【B站保姆级视频教程：Jetson配置YOLOv11环境（九）TensorRT导出参数详解】

Jetson配置YOLOv11环境（9）TensorRT导出参数详解

0. TensorRT 导出参数简介

ultralytics官方导出参数简介

论据	类型	默认值	说明
format	str	engine	导出模型的目标格式，定义与各种部署环境的兼容性。
imgsz	int 或 tuple	640	模型输入所需的图像尺寸。对于正方形图像，可以是一个整数，或者是一个元组 (height, width) 了解具体尺寸。
half	bool	False	启用 FP16（半精度）量化，在支持的硬件上减小模型大小并可能加快推理速度。
int8	bool	False	激活 INT8 量化，进一步压缩模型并加快推理速度，同时将精度损失降至最低，主要用于边缘设备。
dynamic	bool	False	允许动态输入尺寸，提高了处理不同图像尺寸的灵活性。
simplify	bool	True	简化模型图 onnxslim这可能会提高性能和兼容性。
workspace	float 或 None	None	为TensorRT 优化设置最大工作区大小（GiB），以平衡内存使用和性能；使用 None TensorRT 进行自动分配，最高可达设备最大值。
nms	bool	False	增加了非最大值抑制 (NMS)，这对精确高效的检测后处理至关重要。
batch	int	1	指定导出模型的批量推理大小，或导出模型将同时处理的图像的最大数量。 predict 模式。
data	str	coco8.yaml	通往 数据集 配置文件（默认： coco8.yaml)，对量化至关

1. TensorRT 导出参数设置

1.1 [imgsz参数]

imgsz=(height, width) ：设置与相机输入分辨率一致的推理尺寸。以imgsz=480,640为例：

1.1.1 导出

yolo export model=yolo11n.pt format=engine imgsz=480,640

导出完成

TensorRT: export success ✅ 354.2s, saved as 'yolo11n.engine' (12.7 MB)

Export complete (364.2s)
Results saved to /home/nx/Desktop/yolov11/ultralytics-main
Predict:         yolo predict task=detect model=yolo11n.engine imgsz=480,640  
Validate:        yolo val task=detect model=yolo11n.engine imgsz=480,640 data=/usr/src/ultralytics/ultralytics/cfg/datasets/coco.yaml  WARNING ⚠️ non-PyTorch val requires square images, 'imgsz=[480, 640]' will not work. Use export 'imgsz=640' if val is required.
Visualize:       https://netron.app
💡 Learn more at https://docs.ultralytics.com/modes/export

1.1.2 推理

yolo predict task=detect model=weights/imgsz_480x640/yolo11n.engine imgsz=480,640 source=videos/街道.mp4

实测结果：13.0ms

(pytorch) nx@nx-desktop:~/Desktop/yolov11/ultralytics-main$ yolo predict task=detect model=weights/imgsz_480x640/yolo11n.engine imgsz=480,640 source=videos/街道.mp4
Ultralytics 8.3.70 🚀 Python-3.8.20 torch-2.1.0a0+41361538.nv23.06 CUDA:0 (Xavier, 6833MiB)
Loading weights/imgsz_480x640/yolo11n.engine for TensorRT inference...
[02/07/2025-20:33:59] [TRT] [I] Loaded engine size: 12 MiB
[02/07/2025-20:34:06] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +343, GPU +323, now: CPU 691, GPU 3017 (MiB)
[02/07/2025-20:34:06] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +12, now: CPU 0, GPU 12 (MiB)
[02/07/2025-20:34:06] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +0, now: CPU 678, GPU 3017 (MiB)
[02/07/2025-20:34:06] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +15, now: CPU 0, GPU 27 (MiB)

video 1/1 (frame 1/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 12.9ms
video 1/1 (frame 2/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.1ms
video 1/1 (frame 3/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 4/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 5/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 6/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.1ms
video 1/1 (frame 7/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 8/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.1ms
video 1/1 (frame 9/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 10/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms
video 1/1 (frame 11/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 13.0ms

