数据集放在 datasets/coco_minitrain_10k数据集目录结构如下:
datasets/└── coco_mintrain_10k/├── annotations/│ ├── instances_train2017.json│ ├── instances_val2017.json│ ├── ... (其他标注文件)├── train2017/│ ├── 000000000001.jpg│ ├── ... (其他训练图像)├── val2017/│ ├── 000000000001.jpg│ ├── ... (其他验证图像)└── test2017/├── 000000000001.jpg├── ... (其他测试图像)
conda creaet -n yolo11_py310 python=3.10conda activate yolo11_py310pip install -U -r train/requirements.txt
先下载预训练权重:
bash 0_download_wgts.sh
执行预测测试:
bash 1_run_predict_yolo11.sh
预测结果保存在 runs 文件夹下,效果如下:

已经准备好一键训练肩膀,直接执行训练脚本:
bash 2_run_train_yolo11.sh
其中其作用的代码很简单,就在 train/train_yolo11.py 中,如下:
# Load a modelmodel = YOLO(curr_path + "/wgts/yolo11n.pt")# Train the modeltrain_results = model.train(data= curr_path + "/cfg/coco128.yaml", # path to dataset YAMLepochs=100, # number of training epochsimgsz=640, # training image sizedevice="0", # device to run on, i.e. device=0 or device=0,1,2,3 or device=cpu)# Evaluate model performance on the validation setmetrics = model.val()
主要就是配置一下训练参数,如数据集路径、训练轮数、显卡ID、图片大小等,然后执行训练即可
训练完成后,训练日志会在 runs/train 文件夹下,比如训练中 val 预测图片如下:

这样就完成了算法训练
使用 TensorRT 进行算法部署
直接执行一键导出ONNX脚本:
bash 3_run_export_onnx.sh
在脚本中已经对ONNX做了sim的简化
生成的ONNX以及_simONNX模型保存在wgts文件夹下
直接去NVIDIA的官网下载(https://developer.nvidia.com/tensorrt/download)对应版本的tensorrt TAR包,解压基本步骤如下:
tar zxvf TensorRT-xxx-.tar.gz# 软链trtexecsudo ln -s /path/to/TensorRT/bin/trtexec /usr/local/bin# 验证一下trtexec --help# 安装trt的python接口cd pythonpip install tensorrt-xxx.whl
直接执行一键生成trt模型引擎的脚本:
bash 4_build_trt_engine.sh
正常会在wgts路径下生成yolo11n.engine,并有类似如下的日志:
[10/02/2024-21:28:48] [V] === Explanations of the performance metrics ===[10/02/2024-21:28:48] [V] Total Host Walltime: the host walltime from when the first query (after warmups) is enqueued to when the last query is completed.[10/02/2024-21:28:48] [V] GPU Compute Time: the GPU latency to execute the kernels for a query.[10/02/2024-21:28:48] [V] Total GPU Compute Time: the summation of the GPU Compute Time of all the queries. If this is significantly shorter than Total Host Walltime, the GPU may be under-utilized because of host-side overheads or data transfers.[10/02/2024-21:28:48] [V] Throughput: the observed throughput computed by dividing the number of queries by the Total Host Walltime. If this is significantly lower than the reciprocal of GPU Compute Time, the GPU may be under-utilized because of host-side overheads or data transfers.[10/02/2024-21:28:48] [V] Enqueue Time: the host latency to enqueue a query. If this is longer than GPU Compute Time, the GPU may be under-utilized.[10/02/2024-21:28:48] [V] H2D Latency: the latency for host-to-device data transfers for input tensors of a single query.[10/02/2024-21:28:48] [V] D2H Latency: the latency for device-to-host data transfers for output tensors of a single query.[10/02/2024-21:28:48] [V] Latency: the summation of H2D Latency, GPU Compute Time, and D2H Latency. This is the latency to infer a single query.[10/02/2024-21:28:48] [I]&&&& PASSED TensorRT.trtexec [TensorRT v100500] [b18] # trtexec --onnx=../wgts/yolo11n_sim.onnx --saveEngine=../wgts/yolo11n.engine --fp16 --verbose
直接执行一键推理脚本:
bash 5_infer_trt.sh
实际的trt推理脚本在 deploy/infer_trt.py推理成功会有如下日志:
------ trt infer success! ------
推理结果保存在 deploy/output.jpg
如下:

好文章,需要你的鼓励
研究人员意外发现,标准MOSFET晶体管可同时模拟神经元和突触行为,形成"神经突触随机存取存储器"(NSRAM)。该技术仅需一至两个晶体管即可实现传统需数十乃至数百个元件才能完成的神经信号处理,且与现有硅基制造工艺完全兼容,良率达100%。未来有望应用于边缘AI及高能效神经形态芯片,长远或可挑战GPU地位。
复旦大学团队提出ICWM框架,让机器人在任务前通过随机探测动作自主感知当前视角和物理配置,无需重新训练即可适应新摄像头角度,真实机器人测试成功率最高提升175%。
本文提出一种评估人工智能风险的新方法,借鉴生态学与演化论视角,从理论生态模型中推导出三项风险指标,涵盖种群模型与生态系统模型。研究旨在为AI治理策略提供量化工具,并对分析局限性及政策改进方向进行了深入探讨,为构建更科学的AI风险评估体系提供参考框架。
西安交通大学团队提出Fast-LeWM,用动作前缀并行预测替代逐步推演,将AI世界模型规划速度提升近4倍,同时平均成功率从85.8%提升至90.5%。