1、自然语言处理中的语义理解：从 BERT 到 LLaMA 的上下文建模技术cid:link_12、AI模型的边缘部署：TensorRT与ONNX Runtime的优化流程与性能调优cid:link_23、生成式AI服务的高并发架构设计：缓存策略与动态资源调度实践cid:link_34、稀疏Transformer的设计与应用：降低计算复杂度的核心技术解析cid:link_05、从卖工具到卖效果：RaaS如何重塑To B市场格局？cid:link_46、运营商 4A 安全平台：刚需逻辑与资深适配厂商全景分析cid:link_57、边缘安全访问技术：守护数字边界的新型安全屏障cid:link_68、Pi0 网络架构cid:link_79、开源英文图文数据集介绍cid:link_810、智能驾驶的 “最后一公里”：数据闭环与城市复杂场景的攻克cid:link_911、联邦学习：在数据隐私保护下，AI协同训练的实践与挑战cid:link_1012、一粒玉米种子获赔5000万元！种子“数字身份证”斩断侵权黑手cid:link_1113、私有云安全资源池：构建企业云端体系化防护的核心路径cid:link_12

OrangePi AI Studio Pro基于MindYolo实现YOLOv8模型训练及验证OrangePi AI Studio Pro是基于 2 个昇腾 310P 处理器的新一代高性能推理解析卡，提供基础通用算力+超强AI算力，整合了训练和推理的全部底层软件栈，实现训推一体。其中AI半精度FP16算力约为176TFLOPS，整数Int8精度可达352TOPS。本章将介绍如何在昇腾310上基于mindyolo实现YOLOv8模型的训练及验证。一、环境准备首先检查昇腾310P的NPU驱动，在命令行中输入：npu-smi info，可以看到两块昇腾310P的AICore的利用率和内存的占用情况。+--------------------------------------------------------------------------------------------------------+ | npu-smi v1.0 Version: 24.1.rc4.b999 | +-------------------------------+-----------------+------------------------------------------------------+ | NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page) | | Chip Device | Bus-Id | AICore(%) Memory-Usage(MB) | +===============================+=================+======================================================+ | 30208 310P1 | OK | NA 41 0 / 0 | | 0 0 | 0000:77:00.0 | 0 1416 / 89608 | +-------------------------------+-----------------+------------------------------------------------------+ | 30208 310P1 | OK | NA 40 0 / 0 | | 1 1 | 0000:77:00.0 | 0 1622 / 89085 | +===============================+=================+======================================================+ +-------------------------------+-----------------+------------------------------------------------------+ | NPU Chip | Process id | Process name | Process memory(MB) | +===============================+=================+======================================================+ | No running processes found in NPU 30208 | +===============================+=================+======================================================+之后升级CANN的版本以及更新MindSpore，可以参考我的另一篇文章：如何在OrangePi Studio Pro上升级CANN以及的Pytorch和MindSpore，升级完成后，检查MindSpore的安装情况，我使用的版本是2.7.0。source /usr/local/Ascend/ascend-toolkit/set_env.sh python3 -c "import mindspore;mindspore.set_context(device_target='Ascend');mindspore.run_check()" [WARNING] ME(1621400:139701939115840,MainProcess):2025-09-24-10:46:21.978.000 [mindspore/context.py:1412] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead. MindSpore version: 2.7.0 [WARNING] GE_ADPT(1621400,7f0e18710640,python3):2025-09-24-10:46:23.323.570 [mindspore/ops/kernel/ascend/acl_ir/op_api_exec.cc:169] GetAscendDefaultCustomPath] Checking whether the so exists or if permission to access it is available: /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize_vision/op_api/lib/libcust_opapi.so The result of multiplication calculation is correct, MindSpore has been installed on platform [Ascend] successfully! 克隆mindyolo仓库，我们使用由天津大学发布的无人机视觉挑战赛数据集VisDrone-Dataset进行模型的训练及验证。git clone https://github.com/mindspore-lab/mindyolo.git正克隆到 'mindyolo'... remote: Enumerating objects: 3505, done. remote: Counting objects: 100% (157/157), done. remote: Compressing objects: 100% (69/69), done. remote: Total 3505 (delta 114), reused 88 (delta 88), pack-reused 3348 (from 2) 接收对象中: 100% (3505/3505), 6.74 MiB | 8.91 MiB/s, 完成. 处理 delta 中: 100% (2048/2048), 完成.我们将下载后的数据集首先转换成YOLO格式，具体的转换教程可以参考网上的公开资料，经过转换后的visdrone数据集包括以下内容：visdrone ├── train │ ├── images │ │ ├── 000001.jpg │ │ ├── 000002.jpg │ │ ├── ... │ │ └── ... │ └── labels │ ├── 000001.txt │ ├── 000002.txt │ ├── ... │ └── ... └── val ├── images │ ├── 000001.jpg │ ├── 000002.jpg │ ├── ... │ └── ... └── labels ├── 000001.txt ├── 000001.txt ├── ... └── ... 二、数据格式转换由于mindyolo中的train过程使用的数据是yolo格式，而eval过程使用coco数据集中的json文件，因此需要再增加coco格式的标注文件instances_train2017.json、instances_val2017.json以及train.txt和val.txt文件，经过转换后的visdrone数据集包括以下内容：visdrone_COCO_format ├── train.txt ├── val.txt ├── train │ ├── images │ │ ├── 000001.jpg │ │ ├── 000002.jpg │ │ ├── ... │ │ └── ... │ └── labels │ ├── 000001.txt │ ├── 000002.txt │ ├── ... │ └── ... ├── annotations │ ├── instances_train2017.json │ └── instances_val2017.json └── val ├── images │ ├── 000001.jpg │ ├── 000002.jpg │ ├── ... │ └── ... └── labels ├── 000001.txt ├── 000001.txt ├── ... └── ... 我们先把YOLO格式的数据集转换为COCO格式，在mindyolo中实现yolov5_yaml_to_coco.py脚本，具体代码如下：# -*- encoding: utf-8 -*- # @Author: SWHL # @Contact: liekkaskono@163.com import argparse import glob import json import os import shutil import time from pathlib import Path import cv2 import yaml from tqdm import tqdm def read_txt(txt_path): with open(str(txt_path), "r", encoding="utf-8") as f: data = list(map(lambda x: x.rstrip("\n"), f)) return data def mkdir(dir_path): Path(dir_path).mkdir(parents=True, exist_ok=True) def verify_exists(file_path): file_path = Path(file_path).resolve() if not file_path.exists(): raise FileNotFoundError(f"The {file_path} is not exists!!!") class YOLOV5CFG2COCO: def __init__(self, yaml_path): verify_exists(yaml_path) with open(yaml_path, "r", encoding="UTF-8") as f: self.data_cfg = yaml.safe_load(f) self.root_dir = Path(yaml_path).parent.parent self.root_data_dir = Path(self.data_cfg.get("path")) self.train_path = self._get_data_dir("train") self.val_path = self._get_data_dir("val") nc = self.data_cfg["nc"] if "names" in self.data_cfg: self.names = self.data_cfg.get("names") else: # assign class names if missing self.names = [f"class{i}" for i in range(self.data_cfg["nc"])] assert ( len(self.names) == nc ), f"{len(self.names)} names found for nc={nc} dataset in {yaml_path}" # 构建COCO格式目录 self.dst = self.root_dir / f"{Path(self.root_data_dir).stem}_COCO_format" self.coco_train = "train/images" self.coco_val = "val/images" self.coco_annotation = "annotations" self.coco_train_json = ( self.dst / self.coco_annotation / f"instances_train2017.json" ) self.coco_val_json = ( self.dst / self.coco_annotation / f"instances_val2017.json" ) mkdir(self.dst) mkdir(self.dst / self.coco_train) mkdir(self.dst / self.coco_val) mkdir(self.dst / self.coco_annotation) # 构建json内容结构 self.type = "instances" self.categories = [] self._get_category() self.annotation_id = 1 cur_year = time.strftime("%Y", time.localtime(time.time())) self.info = { "year": int(cur_year), "version": "1.0", "description": "For object detection", "date_created": cur_year, } self.licenses = [ { "id": 1, "name": "Apache License v2.0", "url": "https://choosealicense.com/licenses/apache-2.0/", } ] def _get_data_dir(self, mode): data_dir = self.data_cfg.get(mode) if data_dir: if isinstance(data_dir, str): full_path = [str(self.root_data_dir / data_dir)] elif isinstance(data_dir, list): full_path = [str(self.root_data_dir / one_dir) for one_dir in data_dir] else: raise TypeError(f"{data_dir} is not str or list.") else: raise ValueError(f"{mode} dir is not in the yaml.") return full_path def _get_category(self): for i, category in enumerate(self.names, start=1): self.categories.append( { "supercategory": category, "id": i, "name": category, } ) def generate(self): self.train_files = self.get_files(self.train_path) self.valid_files = self.get_files(self.val_path) train_dest_dir = Path(self.dst) / self.coco_train self.gen_dataset( self.train_files, train_dest_dir, self.coco_train_json, mode="train" ) val_dest_dir = Path(self.dst) / self.coco_val self.gen_dataset(self.valid_files, val_dest_dir, self.coco_val_json, mode="val") print(f"The output directory is: {self.dst}") def get_files(self, path): IMG_FORMATS = ["bmp", "dng", "jpeg", "jpg", "mpo", "png", "tif", "tiff", "webp"] f = [] for p in path: p = Path(p) if p.is_dir(): f += glob.glob(str(p / "**" / "*.*"), recursive=True) elif p.is_file(): # file with open(p, "r", encoding="utf-8") as t: t = t.read().strip().splitlines() parent = str(p.parent) + os.sep f += [ x.replace("./", parent) if x.startswith("./") else x for x in t ] else: raise FileExistsError(f"{p} does not exist") im_files = sorted( x.replace("/", os.sep) for x in f if x.split(".")