{"id":52960,"date":"2025-08-11T21:59:18","date_gmt":"2025-08-11T13:59:18","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/52960.html"},"modified":"2025-08-11T21:59:18","modified_gmt":"2025-08-11T13:59:18","slug":"autudl%e5%8d%8e%e4%b8%ba%e6%98%87%e8%85%be%e7%b3%bb%e5%88%97npu%e7%ae%80%e4%bb%8b%e5%92%8c%e9%83%a8%e7%bd%b2%e6%8e%a8%e7%90%86yolo11-yolov8-yolov5%e5%88%86%e5%89%b2%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/52960.html","title":{"rendered":"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b"},"content":{"rendered":"

0.\u914d\u7f6eAutudl<\/h3>\n

\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u4e0b\u9762\u56fe\u7247\u662f\u6211\u6240\u79df\u7684\u6607\u817e\u5361\u548c\u5177\u4f53\u73af\u5883\u7248\u672c，\u592a\u5177\u4f53\u7684\u5c31\u4e0d\u8bf4\u4e86，\u6709\u9700\u8981\u7684\u8bdd\u6211\u5355\u72ec\u51fa\u4e00\u671fAutudl\u79df\u663e\u5361\u7684\u6559\u7a0b，\u4e3b\u8981\u662f\u4e3a\u4e86\u5b66\u4e60\u6607\u817e\u73af\u5883\u5982\u4f55\u8fd0\u884cYolo\u7cfb\u5217\u6a21\u578b\u3002<\/p>\n

<\/p>\n

0.1\u534e\u4e3a\u6607\u817e\u82af\u7247（Ascend）\u7b80\u4ecb<\/h3>\n

1.Ascend 310（\u00a0\u8fb9\u7f18\u63a8\u7406 SoC）<\/h4>\n

\u00a0 \u00a0 \u00a0\u5728\u4ec5 8\u202fW \u7684\u529f\u8017\u4e0b，\u652f\u6301 FP16 \u8fbe 8\u202fTOPS，INT8 \u8fbe 16\u202fTOPS\u3002 \u00a0 \u00a0 \u00a0 \u652f\u6301\u591a\u901a\u9053\u9ad8\u6e05\u89c6\u9891\u5904\u7406：\u6700\u9ad8\u53ef\u5b9e\u73b0 16 \u8def 1080p H.264\/H.265 \u89e3\u7801\u30011 \u8def 1080p \u89c6\u9891\u7f16\u7801，\u4ee5\u53ca JPEG\/PNG \u7f16\u89e3\u7801\u529f\u80fd\u3002 \u00a0 \u00a0 \u00a0 \u00a0 \u5185\u542b 2 \u4e2a Da\u202fVinci Max AI \u6838\u5fc3\u30018 \u4e2a ARM Cortex\u2011A55 CPU \u6838\u5fc3\u548c 8\u202fMB \u7247\u4e0a\u7f13\u5b58\u3002<\/p>\n

2.Ascend 910（\u6570\u636e\u4e2d\u5fc3\u7ea7\u8bad\u7ec3 NPU）<\/h4>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u4e8e 2019 \u5e74\u53d1\u5e03，\u4ee3\u53f7 Ascend\u2011Max，\u57fa\u4e8e Da\u202fVinci \u67b6\u6784，\u62e5\u6709 32 \u4e2a Da\u202fVinci Max AI \u6838\u5fc3，FP16 \u8fbe 256\u202fTFLOPS，INT8 \u8fbe 512\u202fTOPS\u3002 \u00a0 \u00a0 \u00a0 \u00a0 \u914d\u5907\u9ad8\u5e26\u5bbd\u4e92\u8054（NoC Mesh 1024 \u4f4d，HBM2E \u5e26\u5bbd 1.2\u202fTB\/s，350\u202fW \u529f\u8017）<\/p>\n

3.Ascend 910C（2025 \u5e74\u4e3b\u6253\u4ea7\u54c1）<\/h4>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u57fa\u4e8e SMIC \u7684 7\u202fnm (N+2) \u5de5\u827a，\u91c7\u7528\u53cc 910B Die \u5c01\u88c5\u8bbe\u8ba1，\u62e5\u6709\u7ea6 53 \u4ebf\u6676\u4f53\u7ba1\u3002 \u00a0 \u00a0 \u00a0 \u00a0 \u5b9e\u73b0 FP16 \u7ea6 800\u202fTFLOPS \u6027\u80fd，\u642d\u8f7d\u9ad8\u8fbe 128\u202fGB \u7684 HBM3 \u5b58\u50a8（\u9886\u5148\u4e8e NVIDIA H100 \u7684 80\u202fGB）\u3002\u63a8\u6d4b\u6027\u80fd\u7ea6\u4e3a NVIDIA H100 \u63a8\u7406\u6027\u80fd\u7684 60%<\/p>\n

0.2MindSpore\u6846\u67b6<\/h3>\n

\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0MindSpore \u662f\u534e\u4e3a\u81ea\u4e3b\u7814\u53d1\u7684\u5168\u573a\u666f AI \u8ba1\u7b97\u6846\u67b6，\u8986\u76d6\u4e91\u3001\u8fb9\u3001\u7aef\u591a\u79cd\u90e8\u7f72\u73af\u5883\u3002\u5411\u4e0a\u652f\u6301\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u79d1\u5b66\u8ba1\u7b97\u7b49\u591a\u7c7b AI \u5e94\u7528，\u5411\u4e0b\u901a\u8fc7 CANN \u5bf9\u63a5\u6607\u817e AI \u5904\u7406\u5668，\u5b9e\u73b0\u8f6f\u786c\u4ef6\u534f\u540c\u4f18\u5316\u3002 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u63d0\u4f9b\u7aef\u5230\u7aef\u7684\u6a21\u578b\u5f00\u53d1\u80fd\u529b，\u5305\u62ec\u6570\u636e\u5904\u7406\u3001\u6a21\u578b\u6784\u5efa\u3001\u8bad\u7ec3\u3001\u63a8\u7406\u4e0e\u90e8\u7f72，\u5185\u7f6e\u81ea\u52a8\u5e76\u884c\u3001\u7b97\u5b50\u878d\u5408\u3001\u6df7\u5408\u7cbe\u5ea6\u7b49\u4f18\u5316\u673a\u5236，\u663e\u8457\u63d0\u5347\u6027\u80fd\u4e0e\u7b97\u529b\u5229\u7528\u7387\u3002\u5176\u529f\u80fd\u5b9a\u4f4d\u7c7b\u4f3c\u4e8e TensorFlow \u6216 PyTorch，\u5728\u6607\u817e\u751f\u6001\u4e2d\u627f\u62c5\u6838\u5fc3 AI \u6846\u67b6\u7684\u89d2\u8272\u3002<\/p>\n

