trousers
558: Truck
559: Trumpet
560: Turkey
561: Turtle
562: Umbrella
563: Unicycle
564: Van
565: Vase
566: Vegetable
567: Vehicle
568: Vehicle registration plate
569: Violin
570: Volleyball (Ball)
571: Waffle
572: Waffle iron
573: Wall clock
574: Wardrobe
575: Washing machine
576: Waste container
577: Watch
578: Watercraft
579: Watermelon
580: Weapon
581: Whale
582: Wheel
583: Wheelchair
584: Whisk
585: Whiteboard
586: Willow
587: Window
588: Window blind
589: Wine
590: Wine glass
591: Wine rack
592: Winter melon
593: Wok
594: Woman
595: Wood-burning stove
596: Woodpecker
597: Worm
598: Wrench
599: Zebra
600: Zucchini
# Download script/URL (optional) ---------------------------------------------------------------------------------------
download: |
import warnings
from ultralytics.utils import LOGGER, SETTINGS, Path
from ultralytics.utils.checks import check_requirements
check_requirements("fiftyone")
import fiftyone as fo
import fiftyone.zoo as foz
name = "open-images-v7"
fo.config.dataset_zoo_dir = Path(SETTINGS["datasets_dir"]) / "fiftyone" / name
fraction = 1.0 # fraction of full dataset to use
LOGGER.warning("Open Images V7 dataset requires at least **561 GB of free space. Starting download...")
for split in "train", "validation": # 1743042 train, 41620 val images
train = split == "train"
# Load Open Images dataset
dataset = foz.load_zoo_dataset(
name,
split=split,
label_types=["detections"],
max_samples=round((1743042 if train else 41620) * fraction),
)
# Define classes
if train:
classes = dataset.default_classes # all classes
# classes = dataset.distinct('ground_truth.detections.label') # only observed classes
# Export to YOLO format
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning, module="fiftyone.utils.yolo")
dataset.export(
export_dir=str(Path(SETTINGS["datasets_dir"]) / name),
dataset_type=fo.types.YOLOv5Dataset,
label_field="ground_truth",
split="val" if split == "validation" else split,
classes=classes,
overwrite=train,
)Использование#
При первом использовании Open Images V7 загружается автоматически. Для 1 743 042 обучающих и 41 620 валидационных изображений требуется около 561 ГБ свободного места на диске. Скрипт загрузки устанавливает пакет fiftyone, чтобы скачать изображения, и преобразует разметку в формат YOLO; в зависимости от подключения и оборудования это может занять много времени.
Чтобы обучить модель YOLO26n на датасете Open Images V7 в течение 100 эпох с размером изображения 640, используй следующие фрагменты кода.
from ultralytics import YOLO
# Загрузить модель YOLO26n, предобученную на COCO
model = YOLO("yolo26n.pt")
# Обучи модель на наборе данных Open Images V7
results = model.train(data="open-images-v7.yaml", epochs=100, imgsz=640)Примеры изображений и аннотаций#
На примере ниже показаны аннотации ограничивающих рамок, связей и масок, которые Open Images V7 накладывает на одно изображение:
Цитирование и благодарности#
Если ты используешь набор данных Open Images V7 в исследовательской или практической работе, процитируй следующую статью:
@article{OpenImages,
author = {Alina Kuznetsova and Hassan Rom and Neil Alldrin and Jasper Uijlings and Ivan Krasin and Jordi Pont-Tuset and Shahab Kamali and Stefan Popov and Matteo Malloci and Alexander Kolesnikov and Tom Duerig and Vittorio Ferrari},
title = {The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale},
year = {2020},
journal = {IJCV}
}Мы благодарим команду Google AI за создание и поддержку набора данных Open Images V7. Подробнее о наборе данных см. на официальном сайте Open Images V7.
Часто задаваемые вопросы#
Набор данных Open Images V7 используется для обучения и оценки моделей обнаружения объектов в машинном обучении и компьютерном зрении. Конфигурация Ultralytics включает 1,743,042 обучающих и 41,620 валидационных изображений с 601 классом объектов, а исходный выпуск также содержит маски сегментации, визуальные связи, локализованные описания и метки на уровне точек для более широкого спектра исследований.
