NEW Classification Datasets: imagenet10, imagenet100, imagenet1000

[glenn-jocher]

All, FYI I've created 3 new classification datasets for use with debugging/tests/benchmarks: imagenet10, imagenet100, imagenet1000.

These are super small versions of imagenet that train/val in seconds with only 1 image per class (with all 1000 classes, only 100 classes, and only 10 classes). They are reduced in size to imgsz=160 and compressed with PIL so that i.e. imagenet10 is only 70kB, imagenet1000 is only 7MB.

CI

I've migrated our YOLOv8 CI, tests and benchmarks to use them, i.e. benchmarks are now on imagenet100 (top5 accuracy is 0.71 for YOLOv8n-cls on imagenet100).
https://github.com/ultralytics/ultralytics/actions/runs/4223103056/jobs/7332424420

Download

You can download them here:
https://github.com/ultralytics/yolov5/releases/download/v1.0/imagenet10.zip
https://github.com/ultralytics/yolov5/releases/download/v1.0/imagenet100.zip
https://github.com/ultralytics/yolov5/releases/download/v1.0/imagenet1000.zip

Usage

pip install ultralytics

yolo train model=yolov8n-cls.pt data=imagenet100

or

from ultralytics import YOLO

model = YOLO('yolov8n-cls.pt')
results = model.train(data='imagenet100', imgsz=160)