0.3CANN\u8ba1\u7b97\u67b6\u6784（Compute Architecture for Neural Networks）<\/h3>\n

\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u534e\u4e3a\u9762\u5411AI\u573a\u666f\u63a8\u51fa\u7684\u5f02\u6784\u8ba1\u7b97\u67b6\u6784 CANN，\u4f5c\u4e3a\u6607\u817eAI\u5904\u7406\u5668\u7684\u8f6f\u4ef6\u9a71\u52a8\u4e0e\u7b97\u5b50\u52a0\u901f\u5e73\u53f0，\u5411\u4e0a\u517c\u5bb9\u591a\u79cd\u4e3b\u6d41AI\u6846\u67b6（\u5982 MindSpore\u3001PyTorch\u3001TensorFlow \u7b49），\u5411\u4e0b\u5bf9\u63a5\u6607\u817e\u7cfb\u5217AI\u5904\u7406\u5668\u7684\u7b97\u5b50\u5e93\u3001\u8fd0\u884c\u65f6\u4e0e\u8c03\u5ea6\u5c42，\u53d1\u6325\u627f\u4e0a\u542f\u4e0b\u7684\u5173\u952e\u4f5c\u7528\u3002CANN \u5c4f\u853d\u4e86\u5e95\u5c42\u786c\u4ef6\u5dee\u5f02，\u4f7f\u5f00\u53d1\u8005\u80fd\u591f\u5728\u4e0d\u5173\u5fc3\u786c\u4ef6\u7ec6\u8282\u7684\u60c5\u51b5\u4e0b\u65e0\u7f1d\u4f7f\u7528\u4e3b\u6d41\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u8fdb\u884c\u6a21\u578b\u5f00\u53d1\u4e0e\u90e8\u7f72\u3002<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u540c\u65f6，CANN \u9488\u5bf9\u591a\u6837\u5316\u7684\u5e94\u7528\u573a\u666f\u63d0\u4f9b\u591a\u5c42\u6b21\u7f16\u7a0b\u63a5\u53e3（\u5982\u7b97\u5b50\u7ea7\u3001\u56fe\u7ea7\u3001\u5e94\u7528\u7ea7），\u652f\u6301\u7528\u6237\u5feb\u901f\u6784\u5efa\u548c\u4f18\u5316\u57fa\u4e8e\u6607\u817e\u5e73\u53f0\u7684AI\u5e94\u7528\u4e0e\u4e1a\u52a1，\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6267\u884c\u6548\u7387\u4e0e\u786c\u4ef6\u7b97\u529b\u5229\u7528\u7387，\u5176\u529f\u80fd\u5b9a\u4f4d\u7c7b\u4f3c\u4e8e NVIDIA \u7684 CUDA \u5e73\u53f0\u3002<\/p>\n

\u603b\u7ed3：CANN\u7528\u6765\u5c4f\u853d\u5e95\u5c42\u786c\u4ef6\u5dee\u5f02，\u4f7f\u5f97\u7528\u6237\u80fd\u591f\u65e0\u7f1d\u4f7f\u7528Pytorch\u7b49\u4e3b\u6d41\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u8fdb\u884c\u5f00\u53d1<\/p>\n

\u2f2f\u4f5c\u6d41\u7a0b：
\n ONNX
\n \u2192
\n ATC \u8f6c\u6362
\n \u2192
\n ACLruntime \u8c03\u2f64
\n \u2192 \u63a8\u7406\u7ed3\u679c<\/p>\n

1\u914d\u7f6e\u73af\u5883<\/h2>\n

1.1\u521b\u5efa\u4e00\u4e2ayolo\u73af\u5883\u5e76\u8fdb\u5165<\/h3>\n

conda create -n yolo11 python=3.10
\nconda activate yolo11 <\/p>\n

\u00a01.2\u5b89\u88c5yolo\u6240\u9700\u8981\u7684\u5e93<\/h3>\n

pip install ultralytics <\/p>\n

1.3\u5b89\u88c5ais_bench\u63a8\u7406\u5de5\u5177，\u548caclruntime<\/h3>\n

\u6839\u636e\u81ea\u5df1\u4f7f\u7528\u7684python\u3001\u7248\u672c\u7cfb\u7edf\u548c\u67b6\u6784\u9009\u62e9\u5177\u4f53\u7684whl\u6587\u4ef6<\/p>\n

tools: Ascend tools – Gitee.com<\/span>\"https:\/\/gitee.com\/ascend\/tools\/tree\/master\/ais-bench_workload\/tool\/ais_bench#%E4%B8%8B%E8%BD%BDwhl%E5%8C%85%E5%AE%89%E8%A3%85<\/span><\/span><\/p>\n

\u00a0\u4e0b\u8f7d\u5230\u672c\u5730\u540e，\u5b89\u88c5\u4e24\u4e2awhl\u6587\u4ef6<\/p>\n

1.3.1\u5b89\u88c5ais_bench<\/h4>\n

pip install ais_bench-0.0.2-py3-none-any.whl -i https:\/\/mirrors.tuna.tsinghua.edu.cn\/pypi\/web\/simple <\/p>\n

1.3.2\u5b89\u88c5aclruntime<\/h4>\n

pip install aclruntime-0.0.2-cp310-cp310-linux_aarch64.whl <\/p>\n

\u540e\u9762\u542f\u52a8\u53ef\u80fd\u4f1a\u62a5\u4e00\u4e9b\u9519\u8bef，\u8ba9ai\u89e3\u51b3\u4e00\u4e0b，\u5f88\u597d\u89e3\u51b3<\/p>\n