Конфигурация Ultralytics
open-images-v7.yamlвключает 601 класс объектов, 1,743,042 обучающих и 41,620 валидационных изображений, без тестовой выборки. Полная исходная версия содержит около 9 миллионов изображений; в документации Google округлено до 600 число классов, для которых можно задать ограничивающие рамки, но официальный список классов содержит 601 позицию.Для Open Images V7 требуется около 561 ГБ свободного места на диске; загрузка начинается автоматически при первой тренировке с
data="open-images-v7.yaml". Скрипт загрузки устанавливает пакетfiftyoneи преобразует аннотации в формат YOLO. Если тебе нужен набор поменьше, посмотри обзор наборов данных для обнаружения объектов.Для обучения модели YOLO26 на наборе данных Open Images V7 можно использовать команды Python и CLI. Ниже приведён пример обучения модели YOLO26n в течение 100 эпох с размером изображения 640:
Пример обученияfrom ultralytics import YOLO # Загрузить модель YOLO26n, предобученную на COCO model = YOLO("yolo26n.pt") # Обучи модель на наборе данных Open Images V7 results = model.train(data="open-images-v7.yaml", epochs=100, imgsz=640)Подробнее об аргументах и настройках см. на странице Обучение.
Да. Пять предварительно обученных моделей Open Images V7 (
yolov8n-oiv7.pt–yolov8x-oiv7.pt, с показателями от 18,4 до 36,3 mAP) сразу обнаруживают все 601 класс, поэтому ты можешь запускать предсказание на собственных изображениях без загрузки 561 ГБ данных.Open Images V7 крупнее и разнообразнее: 601 класс объектов и 1,743,042 обучающих изображения против 80 классов и 118,287 обучающих изображений в COCO. COCO остаётся стандартным бенчмарком для сравнения детекторов, а Open Images V7 лучше подходит для обнаружения объектов с большим словарём классов и предварительного обучения моделей, которым нужен широкий охват категорий.
Комментарии38
Glenn JocherUltralytics Hey Ultralytics Users!
Exciting news! 🎉 We've added the Open Images V7 dataset to our collection. It's perfect for enhancing your YOLO models across various applications.
What's in it for you?
🌍 Diverse range of categories. 🎯 Accurate annotations for better training. 💡 Ideal for AI/ML advancements in computer vision. Check out our new Open Images V7 page in the docs for all the details.
Questions or need help? Just ask us! We're here to support your YOLO journey.
Happy modeling! 🤖👀
6DevanshL HI,can we finetune this OpenImages v7 dataset only for two classes car,person on yolov2 model.
1Glenn JocherUltralytics Yes, but not in the current Ultralytics package with
YOLOv2: if you only needcarandperson, you can filter Open Images V7 to those 2 classes and train a supported model like Ultralytics YOLO26 on that subset; forYOLOv2you’d need a separate legacy implementation. The Open Images V7 docs show the dataset layout and download setup, and feel free to keep follow-up questions here so the Ultralytics team can see them.