[-1].lower() in IMG_FORMATS ) return im_files def gen_dataset(self, img_paths, target_img_path, target_json, mode): """ https://cocodataset.org/#format-data """ images = [] annotations = [] sa, sb = ( os.sep + "images" + os.sep, os.sep + "labels" + os.sep, ) # /images/, /labels/ substrings for img_id, img_path in enumerate(tqdm(img_paths, desc=mode), 1): label_path = sb.join(img_path.rsplit(sa, 1)).rsplit(".", 1)[0] + ".txt" img_path = Path(img_path) verify_exists(img_path) imgsrc = cv2.imread(str(img_path)) height, width = imgsrc.shape[:2] dest_file_name = f"{img_id:012d}.jpg" save_img_path = target_img_path / dest_file_name if img_path.suffix.lower() == ".jpg": shutil.copyfile(img_path, save_img_path) else: cv2.imwrite(str(save_img_path), imgsrc) images.append( { "date_captured": "2021", "file_name": dest_file_name, "id": img_id, "height": height, "width": width, } ) if Path(label_path).exists(): new_anno = self.read_annotation(label_path, img_id, height, width) if len(new_anno) > 0: annotations.extend(new_anno) else: raise ValueError(f"{label_path} is empty") else: raise FileNotFoundError(f"{label_path} not exists") json_data = { "info": self.info, "images": images, "licenses": self.licenses, "type": self.type, "annotations": annotations, "categories": self.categories, } with open(target_json, "w", encoding="utf-8") as f: json.dump(json_data, f, ensure_ascii=False) def read_annotation(self, txt_file, img_id, height, width): annotation = [] all_info = read_txt(txt_file) for label_info in all_info: # 遍历一张图中不同标注对象 label_info = label_info.split(" ") if len(label_info) < 5: continue category_id, vertex_info = label_info[0], label_info[1:] segmentation, bbox, area = self._get_annotation(vertex_info, height, width) annotation.append( { "segmentation": segmentation, "area": area, "iscrowd": 0, "image_id": img_id, "bbox": bbox, "category_id": int(category_id) + 1, "id": self.annotation_id, } ) self.annotation_id += 1 return annotation @staticmethod def _get_annotation(vertex_info, height, width): cx, cy, w, h = [float(i) for i in vertex_info] cx = cx * width cy = cy * height box_w = w * width box_h = h * height x0 = max(cx - box_w / 2, 0) y0 = max(cy - box_h / 2, 0) x1 = min(x0 + box_w, width) y1 = min(y0 + box_h, height) segmentation = [[x0, y0, x1, y0, x1, y1, x0, y1]] bbox = [x0, y0, box_w, box_h] area = box_w * box_h return segmentation, bbox, area def main(): parser = argparse.ArgumentParser("Datasets converter from YOLOV5 to COCO") parser.add_argument( "--yaml_path", type=str, default="dataset/YOLOV5_yaml/sample.yaml", help="Dataset cfg file", ) args = parser.parse_args() converter = YOLOV5CFG2COCO(args.yaml_path) converter.generate() if __name__ == "__main__": main() 之后在mindyolo目录下创建YOLO格式的配置文件visdrone.yaml：# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] path: /root/workspace/dataset/visdrone # dataset root dir (absolute path) train: train/images # train images (relative to 'path') val: val/images # val images (relative to 'path') test: # test images (optional) nc: 12 # Classes，类别 names: 0: ignored regions 1: pedestrian 2: people 3: bicycle 4: car 5: van 6: truck 7: tricycle 8: awning-tricycle 9: bus 10: motor 11: others在终端中运行如下命令将YOLO格式的数据集转换为COCO格式：python3 yolov5_yaml_to_coco.py --yaml_path visdrone.yamltrain: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6471/6471 [01:13<00:00, 88.07it/s] val: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 548/548 [00:03<00:00, 148.22it/s] The output directory is: visdrone_COCO_format再创建coco2yolo.py的Python脚本将COCO格式的标注文件.json导出为labels文件夹中YOLO格式的标注文件.txt：import json import os import argparse parser = argparse.ArgumentParser(description='Test yolo data.') parser.add_argument('-j', help='JSON file', dest='json', required=True) parser.add_argument('-o', help='path to output folder', dest='out',required=True) args = parser.parse_args() json_file = args.json output = args.out class COCO2YOLO: def __init__(self): self._check_file_and_dir(json_file, output) self.labels = json.load(open(json_file, 'r', encoding='utf-8')) self.coco_id_name_map = self._categories() self.coco_name_list = list(self.coco_id_name_map.values()) print("total images", len(self.labels['images'])) print("total categories", len(self.labels['categories'])) print("total labels", len(self.labels['annotations'])) def _check_file_and_dir(self, file_path, dir_path): if not os.path.exists(file_path): raise ValueError("file not found") if not os.path.exists(dir_path): os.makedirs(dir_path) def _categories(self): categories = {} for cls in self.labels['categories']: categories[cls['id']] = cls['name'] return categories def _load_images_info(self): images_info = {} for image in self.labels['images']: id = image['id'] file_name = image['file_name'] if file_name.find('\\') > -1: file_name = file_name[file_name.index('\\')+1:] w = image['width'] h = image['height'] images_info[id] = (file_name, w, h) return images_info def _bbox_2_yolo(self, bbox, img_w, img_h): x, y, w, h = bbox[0], bbox[1], bbox[2], bbox[3] centerx = bbox[0] + w / 2 centery = bbox[1] + h / 2 dw = 1 / img_w dh = 1 / img_h centerx *= dw w *= dw centery *= dh h *= dh return centerx, centery, w, h def _convert_anno(self, images_info): anno_dict = dict() for anno in self.labels['annotations']: bbox = anno['bbox'] image_id = anno['image_id'] category_id = anno['category_id'] image_info = images_info.get(image_id) image_name = image_info[0] img_w = image_info[1] img_h = image_info[2] yolo_box = self._bbox_2_yolo(bbox, img_w, img_h) anno_info = (image_name, category_id, yolo_box) anno_infos = anno_dict.get(image_id) if not anno_infos: anno_dict[image_id] = [anno_info] else: anno_infos.append(anno_info) anno_dict[image_id] = anno_infos return anno_dict def save_classes(self): sorted_classes = list(map(lambda x: x['name'], sorted(self.labels['categories'], key=lambda x: x['id']))) print('coco names', sorted_classes) with open('coco.names', 'w', encoding='utf-8') as f: for cls in sorted_classes: f.write(cls + '\n') f.close() def coco2yolo(self): print("loading image info...") images_info = self._load_images_info() print("loading done, total images", len(images_info)) print("start converting...") anno_dict = self._convert_anno(images_info) print("converting done, total labels", len(anno_dict)) print("saving txt file...") self._save_txt(anno_dict) print("saving done") def _save_txt(self, anno_dict): for k, v in anno_dict.items(): file_name = os.path.splitext(v[0][0])[0] + ".txt" with open(os.path.join(output, file_name), 'w', encoding='utf-8') as f: print(k, v) for obj in v: cat_name = self.coco_id_name_map.get(obj[1]) category_id = self.coco_name_list.index(cat_name) box = ['{:.6f}'.format(x) for x in obj[2]] box = ' '.join(box) line = str(category_id) + ' ' + box f.write(line + '\n') if __name__ == '__main__': c2y = COCO2YOLO() c2y.coco2yolo() 在终端中切换到mindyolo目录下依次运行如下命令导出instances_train2017.json和instances_val2017.json文件对应的YOLO格式的标注文件到labels文件夹中：python3 coco2yolo.py -j ./visdrone_COCO_format/annotations/instances_train2017.json -o ./visdrone_COCO_format/train/labelspython3 coco2yolo.py -j ./visdrone_COCO_format/annotations/instances_val2017.json -o ./visdrone_COCO_format/val/labels最后创建generate_txt.sh脚本在COCO数据集目录下生成train.txt和val.txt，指定训练图片和验证图片的在数据集中的相对路径：#!/bin/bash # 检查是否提供了数据集路径参数 if [ $# -eq 0 ]; then echo "Usage: $0 <dataset_path>" echo "Example: $0 /path/to/visdrone" exit 1 fi # 获取数据集路径 DATASET_PATH="$1" # 检查数据集路径是否存在 if [ ! -d "$DATASET_PATH" ]; then echo "Error: Dataset path '$DATASET_PATH' does not exist." exit 1 fi # 定义训练和验证图片目录 TRAIN_DIR="$DATASET_PATH/train/images" VAL_DIR="$DATASET_PATH/val/images" # 检查训练和验证目录是否存在 if [ ! -d "$TRAIN_DIR" ]; then echo "Error: Train directory '$TRAIN_DIR' does not exist." exit 1 fi if [ ! -d "$VAL_DIR" ]; then echo "Error: Validation directory '$VAL_DIR' does not exist." exit 1 fi # 生成 train.txt TRAIN_TXT="$DATASET_PATH/train.txt" ls "$TRAIN_DIR" | grep '\.jpg$' | sort | sed 's/^/\.\/train\/images\//' > "$TRAIN_TXT" echo "Generated $TRAIN_TXT" # 生成 val.txt VAL_TXT="$DATASET_PATH/val.txt" ls "$VAL_DIR" | grep '\.jpg$' | sort | sed 's/^/\.\/val\/images\//' > "$VAL_TXT" echo "Generated $VAL_TXT" echo "Successfully generated train.txt and val.txt in $DATASET_PATH" 在终端中运行generate_txt.sh，并传入前面COCO数据集的路径：chmod +x generate_txt.sh ./generate_txt.sh visdrone_COCO_formatGenerated visdrone_COCO_format/train.txt Generated visdrone_COCO_format/val.txt Successfully generated train.txt and val.txt in visdrone_COCO_format最终生成的visdrone_COCO_format数据集的格式如下，可以直接用于MindYOLOv8模型的训练：visdrone_COCO_format ├── train.txt ├── val.txt ├── train │ ├── images │ │ ├── 000001.jpg │ │ ├── 000002.jpg │ │ ├── ... │ │ └── ... │ └── labels │ ├── 000001.txt │ ├── 000002.txt │ ├── ... │ └── ... ├── annotations │ ├── instances_train2017.json │ └── instances_val2017.json └── val ├── images │ ├── 000001.jpg │ ├── 000002.jpg │ ├── ... │ └── ... └── labels ├── 000001.txt ├── 000001.txt ├── ... └── ... 