NEW_LIB=~\/autodl-tmp\/libstdcpp_arm64\/usr\/lib\/aarch64-linux-gnu\/libstdc++.so.6
\nln -sf \/usr\/lib\/aarch64-linux-gnu\/libstdc++.so.6 \/root\/miniconda3\/lib\/libstdc++.so.6 <\/p>\n

2\u51c6\u5907\u6e90\u7801\u548c\u6a21\u578b<\/h2>\n

\u6a21\u578b\u53ef\u4ee5\u662f\u81ea\u5df1\u7684，\u4e5f\u53ef\u4ee5\u4ece\u5b98\u7f51\u4e0a\u4e0b\u8f7d<\/p>\n

2.1\u4e0b\u8f7d\u6e90\u7801<\/h3>\n

https:\/\/github.com\/ultralytics\/ultralytics<\/p>\n

2.2\u4e0b\u8f7d\u6a21\u578b\u6743\u91cd<\/h3>\n

Ultralytics YOLO11 – Ultralytics YOLO \u6587\u6863<\/p>\n

3\u8f6c\u6362\u6a21\u578b<\/h2>\n

3.1\u5c06.pt\u683c\u5f0f\u8f6c\u6362\u4e3a.ONNX\u683c\u5f0f<\/h3>\n

from ultralytics import YOLO<\/p>\n

# Load a model
\nmodel = YOLO(r"\/home\/ultralytics-main\/runs\/train\/yolo11n\/weights\/best.pt")<\/p>\n

# Export onnx,\u4e00\u5b9a\u8981\u8bbe\u7f6eopset=11 \u4e0d\u7136\u4f1a\u62a5\u9519
\nmodel.export(format="onnx",opset=11) <\/p>\n

3.1.1\u67e5\u770bONNX\u2f79\u7edc\u7ed3\u6784\u548c\u8f93\u2f0a\u8f93\u51fashape<\/h4>\n

Netron<\/p>\n

\u8fdb\u5165\u7f51\u7ad9\u540e，\u4e0a\u4f20ONNX\u6a21\u578b\u6587\u4ef6<\/p>\n

<\/p>\n

3.2\u5c06.ONNX\u683c\u5f0f\u8f6c\u6362\u4e3a.om\u683c\u5f0f\u00a0<\/h3>\n

atc \\\\
\n –model=\/root\/autodl-tmp\/ultralytics-main\/runs\/train\/exp4\/weights\/best.onnx \\\\
\n –framework=5 \\\\
\n –output=yolo11s \\\\
\n –input_format=NCHW \\\\
\n –input_shape="images:1,3,640,640" \\\\
\n –soc_version=Ascend910B2 <\/p>\n

\u53c2\u6570\u5206\u522b\u662f： –model \u9700\u8981\u8f6c\u6362\u6a21\u578b\u7684\u8def\u5f84<\/p>\n

–framework \u6846\u67b6，ONNX\u5c31\u75285<\/p>\n

–output \u00a0\u8f93\u51fa.om\u6587\u4ef6\u7684\u540d\u5b57<\/p>\n

–input_shape \u6a21\u578b\u7684\u8f93\u5165shape，640 640\u662f\u8bad\u7ec3\u65f6\u56fe\u7247\u7684\u5927\u5c0f，\u6839\u636e\u9700\u8981\u4fee\u6539<\/p>\n

–soc_version NPU\u7684\u578b\u53f7<\/p>\n

\u6839\u636e\u4e0b\u9762\u7684\u6307\u4ee4\u67e5\u8be2npu\u578b\u53f7\u540e，\u66ff\u6362910B2<\/p>\n

npu-smi info #\u7528\u8fd9\u4e2a\u6307\u4ee4\u53ef\u4ee5\u67e5\u8be2npu\u578b\u53f7 <\/p>\n

4.\u63a8\u7406\u6e90\u7801<\/h3>\n

\u5728ultralytics\u8def\u5f84\u4e0b\u521b\u5efarun_seg.py<\/p>\n

\u5c06\u4e0b\u8ff0\u4ee3\u7801\u590d\u5236\u8fdb\u53bb<\/p>\n

import argparse
\nimport time
\nimport cv2
\nimport numpy as np
\nimport os<\/p>\n

from ais_bench.infer.interface import InferSession<\/p>\n

class YOLO:
\n """YOLO segmentation model class for handling inference"""<\/p>\n

def __init__(self, om_model, imgsz=(640, 640), device_id=0, model_ndtype=np.single, mode="static", postprocess_type="v8", aipp=False):
\n """
\n Initialization.<\/p>\n

Args:
\n om_model (str): Path to the om model.
\n """<\/p>\n

# \u6784\u5efaais_bench\u63a8\u7406\u5f15\u64ce
\n self.session = InferSession(device_id=device_id, model_path=om_model)<\/p>\n

# Numpy dtype: support both FP32(np.single) and FP16(np.half) om model
\n self.ndtype = model_ndtype
\n self.mode = mode
\n self.postprocess_type = postprocess_type
\n self.aipp = aipp <\/p>\n

self.model_height, self.model_width = imgsz[0], imgsz[1] # \u56fe\u50cfresize\u5927\u5c0f<\/p>\n

def __call__(self, im0, conf_threshold=0.4, iou_threshold=0.45):
\n """
\n The whole pipeline: pre-process -> inference -> post-process.<\/p>\n

Args:
\n im0 (Numpy.ndarray): original input image.
\n conf_threshold (float): confidence threshold for filtering predictions.
\n iou_threshold (float): iou threshold for NMS.<\/p>\n

Returns:
\n boxes (List): list of bounding boxes.
\n """
\n # \u524d\u5904\u7406Pre-process
\n t1 = time.time()
\n im, ratio, (pad_w, pad_h) = self.preprocess(im0)
\n pre_time = round(time.time() – t1, 3)<\/p>\n