nandiniigarg I am trying to train the YOLOv8m model using python for selected classes. Here's my YAML file:
# Ultralytics YOLO 🚀, AGPL-3.0 license # Open Images v7 dataset https://storage.googleapis.com/openimages/web/index.html by Google # Documentation: https://docs.ultralytics.com/datasets/detect/open-images-v7/ # Example usage: yolo train data=open-images-v7.yaml # parent # ├── ultralytics # └── datasets # └── open-images-v7 ← downloads here (561 GB) # Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] path: fiftyone/open-images-v7 # dataset root dir train: train # train images (relative to 'path') 1743042 images val: validation # val images (relative to 'path') 41620 images test: # test images (optional) # Classes names: 0: Bathroom accessory 1: Bathroom cabinet 2: Bathtub 3: Bed 4: Bench ..... 159: Wine glass 160: Wine rack 161: Wood-burning stove # Download script/URL (optional) --------------------------------------------------------------------------------------- download: | from ultralytics.utils import LOGGER, SETTINGS, Path, is_ubuntu, get_ubuntu_version from ultralytics.utils.checks import check_requirements, check_version check_requirements('fiftyone') if is_ubuntu() and check_version(get_ubuntu_version(), '>=22.04'): # Ubuntu>=22.04 patch https://github.com/voxel51/fiftyone/issues/2961#issuecomment-1666519347 check_requirements('fiftyone-db-ubuntu2204') import fiftyone as fo import fiftyone.zoo as foz import warnings classes_list = ["Bathroom accessory","Bathroom cabinet","Bathtub","Bed", "Bench", "Bidet", "Blender", "Book", "Bookcase", "Bowl", "Box", "Bronze sculpture> name = 'open-images-v7' fraction = 1.0 # fraction of full dataset to use LOGGER.warning('WARNING ⚠️ Open Images V7 dataset requires at least **561 GB of free space. Starting download...') for split in 'train', 'validation': # 1743042 train, 41620 val images train = split == 'train' # Load Open Images dataset dataset = foz.load_zoo_dataset(name, split=split, label_types=['detections'], classes = classes_list, dataset_dir=Path(SETTINGS['datasets_dir']) / 'fiftyone' / name, max_samples=round((1743042 if train else 41620) * fraction)) # Define classes if train: classes = dataset.default_classes # all classes # classes = dataset.distinct('ground_truth.detections.label') # only observed classes # Export to YOLO format with warnings.catch_warnings(): warnings.filterwarnings("ignore", category=UserWarning, module="fiftyone.utils.yolo") dataset.export(export_dir=str(Path(SETTINGS['datasets_dir']) / name), dataset_type=fo.types.YOLOv5Dataset, label_field='ground_truth', split='val' if split == 'validation' else split, classes=classes, overwrite=train)I am running into this error after the first epoch:
Starting training for 100 epochs... Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size 1/100 7.34G 0 83.18 0 0 640: 7%|▋ | 1794/27205 [13:28<1:51:55, 3.78it/s]^[[B^[[B^[[B^[[B^[[B^[[B 1/100 7.34G 0 5.485 0 0 640: 100%|██████████| 27205/27205 [1:52:47<00:00, 4.02it/s] Class Images Instances Box(P R mAP50 mAP50-95): 100%|██████████| 376/376 [02:20<00:00, 2.67it/s] Traceback (most recent call last): File "/home/nirwana/nandini/open_images_test/ultralyticsTrial/train.py", line 7, in <module> results = model.train(data="open-images-v7.yaml", epochs=100, imgsz=640, device=0, patience=5, save=True, verbose=True, plots=True) File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/engine/model.py", line 813, in train self.trainer.train() File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/engine/trainer.py", line 206, in train self._do_train(world_size) File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/engine/trainer.py", line 430, in _do_train self.metrics, self.fitness = self.validate() File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/engine/trainer.py", line 575, in validate metrics = self.validator(self) File "/home/nirwana/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/engine/validator.py", line 193, in __call__ stats = self.get_stats() File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/models/yolo/detect/val.py", line 181, in get_stats stats = {k: torch.cat(v, 0).cpu().numpy() for k, v in self.stats.items()} # to numpy File "/home/nirwana/.local/lib/python3.10/site-packages/ultralytics/models/yolo/detect/val.py", line 181, in <dictcomp> stats = {k: torch.cat(v, 0).cpu().numpy() for k, v in self.stats.items()} # to numpy RuntimeError: torch.cat(): expected a non-empty list of TensorsI cannot figure out how to fix this error.
1Glenn JocherUltralytics @nandiniigarg it looks like the error might be due to an empty list of tensors during validation. Please ensure that your dataset and annotations are correctly formatted and that the selected classes have sufficient instances. Also, verify if the issue persists with the latest version of the Ultralytics package. For more details, you can refer to the Open Images V7 documentation.
safwene-bahroun(изменено) HELLO @glenn-jocher, i have the same error . Can you help me to fix it ?