三、模型训练MindYOLO支持yaml文件继承机制，因此新编写的配置文件只需要继承MindYOLO提供的原生yaml文件现有配置文件：在configs目录下编写MindYOLO数据集的yaml配置文件，指定训练图片和验证图片的路径以及模型的类别标签：data: dataset_name: visdrone_COCO_format train_set: /root/workspace/mindyolo/visdrone_COCO_format/train.txt val_set: /root/workspace/mindyolo/visdrone_COCO_format/val.txt test_set: /root/workspace/mindyolo/visdrone_COCO_format/val.txt nc: 12 # class names names: ['ignored regions', 'pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor', 'others' ] train_transforms: [] test_transforms: [] 修改configs/yolov8s.yaml文件，注释掉原有的coco.yaml配置文件，指定我们自己的数据集，同时添加epochs、img_size、per_batch_size、multi-stage data augment等自定义训练参数：__BASE__: [ # '../coco.yaml', '../visdrone.yaml', './hyp.scratch.low.yaml', './yolov8-base.yaml' ] overflow_still_update: False network: depth_multiple: 0.33 # scales module repeats width_multiple: 0.50 # scales convolution channels max_channels: 1024 epochs: 10 img_size: 1024 per_batch_size: 16 data: num_parallel_workers: 8 # multi-stage data augment train_transforms: { stage_epochs: [ 5, 5 ], trans_list: [ [ { func_name: mosaic, prob: 1.0 }, { func_name: resample_segments }, { func_name: random_perspective, prob: 1.0, degrees: 0.0, translate: 0.1, scale: 0.5, shear: 0.0 }, {func_name: albumentations}, {func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4}, {func_name: fliplr, prob: 0.5}, {func_name: label_norm, xyxy2xywh_: True}, {func_name: label_pad, padding_size: 160, padding_value: -1}, {func_name: image_norm, scale: 255.}, {func_name: image_transpose, bgr2rgb: True, hwc2chw: True} ], [ {func_name: letterbox, scaleup: True}, {func_name: resample_segments}, {func_name: random_perspective, prob: 1.0, degrees: 0.0, translate: 0.1, scale: 0.5, shear: 0.0}, {func_name: albumentations}, {func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4}, {func_name: fliplr, prob: 0.5}, {func_name: label_norm, xyxy2xywh_: True}, {func_name: label_pad, padding_size: 160, padding_value: -1}, {func_name: image_norm, scale: 255.}, {func_name: image_transpose, bgr2rgb: True, hwc2chw: True} ]] } test_transforms: [ {func_name: letterbox, scaleup: False, only_image: True}, {func_name: image_norm, scale: 255.}, {func_name: image_transpose, bgr2rgb: True, hwc2chw: True} ] 在终端中运行train.py进行模型训练，指定模型的配置文件以及使用昇腾NPU：python3 train.py --config ./configs/yolov8/yolov8s.yaml --device_target Ascend默认是跑在0卡上也可以在环境变量中指定DEVICE_ID让模型的训练代码跑在1卡上：import os os.setenv("DEVICE_ID", 1) 如果不想设置环境变量也可以修改mindyolo\mindyolo\utils\utils.py中默认的参数：import os import random import yaml import cv2 from datetime import datetime import numpy as np import mindspore as ms from mindspore import ops, Tensor, nn from mindspore.communication.management import get_group_size, get_rank, init from mindspore import ParallelMode from mindyolo.utils import logger def set_seed(seed=2): np.random.seed(seed) random.seed(seed) ms.set_seed(seed) def set_default(args): # Set Context ms.set_context(mode=args.ms_mode) ms.set_recursion_limit(args.max_call_depth) if args.ms_mode == 0: ms.set_context(jit_config={"jit_level": "O2"}) if args.device_target == "Ascend": ms.set_device("Ascend", int(os.getenv("DEVICE_ID", 1))) ... 2025-10-23 14:48:02,364 [INFO] parse_args: 2025-10-23 14:48:02,364 [INFO] task detect 2025-10-23 14:48:02,364 [INFO] device_target Ascend 2025-10-23 14:48:02,364 [INFO] save_dir ./runs/2025.10.23-14.48.02 2025-10-23 14:48:02,364 [INFO] log_level INFO 2025-10-23 14:48:02,364 [INFO] is_parallel False 2025-10-23 14:48:02,364 [INFO] ms_mode 0 2025-10-23 14:48:02,364 [INFO] max_call_depth 2000 2025-10-23 14:48:02,364 [INFO] ms_amp_level O0 2025-10-23 14:48:02,364 [INFO] keep_loss_fp32 True 2025-10-23 14:48:02,364 [INFO] anchor_base False 2025-10-23 14:48:02,364 [INFO] ms_loss_scaler static 2025-10-23 14:48:02,364 [INFO] ms_loss_scaler_value 1024.0 2025-10-23 14:48:02,364 [INFO] ms_jit True 2025-10-23 14:48:02,364 [INFO] ms_enable_graph_kernel False 2025-10-23 14:48:02,364 [INFO] ms_datasink False 2025-10-23 14:48:02,364 [INFO] overflow_still_update False 2025-10-23 14:48:02,364 [INFO] clip_grad False 2025-10-23 14:48:02,364 [INFO] clip_grad_value 10.0 2025-10-23 14:48:02,364 [INFO] ema True 2025-10-23 14:48:02,364 [INFO] weight 2025-10-23 14:48:02,364 [INFO] ema_weight 2025-10-23 14:48:02,364 [INFO] freeze [] 2025-10-23 14:48:02,364 [INFO] epochs 10 2025-10-23 14:48:02,364 [INFO] per_batch_size 16 2025-10-23 14:48:02,364 [INFO] img_size 1024 2025-10-23 14:48:02,364 [INFO] nbs 64 2025-10-23 14:48:02,364 [INFO] accumulate 1 2025-10-23 14:48:02,364 [INFO] auto_accumulate False 2025-10-23 14:48:02,364 [INFO] log_interval 100 2025-10-23 14:48:02,364 [INFO] single_cls False 2025-10-23 14:48:02,364 [INFO] sync_bn False 2025-10-23 14:48:02,364 [INFO] keep_checkpoint_max 100 2025-10-23 14:48:02,364 [INFO] run_eval False 2025-10-23 14:48:02,364 [INFO] run_eval_interval 1 2025-10-23 14:48:02,364 [INFO] conf_thres 0.001 2025-10-23 14:48:02,364 [INFO] iou_thres 0.7 2025-10-23 14:48:02,364 [INFO] conf_free True 2025-10-23 14:48:02,364 [INFO] rect False 2025-10-23 14:48:02,364 [INFO] nms_time_limit 20.0 2025-10-23 14:48:02,364 [INFO] recompute False 2025-10-23 14:48:02,364 [INFO] recompute_layers 0 2025-10-23 14:48:02,364 [INFO] seed 2 2025-10-23 14:48:02,364 [INFO] summary True 2025-10-23 14:48:02,364 [INFO] profiler False 2025-10-23 14:48:02,364 [INFO] profiler_step_num 1 2025-10-23 14:48:02,364 [INFO] opencv_threads_num 0 2025-10-23 14:48:02,364 [INFO] strict_load True 2025-10-23 14:48:02,364 [INFO] enable_modelarts False 2025-10-23 14:48:02,364 [INFO] data_url 2025-10-23 14:48:02,364 [INFO] ckpt_url 2025-10-23 14:48:02,364 [INFO] multi_data_url 2025-10-23 14:48:02,364 [INFO] pretrain_url 2025-10-23 14:48:02,364 [INFO] train_url 2025-10-23 14:48:02,364 [INFO] data_dir /cache/data/ 2025-10-23 14:48:02,364 [INFO] ckpt_dir /cache/pretrain_ckpt/ 2025-10-23 14:48:02,364 [INFO] data.dataset_name result 2025-10-23 14:48:02,364 [INFO] data.train_set /root/workspace/mindyolo/visdrone_COCO_format/train.txt 2025-10-23 14:48:02,364 [INFO] data.val_set /root/workspace/mindyolo/visdrone_COCO_format/val.txt 2025-10-23 14:48:02,364 [INFO] data.test_set /root/workspace/mindyolo/visdrone_COCO_format/val.txt 2025-10-23 14:48:02,364 [INFO] data.nc 12 2025-10-23 14:48:02,364 [INFO] data.names ['ignored regions', 'pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor', 'others'] 2025-10-23 14:48:02,364 [INFO] train_transforms.stage_epochs [5, 5] 2025-10-23 14:48:02,364 [INFO] train_transforms.trans_list [[{'func_name': 'mosaic', 'prob': 1.0}, {'func_name': 'resample_segments'}, {'func_name': 'random_perspective', 'prob': 1.0, 'degrees': 0.0, 'translate': 0.1, 'scale': 0.5, 'shear': 0.0}, {'func_name': 'albumentations'}, {'func_name': 'hsv_augment', 'prob': 1.0, 'hgain': 0.015, 'sgain': 0.7, 'vgain': 0.4}, {'func_name': 'fliplr', 'prob': 0.5}, {'func_name': 'label_norm', 'xyxy2xywh_': True}, {'func_name': 'label_pad', 'padding_size': 160, 'padding_value': -1}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}], [{'func_name': 'letterbox', 'scaleup': True}, {'func_name': 'resample_segments'}, {'func_name': 'random_perspective', 'prob': 1.0, 'degrees': 0.0, 'translate': 0.1, 'scale': 0.5, 'shear': 0.0}, {'func_name': 'albumentations'}, {'func_name': 'hsv_augment', 'prob': 1.0, 'hgain': 0.015, 'sgain': 0.7, 'vgain': 0.4}, {'func_name': 'fliplr', 'prob': 0.5}, {'func_name': 'label_norm', 'xyxy2xywh_': True}, {'func_name': 'label_pad', 'padding_size': 160, 'padding_value': -1}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}]] 2025-10-23 14:48:02,364 [INFO] data.test_transforms [{'func_name': 'letterbox', 'scaleup': False, 'only_image': True}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}] 2025-10-23 14:48:02,364 [INFO] data.num_parallel_workers 8 2025-10-23 14:48:02,364 [INFO] optimizer.optimizer momentum 2025-10-23 14:48:02,364 [INFO] optimizer.lr_init 0.01 2025-10-23 14:48:02,364 [INFO] optimizer.momentum 0.937 2025-10-23 14:48:02,364 [INFO] optimizer.nesterov True 2025-10-23 14:48:02,364 [INFO] optimizer.loss_scale 1.0 2025-10-23 14:48:02,364 [INFO] optimizer.warmup_epochs 3 2025-10-23 14:48:02,364 [INFO] optimizer.warmup_momentum 0.8 2025-10-23 14:48:02,364 [INFO] optimizer.warmup_bias_lr 0.1 2025-10-23 14:48:02,364 [INFO] optimizer.min_warmup_step 1000 2025-10-23 14:48:02,364 [INFO] optimizer.group_param yolov8 2025-10-23 14:48:02,364 [INFO] optimizer.gp_weight_decay 0.0005 2025-10-23 14:48:02,364 [INFO] optimizer.start_factor 1.0 2025-10-23 14:48:02,364 [INFO] optimizer.end_factor 0.01 2025-10-23 14:48:02,364 [INFO] optimizer.epochs 10 2025-10-23 14:48:02,364 [INFO] optimizer.nbs 64 2025-10-23 14:48:02,364 [INFO] optimizer.accumulate 1 2025-10-23 14:48:02,364 [INFO] optimizer.total_batch_size 16 2025-10-23 14:48:02,364 [INFO] loss.name YOLOv8Loss 2025-10-23 14:48:02,364 [INFO] loss.box 7.5 2025-10-23 14:48:02,364 [INFO] loss.cls 0.5 2025-10-23 14:48:02,364 [INFO] loss.dfl 1.5 2025-10-23 14:48:02,364 [INFO] loss.reg_max 16 2025-10-23 14:48:02,364 [INFO] network.model_name yolov8 2025-10-23 14:48:02,364 [INFO] network.nc 80 2025-10-23 14:48:02,364 [INFO] network.reg_max 16 2025-10-23 14:48:02,364 [INFO] network.stride [8, 16, 32] 2025-10-23 14:48:02,364 [INFO] network.backbone [[-1, 1, 'ConvNormAct', [64, 3, 2]], [-1, 1, 'ConvNormAct', [128, 3, 2]], [-1, 3, 'C2f', [128, True]], [-1, 1, 'ConvNormAct', [256, 3, 2]], [-1, 6, 