# \u63a8\u7406 inference
\n t2 = time.time()
\n preds = self.session.infer([im], mode=self.mode) # mode\u6709\u52a8\u6001"dymshape"\u548c\u9759\u6001"static"\u7b49
\n det_time = round(time.time() – t2, 3)<\/p>\n

# \u540e\u5904\u7406Post-process
\n t3 = time.time()
\n if self.postprocess_type == "v5":
\n boxes, segments, masks = self.postprocess_v5(preds,
\n im0=im0,
\n ratio=ratio,
\n pad_w=pad_w,
\n pad_h=pad_h,
\n conf_threshold=conf_threshold,
\n iou_threshold=iou_threshold,
\n )<\/p>\n

elif self.postprocess_type == "v8":
\n boxes, segments, masks = self.postprocess_v8(preds,
\n im0=im0,
\n ratio=ratio,
\n pad_w=pad_w,
\n pad_h=pad_h,
\n conf_threshold=conf_threshold,
\n iou_threshold=iou_threshold,
\n )<\/p>\n

else:
\n boxes=[], segments=[], masks=[]<\/p>\n

post_time = round(time.time() – t3, 3)<\/p>\n

return boxes, segments, masks, (pre_time, det_time, post_time)<\/p>\n

# \u524d\u5904\u7406，\u5305\u62ec：resize, pad, \u5176\u4e2dHWC to CHW，BGR to RGB，\u5f52\u4e00\u5316，\u589e\u52a0\u7ef4\u5ea6CHW -> BCHW\u53ef\u9009\u62e9\u662f\u5426\u5f00\u542fAIPP\u52a0\u901f\u5904\u7406
\n def preprocess(self, img):
\n """
\n Pre-processes the input image.<\/p>\n

Args:
\n img (Numpy.ndarray): image about to be processed.<\/p>\n

Returns:
\n img_process (Numpy.ndarray): image preprocessed for inference.
\n ratio (tuple): width, height ratios in letterbox.
\n pad_w (float): width padding in letterbox.
\n pad_h (float): height padding in letterbox.
\n """
\n # Resize and pad input image using letterbox() (Borrowed from Ultralytics)
\n shape = img.shape[:2] # original image shape
\n new_shape = (self.model_height, self.model_width)
\n r = min(new_shape[0] \/ shape[0], new_shape[1] \/ shape[1])
\n ratio = r, r
\n new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
\n pad_w, pad_h = (new_shape[1] – new_unpad[0]) \/ 2, (new_shape[0] – new_unpad[1]) \/ 2 # wh padding
\n if shape[::-1] != new_unpad: # resize
\n img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)<\/p>\n

top, bottom = int(round(pad_h – 0.1)), int(round(pad_h + 0.1))
\n left, right = int(round(pad_w – 0.1)), int(round(pad_w + 0.1))
\n img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=(114, 114, 114)) # \u586b\u5145<\/p>\n

# \u662f\u5426\u5f00\u542faipp\u52a0\u901f\u9884\u5904\u7406，\u9700atc\u4e2d\u5b8c\u6210
\n if self.aipp:
\n return img, ratio, (pad_w, pad_h)<\/p>\n

# Transforms: HWC to CHW -> BGR to RGB -> div(255) -> contiguous -> add axis(optional)
\n img = np.ascontiguousarray(np.einsum('HWC->CHW', img)[::-1], dtype=self.ndtype) \/ 255.0
\n img_process = img[None] if len(img.shape) == 3 else img
\n return img_process, ratio, (pad_w, pad_h)<\/p>\n

# YOLOv5\/6\/7\u901a\u7528\u540e\u5904\u7406，\u5305\u62ec：\u9608\u503c\u8fc7\u6ee4\u4e0eNMS+masks\u5904\u7406
\n def postprocess_v5(self, preds, im0, ratio, pad_w, pad_h, conf_threshold, iou_threshold, nm=32):
\n """
\n Post-process the prediction.<\/p>\n

Args:
\n preds (Numpy.ndarray): predictions come from ort.session.run().
\n im0 (Numpy.ndarray): [h, w, c] original input image.
\n ratio (tuple): width, height ratios in letterbox.
\n pad_w (float): width padding in letterbox.
\n pad_h (float): height padding in letterbox.
\n conf_threshold (float): conf threshold.
\n iou_threshold (float): iou threshold.
\n nm (int): the number of masks.<\/p>\n

Returns:
\n boxes (List): list of bounding boxes.
\n segments (List): list of segments.
\n masks (np.ndarray): [N, H, W], output masks.
\n """
\n # (Batch_size, Num_anchors, xywh_score_conf_cls), v5\u548cv6_1.0\u7684[…, 4]\u662f\u7f6e\u4fe1\u5ea6\u5206\u6570，v8v9\u91c7\u7528\u7c7b\u522b\u91cc\u9762\u6700\u5927\u7684\u6982\u7387\u4f5c\u4e3a\u7f6e\u4fe1\u5ea6score
\n x, protos = preds[0], preds[1] # \u4e0ebbox\u533a\u522b：Two outputs: \u68c0\u6d4b\u5934\u7684\u8f93\u51fa(1, 8400*3, 117), \u5206\u5272\u5934\u7684\u8f93\u51fa(1, 32, 160, 160)<\/p>\n

# Predictions filtering by conf-threshold
\n x = x[x[…, 4] > conf_threshold]<\/p>\n

# Create a new matrix which merge these(box, score, cls, nm) into one
\n # For more details about `numpy.c_()`: https:\/\/numpy.org\/doc\/1.26\/reference\/generated\/numpy.c_.html
\n x = np.c_[x[…, :4], x[…, 4], np.argmax(x[…, 5:-nm], axis=-1), x[…, -nm:]]<\/p>\n

# NMS filtering
\n # \u7ecf\u8fc7NMS\u540e\u7684\u503c, np.array([[x, y, w, h, conf, cls, nm], …]), shape=(-1, 4 + 1 + 1 + 32)
\n x = x[cv2.dnn.NMSBoxes(x[:, :4], x[:, 4], conf_threshold, iou_threshold)]<\/p>\n