Glenn JocherUltralytics @safwene-bahroun please provide a https://docs.ultralytics.com/help/minimum-reproducible-example/ to allow us to help you better.
safwene-bahroun path: ../datasets/open-images-v7 train: C:\Users\GIGABYTE\datasets\open-images-v7\images\open-images-v7\train val: C:\Users\GIGABYTE\datasets\open-images-v7\images\open-images-v7\validation nc: 36 names: 0: Person 1: Man 2: Woman 3: Boy 4: Belt 5: Clothing 6: Coat 7: Cowboy hat 8: Dress 9: Fedora 10: Footwear 11: Glasses 12: Glove 13: Goggles 14: Hat 15: High heels 16: Jeans 17: Miniskirt 18: Scarf 19: Shirt 20: Shorts 21: Skirt 22: Sock 23: Sombrero 24: Suit 25: Sun hat 26: Sunglasses 27: Swim cap 28: Swimwear 29: Tie 30: Trousers 31: Cosmetics 32: Face powder 33: Hair spray 34: Lipstick 35: Perfume download: | from ultralytics.utils import LOGGER, SETTINGS, Path, is_ubuntu, get_ubuntu_version from ultralytics.utils.checks import check_requirements, check_version import fiftyone as fo import fiftyone.zoo as foz from fiftyone import ViewField as F # Import F for filtering import warnings
Define the list of classes
classes = [ "Person", "Man", "Woman", "Boy", "Belt", "Clothing", "Coat", "Cowboy hat", "Dress", "Fedora", "Footwear", "Glasses", "Glove", "Goggles", "Hat", "High heels", "Jeans", "Miniskirt", "Scarf", "Shirt", "Shorts", "Skirt", "Sock", "Sombrero", "Suit", "Sun hat", "Sunglasses", "Swim cap", "Swimwear", "Tie", "Trousers", "Cosmetics", "Face powder", "Hair spray", "Lipstick", "Perfume" ]
Dataset name and directory setup
name = "open-images-v7" fo.config.dataset_zoo_dir = Path(SETTINGS["datasets_dir"]) / name / "images"
Log the start of the download process
LOGGER.warning('Open Images V7 dataset: Starting download...')
Loop through train and validation splits
for split in ['train', 'validation']: # Load Open Images dataset dataset = foz.load_zoo_dataset( name, split=split, classes=classes, label_types=['detections'], max_samples=5000 )
# Define export directory export_dir = Path(SETTINGS['datasets_dir']) / name / "images" / ('train' if split == 'train' else 'validation') # Filter dataset to include only the specified classes filtered_dataset = dataset.match(F("ground_truth.detections.label").is_in(classes)) # Export to YOLO format with warnings.catch_warnings(): warnings.filterwarnings("ignore", category=UserWarning, module="fiftyone.utils.yolo") filtered_dataset.export( export_dir=str(export_dir), dataset_type=fo.types.YOLOv5Dataset, label_field='ground_truth', split=split, classes=classes, overwrite=True )Glenn JocherUltralytics The
RuntimeError: torch.cat(): expected a non-empty list of Tensorstypically occurs when no validation predictions are generated. For Open Images V7 training with custom classes, please verify:- Class Consistency: Ensure your 36 selected classes exist in both train/val splits. Some fine-grained categories like "Man"/"Woman" might have limited val instances.
- Path Formatting: Use forward slashes in paths:
train: datasets/open-images-v7/images/train - Annotations: Confirm your filtered dataset contains labels by inspecting
labels/valfolder
For reference, see the Open Images V7 dataset setup guide in our documentation. Let us know if these checks resolve the validation issue.
1
danielbob32 Hi Glenn,
I noticed that while FiftyOne has support for YOLOv5Dataset, it doesn’t currently support instance segmentation, which I needed for training a custom YOLO model on a new class (lemons - don’t ask 😄).