'C2f', [256, True]], [-1, 1, 'ConvNormAct', [512, 3, 2]], [-1, 6, 'C2f', [512, True]], [-1, 1, 'ConvNormAct', [1024, 3, 2]], [-1, 3, 'C2f', [1024, True]], [-1, 1, 'SPPF', [1024, 5]]] 2025-10-23 14:48:02,364 [INFO] network.head [[-1, 1, 'Upsample', ['None', 2, 'nearest']], [[-1, 6], 1, 'Concat', [1]], [-1, 3, 'C2f', [512]], [-1, 1, 'Upsample', ['None', 2, 'nearest']], [[-1, 4], 1, 'Concat', [1]], [-1, 3, 'C2f', [256]], [-1, 1, 'ConvNormAct', [256, 3, 2]], [[-1, 12], 1, 'Concat', [1]], [-1, 3, 'C2f', [512]], [-1, 1, 'ConvNormAct', [512, 3, 2]], [[-1, 9], 1, 'Concat', [1]], [-1, 3, 'C2f', [1024]], [[15, 18, 21], 1, 'YOLOv8Head', ['nc', 'reg_max', 'stride']]] 2025-10-23 14:48:02,364 [INFO] network.depth_multiple 0.33 2025-10-23 14:48:02,364 [INFO] network.width_multiple 0.5 2025-10-23 14:48:02,364 [INFO] network.max_channels 1024 2025-10-23 14:48:02,364 [INFO] config ./configs/yolov8/yolov8s.yaml 2025-10-23 14:48:02,364 [INFO] rank 0 2025-10-23 14:48:02,364 [INFO] rank_size 1 2025-10-23 14:48:02,364 [INFO] total_batch_size 16 2025-10-23 14:48:02,364 [INFO] callback [] 2025-10-23 14:48:02,364 [INFO] 2025-10-23 14:48:02,365 [INFO] Please check the above information for the configurations 2025-10-23 14:48:02,441 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 14:48:02,451 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 14:48:02,528 [INFO] number of network params, total: 11.160279M, trainable: 11.140228M [WARNING] GE_ADPT(336686,7ff4350e8740,python3):2025-10-23-14:48:13.472.732 [mindspore/ops/kernel/ascend/acl_ir/op_api_exec.cc:169] GetAscendDefaultCustomPath] Checking whether the so exists or if permission to access it is available: /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize_vision/op_api/lib/libcust_opapi.so 2025-10-23 14:48:14,547 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 14:48:14,558 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 14:48:14,646 [INFO] number of network params, total: 11.160279M, trainable: 11.140228M .2025-10-23 14:48:30,416 [INFO] ema_weight not exist, default pretrain weight is currently used. 2025-10-23 14:48:30,421 [INFO] No dataset cache available, caching now... Scanning images: 0%| | 0/6471 [00:00<?, ?it/s]WARNING ⚠️ /root/workspace/mindyolo/visdrone_COCO_format/train/images/000000000335.jpg: 1 duplicate labels removed Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache' images and labels... 397 found, 0 missing, 0 empty, 0 corrupted: 6%|████ | 397/6471 [00:00<00:01, 3960.91it/s]WARNING ⚠️ /root/workspace/mindyolo/visdrone_COCO_format/train/images/000000000427.jpg: 1 duplicate labels removed Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache' images and labels... 1261 found, 0 missing, 0 empty, 0 corrupted: 19%|████████████▋ | 1261/6471 [00:00<00:01, 4238.38it/s]WARNING ⚠️ /root/workspace/mindyolo/visdrone_COCO_format/train/images/000000001492.jpg: 1 duplicate labels removed Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache' images and labels... 3866 found, 0 missing, 0 empty, 0 corrupted: 60%|██████████████████████████████████████▊ | 3866/6471 [00:00<00:00, 4332.85it/s]WARNING ⚠️ /root/workspace/mindyolo/visdrone_COCO_format/train/images/000000003868.jpg: 1 duplicate labels removed Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache' images and labels... 5607 found, 0 missing, 0 empty, 0 corrupted: 87%|████████████████████████████████████████████████████████▎ | 5607/6471 [00:01<00:00, 4337.04it/s]WARNING ⚠️ /root/workspace/mindyolo/visdrone_COCO_format/train/images/000000005742.jpg: 1 duplicate labels removed Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache' images and labels... 6471 found, 0 missing, 0 empty, 0 corrupted: 100%|█████████████████████████████████████████████████████████████████| 6471/6471 [00:01<00:00, 4307.45it/s] 2025-10-23 14:48:32,028 [INFO] New cache created: /root/workspace/mindyolo/visdrone_COCO_format/train.cache.npy 2025-10-23 14:48:32,029 [INFO] Dataset caching success. 2025-10-23 14:48:32,051 [INFO] Dataloader num parallel workers: [8] 2025-10-23 14:48:32,135 [INFO] Dataset Cache file hash/version check success. 2025-10-23 14:48:32,135 [INFO] Load dataset cache from [/root/workspace/mindyolo/visdrone_COCO_format/train.cache.npy] success. Scanning '/root/workspace/mindyolo/visdrone_COCO_format/train.cache.npy' images and labels... 6471 found, 0 missing, 0 empty, 0 corrupted: 100%|███████████████████████████████████████████████████████████████████████| 6471/6471 [00:00<?, ?it/s] 2025-10-23 14:48:32,157 [INFO] Dataloader num parallel workers: [8] 2025-10-23 14:48:32,273 [INFO] Registry(name=callback, total=4) 2025-10-23 14:48:32,273 [INFO] (0): YoloxSwitchTrain in mindyolo/utils/callback.py 2025-10-23 14:48:32,273 [INFO] (1): EvalWhileTrain in mindyolo/utils/callback.py 2025-10-23 14:48:32,273 [INFO] (2): SummaryCallback in mindyolo/utils/callback.py 2025-10-23 14:48:32,273 [INFO] (3): ProfilerCallback in mindyolo/utils/callback.py 2025-10-23 14:48:32,273 [INFO] 2025-10-23 14:48:32,276 [INFO] got 1 active callback as follows: 2025-10-23 14:48:32,276 [INFO] SummaryCallback() 2025-10-23 14:48:32,276 [WARNING] The first epoch will be compiled for the graph, which may take a long time; You can come back later :). albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success [INFO] albumentations load success [INFO] albumentations load success [INFO] albumentations load success [INFO] albumentations load success [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8)) [INFO] albumentations load success .........2025-10-23 14:52:54,293 [INFO] Epoch 1/10, Step 100/404, imgsize (1024, 1024), loss: 5.5585, lbox: 3.2052, lcls: 0.4855, dfl: 1.8678, cur_lr: 0.09257426112890244 2025-10-23 14:52:55,203 [INFO] Epoch 1/10, Step 100/404, step time: 2629.27 ms 2025-10-23 14:55:40,115 [INFO] Epoch 1/10, Step 200/404, imgsize (1024, 1024), loss: 4.5693, lbox: 2.5884, lcls: 0.4230, dfl: 1.5578, cur_lr: 0.08514851331710815 2025-10-23 14:55:40,138 [INFO] Epoch 1/10, Step 200/404, step time: 1649.36 ms 2025-10-23 14:58:25,055 [INFO] Epoch 1/10, Step 300/404, imgsize (1024, 1024), loss: 3.9681, lbox: 2.1428, lcls: 0.3853, dfl: 1.4400, cur_lr: 0.07772277295589447 2025-10-23 14:58:25,078 [INFO] Epoch 1/10, Step 300/404, step time: 1649.39 ms 2025-10-23 15:01:10,020 [INFO] Epoch 1/10, Step 400/404, imgsize (1024, 1024), loss: 3.6795, lbox: 2.0528, lcls: 0.3339, dfl: 1.2929, cur_lr: 0.07029703259468079 2025-10-23 15:01:10,044 [INFO] Epoch 1/10, Step 400/404, step time: 1649.65 ms 2025-10-23 15:01:17,111 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-1_404.ckpt 2025-10-23 15:01:17,111 [INFO] Epoch 1/10, epoch time: 12.75 min. 2025-10-23 15:04:02,010 [INFO] Epoch 2/10, Step 100/404, imgsize (1024, 1024), loss: 3.5361, lbox: 1.9678, lcls: 0.3183, dfl: 1.2500, cur_lr: 0.062162574380636215 2025-10-23 15:04:02,018 [INFO] Epoch 2/10, Step 100/404, step time: 1649.07 ms 2025-10-23 15:06:46,939 [INFO] Epoch 2/10, Step 200/404, imgsize (1024, 1024), loss: 3.3767, lbox: 1.8395, lcls: 0.3042, dfl: 1.2329, cur_lr: 0.05465514957904816 2025-10-23 15:06:46,947 [INFO] Epoch 2/10, Step 200/404, step time: 1649.28 ms 2025-10-23 15:09:31,885 [INFO] Epoch 2/10, Step 300/404, imgsize (1024, 1024), loss: 3.3604, lbox: 1.8753, lcls: 0.3134, dfl: 1.1718, cur_lr: 0.0471477210521698 2025-10-23 15:09:31,894 [INFO] Epoch 2/10, Step 300/404, step time: 1649.46 ms 2025-10-23 15:12:16,806 [INFO] Epoch 2/10, Step 400/404, imgsize (1024, 1024), loss: 3.2902, lbox: 1.8262, lcls: 0.2795, dfl: 1.1846, cur_lr: 0.03964029625058174 2025-10-23 15:12:16,814 [INFO] Epoch 2/10, Step 400/404, step time: 1649.20 ms 2025-10-23 15:12:23,860 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-2_404.ckpt 2025-10-23 15:12:23,860 [INFO] Epoch 2/10, epoch time: 11.11 min. 2025-10-23 15:15:08,782 [INFO] Epoch 3/10, Step 100/404, imgsize (1024, 1024), loss: 3.3220, lbox: 1.7991, lcls: 0.3124, dfl: 1.2106, cur_lr: 0.031090890988707542 2025-10-23 15:15:08,791 [INFO] Epoch 3/10, Step 100/404, step time: 1649.30 ms 2025-10-23 15:17:53,703 [INFO] Epoch 3/10, Step 200/404, imgsize (1024, 1024), loss: 3.1162, lbox: 1.6879, lcls: 0.2824, dfl: 1.1460, cur_lr: 0.02350178174674511 2025-10-23 15:17:53,711 [INFO] Epoch 3/10, Step 200/404, step time: 1649.20 ms 2025-10-23 15:20:38,631 [INFO] Epoch 3/10, Step 300/404, imgsize (1024, 1024), loss: 3.0332, lbox: 1.6024, lcls: 0.2703, dfl: 1.1605, cur_lr: 0.015912672504782677 2025-10-23 15:20:38,639 [INFO] Epoch 3/10, Step 300/404, step time: 1649.28 ms 2025-10-23 15:23:23,580 [INFO] Epoch 3/10, Step 400/404, imgsize (1024, 1024), loss: 3.1371, lbox: 1.7095, lcls: 0.2808, dfl: 1.1469, cur_lr: 0.008323564194142818 2025-10-23 15:23:23,589 [INFO] Epoch 3/10, Step 400/404, step time: 1649.49 ms 2025-10-23 15:23:30,617 