1.2 [half参数]

half=True：启用 FP16（半精度）量化，在支持的硬件上减小模型大小并可能加快推理速度。

1.2.0 Int8 vs FP16

对推理速度无极致的要求时，建议选择fp16即可。以下是使用 JetPack 6.0 (L4T 36.3) Ubuntu 22.04.4 LTS 测试。

精度	评估测试	平均 (ms)	min | max (ms)	mAPval 50(B)	mAPval 50-95(B)	batch	尺寸 （像素）
FP32	Predict	6.11	6.10 | 6.29			8	640
FP32	COCOval	6.17		0.52	0.37	1	640
FP16	Predict	3.18	3.18 | 3.20			8	640
FP16	COCOval	3.19		0.52	0.37	1	640
INT8	Predict	2.30	2.29 | 2.35			8	640
INT8	COCOval	2.32		0.46	0.32	1	640

• FP16（半精度）：
  • 直接降低模型权重和激活值的精度（从FP32转为FP16），不涉及量化过程。
  • 无需校准数据集，仅需在TensorRT引擎构建时启用FP16模式，由TensorRT自动完成精度转换。
  • 优势：计算速度提升（GPU的FP16算力通常更高），显存占用减少约50%。
• INT8（整数量化）：
  • 需要校准数据集（通常数百张代表性图片）统计激活值分布，生成动态范围校准表。
  • 必须通过校准器（如IInt8EntropyCalibrator2）完成量化，否则无法正确推理。
  • 优势：进一步加速推理（适用于对延迟敏感的场景），但可能引入精度损失。

1.2.1 导出

yolo export model=yolo11n.pt format=engine imgsz=480,640 half=True

导出完成

TensorRT: export success ✅ 1039.7s, saved as 'yolo11n.engine' (7.1 MB)

Export complete (1049.0s)
Results saved to /home/nx/Desktop/yolov11/ultralytics-main
Predict:         yolo predict task=detect model=yolo11n.engine imgsz=480,640 half 
Validate:        yolo val task=detect model=yolo11n.engine imgsz=480,640 data=/usr/src/ultralytics/ultralytics/cfg/datasets/coco.yaml half WARNING ⚠️ non-PyTorch val requires square images, 'imgsz=[480, 640]' will not work. Use export 'imgsz=640' if val is required.
Visualize:       https://netron.app
💡 Learn more at https://docs.ultralytics.com/modes/export

1.2.2 推理

yolo predict task=detect model=weights/half/yolo11n.engine imgsz=480,640 source=videos/街道.mp4

实测结果：7.3ms

(pytorch) nx@nx-desktop:~/Desktop/yolov11/ultralytics-main$ yolo predict task=detect model=weights/half/yolo11n.engine imgsz=480,640 source=videos/街道.mp4
Ultralytics 8.3.70 🚀 Python-3.8.20 torch-2.1.0a0+41361538.nv23.06 CUDA:0 (Xavier, 6833MiB)
Loading weights/half/yolo11n.engine for TensorRT inference...
[02/07/2025-21:08:12] [TRT] [I] Loaded engine size: 7 MiB
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +261, GPU +426, now: CPU 609, GPU 3185 (MiB)
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +82, GPU +151, now: CPU 691, GPU 3336 (MiB)
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +5, now: CPU 0, GPU 5 (MiB)
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +13, now: CPU 684, GPU 3333 (MiB)
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] Init cuDNN: CPU +0, GPU +4, now: CPU 684, GPU 3337 (MiB)
[02/07/2025-21:08:14] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +9, now: CPU 0, GPU 14 (MiB)

video 1/1 (frame 1/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 9.1ms
video 1/1 (frame 2/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 8.0ms
video 1/1 (frame 3/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 4/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 5/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.5ms
video 1/1 (frame 6/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 7/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 8/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 9/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 10/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 11/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.7ms
video 1/1 (frame 12/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 8.9ms
video 1/1 (frame 13/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 14/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 15/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 16/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 17/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 18/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 19/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms
video 1/1 (frame 20/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.4ms
video 1/1 (frame 21/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 7.3ms

