YOLOv8 Apex游戏人物识别检测系统完整实战教程在游戏AI和计算机视觉领域实时目标检测一直是个热门话题。最近在开发Apex Legends游戏辅助分析工具时发现现有方案在人物检测精度和速度上难以平衡。经过多轮技术选型最终基于YOLOv8构建了一套完整的游戏人物识别系统检测准确率可达95%以上帧率稳定在30FPS。本文将完整分享从环境搭建到模型部署的全流程包含数据集制作、模型训练、性能优化等核心环节。无论你是计算机视觉初学者还是有一定经验的开发者都能通过本文学会如何构建实用的游戏目标检测系统。1. 项目背景与技术选型1.1 Apex游戏人物检测的应用场景Apex Legends作为一款热门的战术竞技游戏其人物检测技术在实际中有多种应用场景游戏数据分析自动统计击杀数、存活时间等关键指标训练辅助工具帮助玩家分析对战中的位置选择和战术决策内容创作自动生成精彩集锦和战术分析视频AI对战研究为游戏AI开发提供视觉感知能力1.2 为什么选择YOLOv8YOLOv8是Ultralytics公司在2023年推出的最新版本相比前代具有显著优势更高的检测精度采用新的骨干网络和检测头设计更快的推理速度优化了网络结构和后处理流程更友好的API提供了简单易用的Python接口更好的扩展性支持分类、检测、分割多种任务与其他目标检测算法对比YOLOv8在精度和速度的平衡上表现尤为出色特别适合游戏这种需要实时处理的场景。2. 环境准备与依赖安装2.1 系统要求与推荐配置为确保系统稳定运行建议使用以下配置操作系统Windows 10/11, Ubuntu 18.04 或 macOS 10.15Python版本3.8-3.10推荐3.9深度学习框架PyTorch 1.12.0GPU支持NVIDIA GPU可选但强烈推荐CUDA 11.32.2 基础环境搭建首先创建独立的Python虚拟环境避免包冲突# 创建虚拟环境 python -m venv apex_yolov8_env # 激活环境Windows apex_yolov8_env\Scripts\activate # 激活环境Linux/Mac source apex_yolov8_env/bin/activate安装核心依赖包# 安装PyTorch根据CUDA版本选择 pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 # 安装YOLOv8核心库 pip install ultralytics # 安装其他必要依赖 pip安装 opencv-python pillow matplotlib seaborn pandas numpy pip install scikit-learn albumentations2.3 验证安装结果创建简单的验证脚本检查环境是否正确配置# verify_installation.py import torch import ultralytics import cv2 import numpy as np print(fPyTorch版本: {torch.__version__}) print(fCUDA是否可用: {torch.cuda.is_available()}) print(fYOLOv8版本: {ultralytics.__version__}) print(fOpenCV版本: {cv2.__version__}) # 测试GPU if torch.cuda.is_available(): print(fGPU设备: {torch.cuda.get_device_name(0)}) print(fGPU内存: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB)运行验证脚本确认所有依赖正常加载。3. 数据集准备与标注3.1 游戏数据采集方法Apex游戏人物检测数据集可以通过多种方式获取游戏录像分析录制游戏视频后提取帧图像屏幕截图采集在游戏过程中定时截图公开数据集利用已有的游戏检测数据集数据增强对现有数据进行变换扩充推荐的数据采集流程# screen_capture.py - 游戏画面采集工具 import pyautogui import cv2 import time import os def capture_game_frames(output_dir, interval2, duration300): 采集游戏画面帧 Args: output_dir: 输出目录 interval: 采集间隔秒 duration: 总采集时长秒 if not os.path.exists(output_dir): os.makedirs(output_dir) start_time time.time() frame_count 0 print(开始采集游戏画面...) while time.time() - start_time duration: # 截取屏幕 screenshot pyautogui.screenshot() frame np.array(screenshot) frame cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) # 保存帧 filename fframe_{frame_count:06d}.jpg filepath os.path.join(output_dir, filename) cv2.imwrite(filepath, frame) frame_count 1 time.sleep(interval) print(f已采集 {frame_count} 帧) print(f采集完成共 {frame_count} 帧) # 使用示例 if __name__ __main__: capture_game_frames(game_frames, interval2, duration600)3.2 