# \u91cd\u65b0\u7f29\u653e\u8fb9\u754c\u6846，\u4e3a\u753b\u56fe\u505a\u51c6\u5907
\n if len(x) > 0:
\n # Bounding boxes format change: cxcywh -> xyxy
\n x[…, [0, 1]] -= x[…, [2, 3]] \/ 2
\n x[…, [2, 3]] += x[…, [0, 1]]<\/p>\n

# Rescales bounding boxes from model shape(model_height, model_width) to the shape of original image
\n x[…, :4] -= [pad_w, pad_h, pad_w, pad_h]
\n x[…, :4] \/= min(ratio)<\/p>\n

# Bounding boxes boundary clamp
\n x[…, [0, 2]] = x[:, [0, 2]].clip(0, im0.shape[1])
\n x[…, [1, 3]] = x[:, [1, 3]].clip(0, im0.shape[0])<\/p>\n

# \u4e0ebbox\u533a\u522b：\u589e\u52a0masks\u5904\u7406
\n # Process masks
\n masks = self.process_mask(protos[0], x[:, 6:], x[:, :4], im0.shape)
\n # Masks -> Segments(contours)
\n segments = self.masks2segments(masks)<\/p>\n

return x[…, :6], segments, masks # boxes, segments, masks
\n else:
\n return [], [], []<\/p>\n

def postprocess_v8(self, preds, im0, ratio, pad_w, pad_h, conf_threshold, iou_threshold):
\n x, protos = preds[0], preds[1] # x: (1, 37, 8400), protos: (1, 32, 160, 160)<\/p>\n

# \u7edf\u4e00\u4e3a (8400, 37)
\n if x.ndim == 3:
\n if x.shape[1] < x.shape[2]: # (1, 37, 8400)
\n x = np.einsum('bcn->bnc', x)[0] # -> (8400, 37)
\n else: # (1, 8400, 37)
\n x = x[0]
\n else:
\n raise ValueError(f'unexpected pred0 shape: {x.shape}')<\/p>\n

# \u52a8\u6001\u786e\u5b9a nm（\u6b64\u6a21\u578b\u5c31\u662f 32）
\n nm = protos.shape[1] if (protos.ndim == 4 and protos.shape[1] in (16, 32, 64)) else 32<\/p>\n

# \u53d6\u7c7b\u522b\u5206\u652f\u5e76\u6309\u9700\u505a sigmoid
\n cls_blob = x[:, 4:-nm] # \u7ef4\u5ea6\u5e94\u4e3a (8400, num_classes)；\u4f60\u7684\u6a21\u578b num_classes=1
\n if cls_blob.size and (cls_blob.max() > 1 or cls_blob.min() < 0):
\n cls_scores = 1.0 \/ (1.0 + np.exp(-cls_blob))
\n else:
\n cls_scores = cls_blob<\/p>\n

if cls_scores.size == 0:
\n return [], [], []<\/p>\n

scores = cls_scores.max(axis=-1)
\n clses = cls_scores.argmax(axis=-1)<\/p>\n

keep = scores > conf_threshold
\n if not np.any(keep):
\n return [], [], []<\/p>\n

# \u62fc\u63a5\u4e3a\u7edf\u4e00\u6570\u7ec4
\n x = np.c_[x[keep, :4], scores[keep], clses[keep], x[keep, -nm:]] # [cx,cy,w,h,score,cls, 32-vec]<\/p>\n

# === NMS \u524d\u5148\u628a cx,cy,w,h -> x,y,w,h（\u5de6\u4e0a+\u5bbd\u9ad8）===
\n tlwh = np.c_[x[:, 0] – x[:, 2] \/ 2,
\n x[:, 1] – x[:, 3] \/ 2,
\n x[:, 2],
\n x[:, 3]]<\/p>\n

idxs = cv2.dnn.NMSBoxes(tlwh.tolist(), x[:, 4].astype(float).tolist(),
\n conf_threshold, iou_threshold)
\n if len(idxs) == 0:
\n return [], [], []
\n idxs = np.array(idxs).reshape(-1)
\n x = x[idxs]<\/p>\n

# \u8f6c\u56de xyxy \u5e76\u6620\u5c04\u56de\u539f\u56fe
\n x[…, [0, 1]] -= x[…, [2, 3]] \/ 2
\n x[…, [2, 3]] += x[…, [0, 1]]<\/p>\n

x[:, :4] -= [pad_w, pad_h, pad_w, pad_h]
\n x[:, :4] \/= min(ratio)
\n x[:, [0, 2]] = x[:, [0, 2]].clip(0, im0.shape[1])
\n x[:, [1, 3]] = x[:, [1, 3]].clip(0, im0.shape[0])<\/p>\n

# \u5904\u7406 mask
\n masks = self.process_mask(protos[0], x[:, 6:], x[:, :4], im0.shape)
\n segments = self.masks2segments(masks)
\n return x[:, :6], segments, masks<\/p>\n

@staticmethod
\n def masks2segments(masks):
\n """
\n It takes a list of masks(n,h,w) and returns a list of segments(n,xy) (Borrowed from
\n https:\/\/github.com\/ultralytics\/ultralytics\/blob\/465df3024f44fa97d4fad9986530d5a13cdabdca\/ultralytics\/utils\/ops.py#L750)<\/p>\n

Args:
\n masks (numpy.ndarray): the output of the model, which is a tensor of shape (batch_size, 160, 160).<\/p>\n