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-3_404.ckpt 2025-10-23 15:23:30,617 [INFO] Epoch 3/10, epoch time: 11.11 min. 2025-10-23 15:26:15,527 [INFO] Epoch 4/10, Step 100/404, imgsize (1024, 1024), loss: 3.2965, lbox: 1.8179, lcls: 0.2614, dfl: 1.2172, cur_lr: 0.007029999978840351 2025-10-23 15:26:15,535 [INFO] Epoch 4/10, Step 100/404, step time: 1649.18 ms 2025-10-23 15:29:00,451 [INFO] Epoch 4/10, Step 200/404, imgsize (1024, 1024), loss: 3.1855, lbox: 1.7697, lcls: 0.2504, dfl: 1.1654, cur_lr: 0.007029999978840351 2025-10-23 15:29:00,459 [INFO] Epoch 4/10, Step 200/404, step time: 1649.24 ms 2025-10-23 15:31:45,369 [INFO] Epoch 4/10, Step 300/404, imgsize (1024, 1024), loss: 2.9900, lbox: 1.6270, lcls: 0.2307, dfl: 1.1323, cur_lr: 0.007029999978840351 2025-10-23 15:31:45,378 [INFO] Epoch 4/10, Step 300/404, step time: 1649.18 ms 2025-10-23 15:34:30,277 [INFO] Epoch 4/10, Step 400/404, imgsize (1024, 1024), loss: 3.1742, lbox: 1.7506, lcls: 0.2590, dfl: 1.1646, cur_lr: 0.007029999978840351 2025-10-23 15:34:30,285 [INFO] Epoch 4/10, Step 400/404, step time: 1649.07 ms 2025-10-23 15:34:37,315 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-4_404.ckpt 2025-10-23 15:34:37,316 [INFO] Epoch 4/10, epoch time: 11.11 min. 2025-10-23 15:37:22,195 [INFO] Epoch 5/10, Step 100/404, imgsize (1024, 1024), loss: 2.9632, lbox: 1.6123, lcls: 0.2424, dfl: 1.1085, cur_lr: 0.006039999891072512 2025-10-23 15:37:22,204 [INFO] Epoch 5/10, Step 100/404, step time: 1648.88 ms 2025-10-23 15:40:07,094 [INFO] Epoch 5/10, Step 200/404, imgsize (1024, 1024), loss: 2.7776, lbox: 1.4777, lcls: 0.2025, dfl: 1.0975, cur_lr: 0.006039999891072512 2025-10-23 15:40:07,103 [INFO] Epoch 5/10, Step 200/404, step time: 1648.99 ms 2025-10-23 15:42:52,021 [INFO] Epoch 5/10, Step 300/404, imgsize (1024, 1024), loss: 2.7209, lbox: 1.4253, lcls: 0.2130, dfl: 1.0826, cur_lr: 0.006039999891072512 2025-10-23 15:42:52,029 [INFO] Epoch 5/10, Step 300/404, step time: 1649.26 ms 2025-10-23 15:45:36,965 [INFO] Epoch 5/10, Step 400/404, imgsize (1024, 1024), loss: 2.7360, lbox: 1.4817, lcls: 0.2157, dfl: 1.0387, cur_lr: 0.006039999891072512 2025-10-23 15:45:36,973 [INFO] Epoch 5/10, Step 400/404, step time: 1649.44 ms 2025-10-23 15:45:44,037 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-5_404.ckpt 2025-10-23 15:45:44,037 [INFO] Epoch 5/10, epoch time: 11.11 min. 2025-10-23 15:48:28,914 [INFO] Epoch 6/10, Step 100/404, imgsize (1024, 1024), loss: 2.6675, lbox: 1.4472, lcls: 0.2042, dfl: 1.0161, cur_lr: 0.005049999803304672 2025-10-23 15:48:28,923 [INFO] Epoch 6/10, Step 100/404, step time: 1648.85 ms 2025-10-23 15:51:13,798 [INFO] Epoch 6/10, Step 200/404, imgsize (1024, 1024), loss: 2.7114, lbox: 1.4235, lcls: 0.1986, dfl: 1.0893, cur_lr: 0.005049999803304672 2025-10-23 15:51:13,807 [INFO] Epoch 6/10, Step 200/404, step time: 1648.84 ms 2025-10-23 15:53:58,688 [INFO] Epoch 6/10, Step 300/404, imgsize (1024, 1024), loss: 2.6783, lbox: 1.4169, lcls: 0.1985, dfl: 1.0629, cur_lr: 0.005049999803304672 2025-10-23 15:53:58,697 [INFO] Epoch 6/10, Step 300/404, step time: 1648.90 ms 2025-10-23 15:56:43,578 [INFO] Epoch 6/10, Step 400/404, imgsize (1024, 1024), loss: 2.7539, lbox: 1.4734, lcls: 0.2037, dfl: 1.0768, cur_lr: 0.005049999803304672 2025-10-23 15:56:43,586 [INFO] Epoch 6/10, Step 400/404, step time: 1648.89 ms 2025-10-23 15:56:50,613 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-6_404.ckpt 2025-10-23 15:56:50,613 [INFO] Epoch 6/10, epoch time: 11.11 min. 2025-10-23 15:59:35,561 [INFO] Epoch 7/10, Step 100/404, imgsize (1024, 1024), loss: 2.9109, lbox: 1.6203, lcls: 0.2210, dfl: 1.0696, cur_lr: 0.00406000018119812 2025-10-23 15:59:35,569 [INFO] Epoch 7/10, Step 100/404, step time: 1649.56 ms 2025-10-23 16:02:20,470 [INFO] Epoch 7/10, Step 200/404, imgsize (1024, 1024), loss: 2.6941, lbox: 1.4727, lcls: 0.2068, dfl: 1.0147, cur_lr: 0.00406000018119812 2025-10-23 16:02:20,479 [INFO] Epoch 7/10, Step 200/404, step time: 1649.10 ms 2025-10-23 16:05:05,384 [INFO] Epoch 7/10, Step 300/404, imgsize (1024, 1024), loss: 2.8098, lbox: 1.4810, lcls: 0.2188, dfl: 1.1101, cur_lr: 0.00406000018119812 2025-10-23 16:05:05,391 [INFO] Epoch 7/10, Step 300/404, step time: 1649.12 ms 2025-10-23 16:07:50,302 [INFO] Epoch 7/10, Step 400/404, imgsize (1024, 1024), loss: 2.8426, lbox: 1.5529, lcls: 0.2108, dfl: 1.0788, cur_lr: 0.00406000018119812 2025-10-23 16:07:50,310 [INFO] Epoch 7/10, Step 400/404, step time: 1649.18 ms 2025-10-23 16:07:57,341 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-7_404.ckpt 2025-10-23 16:07:57,342 [INFO] Epoch 7/10, epoch time: 11.11 min. 2025-10-23 16:10:42,225 [INFO] Epoch 8/10, Step 100/404, imgsize (1024, 1024), loss: 2.4095, lbox: 1.2257, lcls: 0.1704, dfl: 1.0134, cur_lr: 0.0030700000934302807 2025-10-23 16:10:42,233 [INFO] Epoch 8/10, Step 100/404, step time: 1648.92 ms 2025-10-23 16:13:27,126 [INFO] Epoch 8/10, Step 200/404, imgsize (1024, 1024), loss: 2.6034, lbox: 1.3788, lcls: 0.1872, dfl: 1.0374, cur_lr: 0.0030700000934302807 2025-10-23 16:13:27,134 [INFO] Epoch 8/10, Step 200/404, step time: 1649.00 ms 2025-10-23 16:16:12,032 [INFO] Epoch 8/10, Step 300/404, imgsize (1024, 1024), loss: 2.6074, lbox: 1.3916, lcls: 0.1787, dfl: 1.0371, cur_lr: 0.0030700000934302807 2025-10-23 16:16:12,041 [INFO] Epoch 8/10, Step 300/404, step time: 1649.07 ms 2025-10-23 16:18:56,946 [INFO] Epoch 8/10, Step 400/404, imgsize (1024, 1024), loss: 2.8867, lbox: 1.4981, lcls: 0.2189, dfl: 1.1697, cur_lr: 0.0030700000934302807 2025-10-23 16:18:56,954 [INFO] Epoch 8/10, Step 400/404, step time: 1649.13 ms 2025-10-23 16:19:03,973 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-8_404.ckpt 2025-10-23 16:19:03,973 [INFO] Epoch 8/10, epoch time: 11.11 min. 2025-10-23 16:21:48,883 [INFO] Epoch 9/10, Step 100/404, imgsize (1024, 1024), loss: 2.8544, lbox: 1.6248, lcls: 0.2181, dfl: 1.0115, cur_lr: 0.0020800000056624413 2025-10-23 16:21:48,891 [INFO] Epoch 9/10, Step 100/404, step time: 1649.18 ms 2025-10-23 16:24:33,791 [INFO] Epoch 9/10, Step 200/404, imgsize (1024, 1024), loss: 2.9393, lbox: 1.6026, lcls: 0.2223, dfl: 1.1145, cur_lr: 0.0020800000056624413 2025-10-23 16:24:33,799 [INFO] Epoch 9/10, Step 200/404, step time: 1649.08 ms 2025-10-23 16:27:18,695 [INFO] Epoch 9/10, Step 300/404, imgsize (1024, 1024), loss: 2.4632, lbox: 1.2884, lcls: 0.1701, dfl: 1.0047, cur_lr: 0.0020800000056624413 2025-10-23 16:27:18,703 [INFO] Epoch 9/10, Step 300/404, step time: 1649.04 ms 2025-10-23 16:30:03,567 [INFO] Epoch 9/10, Step 400/404, imgsize (1024, 1024), loss: 2.7216, lbox: 1.4867, lcls: 0.2002, dfl: 1.0346, cur_lr: 0.0020800000056624413 2025-10-23 16:30:03,575 [INFO] Epoch 9/10, Step 400/404, step time: 1648.72 ms 2025-10-23 16:30:10,627 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-9_404.ckpt 2025-10-23 16:30:10,627 [INFO] Epoch 9/10, epoch time: 11.11 min. 2025-10-23 16:32:55,537 [INFO] Epoch 10/10, Step 100/404, imgsize (1024, 1024), loss: 2.5899, lbox: 1.4239, lcls: 0.1668, dfl: 0.9992, cur_lr: 0.0010900000343099236 2025-10-23 16:32:55,545 [INFO] Epoch 10/10, Step 100/404, step time: 1649.18 ms 2025-10-23 16:35:40,433 [INFO] Epoch 10/10, Step 200/404, imgsize (1024, 1024), loss: 2.5535, lbox: 1.3745, lcls: 0.1813, dfl: 0.9976, cur_lr: 0.0010900000343099236 2025-10-23 16:35:40,441 [INFO] Epoch 10/10, Step 200/404, step time: 1648.95 ms 2025-10-23 16:38:25,358 [INFO] Epoch 10/10, Step 300/404, imgsize (1024, 1024), loss: 2.4509, lbox: 1.2441, lcls: 0.1717, dfl: 1.0351, cur_lr: 0.0010900000343099236 2025-10-23 16:38:25,366 [INFO] Epoch 10/10, Step 300/404, step time: 1649.25 ms 2025-10-23 16:41:10,260 [INFO] Epoch 10/10, Step 400/404, imgsize (1024, 1024), loss: 2.6832, lbox: 1.4217, lcls: 0.1896, dfl: 1.0719, cur_lr: 0.0010900000343099236 2025-10-23 16:41:10,268 [INFO] Epoch 10/10, Step 400/404, step time: 1649.02 ms 2025-10-23 16:41:17,324 [INFO] Saving model to ./runs/2025.10.23-14.48.02/weights/yolov8s-10_404.ckpt 2025-10-23 16:41:17,324 [INFO] Epoch 10/10, epoch time: 11.11 min. 2025-10-23 16:41:17,742 [INFO] End Train. 2025-10-23 16:41:18,446 [INFO] Training completed.