YOLOv8 Image Classifier

Using YOLOv8 to classifying an entire image into one of a set of predefined classes.

Model	size (pixels)	acc top1	acc top5	Speed CPU ONNX (ms)	Speed A100 TensorRT (ms)	params (M)	FLOPs (B) at 640
YOLOv8n-cls	224	66.6	87.0	12.9	0.31	2.7	4.3
YOLOv8s-cls	224	72.3	91.1	23.4	0.35	6.4	13.5
YOLOv8m-cls	224	76.4	93.2	85.4	0.62	17.0	42.7
YOLOv8l-cls	224	78.0	94.1	163.0	0.87	37.5	99.7
YOLOv8x-cls	224	78.4	94.3	232.0	1.01	57.4	154.8

Docker Environment

• PyTorch Jupyter

docker run --gpus all -ti --rm \
    -v $(pwd):/opt/app -p 8888:8888 \
    --name pytorch-jupyter \
    pytorch-jupyter:latest

Training Results

• Flower Dataset (12,297 files / 48 classes)

	top1_acc	top5_acc	validation [it/s]	inference [ms]	prediction preprocess [ms]	prediction inference [ms]	prediction postprocess [ms]	weights [MB]	weights (ONNX) [MB]	weights (TensorRT) [MB]
YOLOv8n	0.792	0.954	86.50	0.2	3.600000	2.733333	0.033333	2.935699	5.731391	7.929402
YOLOv8s	0.832	0.965	91.72	0.2	5.333333	4.500000	0.100000	9.890900	19.618174	26.030774
YOLOv8m	0.858	0.969	87.90	0.4	0.466667	5.000000	0.066667	30.324455	60.397784	63.700307
YOLOv8l	0.850	0.973	72.36	0.8	0.500000	7.300000	0.066667	69.331619	138.311538	166.345273
YOLOv8x	0.861	0.971	53.22	1.1	2.166667	14.400000	0.100000	107.382278	214.371352	225.140141

Model Export Options

Format	format Argument	Model	Metadata	Arguments
PyTorch	-	yolov8n-cls.pt	✅	-
TorchScript	torchscript	yolov8n-cls.torchscript	✅	imgsz, optimize
ONNX	onnx	yolov8n-cls.onnx	✅	imgsz, half, dynamic, simplify, opset
OpenVINO	openvino	yolov8n-cls_openvino_model/	✅	imgsz, half
TensorRT	engine	yolov8n-cls.engine	✅	imgsz, half, dynamic, simplify, workspace
CoreML	coreml	yolov8n-cls.mlpackage	✅	imgsz, half, int8, nms
TF SavedModel	saved_model	yolov8n-cls_saved_model/	✅	imgsz, keras
TF GraphDef	pb	yolov8n-cls.pb	❌	imgsz
TF Lite	tflite	yolov8n-cls.tflite	✅	imgsz, half, int8
TF Edge TPU	edgetpu	yolov8n-cls_edgetpu.tflite	✅	imgsz
TF.js	tfjs	yolov8n-cls_web_model/	✅	imgsz
PaddlePaddle	paddle	yolov8n-cls_paddle_model/	✅	imgsz
ncnn	ncnn	yolov8n-cls_ncnn_model/	✅	imgsz, half