Returns:
\n segments (List): list of segment masks.
\n """
\n segments = []
\n for x in masks.astype('uint8'):
\n c = cv2.findContours(x, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)[0] # CHAIN_APPROX_SIMPLE \u8be5\u51fd\u6570\u7528\u4e8e\u67e5\u627e\u4e8c\u503c\u56fe\u50cf\u4e2d\u7684\u8f6e\u5ed3\u3002
\n if c:
\n # \u8fd9\u6bb5\u4ee3\u7801\u7684\u76ee\u7684\u662f\u627e\u5230\u56fe\u50cfx\u4e2d\u7684\u6700\u5916\u5c42\u8f6e\u5ed3，\u5e76\u4ece\u4e2d\u9009\u62e9\u6700\u957f\u7684\u8f6e\u5ed3，\u7136\u540e\u5c06\u5176\u8f6c\u6362\u4e3aNumPy\u6570\u7ec4\u7684\u5f62\u5f0f\u3002
\n c = np.array(c[np.array([len(x) for x in c]).argmax()]).reshape(-1, 2)
\n else:
\n c = np.zeros((0, 2)) # no segments found
\n segments.append(c.astype('float32'))
\n return segments<\/p>\n

def process_mask(self, protos, masks_in, bboxes, im0_shape):
\n c, mh, mw = protos.shape
\n masks = np.matmul(masks_in, protos.reshape((c, -1))).reshape((-1, mh, mw)) # [n, mh, mw]<\/p>\n

# \u2605 \u5148 sigmoid，\u628a\u7ebf\u6027\u53e0\u52a0\u7684 logits \u8f6c\u6210\u6982\u7387
\n masks = 1.0 \/ (1.0 + np.exp(-masks))<\/p>\n

masks = np.transpose(masks, (1, 2, 0)) # -> [mh, mw, n]
\n masks = np.ascontiguousarray(masks)
\n masks = self.scale_mask(masks, im0_shape) # -> [H, W, n]
\n masks = np.einsum('HWN -> NHW', masks) # -> [n, H, W]
\n masks = self.crop_mask(masks, bboxes) # \u88c1\u526a\u5230\u6846\u5185
\n return masks > 0.5<\/p>\n

@staticmethod
\n def scale_mask(masks, im0_shape, ratio_pad=None):
\n """
\n Takes a mask, and resizes it to the original image size. (Borrowed from
\n https:\/\/github.com\/ultralytics\/ultralytics\/blob\/465df3024f44fa97d4fad9986530d5a13cdabdca\/ultralytics\/utils\/ops.py#L305)<\/p>\n

Args:
\n masks (np.ndarray): resized and padded masks\/images, [h, w, num]\/[h, w, 3].
\n im0_shape (tuple): the original image shape.
\n ratio_pad (tuple): the ratio of the padding to the original image.<\/p>\n

Returns:
\n masks (np.ndarray): The masks that are being returned.
\n """
\n im1_shape = masks.shape[:2]
\n if ratio_pad is None: # calculate from im0_shape
\n gain = min(im1_shape[0] \/ im0_shape[0], im1_shape[1] \/ im0_shape[1]) # gain = old \/ new
\n pad = (im1_shape[1] – im0_shape[1] * gain) \/ 2, (im1_shape[0] – im0_shape[0] * gain) \/ 2 # wh padding
\n else:
\n pad = ratio_pad[1]<\/p>\n

# Calculate tlbr of mask
\n top, left = int(round(pad[1] – 0.1)), int(round(pad[0] – 0.1)) # y, x
\n bottom, right = int(round(im1_shape[0] – pad[1] + 0.1)), int(round(im1_shape[1] – pad[0] + 0.1))
\n if len(masks.shape) < 2:
\n raise ValueError(f'"len of masks shape" should be 2 or 3, but got {len(masks.shape)}')
\n masks = masks[top:bottom, left:right]
\n masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]),
\n interpolation=cv2.INTER_LINEAR) # INTER_CUBIC would be better
\n if len(masks.shape) == 2:
\n masks = masks[:, :, None]
\n return masks<\/p>\n

@staticmethod
\n def crop_mask(masks, boxes):
\n """
\n It takes a mask and a bounding box, and returns a mask that is cropped to the bounding box. (Borrowed from
\n https:\/\/github.com\/ultralytics\/ultralytics\/blob\/465df3024f44fa97d4fad9986530d5a13cdabdca\/ultralytics\/utils\/ops.py#L599)<\/p>\n

Args:
\n masks (Numpy.ndarray): [n, h, w] tensor of masks.
\n boxes (Numpy.ndarray): [n, 4] tensor of bbox coordinates in relative point form.<\/p>\n

Returns:
\n (Numpy.ndarray): The masks are being cropped to the bounding box.
\n """
\n n, h, w = masks.shape
\n x1, y1, x2, y2 = np.split(boxes[:, :, None], 4, 1)
\n r = np.arange(w, dtype=x1.dtype)[None, None, :]
\n c = np.arange(h, dtype=x1.dtype)[None, :, None]
\n return masks * ((r >= x1) * (r < x2) * (c >= y1) * (c < y2))<\/p>\n

if __name__ == '__main__':
\n # Create an argument parser to handle command-line arguments
\n parser = argparse.ArgumentParser()
\n parser.add_argument('–seg_model', type=str, default=r"yolov8s-seg.om", help='Path to OM model')
\n parser.add_argument('–source', type=str, default=r'images', help='Path to input image')
\n parser.add_argument('–out_path', type=str, default=r'results', help='\u7ed3\u679c\u4fdd\u5b58\u6587\u4ef6\u5939')
\n parser.add_argument('–imgsz_seg', type=tuple, default=(640, 640), help='Image input size')
\n parser.add_argument('–classes', type=list, default=['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
\n 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
\n 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
\n 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
\n 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich',
\n 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
\n 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven',
\n 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'], help='\u7c7b\u522b')<\/p>\n

parser.add_argument('–conf', type=float, default=0.7, help='Confidence threshold')
\n parser.add_argument('–iou', type=float, default=0.7, help='NMS IoU threshold')
\n parser.add_argument('–device_id', type=int, default=0, help='device id')
\n parser.add_argument('–mode', default='static', help='om\u662f\u52a8\u6001dymshape\u6216\u9759\u6001static')
\n parser.add_argument('–model_ndtype', default=np.single, help='om\u662ffp32\u6216fp16')
\n parser.add_argument('–postprocess_type', type=str, default='v8', help='\u540e\u5904\u7406\u65b9\u5f0f, \u5bf9\u5e94v5\/v8\u4e24\u79cd\u540e\u5904\u7406')
\n parser.add_argument('–aipp', default=False, action='store_true', help='\u662f\u5426\u5f00\u542faipp\u52a0\u901fYOLO\u9884\u5904\u7406, \u9700atc\u4e2d\u5b8c\u6210om\u96c6\u6210')
\n args = parser.parse_args()<\/p>\n