平均每个epoch耗时约10min左右，在训练过程中我们也可以查看AI Core的利用率以及内存的占用情况：npu-smi info+--------------------------------------------------------------------------------------------------------+ | npu-smi v1.0 Version: 24.1.rc4.b999 | +-------------------------------+-----------------+------------------------------------------------------+ | NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page) | | Chip Device | Bus-Id | AICore(%) Memory-Usage(MB) | +===============================+=================+======================================================+ | 30208 310P1 | OK | NA 52 11372 / 11372 | | 0 0 | 0000:77:00.0 | 99 24288/ 89608 | +-------------------------------+-----------------+------------------------------------------------------+ | 30208 310P1 | OK | NA 42 0 / 0 | | 1 1 | 0000:77:00.0 | 0 1576 / 89085 | +===============================+=================+======================================================+ +-------------------------------+-----------------+------------------------------------------------------+ | NPU Chip | Process id | Process name | Process memory(MB) | +===============================+=================+======================================================+ | 30208 0 | 336686 | python3 | 22835 | +===============================+=================+======================================================+四、模型验证这里我们仅训练了10个epoch进行模型的验证，可以看到模型的精度和召回率如下：python3 test.py --config ./configs/yolov8/yolov8s.yaml --device_target Ascend --weight ./runs/2025.10.23-14.48.02/weights/yolov8s-10_404.ckpt2025-10-23 16:46:18,824 [INFO] parse_args: 2025-10-23 16:46:18,824 [INFO] task detect 2025-10-23 16:46:18,824 [INFO] device_target Ascend 2025-10-23 16:46:18,824 [INFO] ms_mode 0 2025-10-23 16:46:18,824 [INFO] ms_amp_level O0 2025-10-23 16:46:18,824 [INFO] ms_enable_graph_kernel False 2025-10-23 16:46:18,824 [INFO] precision_mode None 2025-10-23 16:46:18,824 [INFO] weight ./runs/2025.10.23-14.48.02/weights/yolov8s-10_404.ckpt 2025-10-23 16:46:18,824 [INFO] per_batch_size 16 2025-10-23 16:46:18,824 [INFO] img_size 1024 2025-10-23 16:46:18,824 [INFO] single_cls False 2025-10-23 16:46:18,824 [INFO] rect False 2025-10-23 16:46:18,824 [INFO] exec_nms True 2025-10-23 16:46:18,824 [INFO] nms_time_limit 60.0 2025-10-23 16:46:18,824 [INFO] conf_thres 0.001 2025-10-23 16:46:18,824 [INFO] iou_thres 0.7 2025-10-23 16:46:18,824 [INFO] conf_free True 2025-10-23 16:46:18,824 [INFO] seed 2 2025-10-23 16:46:18,824 [INFO] log_level INFO 2025-10-23 16:46:18,824 [INFO] save_dir ./runs_test/2025.10.23-16.46.18 2025-10-23 16:46:18,824 [INFO] enable_modelarts False 2025-10-23 16:46:18,824 [INFO] data_url 2025-10-23 16:46:18,824 [INFO] ckpt_url 2025-10-23 16:46:18,824 [INFO] train_url 2025-10-23 16:46:18,824 [INFO] data_dir /cache/data/ 2025-10-23 16:46:18,824 [INFO] is_parallel False 2025-10-23 16:46:18,824 [INFO] ckpt_dir /cache/pretrain_ckpt/ 2025-10-23 16:46:18,824 [INFO] data.dataset_name result 2025-10-23 16:46:18,824 [INFO] data.train_set /root/workspace/mindyolo/visdrone_COCO_format/train.txt 2025-10-23 16:46:18,824 [INFO] data.val_set /root/workspace/mindyolo/visdrone_COCO_format/val.txt 2025-10-23 16:46:18,824 [INFO] data.test_set /root/workspace/mindyolo/visdrone_COCO_format/val.txt 2025-10-23 16:46:18,824 [INFO] data.nc 12 2025-10-23 16:46:18,824 [INFO] data.names ['ignored regions', 'pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor', 'others'] 2025-10-23 16:46:18,824 [INFO] train_transforms.stage_epochs [5, 5] 2025-10-23 16:46:18,824 [INFO] train_transforms.trans_list [[{'func_name': 'mosaic', 'prob': 1.0}, {'func_name': 'resample_segments'}, {'func_name': 'random_perspective', 'prob': 1.0, 'degrees': 0.0, 'translate': 0.1, 'scale': 0.5, 'shear': 0.0}, {'func_name': 'albumentations'}, {'func_name': 'hsv_augment', 'prob': 1.0, 'hgain': 0.015, 'sgain': 0.7, 'vgain': 0.4}, {'func_name': 'fliplr', 'prob': 0.5}, {'func_name': 'label_norm', 'xyxy2xywh_': True}, {'func_name': 'label_pad', 'padding_size': 160, 'padding_value': -1}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}], [{'func_name': 'letterbox', 'scaleup': True}, {'func_name': 'resample_segments'}, {'func_name': 'random_perspective', 'prob': 1.0, 'degrees': 0.0, 'translate': 0.1, 'scale': 0.5, 'shear': 0.0}, {'func_name': 'albumentations'}, {'func_name': 'hsv_augment', 'prob': 1.0, 'hgain': 0.015, 'sgain': 0.7, 'vgain': 0.4}, {'func_name': 'fliplr', 'prob': 0.5}, {'func_name': 'label_norm', 'xyxy2xywh_': True}, {'func_name': 'label_pad', 'padding_size': 160, 'padding_value': -1}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}]] 2025-10-23 16:46:18,824 [INFO] data.test_transforms [{'func_name': 'letterbox', 'scaleup': False, 'only_image': True}, {'func_name': 'image_norm', 'scale': 255.0}, {'func_name': 'image_transpose', 'bgr2rgb': True, 'hwc2chw': True}] 2025-10-23 16:46:18,824 [INFO] data.num_parallel_workers 8 2025-10-23 16:46:18,824 [INFO] optimizer.optimizer momentum 2025-10-23 16:46:18,824 [INFO] optimizer.lr_init 0.01 2025-10-23 16:46:18,824 [INFO] optimizer.momentum 0.937 2025-10-23 16:46:18,824 [INFO] optimizer.nesterov True 2025-10-23 16:46:18,824 [INFO] optimizer.loss_scale 1.0 2025-10-23 16:46:18,824 [INFO] optimizer.warmup_epochs 3 2025-10-23 16:46:18,824 [INFO] optimizer.warmup_momentum 0.8 2025-10-23 16:46:18,824 [INFO] optimizer.warmup_bias_lr 0.1 2025-10-23 16:46:18,824 [INFO] optimizer.min_warmup_step 1000 2025-10-23 16:46:18,824 [INFO] optimizer.group_param yolov8 2025-10-23 16:46:18,824 [INFO] optimizer.gp_weight_decay 0.0005 2025-10-23 16:46:18,824 [INFO] optimizer.start_factor 1.0 2025-10-23 16:46:18,824 [INFO] optimizer.end_factor 0.01 2025-10-23 16:46:18,824 [INFO] loss.name YOLOv8Loss 2025-10-23 16:46:18,824 [INFO] loss.box 7.5 2025-10-23 16:46:18,824 [INFO] loss.cls 0.5 2025-10-23 16:46:18,824 [INFO] loss.dfl 1.5 2025-10-23 16:46:18,824 [INFO] loss.reg_max 16 2025-10-23 16:46:18,824 [INFO] epochs 10 2025-10-23 16:46:18,824 [INFO] sync_bn True 2025-10-23 16:46:18,824 [INFO] anchor_base False 2025-10-23 16:46:18,824 [INFO] opencv_threads_num 0 2025-10-23 16:46:18,824 [INFO] network.model_name yolov8 2025-10-23 16:46:18,824 [INFO] network.nc 80 2025-10-23 16:46:18,824 [INFO] network.reg_max 16 2025-10-23 16:46:18,824 [INFO] network.stride [8, 16, 32] 2025-10-23 16:46:18,824 [INFO] network.backbone [[-1, 1, 'ConvNormAct', [64, 3, 2]], [-1, 1, 'ConvNormAct', [128, 3, 2]], [-1, 3, 'C2f', [128, True]], [-1, 1, 'ConvNormAct', [256, 3, 2]], [-1, 6, 'C2f', [256, True]], [-1, 1, 'ConvNormAct', [512, 3, 2]], [-1, 6, 'C2f', [512, True]], [-1, 1, 'ConvNormAct', [1024, 3, 2]], [-1, 3, 'C2f', [1024, True]], [-1, 1, 'SPPF', [1024, 5]]] 2025-10-23 16:46:18,824 [INFO] network.head [[-1, 1, 'Upsample', ['None', 2, 'nearest']], [[-1, 6], 1, 'Concat', [1]], [-1, 3, 'C2f', [512]], [-1, 1, 'Upsample', ['None', 2, 'nearest']], [[-1, 4], 1, 'Concat', [1]], [-1, 3, 'C2f', [256]], [-1, 1, 'ConvNormAct', [256, 3, 2]], [[-1, 12], 1, 'Concat', [1]], [-1, 3, 'C2f', [512]], [-1, 1, 'ConvNormAct', [512, 3, 2]], [[-1, 9], 1, 'Concat', [1]], [-1, 3, 'C2f', [1024]], [[15, 18, 21], 1, 'YOLOv8Head', ['nc', 'reg_max', 'stride']]] 2025-10-23 16:46:18,824 [INFO] network.depth_multiple 0.33 2025-10-23 16:46:18,824 [INFO] network.width_multiple 0.5 2025-10-23 16:46:18,824 [INFO] network.max_channels 1024 2025-10-23 16:46:18,824 [INFO] overflow_still_update False 2025-10-23 16:46:18,824 [INFO] config ./configs/yolov8/yolov8s.yaml 2025-10-23 16:46:18,824 [INFO] rank 0 2025-10-23 16:46:18,824 [INFO] rank_size 1 2025-10-23 16:46:18,824 [INFO] 2025-10-23 16:46:18,898 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 16:46:18,909 [WARNING] Parse Model, args: nearest, keep str type 2025-10-23 16:46:18,984 [INFO] number of network params, total: 11.160279M, trainable: 11.140228M [WARNING] GE_ADPT(540183,7efcd8e26740,python3):2025-10-23-16:46:22.493.658 [mindspore/ops/kernel/ascend/acl_ir/op_api_exec.cc:169] GetAscendDefaultCustomPath] Checking whether the so exists or if permission to access it is available: /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize_vision/op_api/lib/libcust_opapi.so 2025-10-23 16:46:23,434 [INFO] Load checkpoint from [./runs/2025.10.23-14.48.02/weights/yolov8s-10_404.ckpt] success. 2025-10-23 16:46:23,437 [INFO] No dataset cache available, caching now... Scanning '/root/workspace/mindyolo/visdrone_COCO_format/val.cache' images and labels... 548 found, 0 missing, 0 empty, 0 corrupted: 100%|█████████████████████████████████████████████████████████████████████████████| 548/548 [00:00<00:00, 3754.44it/s] 2025-10-23 16:46:23,595 [INFO] New cache created: /root/workspace/mindyolo/visdrone_COCO_format/val.cache.npy 2025-10-23 16:46:23,595 [INFO] Dataset caching success. 