After digging around, I couldn’t find a built-in API to export data in the format required by the segmentation models. So, I ended up creating a custom pipeline to handle the conversion. After a bit of fine-tuning, it worked quite well.
I’m wondering if this is something you’d be interested in adding here, and if so, what’s the best way to share it?
Also, I’d appreciate it if you could help me verify that my polygon annotations are formatted correctly.
Thanks! Daniel
ABINAYA200264 Hello I'm Abinaya I want selective classes in Open Images v7 dataset I use class Id to extract spacific classes but it not work it detect all the the defult classes I want only selected class, I don't have training spcilities please can some one help me
laila-e Install necessary libraries
!pip install fiftyone
import fiftyone as fo import fiftyone.zoo as foz
Define the classes you want to include
classes = ["Broccoli", "Carrot"]
Load the Open Images V7 dataset
dataset = foz.load_zoo_dataset( "open-images-v7", split="train", label_types=["detections"], classes=classes, max_samples=1000, # Adjust this number based on your needs dataset_dir="/content/open-images-v7" )
Export the dataset to YOLO format
dataset.export( export_dir="/content/open-images-v7-yolo", dataset_type=fo.types.YOLOv5Dataset, label_field="detections", classes=classes, split="train" ) I try this code and it works well but I need labels for detection and for visual relationship and I do not know how do that
Glenn JocherUltralytics @laila-e to include labels for both detection and visual relationships in your dataset, you'll need to adjust the
label_typesparameter in theload_zoo_datasetfunction to include both "detections" and "relationships". For more detailed guidance, please refer to the Open Images V7 documentation.
MiguelTucto Hello! It's my first time training with yolov8. I need help about some processes. I've downloaded 3126 images from openimagesv7. These images are between fruits and vegetables. I trained with 50 epochs, but it stopped in 19 because It didn't present positive changes. After that, I've used the best.pt model exported to detect. When I use that model, It don't recognize how I expected the fruits and vegetable that I've trained. I need add more epochs? or maybe the quantity of fruits and vegetable have to be similar? Thank you in advance!
Glenn JocherUltralytics @MiguelTucto it sounds like your model might benefit from more training data or balanced class representation. Consider increasing the number of images for each class and possibly extending the training epochs. You can also experiment with data augmentation to enhance model performance. For more detailed guidance, you might want to explore the Open Images V7 documentation.
1MiguelTucto Thanks dude! but I was thinking in two possibilities. The first one is to apply data augmentation and all images are between 2000 images. The second one is reduce all of images to 500 images. Which one will work better?
Glenn JocherUltralytics Applying data augmentation to your 2000 images would likely yield better results than reducing the dataset to 500 images. Augmentation can enhance model robustness by introducing variability, which is beneficial for training.
davidhuangal For the pretrained Oi7 models, what hyperparameters were used? Was it the
default.yamlparameters? Or were things changed, like the number of epochs, etc.? I would like to train some YOLOv9 and YOLOv10 models on the dataset and wanted to make sure I was following the same training design as the YOLOv8 pretrained Oi7 models.Glenn JocherUltralytics(изменено) @davidhuangal the pretrained YOLOv8 models for the Open Images V7 dataset were trained using the
default.yamlparameters, exceptepochswas lowered to 30 since OIv7 is much larger than COCO. For your YOLOv9 and YOLOv10 models, you can start with these default settings and adjust as needed based on your specific requirements. If you encounter any issues or need further customization, please refer to the Ultralytics documentation for detailed guidance.1Glenn JocherUltralytics @davidhuangal ah one small change, epochs was lowered to 30 since OIv7 is so much larger than COCO. All else should be identical.
1
kaymanas I have downloaded and used the Ultralytics Open Images V7 Pretrained Models and they work great. However when I export to CoreML to use with XCode, the results are very different and much worse. There are many many more predictions and they're way off, the image attached predicted Snake, Treadmill, dumbell, Tent and many more at 99-100% confidence.
See the images below for example of the output with the model as is and then converted to mlpackage format.
Is there something wrong with the conversion process?