# \u521b\u5efa\u7ed3\u679c\u4fdd\u5b58\u6587\u4ef6\u5939
\n if not os.path.exists(args.out_path):
\n os.mkdir(args.out_path)<\/p>\n

print('\u5f00\u59cb\u8fd0\u884c：')
\n # Build model
\n seg_model = YOLO(args.seg_model, args.imgsz_seg, args.device_id, args.model_ndtype, args.mode, args.postprocess_type, args.aipp)
\n color_palette = np.random.uniform(0, 255, size=(len(args.classes), 3)) # \u4e3a\u6bcf\u4e2a\u7c7b\u522b\u751f\u6210\u8c03\u8272\u677f<\/p>\n

for i, img_name in enumerate(os.listdir(args.source)):
\n try:
\n t1 = time.time()
\n # Read image by OpenCV
\n img = cv2.imread(os.path.join(args.source, img_name))<\/p>\n

# \u68c0\u6d4bInference
\n boxes, segments, _ , (pre_time, det_time, post_time) = seg_model(img, conf_threshold=args.conf, iou_threshold=args.iou)
\n print('{}\/{} ==>\u603b\u8017\u65f6\u95f4: {:.3f}s, \u5176\u4e2d, \u9884\u5904\u7406: {:.3f}s, \u63a8\u7406: {:.3f}s, \u540e\u5904\u7406: {:.3f}s, \u8bc6\u522b{}\u4e2a\u76ee\u6807'.format(i+1, len(os.listdir(args.source)), time.time() – t1, pre_time, det_time, post_time, len(boxes)))<\/p>\n

# Draw rectangles and polygons
\n im_canvas = img.copy()
\n # \u5728\u7ed8\u5236\u5faa\u73af\u91cc\u52a0\u5065\u58ee\u6027\u5224\u65ad
\n for (*box, conf, cls_), segment in zip(boxes, segments):
\n # segment \u53ef\u80fd\u4e3a\u7a7a\u6216\u5f88\u77ed；\u4e5f\u53ef\u80fd\u662f float，\u9700\u8981\u8f6c int32
\n if segment is None or len(segment) < 3:
\n continue
\n seg = np.round(segment).astype(np.int32).reshape(-1, 1, 2) # -> Nx1x2, int32<\/p>\n

# \u5148\u753b\u8fb9，\u518d\u586b\u5145
\n cv2.polylines(img, [seg], True, (255, 255, 255), 2)
\n cv2.fillPoly(im_canvas, [seg], (255, 0, 0))<\/p>\n

# \u753b bbox \u548c\u6807\u7b7e
\n x1, y1, x2, y2 = map(int, box[:4])
\n cls_i = int(cls_)
\n cv2.rectangle(img, (x1, y1), (x2, y2), color_palette[cls_i], 1, cv2.LINE_AA)
\n label = args.classes[cls_i] if 0 <= cls_i < len(args.classes) else f'cls{cls_i}'
\n cv2.putText(img, f'{label}: {conf:.3f}', (x1, max(0, y1 – 9)),
\n cv2.FONT_HERSHEY_SIMPLEX, 0.7, color_palette[cls_i], 2, cv2.LINE_AA)<\/p>\n

# Mix image
\n img = cv2.addWeighted(im_canvas, 0.3, img, 0.7, 0)<\/p>\n

cv2.imwrite(os.path.join(args.out_path, img_name), img)<\/p>\n

except Exception as e:
\n print(e)<\/p>\n

5.\u542f\u52a8\u547d\u4ee4\u00a0<\/h3>\n

python ~\/autodl-tmp\/ultralytics-main\/run_seg.py \\\\
\n–seg_model \/root\/autodl-tmp\/ultralytics-main\/yolo11n-seg1.om \\\\
\n–source \/root\/autodl-tmp\/ultralytics-main\/test \\\\
\n–classes [person] \\\\
\n–postprocess_type v8 <\/p>\n

–classes \u662f\u6a21\u578b\u68c0\u6d4b\u6807\u7b7e\u540d<\/p>\n

–source \u63a8\u7406\u6587\u4ef6\u6240\u5728\u6587\u4ef6\u5939<\/p>\n

–det_model om\u6a21\u578b\u53c2\u6570\u8def\u5f84<\/p>\n

–postprocess_type yolov8 v9 v11\u90fd\u7528v8，\u4f7f\u7528\u7684yolov5\u7684\u8bdd\u6539\u4e3ayolov5<\/p>\n