2025-10-23 16:46:23,597 [INFO] Dataloader num parallel workers: [8] 2025-10-23 16:46:23,607 [WARNING] unable to load fast_coco_eval api, use normal one instead Warning: tiling offset out of range, index: 32 ..2025-10-23 16:46:55,297 [INFO] Sample 35/1, time cost: 30512.14 ms. 2025-10-23 16:46:57,108 [INFO] Sample 35/2, time cost: 1722.38 ms. 2025-10-23 16:46:58,628 [INFO] Sample 35/3, time cost: 1420.95 ms. 2025-10-23 16:47:00,538 [INFO] Sample 35/4, time cost: 1809.91 ms. 2025-10-23 16:47:02,502 [INFO] Sample 35/5, time cost: 1865.30 ms. 2025-10-23 16:47:04,321 [INFO] Sample 35/6, time cost: 1718.46 ms. 2025-10-23 16:47:06,724 [INFO] Sample 35/7, time cost: 2303.35 ms. 2025-10-23 16:47:08,940 [INFO] Sample 35/8, time cost: 2117.25 ms. 2025-10-23 16:47:11,018 [INFO] Sample 35/9, time cost: 1978.46 ms. 2025-10-23 16:47:13,101 [INFO] Sample 35/10, time cost: 1982.41 ms. 2025-10-23 16:47:14,871 [INFO] Sample 35/11, time cost: 1671.05 ms. 2025-10-23 16:47:17,112 [INFO] Sample 35/12, time cost: 2140.79 ms. 2025-10-23 16:47:19,142 [INFO] Sample 35/13, time cost: 1930.53 ms. 2025-10-23 16:47:20,984 [INFO] Sample 35/14, time cost: 1741.35 ms. 2025-10-23 16:47:23,393 [INFO] Sample 35/15, time cost: 2307.50 ms. 2025-10-23 16:47:25,557 [INFO] Sample 35/16, time cost: 2060.89 ms. 2025-10-23 16:47:27,324 [INFO] Sample 35/17, time cost: 1664.00 ms. 2025-10-23 16:47:29,254 [INFO] Sample 35/18, time cost: 1824.31 ms. 2025-10-23 16:47:31,281 [INFO] Sample 35/19, time cost: 1921.78 ms. 2025-10-23 16:47:33,331 [INFO] Sample 35/20, time cost: 1942.85 ms. 2025-10-23 16:47:35,806 [INFO] Sample 35/21, time cost: 2368.87 ms. 2025-10-23 16:47:38,165 [INFO] Sample 35/22, time cost: 2255.00 ms. 2025-10-23 16:47:40,453 [INFO] Sample 35/23, time cost: 2182.96 ms. 2025-10-23 16:47:42,588 [INFO] Sample 35/24, time cost: 2029.14 ms. 2025-10-23 16:47:44,490 [INFO] Sample 35/25, time cost: 1796.02 ms. 2025-10-23 16:47:46,804 [INFO] Sample 35/26, time cost: 2207.91 ms. 2025-10-23 16:47:49,181 [INFO] Sample 35/27, time cost: 2270.69 ms. 2025-10-23 16:47:50,926 [INFO] Sample 35/28, time cost: 1638.70 ms. 2025-10-23 16:47:53,079 [INFO] Sample 35/29, time cost: 2046.37 ms. 2025-10-23 16:47:55,061 [INFO] Sample 35/30, time cost: 1875.28 ms. 2025-10-23 16:47:57,140 [INFO] Sample 35/31, time cost: 1972.00 ms. 2025-10-23 16:47:59,895 [INFO] Sample 35/32, time cost: 2647.24 ms. 2025-10-23 16:48:02,196 [INFO] Sample 35/33, time cost: 2191.50 ms. 2025-10-23 16:48:04,739 [INFO] Sample 35/34, time cost: 2434.77 ms. ..2025-10-23 16:48:20,509 [INFO] Sample 35/35, time cost: 15723.18 ms. 2025-10-23 16:48:20,509 [INFO] loading annotations into memory... 2025-10-23 16:48:20,639 [INFO] Done (t=0.13s) 2025-10-23 16:48:20,639 [INFO] creating index... 2025-10-23 16:48:20,650 [INFO] index created! 2025-10-23 16:48:20,650 [INFO] Loading and preparing results... 2025-10-23 16:48:21,106 [INFO] DONE (t=0.46s) 2025-10-23 16:48:21,106 [INFO] creating index... 2025-10-23 16:48:21,134 [INFO] index created! 2025-10-23 16:48:21,135 [INFO] Running per image evaluation... 2025-10-23 16:48:21,135 [INFO] Evaluate annotation type *bbox* 2025-10-23 16:48:31,087 [INFO] DONE (t=9.95s). 2025-10-23 16:48:31,087 [INFO] Accumulating evaluation results... 2025-10-23 16:48:31,996 [INFO] DONE (t=0.91s). 2025-10-23 16:48:31,996 [INFO] Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.019 2025-10-23 16:48:31,996 [INFO] Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.036 2025-10-23 16:48:31,996 [INFO] Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.019 2025-10-23 16:48:31,996 [INFO] Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.016 2025-10-23 16:48:31,997 [INFO] Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.024 2025-10-23 16:48:31,997 [INFO] Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.056 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.009 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.049 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.076 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.057 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.106 2025-10-23 16:48:31,997 [INFO] Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.161 2025-10-23 16:48:31,997 [INFO] Speed: 99.0/100.3/199.3 ms inference/NMS/total per 1024x1024 image at batch-size 16; 2025-10-23 16:48:31,997 [INFO] Testing completed, cost 133.18s.使用predict.py测试训练模型参数的结果并进行可视化推理，运行方式如下：python3 examples/finetune_visdrone/predict.py --config ./configs/yolov8/yolov8s.yaml --weight=./runs/2025.10.23-14.48.02/weights/yolov8s-120_404.ckpt --image_path ./visdrone_COCO_format/val/images/000000000001.jpg训练120个epoch后，模型的推理效果如下：五、小结本文详细阐述了在OrangePi AI Studio Pro上基于昇腾310P使用MindYolo框架实现YOLOv8模型训练与验证的完整流程，涵盖环境准备、数据集格式转换、模型训练参数配置及性能评估。

 为方便开发者们参照实操案例进行智果（AgentArts）产品体验特地把收录进案例中心的AgentArts相关案例搬运过来，方便查看后续持续更新~hi~ 来试试构建出彰显个人创意的AI Agent智能体(小助手） 案例标题入口更新时间推荐值案例创作方NEW~ 敏感词猎手：用AgentArts工作流一键扫描文档风险LINK2026/06/11 ⭐⭐⭐⭐⭐官方案例中心【免费版订阅指导】华为云智能体构建平台AgentArts Agent免费版订阅使用指导LINK2025/11/20⭐⭐⭐⭐官方案例中心基于华为开发者空间-Versatile Agent构建实时股票分析助手LINK2026/01/21 ⭐⭐⭐⭐官方案例中心基于华为开发者空间-Versatile Agent构建旅游出行助手LINK2025/10/22⭐⭐⭐⭐官方案例中心华为开发者空间云开发环境+Versatile Agent，构建AI轻量级智能办公助手LINK2025/11/25 ⭐⭐⭐⭐官方案例中心基于华为开发者空间开发平台 MCP资产快速构建AI Agent应用LINK2025/11/12 ⭐⭐⭐官方案例中心基于华为开发者空间开发平台构建We码会议助手LINK2025/10/10 ⭐⭐⭐⭐官方案例中心基于华为开发者空间-云开发环境（容器）与Versatile Agent构建AI自动评分助手 LINK2026/01/12 ⭐⭐⭐⭐⭐贡献用户：tmq244基于华为开发者空间-云开发环境（容器）与Versatile Agent构建AI轻量级智能笑话机器人助手LINK2026/01/29⭐⭐⭐⭐贡献用户：liujxu【案例共创】基于华为云开发者空间-Versatile Agent开发平台构建昇腾C算子开发知识库LINK2025/11/17⭐⭐⭐⭐⭐贡献用户：黄生【案例共创】使用开发者空间 AI Agent+RAG+高德地图MCP开发班车出行助手LINK2025/07/05 ⭐⭐⭐贡献用户：神一样的老师【案例共创】基于华为开发者空间Versatile agent平台快速搭建智能搜索可视化应用LINK2025/11/17 ⭐⭐⭐贡献用户：胡琦【案例共创】基于华为开发者空间-AI Agent开发平台构建旅游规划助手LINK2025/11/17 ⭐⭐⭐贡献用户：yd_272483742【案例共创】基于华为开发者空间开发平台 MCP资产快速构建税务AI助手服务LINK2025/11/17 ⭐⭐⭐贡献用户：小草飞上天【案例共创】基于华为云开发者空间-Versatile Agent开发平台零基础开发购房助手LINK2025/11/17 ⭐⭐⭐贡献用户：CC07   < 华为云智果（AgentArts）智能体平台 官网主页> （点击进入）

当大模型参数向万亿级跨越、多模态应用对算力提出指数级需求时，长期主导 AI 训练的 GPU 架构，正面临功耗、成本与专用性的三重挑战。从谷歌 TPU 的脉动阵列到寒武纪思元的智能处理器，从专注训练的 DPU 到面向边缘推理的 NPU，各类专用 AI 芯片架构如雨后春笋般涌现，试图在算力竞赛中抢占下一代核心赛道。这些新兴架构究竟是对 GPU 的补充，还是颠覆式的替代？它们在并行计算效率、软硬件生态适配、成本控制等维度，与传统 GPU 相比有哪些核心优势与短板？在生成式 AI、自动驾驶、科学计算等不同场景下，又该如何选择最优的算力架构？欢迎分享你的观点：你认为哪种芯片架构最有可能成为 GPU 之后的下一代算力核心？其技术突破点与商业化瓶颈分别是什么？一起探讨 AI 芯片架构的革新方向与产业格局的未来演变。

人工智能论坛10月干货合集在人工智能技术加速渗透千行百业的当下，从生成式 AI 的能力平衡到核心架构的技术迭代，从隐私保护下的协同训练到高效模型的工程实践，每一项突破都在重塑智能应用的边界。本月人工智能板块精选 6 篇深度好文，聚焦行业核心议题：既剖析生成式 AI 创造力、准确性与可控性的三角平衡逻辑，也梳理从 BERT 到 LLaMA 的语义理解技术演进；既解读联邦学习在数据隐私与协同训练间的破局路径，也深入混合专家（MoE）机制、Transformer 架构、扩散模型的原理与实践 —— 全方位覆盖 AI 技术从理论基石到产业落地的关键脉络，为从业者与爱好者提供兼具深度与实用性的技术参考。以下是本月好文的完整收录，邀您一同探索人工智能的前沿动态与发展方向：https://bbs.huaweicloud.com/forum/thread-0231196590272501066-1-1.htmlhttps://bbs.huaweicloud.com/forum/thread-0254196594039800064-1-1.htmlhttps://bbs.huaweicloud.com/forum/thread-0254196590836442058-1-1.htmlhttps://bbs.huaweicloud.com/forum/thread-02107196591551809050-1-1.htmlhttps://bbs.huaweicloud.com/forum/thread-0282196591705025047-1-1.htmlhttps://bbs.huaweicloud.com/forum/thread-0259196591815993058-1-1.html大模型技术核心干货总结本次干货合集聚焦大模型技术体系的核心维度，从技术原理架构到算法优化效率，系统梳理了关键技术脉络与实践路径，为理解大模型技术演进与工程落地提供了全景式参考。技术原理与架构是大模型能力的根基，合集重点拆解了五大核心技术方向。混合专家（MoE）机制作为大模型训练的关键创新，通过将模型拆分为多个专业子模型并动态调度，在提升模型规模的同时控制计算成本，其原理核心在于专家选择策略与负载均衡优化，已成为大规模模型训练的主流架构之一。Transformer架构的演进则勾勒出大模型发展的技术主线，从Attention机制突破传统序列建模瓶颈，到GPT系列通过 decoder-only 架构、预训练-微调范式的持续迭代，实现了语言理解与生成能力的阶梯式提升。扩散模型凭借独特的“加噪-去噪”生成逻辑，在图像等生成任务中展现卓越性能，其数学原理围绕随机过程与概率建模展开，而生成质量优化则聚焦于采样效率提升与细节保真度增强。强化学习的奖励机制设计直击“探索-利用”平衡难题，通过动态调整奖励函数与探索策略，为大模型对齐人类偏好提供了关键技术支撑。多模态大模型的跨模态对齐技术则打破单一模态局限，从早期特征层面的简单融合，演进至语义层面的深度统一，实现了文本、图像等多模态信息的高效交互。算法优化与效率提升是大模型从实验室走向产业化的核心保障，合集覆盖五大关键优化路径。量化压缩技术通过INT4/FP8等低精度格式转换，在大幅降低存储与计算开销的同时，通过量化感知训练等技术保障性能损失可控，成为端侧部署的核心技术。推理效率提升则构建了“算法-硬件”协同体系，算子层面的深度优化、模型剪枝带来的冗余参数精简，结合硬件架构的定制化设计，实现了推理速度的数倍提升。联邦学习的通信效率优化聚焦隐私保护场景下的技术痛点，梯度压缩技术减少数据传输量，异步训练打破同步等待瓶颈，为跨机构数据协作提供了高效解决方案。稀疏Transformer通过结构化稀疏或非结构化稀疏设计，精准削减冗余计算，在保持模型性能的前提下降低计算复杂度，适配更广泛的计算场景。小样本学习中的元学习算法改进则针对数据稀缺问题，通过“学会学习”的范式优化，提升模型在少量样本下的泛化能力，其评估体系的完善更为技术落地提供了科学依据。整体而言，这些技术共同构建了大模型从理论到实践的完整技术链路，原理架构的创新决定了能力上限，而算法效率的优化则拓宽了应用边界，为大模型技术的规模化落地奠定了坚实基础。

自然语言处理中的语义理解：从 BERT 到 LLaMA 的上下文建模技术自然语言处理（NLP）的核心挑战在于实现机器对人类语言深层语义的精准理解。传统方法依赖词法分析和语法规则，但面对“他打破了记录”这类语义歧义时，传统方法难以区分“破坏”与“刷新”的语境差异。随着深度学习技术的突破，以BERT和LLaMA为代表的预训练语言模型，通过上下文建模技术重新定义了语义理解的范式。一、BERT：双向上下文建模的革命性突破BERT（Bidirectional Encoder Representations from Transformers）作为NLP领域的里程碑式模型，其核心创新在于双向上下文建模。传统模型如GPT仅能单向处理文本，而BERT通过Transformer编码器的自注意力机制，首次实现了对文本的双向动态感知。例如，在句子“银行”中，BERT能根据上下文区分“金融银行”与“河流岸边”的语义差异，这种能力源于其独特的预训练任务设计：掩码语言模型（MLM）：随机遮蔽15%的词汇，迫使模型通过上下文预测被遮蔽词。例如，输入“[MASK]在河边钓鱼”，模型需结合“河边”推断出“他”或“她”。下一句预测（NSP）：判断两个句子是否连续，强化模型对语义连贯性的理解。例如，输入“今天天气很好”和“我去了图书馆”，模型需判断二者是否构成合理语境。BERT的架构设计同样体现工程智慧：其输入表示由Token Embeddings（词向量）、Segment Embeddings（句子区分）和Position Embeddings（位置编码）三部分叠加构成。值得注意的是，BERT的位置编码通过学习而非固定三角函数生成，使其能更灵活地捕捉序列位置的语义变化。例如，在“存钱到银行”中，BERT能通过“存钱”与“银行”的共现关系，动态生成更准确的词向量。二、LLaMA：高效参数利用与长上下文建模LLaMA（Large Language Model Meta AI）作为Meta AI推出的开源模型，其技术路线与BERT形成互补。基于Transformer解码器架构，LLaMA通过自回归生成和多头自注意力机制，实现了对长文本的上下文建模。例如，在处理“用户咨询智能客服关于天气和航班的问题”时，LLaMA能通过前文“北京明天天气”推断出后文“航班是否会延误”的关联性。LLaMA的技术优势体现在三方面：高效参数利用：通过优化Transformer架构，LLaMA-65B模型在参数规模仅为GPT-3的1/8时，性能超越Chinchilla-70B。其关键技术包括：预归一化：稳定训练过程，减少梯度爆炸风险。SwiGLU激活函数：提升非线性表达能力，使模型能更精准地捕捉语义关联。旋转位置嵌入（RoPE）：通过相对位置编码，增强模型对长距离依赖的处理能力。多语言支持：LLaMA在预训练阶段使用包含20种语言的Wikipedia数据集，使其能直接处理“查询巴黎天气”或“翻译中文古诗”等跨语言任务。灵活规模选择：提供7B、13B、33B、65B四种参数规模，开发者可根据计算资源选择适配模型。例如，在边缘设备上部署LLaMA-7B，在云端使用LLaMA-65B。三、上下文建模技术的演进方向从BERT到LLaMA，上下文建模技术正朝着三个方向演进：更长的上下文窗口：LLaMA-3.1支持128K标记的上下文窗口，能处理完整书籍或长篇报告的语义关联。例如，在法律文档分析中，模型可同时参考“合同条款”与“历史判例”进行推理。多模态融合：LLaMA通过组合式方法集成图像、视频和语音能力。例如，在医疗诊断中，模型可结合“患者CT影像”与“病历文本”进行综合判断。知识增强：通过整合知识图谱（如Wikidata）与神经网络，解决纯数据驱动模型的逻辑推理缺陷。例如，在数学推理任务中，模型可调用“勾股定理”等知识进行步骤推导。四、应用场景的深度拓展上下文建模技术的突破，正在重塑多个行业的交互方式：智能客服：LLaMA模型可实时理解用户意图，例如将“我要改签”与“航班延误通知”关联，提供自动化解决方案。医疗诊断：BERT通过分析电子病历中的上下文信息，辅助医生识别“胸痛”与“心肌梗死”的关联性。代码生成：LLaMA在HumanEval基准测试中，通过理解“生成排序算法”的需求，生成符合上下文的高质量代码。五、未来挑战与技术展望尽管上下文建模技术已取得显著进展，但仍面临两大挑战：事实性错误：纯数据驱动模型可能生成“爱因斯坦发明电灯”等错误信息，需通过知识图谱校验进行修正。伦理与安全：模型可能生成偏见性内容，需通过拒绝采样（RS）和直接偏好优化（DPO）等技术进行后训练校正。未来，随着LLaMA-3等4050亿参数模型的发布，上下文建模技术将进一步渗透到自动驾驶、金融风控等领域。例如，在自动驾驶中，模型可通过理解“前方施工”与“绕行建议”的上下文关联，生成更安全的决策路径。从BERT的双向上下文建模到LLaMA的长文本处理，NLP技术正逐步实现“理解每一个用户意图”的愿景。随着知识增强与多模态融合的深入，语义理解将不再局限于文本，而是成为连接人类与AI的通用语言。