My python code here:
from ultralytics import YOLO
model = YOLO("yolov8x-oiv7.pt") # load model
model.export(format="coreml",nms=True) # convert model
Paula DerrengerUltralytics @kaymanas it sounds like the CoreML conversion might be causing issues with the model's performance. First, ensure you are using the latest version of the Ultralytics package. If the issue persists, try adjusting the
nmsparameter during export to see if it improves the results. Additionally, verify that the input image size and preprocessing steps in your XCode project match those used during training. If the problem continues, please provide more details or share minimal reproducible code snippets for further assistance.kaymanas Thanks @pderrenger. I've updated to the latest Ultralytics package (8.2.61).
I've tested exporting the model with nms=False. When reloaded into python and tested against a test image from the open image dataset it works perfectly. I cannot see the prediction when I load this into XCode unfortunately as there is no preview on the model page.
When I export with nms=True I cannot reload this into the python script as I get this error: "Ultralytics only supports inference of non-pipelined CoreML models exported with 'nms=False', but 'model=640' has an NMS pipeline created by an 'nms=True' export."
So I can only test this in XCode and it has a very poor prediction. Two image results here: https://imgur.com/a/am7oF6Z
Paula DerrengerUltralytics Thanks for updating to the latest version and testing the export options. The error message you're encountering indicates that the current implementation only supports non-pipelined CoreML models with
nms=False. This limitation might be causing the poor predictions in XCode. To address this, ensure that the preprocessing steps in XCode match those used during training. If the issue persists, consider reporting it on our GitHub issues page for further investigation.kaymanas How do I ensure that the preprocessing steps in XCode match those used during training? Sorry I'm new to this, I tried googling but couldn't find any info...
Paula DerrengerUltralytics @kaymanas to ensure preprocessing steps in XCode match those used during training, you need to replicate the same image resizing, normalization, and other preprocessing steps. Typically, this involves resizing images to the same dimensions used during training (e.g., 640x640), normalizing pixel values to the same range (e.g., 0-1), and applying any other transformations like mean subtraction or standard deviation scaling. You can find the exact preprocessing steps in the training script or model documentation. If you need further assistance, please refer to the Ultralytics documentation.
kaymanas Thanks for the help! where do I find the training script or model documentation? This page (the one you linked) doesn't have that information....
Paula DerrengerUltralytics You can find the training script and model documentation in the Ultralytics YOLO repository on GitHub. Specifically, the Ultralytics documentation provides comprehensive details on training, validation, and exporting models. For the Open Images V7 dataset, refer to the dataset's YAML configuration and the training section in the documentation. If you need further assistance, please check the repository's README or the relevant sections in the documentation.
naga08krishna I have converted the dataset into yolo format but all the images are in the training set only. Please tell me how to divide the dataset into training and validate datasets
On Mon, Jun 24, 2024, 22:56 Naga Krishna Reddy @.***> wrote:
Hi sorry to disturb you but please tell help me in this doubt After downloading the images, when I try to convert the dataset into yolo format I am getting below error repeatedly. Please help me in this. [image: image.png] [image: image.png]
And also tell me how to divide the dataset into train and validate datasets.
On Sat, Jun 22, 2024 at 11:35 PM Glenn Jocher @.***> wrote:
Hi @naga08krishna https://github.com/naga08krishna,
To train a YOLO model on only vegetable images from the Open Images V7 dataset, you can create a custom YAML file that includes only the classes you're interested in. Here's how you can do it in Google Colab:
- Create a Custom YAML File: Modify the open-images-v7.yaml to include only the vegetable classes you need. For example, if you want to include "Broccoli" and "Carrot", your YAML file would look like this:
Custom Open Images V7 dataset for vegetable classespath: ../datasets/open-images-v7train: images/trainval: images/val
Classesnames:
0: Broccoli 1: Carrot
- Download and Prepare the Dataset: Use the fiftyone library to download and filter the dataset for your specific classes. Here’s a Colab script to help you get started:
Install necessary libraries
!pip install fiftyone import fiftyone as foimport fiftyone.zoo as foz
Define the classes you want to includeclasses = ["Broccoli", "Carrot"]
Load the Open Images V7 datasetdataset = foz.load_zoo_dataset(
"open-images-v7", split="train", label_types=["detections"], classes=classes, max_samples=1000, # Adjust this number based on your needs dataset_dir="/content/open-images-v7")
Export the dataset to YOLO formatdataset.export(
export_dir="/content/open-images-v7-yolo", dataset_type=fo.types.YOLOv5Dataset, label_field="detections", classes=classes, split="train")
- Train the Model: Use the custom YAML file to train your model.