\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb351\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86\u5728\u6607\u817eAI\u5904\u7406\u5668\u4e0a\u90e8\u7f72YOLO\u7cfb\u5217\u6a21\u578b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u9996\u5148\u6982\u8ff0\u4e86\u534e\u4e3a\u6607\u817e\u82af\u7247\uff08Ascend 310\/910\/910C\uff09\u548cMindSpore\u6846\u67b6\u7684\u7279\u6027\uff0c\u4ee5\u53caCANN\u8ba1\u7b97\u67b6\u6784\u7684\u4f5c\u7528\u3002\u7136\u540e\u8be6\u7ec6\u8bf4\u660e\u4e86\u73af\u5883\u914d\u7f6e\u6b65\u9aa4\uff1a1)\u521b\u5efaconda\u73af\u5883\u5e76\u5b89\u88c5\u4f9d\u8d56\u5e93\uff1b2)\u51c6\u5907YOLO\u6e90\u7801\u548c\u6a21\u578b\u6743\u91cd\uff1b3)\u6a21\u578b\u8f6c\u6362\u8fc7\u7a0b\uff08.pt\u2192.onnx\u2192.om\uff09\uff1b4)\u7f16\u5199\u63a8\u7406\u4ee3\u7801\uff0c\u5305\u62ec\u9884\u5904\u7406\u3001\u63a8\u7406\u548c\u540e\u5904\u7406\u6a21\u5757\uff1b5)\u8fd0\u884c\u63a8\u7406\u547d\u4ee4\u3002\u91cd\u70b9\u4ecb\u7ecd\u4e86\u5982\u4f55\u5229\u7528ais_bench\u5de5\u5177\u8fdb\u884c\u9ad8\u6027\u80fd\u63a8\u7406\uff0c\u5e76\u5904\u7406\u4e86\u53ef\u80fd\u51fa\u73b0\u7684\u73af\u5883\u914d\u7f6e\u95ee\u9898\u3002\u8be5\u65b9\u6848\u5b9e\u73b0\u4e86YOLO\u6a21\u578b\u5728\u6607\u817e<\/p>\n","protected":false},"author":2,"featured_media":52956,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[156,50,3829,86,1665,427,523],"topic":[],"class_list":["post-52960","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-yolo","tag-50","tag-3829","tag-86","tag-1665","tag-427","tag-523"],"yoast_head":"\nAutudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/52960.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb351\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86\u5728\u6607\u817eAI\u5904\u7406\u5668\u4e0a\u90e8\u7f72YOLO\u7cfb\u5217\u6a21\u578b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u9996\u5148\u6982\u8ff0\u4e86\u534e\u4e3a\u6607\u817e\u82af\u7247\uff08Ascend 310\/910\/910C\uff09\u548cMindSpore\u6846\u67b6\u7684\u7279\u6027\uff0c\u4ee5\u53caCANN\u8ba1\u7b97\u67b6\u6784\u7684\u4f5c\u7528\u3002\u7136\u540e\u8be6\u7ec6\u8bf4\u660e\u4e86\u73af\u5883\u914d\u7f6e\u6b65\u9aa4\uff1a1)\u521b\u5efaconda\u73af\u5883\u5e76\u5b89\u88c5\u4f9d\u8d56\u5e93\uff1b2)\u51c6\u5907YOLO\u6e90\u7801\u548c\u6a21\u578b\u6743\u91cd\uff1b3)\u6a21\u578b\u8f6c\u6362\u8fc7\u7a0b\uff08.pt\u2192.onnx\u2192.om\uff09\uff1b4)\u7f16\u5199\u63a8\u7406\u4ee3\u7801\uff0c\u5305\u62ec\u9884\u5904\u7406\u3001\u63a8\u7406\u548c\u540e\u5904\u7406\u6a21\u5757\uff1b5)\u8fd0\u884c\u63a8\u7406\u547d\u4ee4\u3002\u91cd\u70b9\u4ecb\u7ecd\u4e86\u5982\u4f55\u5229\u7528ais_bench\u5de5\u5177\u8fdb\u884c\u9ad8\u6027\u80fd\u63a8\u7406\uff0c\u5e76\u5904\u7406\u4e86\u53ef\u80fd\u51fa\u73b0\u7684\u73af\u5883\u914d\u7f6e\u95ee\u9898\u3002\u8be5\u65b9\u6848\u5b9e\u73b0\u4e86YOLO\u6a21\u578b\u5728\u6607\u817e\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/52960.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2025-08-11T13:59:18+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250811135915-6899f7339d14e.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"13 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/52960.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/52960.html\",\"name\":\"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2025-08-11T13:59:18+00:00\",\"dateModified\":\"2025-08-11T13:59:18+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/52960.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/52960.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/52960.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/52960.html","og_locale":"zh_CN","og_type":"article","og_title":"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb351\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86\u5728\u6607\u817eAI\u5904\u7406\u5668\u4e0a\u90e8\u7f72YOLO\u7cfb\u5217\u6a21\u578b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u9996\u5148\u6982\u8ff0\u4e86\u534e\u4e3a\u6607\u817e\u82af\u7247\uff08Ascend 310\/910\/910C\uff09\u548cMindSpore\u6846\u67b6\u7684\u7279\u6027\uff0c\u4ee5\u53caCANN\u8ba1\u7b97\u67b6\u6784\u7684\u4f5c\u7528\u3002\u7136\u540e\u8be6\u7ec6\u8bf4\u660e\u4e86\u73af\u5883\u914d\u7f6e\u6b65\u9aa4\uff1a1)\u521b\u5efaconda\u73af\u5883\u5e76\u5b89\u88c5\u4f9d\u8d56\u5e93\uff1b2)\u51c6\u5907YOLO\u6e90\u7801\u548c\u6a21\u578b\u6743\u91cd\uff1b3)\u6a21\u578b\u8f6c\u6362\u8fc7\u7a0b\uff08.pt\u2192.onnx\u2192.om\uff09\uff1b4)\u7f16\u5199\u63a8\u7406\u4ee3\u7801\uff0c\u5305\u62ec\u9884\u5904\u7406\u3001\u63a8\u7406\u548c\u540e\u5904\u7406\u6a21\u5757\uff1b5)\u8fd0\u884c\u63a8\u7406\u547d\u4ee4\u3002\u91cd\u70b9\u4ecb\u7ecd\u4e86\u5982\u4f55\u5229\u7528ais_bench\u5de5\u5177\u8fdb\u884c\u9ad8\u6027\u80fd\u63a8\u7406\uff0c\u5e76\u5904\u7406\u4e86\u53ef\u80fd\u51fa\u73b0\u7684\u73af\u5883\u914d\u7f6e\u95ee\u9898\u3002\u8be5\u65b9\u6848\u5b9e\u73b0\u4e86YOLO\u6a21\u578b\u5728\u6607\u817e","og_url":"https:\/\/www.wsisp.com\/helps\/52960.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2025-08-11T13:59:18+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250811135915-6899f7339d14e.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"13 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/52960.html","url":"https:\/\/www.wsisp.com\/helps\/52960.html","name":"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2025-08-11T13:59:18+00:00","dateModified":"2025-08-11T13:59:18+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/52960.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/52960.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/52960.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Autudl\u534e\u4e3a\u6607\u817e\u7cfb\u5217NPU\u7b80\u4ecb\u548c\u90e8\u7f72\u63a8\u7406yolo11 yolov8 yolov5\u5206\u5272\u6a21\u578b"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/52960","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=52960"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/52960\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/52956"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=52960"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=52960"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=52960"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=52960"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}