Skip to content

Model Benchmarking with Ultralytics YOLO

Ultralytics YOLO ecosystem and integrations

Introduction

Once your model is trained and validated, the next logical step is to evaluate its performance in various real-world scenarios. Benchmark mode in Ultralytics YOLOv8 serves this purpose by providing a robust framework for assessing the speed and accuracy of your model across a range of export formats.



Watch: Ultralytics Modes Tutorial: Benchmark

Why Is Benchmarking Crucial?

  • Informed Decisions: Gain insights into the trade-offs between speed and accuracy.
  • Resource Allocation: Understand how different export formats perform on different hardware.
  • Optimization: Learn which export format offers the best performance for your specific use case.
  • Cost Efficiency: Make more efficient use of hardware resources based on benchmark results.

Key Metrics in Benchmark Mode

  • mAP50-95: For object detection, segmentation, and pose estimation.
  • accuracy_top5: For image classification.
  • Inference Time: Time taken for each image in milliseconds.

Supported Export Formats

  • ONNX: For optimal CPU performance
  • TensorRT: For maximal GPU efficiency
  • OpenVINO: For Intel hardware optimization
  • CoreML, TensorFlow SavedModel, and More: For diverse deployment needs.

Tip

  • Export to ONNX or OpenVINO for up to 3x CPU speedup.
  • Export to TensorRT for up to 5x GPU speedup.

Usage Examples

Run YOLOv8n benchmarks on all supported export formats including ONNX, TensorRT etc. See Arguments section below for a full list of export arguments.

Example

from ultralytics.utils.benchmarks import benchmark

# Benchmark on GPU
benchmark(model='yolov8n.pt', data='coco8.yaml', imgsz=640, half=False, device=0)

yolo benchmark model=yolov8n.pt data='coco8.yaml' imgsz=640 half=False device=0

Arguments

Arguments such as model, data, imgsz, half, device, and verbose provide users with the flexibility to fine-tune the benchmarks to their specific needs and compare the performance of different export formats with ease.

Key Default Value Description
model None Specifies the path to the model file. Accepts both .pt and .yaml formats, e.g., "yolov8n.pt" for pre-trained models or configuration files.
data None Path to a YAML file defining the dataset for benchmarking, typically including paths and settings for validation data. Example: "coco8.yaml".
imgsz 640 The input image size for the model. Can be a single integer for square images or a tuple (width, height) for non-square, e.g., (640, 480).
half False Enables FP16 (half-precision) inference, reducing memory usage and possibly increasing speed on compatible hardware. Use half=True to enable.
int8 False Activates INT8 quantization for further optimized performance on supported devices, especially useful for edge devices. Set int8=True to use.
device None Defines the computation device(s) for benchmarking, such as "cpu", "cuda:0", or a list of devices like "cuda:0,1" for multi-GPU setups.
verbose False Controls the level of detail in logging output. A boolean value; set verbose=True for detailed logs or a float for thresholding errors.

Export Formats

Benchmarks will attempt to run automatically on all possible export formats below.

Format format Argument Model Metadata Arguments
PyTorch - yolov8n.pt ✅ -
TorchScript torchscript yolov8n.torchscript ✅ imgsz, optimize
ONNX onnx yolov8n.onnx ✅ imgsz, half, dynamic, simplify, opset
OpenVINO openvino yolov8n_openvino_model/ ✅ imgsz, half, int8
TensorRT engine yolov8n.engine ✅ imgsz, half, dynamic, simplify, workspace
CoreML coreml yolov8n.mlpackage ✅ imgsz, half, int8, nms
TF SavedModel saved_model yolov8n_saved_model/ ✅ imgsz, keras, int8
TF GraphDef pb yolov8n.pb ❌ imgsz
TF Lite tflite yolov8n.tflite ✅ imgsz, half, int8
TF Edge TPU edgetpu yolov8n_edgetpu.tflite ✅ imgsz
TF.js tfjs yolov8n_web_model/ ✅ imgsz, half, int8
PaddlePaddle paddle yolov8n_paddle_model/ ✅ imgsz
NCNN ncnn yolov8n_ncnn_model/ ✅ imgsz, half

See full export details in the Export page.



Created 2023-11-12, Updated 2024-04-18
Authors: glenn-jocher (11), RizwanMunawar (1), Laughing-q (1), maianumerosky (1)

Comments