端侧AI模型的能效优化：针对移动设备的轻量化设计与推理加速在移动设备部署AI模型时，算力、功耗与内存的严格限制成为主要挑战。通过轻量化模型设计与推理加速技术的结合，可在保持精度的同时将模型体积缩小90%、推理延迟降低70%，实现端侧AI的实时响应与长效续航。轻量化模型设计：压缩与剪枝结构化剪枝采用通道剪枝算法（如L1范数筛选），移除卷积层中权重绝对值较小的通道。在MobileNetV3上，通过渐进式剪枝可移除50%的通道，模型体积从12MB降至5MB，且在ImageNet上的Top-1准确率仅下降1.2%。知识蒸馏使用大模型（如ResNet-152）作为教师模型，指导轻量级学生模型（如MobileNet）学习特征分布。实验表明，蒸馏后的MobileNet在CIFAR-100上的准确率提升8%，接近教师模型性能的95%。神经架构搜索（NAS）通过强化学习或遗传算法自动搜索硬件友好的模型结构。例如，MnasNet针对移动端GPU优化，在相同精度下推理速度比MobileNetV2快1.5倍。推理加速技术：硬件协同优化量化感知训练将模型权重从FP32量化为INT8，配合模拟量化训练（QAT）减少精度损失。在骁龙865上，量化后的YOLOv5模型推理速度提升4倍，内存占用减少75%。算子融合与硬件加速合并卷积、偏置与激活操作（如Conv+BN+ReLU→FusedConv），并调用移动端NPU（如苹果神经引擎）或GPU加速。在华为麒麟芯片上，算子融合使ResNet-50的推理延迟从120ms降至35ms。动态批处理与内存复用通过动态调整输入批大小（如从1到8）提升GPU利用率，同时复用中间结果内存。在树莓派4B上，此技术使BERT-base的推理吞吐量提升3倍。实践案例：某人脸识别应用的优化通过剪枝+量化将模型体积从50MB压缩至5MB，结合NPU加速使单帧推理延迟从200ms降至30ms，功耗降低60%。未来，存算一体芯片与自适应计算架构将进一步推动端侧AI的能效突破。

AI训练数据的清洗与增强：基于规则与模型的自动化处理方案在AI模型训练中，数据质量直接决定模型性能上限。面对海量、多源、噪声密集的原始数据，基于规则与模型的自动化清洗与增强方案成为提升数据可用性的关键。通过“规则过滤+模型修正”的双层架构，可实现90%以上数据问题的自动处理，同时降低人工标注成本60%以上。规则驱动的数据清洗：精准过滤噪声结构化规则引擎针对表格数据或文本中的格式错误（如日期格式混乱、数值越界），构建正则表达式与逻辑判断规则库。例如，在金融风控数据中，通过规则过滤掉“年龄>120岁”或“收入为负值”的异常样本，清洗准确率达99%。语义一致性校验利用预训练语言模型（如BERT）检测文本中的语义矛盾。例如，在医疗记录中识别“患者无高血压病史但服用降压药”的矛盾表述，结合规则标记需人工复核的样本。多模态数据对齐对图文配对数据，通过计算图像特征与文本嵌入的余弦相似度，过滤掉相似度低于阈值的样本。在电商商品数据中，此方法可剔除30%的图文不匹配数据。模型驱动的数据增强：生成高质量样本上下文感知的文本增强采用T5等文本生成模型，根据上下文生成同义句或参数化变体。例如，将“用户喜欢红色手机”增强为“用户偏好红色款式的移动设备”，保留语义同时提升数据多样性。可控的图像生成增强结合Stable Diffusion的ControlNet插件，对原始图像进行风格迁移（如卡通化）、几何变换（如旋转）或局部遮挡，生成对抗样本提升模型鲁棒性。在自动驾驶数据中，此方法使目标检测模型在雨天场景下的准确率提升15%。合成数据生成利用GAN或扩散模型生成完全合成的训练数据。例如，在工业缺陷检测中，通过CycleGAN生成不同光照条件下的缺陷图像，解决真实缺陷样本稀缺的问题。实践案例：某NLP模型的优化通过规则清洗去除12%的噪声数据，结合文本增强生成2倍于原始数据的同义样本，最终使模型在少样本场景下的F1值提升18%。未来，随着大模型指令微调技术的发展，自动化清洗与增强方案将向更精准、更高效的方向演进。

生成式AI服务的高并发架构设计：缓存策略与动态资源调度实践生成式AI服务（如大语言模型、文生图）在面对高并发请求时，需解决计算资源密集、响应延迟敏感等挑战。通过分层缓存策略与动态资源调度技术的结合，可实现千级QPS下的稳定服务，同时降低30%以上的算力成本。分层缓存体系：降低重复计算请求级缓存针对重复提问（如“今天天气？”）或热门prompt，采用Redis集群存储输入-输出对。通过哈希算法快速匹配缓存，在某文生图服务中，请求级缓存命中率达45%，直接减少70%的GPU计算量。片段级缓存对长文本生成任务，分解为句子级片段并缓存中间结果。例如，在法律文书生成场景中，缓存常用条款片段（如“保密协议条款”），使生成速度提升2倍，同时保持上下文一致性。模型参数缓存在多模型服务场景中，通过共享内存池缓存模型权重，避免频繁加载。使用NVIDIA Triton推理服务器的模型仓库功能，可将模型加载时间从秒级降至毫秒级。动态资源调度：弹性应对流量波峰基于K8s的GPU弹性伸缩结合Prometheus监控实时QPS，通过自定义指标（如等待队列长度）触发GPU节点扩容。在某大语言模型服务中，波峰期间自动增加30%的GPU资源，确保95%的请求在2秒内响应。异构计算调度将简单任务（如文本分类）调度至CPU节点，复杂任务（如长文本生成）分配至GPU。通过ONNX Runtime的异构执行提供者，在AMD CPU+NVIDIA GPU混合环境中，资源利用率提升25%。优先级队列管理对付费用户或紧急请求设置高优先级队列，采用加权轮询算法分配资源。实验表明，优先级调度可使VIP用户平均等待时间降低80%。实践案例：某文生图平台的优化通过引入请求级缓存与动态GPU扩容，该平台在促销活动期间（QPS从500飙升至3000）保持了99.9%的可用性，同时算力成本下降35%。未来，结合模型量化与存算一体芯片，高并发架构将向更低延迟、更高能效的方向演进。

AI模型的边缘部署：TensorRT与ONNX Runtime的优化流程与性能调优在边缘设备（如手机、IoT终端）部署AI模型时，需平衡模型精度、推理速度与硬件资源限制。TensorRT（NVIDIA）与ONNX Runtime（微软/Linux基金会）作为两大主流优化框架，分别通过硬件加速与跨平台优化，为边缘AI部署提供了高效解决方案。TensorRT：GPU边缘设备的极致优化TensorRT针对NVIDIA Jetson系列等嵌入式GPU设备，通过三步优化流程实现性能突破：模型解析与层融合：自动合并卷积、偏置与激活层（如Conv+ReLU→FusedConv），减少内存访问与计算开销。精度校准与量化：支持FP16/INT8量化，在Jetson AGX Xavier上，ResNet-50的INT8推理速度较FP32提升4倍，精度损失＜1%。内核自动选择：根据硬件架构（如Tensor Core）动态生成最优CUDA内核，在Jetson Nano上实现YOLOv5的22FPS实时检测。调优技巧：启用动态批处理（Dynamic Batching）提升吞吐量，关闭非必要日志输出以减少CPU占用。ONNX Runtime：跨平台的高效部署ONNX Runtime支持ARM CPU、NVIDIA GPU及苹果神经引擎等多硬件，优化流程包括：ONNX模型转换：将PyTorch/TensorFlow模型转为ONNX格式，消除框架差异。执行提供者配置：针对ARM CPU启用NNAPI或OpenVINO执行提供者，在树莓派4B上实现MobileNetV3的15ms延迟。图级优化：通过常量折叠、死代码消除等操作减少计算量，在苹果M1芯片上使BERT推理速度提升30%。调优技巧：启用ORT_DISABLE_ALL_THREADS单线程模式降低延迟，或通过CUDA_EP配置实现GPU-CPU混合推理。技术选型建议NVIDIA硬件优先：选择TensorRT以充分利用Tensor Core与DLA加速。跨平台需求：ONNX Runtime支持ARM、x86及移动端，适合多设备部署场景。混合部署策略：在NVIDIA Jetson上结合TensorRT（GPU部分）与ONNX Runtime（CPU部分），实现资源最大化利用。未来，随着边缘设备算力提升，自动量化、动态图优化等技术将进一步降低部署门槛，推动AI模型从云端向边缘侧的全面迁移。

大模型分布式训练框架对比：Megatron-LM、DeepSpeed 与 FSDP 的技术选型在千亿参数级大模型训练中，分布式框架的技术选型直接影响训练效率与资源利用率。Megatron-LM、DeepSpeed 与 FSDP 作为三大主流方案，分别代表了硬件协同优化、显存效率突破与 PyTorch 原生集成的技术路线。Megatron-LM：极致硬件协同NVIDIA 开发的 Megatron-LM 专注于张量并行与流水线并行的深度融合。其核心优势在于通过列并行与行并行拆分 Transformer 层的 QKV 投影与输出投影，结合 GPipe 调度策略减少流水线气泡。在 NVIDIA A100 集群中，Megatron-LM 可实现 92% 的硬件利用率，但需依赖 NVLink 高速互联，且内存占用随流水线阶段数线性增长。典型应用场景包括 GPT-3、BLOOM 等模型的预训练。DeepSpeed：显存效率革命微软的 DeepSpeed 通过 ZeRO 优化技术重新定义了显存管理。ZeRO-3 将优化器状态、梯度与模型参数分片到不同 GPU，配合梯度累积流水线将空泡率从 30% 降至 15%。在 175B 参数模型训练中，DeepSpeed 可将单卡显存占用从 24GB 降至 5.8GB，并支持 NVMe 卸载实现 10B 参数模型的单机训练。其序列并行技术使长序列（如 10k token）通信效率提升 40%，但需精确控制 CUDA 流时序。FSDP：PyTorch 原生之选作为 PyTorch 2.1+ 的官方方案，FSDP 通过动态参数分片与通信-计算重叠，实现了显存占用与计算效率的平衡。其自动分片阈值机制可对参数数量超过阈值的层进行分片，避免小层通信开销。在 13B 参数模型训练中，FSDP 的通信量较 Megatron-LM 减少 80%，且与 Hugging Face Transformers 库深度集成，成为中小规模团队的首选。技术选型决策树硬件环境优先：配备 NVLink 3.0 的集群选择 Megatron-LM 以发挥硬件极限。显存受限场景：单卡显存＜80GB 时，DeepSpeed ZeRO-3 可训练模型规模提升 3 倍。PyTorch 生态需求：FSDP2 结合 torch.compile 实现编译加速，适合百亿参数模型训练。未来，随着 3D 封装与光互连技术成熟，混合并行策略与自动化调优工具将成为主流。Megatron-LM 与 DeepSpeed 的融合实践（如 Megatron-DeepSpeed）已展现出性能与易用性的双重突破，推动大模型训练进入“小时级”时代。

小样本学习中的元学习算法改进与评估小样本学习（Few-shot Learning）旨在通过极少量标注样本（如每类5-10个）实现模型快速泛化，而元学习（Meta-Learning）作为其核心框架，通过“学习如何学习”的范式，显著提升了模型在新任务上的适应能力。然而，传统元学习算法在跨域迁移、样本噪声等场景下性能受限，改进算法设计与评估体系成为当前研究重点。元学习算法的核心改进方向跨模态元学习针对多模态小样本任务（如图文联合分类），引入跨模态注意力机制，使模型在训练阶段学习模态间共享的元知识。例如，Meta-CMAN算法通过动态调整图文特征的权重分配，在跨模态Few-shot分类任务中准确率提升12%。动态任务适应传统MAML算法采用固定初始化参数，难以适应任务分布差异。改进的Prototypical Networks++通过引入任务自适应原型修正模块，根据当前任务样本动态调整原型中心，在样本分布偏移场景下（如跨数据集测试）将错误率降低18%。噪声鲁棒性增强针对小样本中常见的标注噪声问题，Meta-NoiseNet算法在元训练阶段注入可控噪声，并设计噪声感知损失函数，使模型在含30%噪声的样本上仍保持89%的准确率，较传统方法提升25%。评估体系的优化实践跨域泛化评估传统评估仅在同分布测试集上验证，改进方案引入跨域测试集（如从自然图像迁移到医学影像），要求模型在未见过的数据域上保持性能。实验表明，跨域元学习算法在域外测试集上的准确率衰减较标准方法减少40%。少样本生成评估结合生成模型（如GAN），评估元学习算法在极少量样本下的数据生成能力。例如，Meta-GAN通过元训练学习生成器的快速适应策略，在5-shot条件下生成的样本质量（FID评分）较基线模型提升35%。未来方向：动态元学习与硬件协同下一代元学习将向动态任务分解与硬件友好型设计演进。例如，通过强化学习自动生成任务序列以优化元训练效率，或结合存算一体芯片实现元参数的原位更新。随着小样本学习在医疗诊断、工业质检等领域的落地，其评估体系也将更注重实际场景中的鲁棒性与可解释性。

稀疏Transformer的设计与应用：降低计算复杂度的核心技术解析Transformer模型凭借自注意力机制在NLP、CV等领域取得突破，但其标准实现面临计算复杂度随序列长度二次增长的瓶颈。稀疏Transformer通过重构注意力计算范式，将复杂度从O(n²)降至O(n log n)甚至线性级别，成为处理长序列与大规模模型的核心技术。稀疏化设计的核心策略结构化稀疏模式采用固定稀疏模式限制注意力范围，如滑动窗口注意力通过局部窗口（如512个token）计算注意力，Longformer模型将此模式应用于64k长度序列，解码阶段速度提升3倍。块状稀疏化将序列划分为16×16的块，仅计算块内注意力，BigBird模型结合随机注意力、局部窗口与全局注意力，在保持性能的同时降低计算量。动态稀疏机制Top-k注意力通过评分函数选择相关性最高的k个token，如SpargeAttn在Llama3.1（128K序列）中实现0.54稀疏度，速度达708.1 TOPS且性能无损。路由注意力引入可学习路由网络，动态决定token间的连接关系，适用于多模态场景中跨模态关联的稀疏建模。混合精度量化结合8-bit量化与稀疏化，如SageAttention框架在稀疏注意力计算中采用低精度存储，进一步减少内存占用。实验表明，量化后的稀疏Transformer在图像-视频跨模态检索任务中，推理速度提升40%，精度损失小于1%。工业级应用场景长文本处理在法律文书分析中，稀疏Transformer可处理超长文本（如10万词合同），通过滑动窗口与全局注意力结合，准确提取条款关联关系，推理时间较标准模型缩短65%。多模态建模稀疏表示引导的Transformer在跨模态检索中，通过稀疏化图文注意力矩阵，降低90%的计算开销，同时保持98%以上的检索准确率。例如，在图像-视频检索系统中，用户上传图片后，模型可在毫秒级返回相关视频片段。边缘设备部署结合剪枝与稀疏化，BERT模型在移动端实现40%参数剪枝+INT8量化后，内存占用从1.2GB降至300MB，推理延迟从120ms降至28ms，满足实时语音交互需求。未来方向：自适应稀疏架构下一代稀疏Transformer将向动态模式生成与硬件协同优化演进。例如，通过元学习自动调整稀疏模式以适应不同任务，或结合存算一体芯片实现稀疏注意力矩阵的原位计算。随着模型规模突破万亿参数，稀疏化技术将成为AI大模型走向实用化的关键推手。

联邦学习中的通信效率优化：梯度压缩与异步训练技术实践联邦学习通过分布式训练实现数据隐私保护，但客户端与服务器间的频繁通信成为性能瓶颈。尤其在跨设备、跨机构场景中，网络带宽限制和异构硬件差异导致训练效率低下。梯度压缩与异步训练技术的结合，为联邦学习通信优化提供了高效解决方案。梯度压缩：降低传输数据量传统联邦学习需上传完整梯度向量，通信开销随模型规模线性增长。梯度量化通过减少梯度数值精度（如FP32→INT8）将数据量压缩至1/4，同时结合误差补偿机制（如SignSGD算法）抵消量化误差，实验表明在图像分类任务中可保持98%以上的模型精度。稀疏化压缩则进一步筛选重要梯度（如Top-K梯度），仅上传关键参数，配合局部梯度累积策略，使通信量减少90%以上。异步训练：突破同步等待限制同步联邦学习要求所有客户端完成本地训练后统一聚合，易因设备算力差异或网络延迟导致“拖尾效应”。异步聚合允许客户端随时上传梯度，服务器采用加权平均或动态权重调整策略（如基于客户端数据量的权重分配）更新全局模型。谷歌提出的FedAsync框架在移动端NLP任务中，将单轮训练时间从同步模式的12分钟缩短至4分钟，同时收敛速度提升30%。实践挑战与工程优化实际部署需解决梯度陈旧性（Stale Gradient）问题。通过引入梯度时效性评估（如基于梯度变化率的动态权重衰减）和客户端选择策略（如优先调度高带宽设备），可平衡模型新鲜度与通信效率。此外，结合边缘计算节点进行局部聚合，能进一步减少服务器通信压力。未来方向随着5G/6G网络普及，联邦学习将向“超低延迟、海量设备”场景演进。梯度压缩与异步训练的融合将与区块链技术结合，实现去中心化可信聚合，推动医疗、金融等敏感领域的规模化应用。

AI推理效率提升：算子优化、模型剪枝与硬件协同设计方案随着AI模型参数量突破千亿级，推理阶段的计算效率与能耗成为制约应用落地的关键瓶颈。通过算子优化、模型剪枝与硬件协同设计的三维联动，可实现推理性能的指数级提升，推动AI从实验室走向真实场景。算子优化：挖掘计算内核潜力算子（Operator）是模型执行的基本单元，其效率直接影响推理速度。算子融合技术通过合并相邻算子（如Conv+ReLU→FusedConv），减少内存访问次数，在ResNet等模型中可降低30%的延迟。稀疏化算子针对剪枝后的稀疏权重设计专用内核，如NVIDIA的A100 GPU通过结构化稀疏（2:4模式）实现2倍加速。此外，低精度算子（如FP8/INT4）与硬件指令集（如AMD的FP8指令）的结合，使单次运算能耗降低75%。模型剪枝：结构性去除冗余参数模型剪枝通过移除不重要的权重或通道，平衡精度与计算量。非结构化剪枝随机删除权重，需配合稀疏矩阵存储（如CSR格式）和专用硬件（如Google TPU的稀疏核）；结构化剪枝则按通道/层裁剪，生成规则化模型，可直接部署于现有硬件。实验表明，在BERT模型中，结构化剪枝可减少80%参数，同时保持95%以上的任务准确率。硬件协同设计：定制化加速架构硬件与算法的协同优化是效率突破的关键。存算一体架构（如Mythic AMP芯片）将计算单元嵌入存储器，消除“内存墙”瓶颈，使图像分类推理能耗降低10倍。可重构计算（如Xilinx Versal ACAP）通过动态调整硬件资源，适配不同模型结构。此外，异构计算（CPU+GPU+NPU）的调度优化，可使移动端模型推理速度提升4倍。未来趋势：软硬一体化的全栈优化未来，AI推理效率将依赖“算法-编译器-硬件”的全栈协同。例如，MLIR编译器框架可自动生成针对特定硬件的优化算子，结合自动化剪枝工具（如Intel Neural Compressor），实现从模型设计到部署的无缝加速。这一趋势将推动AI在自动驾驶、实时语音交互等低延迟场景中的规模化应用。