1.3 [dla参数]

device="dla:0" 启用NVIDIA 深度学习加速器(Deep Learning Accelerator)。通过卸载GPU 的任务（将其释放给更密集的进程），DLA 可使模型以更低的功耗运行，同时保持高吞吐量（降速节能！）
注意：dla需配合FP16或INT8一起使用。

ultralytics官方：使用NVIDIA 深度学习加速器 (DLA)
在使用 DLA 导出时，某些层可能不支持在 DLA 上运行，而会回退到GPU 上执行。这种回退会带来额外的延迟，影响整体推理性能。因此，与完全在GPU 上运行的TensorRT 相比，DLA 的主要目的不是减少推理延迟，而是提高吞吐量和能效。

1.3.1 导出

yolo export model=yolo11n.pt format=engine imgsz=480,640 half=True device="dla:0" 

导出完成

TensorRT: export success ✅ 119.3s, saved as 'yolo11n.engine' (7.3 MB)

Export complete (129.7s)
Results saved to /home/nx/Desktop/yolov11/ultralytics-main
Predict:         yolo predict task=detect model=yolo11n.engine imgsz=480,640 half 
Validate:        yolo val task=detect model=yolo11n.engine imgsz=480,640 data=/usr/src/ultralytics/ultralytics/cfg/datasets/coco.yaml half WARNING ⚠️ non-PyTorch val requires square images, 'imgsz=[480, 640]' will not work. Use export 'imgsz=640' if val is required.
Visualize:       https://netron.app
💡 Learn more at https://docs.ultralytics.com/modes/export

1.3.2 推理

yolo predict task=detect model=weights/dla/yolo11n.engine imgsz=480,640 source=videos/街道.mp4

实测结果：27.0ms

(pytorch) nx@nx-desktop:~/Desktop/yolov11/ultralytics-main$ yolo predict task=detect model=weights/dla/yolo11n.engine imgsz=480,640 source=videos/街道.mp4
Ultralytics 8.3.70 🚀 Python-3.8.20 torch-2.1.0a0+41361538.nv23.06 CUDA:0 (Xavier, 6833MiB)
Loading weights/dla/yolo11n.engine for TensorRT inference...
[02/07/2025-22:12:54] [TRT] [I] Loaded engine size: 7 MiB
[02/07/2025-22:12:56] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +261, GPU +425, now: CPU 606, GPU 3086 (MiB)
[02/07/2025-22:12:56] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +7, GPU +0, now: CPU 7, GPU 0 (MiB)
[02/07/2025-22:12:56] [TRT] [I] [MemUsageChange] Init cuBLAS/cuBLASLt: CPU +0, GPU +26, now: CPU 598, GPU 3112 (MiB)
[02/07/2025-22:12:56] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +35, now: CPU 7, GPU 35 (MiB)

video 1/1 (frame 1/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 27.8ms
video 1/1 (frame 2/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 28.0ms
video 1/1 (frame 3/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 27.5ms
video 1/1 (frame 4/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 27.8ms
video 1/1 (frame 5/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 27.9ms
video 1/1 (frame 6/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.3ms
video 1/1 (frame 7/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.3ms
video 1/1 (frame 8/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.4ms
video 1/1 (frame 9/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.4ms
video 1/1 (frame 10/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.3ms
video 1/1 (frame 11/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.3ms
video 1/1 (frame 12/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.5ms
video 1/1 (frame 13/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.9ms
video 1/1 (frame 14/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.8ms
video 1/1 (frame 15/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 26.9ms
video 1/1 (frame 16/601) /home/nx/Desktop/yolov11/ultralytics-main/videos/街道.mp4: 480x640 1 car, 27.0ms

1.3.3 结论

发现使用dla推理速度变慢了，但是能耗并没有显著的降低。

1. 使用dla导出推理
   
2. 不使用dla导出推理
   

2. 总结

追求推理速度与精度的平衡，导出参数可配置为：imgsz=相机输入分辨率，half=True

yolo export model=yolo11n.pt format=engine imgsz=480,640 half=True