Python Example
from ultralytics import YOLO
Load a COCO-pretrained YOLOv8n modelmodel = YOLO("yolov8n.pt")
Train the model on the custom Open Images V7 datasetresults = model.train(data="/content/open-images-v7-yolo/custom-open-images-v7.yaml", epochs=100, imgsz=640)
CLI Example
Train a COCO-pretrained YOLOv8n model on the custom Open Images V7 dataset
yolo detect train data=/content/open-images-v7-yolo/custom-open-images-v7.yaml model=yolov8n.pt epochs=100 imgsz=640
This approach ensures that your training process focuses only on the specified vegetable classes. For more details, you can refer to the Open Images V7 documentation https://docs.ultralytics.com/datasets/detect/open-images-v7/.
Feel free to reach out if you have any more questions! 😊
— Reply to this email directly, view it on GitHub https://github.com/orgs/ultralytics/discussions/7912#discussioncomment-9848232, or unsubscribe https://github.com/notifications/unsubscribe-auth/BAV6QX2UTYJPQL6VRYD25T3ZIW4GJAVCNFSM6AAAAABCR3JU3KVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TQNBYGIZTE . You are receiving this because you were mentioned.Message ID: @.*** com>
Paula DerrengerUltralytics(изменено) Hi @naga08krishna,
Thank you for reaching out! To split your dataset into training and validation sets, you can use the
fiftyonelibrary, which provides a convenient way to handle and manipulate datasets. Here's a step-by-step guide to help you achieve this:-
Install FiftyOne: Ensure you have the
fiftyonelibrary installed. You can install it using pip:pip install fiftyone -
Load and Split the Dataset: Use the following script to load your dataset and split it into training and validation sets. This example assumes you have already converted the dataset into YOLO format and all images are currently in the training set.
import fiftyone as fo import fiftyone.zoo as foz # Load your dataset dataset_dir = "/path/to/your/yolo/dataset" dataset = fo.Dataset.from_dir( dataset_type=fo.types.YOLOv5Dataset, data_path=dataset_dir, labels_path=dataset_dir, include_id=True ) # Split the dataset into training and validation sets train_view = dataset.take(int(0.8 * len(dataset)), seed=51) val_view = dataset.exclude([s.id for s in train_view]) # Export the training set train_view.export( export_dir="/path/to/save/train", dataset_type=fo.types.YOLOv5Dataset, label_field="ground_truth", classes=dataset.default_classes, split="train" ) # Export the validation set val_view.export( export_dir="/path/to/save/val", dataset_type=fo.types.YOLOv5Dataset, label_field="ground_truth", classes=dataset.default_classes, split="val" ) -
Update Your YAML File: Ensure your
open-images-v7.yamlfile points to the correct directories for training and validation images:path: ../datasets/open-images-v7 train: /path/to/save/train/images val: /path/to/save/val/images -
Train Your Model: Now you can train your model using the updated YAML file:
from ultralytics import YOLO # Load a COCO-pretrained YOLO26n model model = YOLO("yolo26n.pt") # Train the model on the Open Images V7 dataset results = model.train(data="open-images-v7.yaml", epochs=100, imgsz=640)Or via CLI:
yolo detect train data=open-images-v7.yaml model=yolo26n.pt epochs=100 imgsz=640
This should help you split your dataset and proceed with training. If you encounter any issues or need further assistance, feel free to ask! 😊
For more detailed information, you can refer to the Open Images V7 documentation.
-