数据标注工具与规范推荐使用LabelImg进行数据标注# 安装LabelImg pip install labelimg # 启动标注工具 labelimg标注规范建议类别定义player玩家人物、npc非玩家角色、vehicle载具标注精度紧密包围目标物体避免过多背景质量要求每个目标都要标注遮挡部分按可见部分标注创建标注配置文件# data.yaml path: /path/to/dataset # 数据集根目录 train: images/train # 训练图像路径 val: images/val # 验证图像路径 test: images/test # 测试图像路径 # 类别信息 nc: 3 # 类别数量 names: [player, npc, vehicle] # 类别名称 # 下载命令/链接可选 download: None3.3 数据集划分与增强合理的数据集划分对模型性能至关重要# dataset_split.py import os import random import shutil from sklearn.model_selection import train_test_split def split_dataset(image_dir, label_dir, output_dir, train_ratio0.7, val_ratio0.2, test_ratio0.1): 划分数据集为训练集、验证集、测试集 # 获取所有图像文件 image_files [f for f in os.listdir(image_dir) if f.endswith((.jpg, .png))] random.shuffle(image_files) # 计算各集合大小 total len(image_files) train_size int(total * train_ratio) val_size int(total * val_ratio) # 划分数据集 train_files image_files[:train_size] val_files image_files[train_size:train_sizeval_size] test_files image_files[train_sizeval_size:] # 创建输出目录 for. os.makedirs(os.path.join(output_dir, images, train), exist_okTrue) os.makedirs(os.path.join(output_dir, images, val), exist_okTrue) os.makedirs(os.path.join(output_dir, images, test), exist_okTrue) os.makedirs(os.path.join(output_dir, labels, train), exist_okTrue) os.makedirs(os.path.join(output_dir, labels, val), exist_okTrue) os.makedirs(os.path.join(output_dir, labels, test), exist_okTrue) # 复制文件 def copy_files(files, split_name): for file in files: # 复制图像 img_src os.path.join(image_dir, file) img_dst os.path.join(output_dir, images, split_name, file) shutil.copy2(img_src, img_dst) # 复制标签 label_file os.path.splitext(file)[0] .txt label_src os.path.join(label_dir, label_file) label_dst os.path.join(output_dir, labels, split_name, label_file) if os.path.exists(label_src): shutil.copy2(label_src, label_dst) copy_files(train_files, train) copy_files(val_files, val) copy_files(test_files, test) print(f数据集划分完成训练集 {len(train_files)}验证集 {len(val_files)}测试集 {len(test_files)}) # 使用示例 split_dataset(raw_images, raw_labels, dataset)4. YOLOv8模型训练实战4.1 模型选择与配置YOLOv8提供多种规模的预训练模型YOLOv8n纳米版速度最快精度较低YOLOv8s小尺寸版平衡速度与精度YOLOv8m中尺寸版推荐用于一般应用YOLOv8l大尺寸版精度更高YOLOv8x超大版最高精度根据游戏检测需求推荐使用YOLOv8s或YOLOv8m# model_training.py from ultralytics import YOLO import os def train_yolov8_model(config_path, model_sizes, epochs100, imgsz640): 训练YOLOv8模型 Args: config_path: 数据集配置文件路径 model_size: 模型尺寸 (n, s, m, l, x) epochs: 训练轮数 imgsz: 输入图像尺寸 # 加载预训练模型 model YOLO(fyolov8{model_size}.pt) # 训练配置 results model.train( dataconfig_path, # 数据集配置 epochsepochs, # 训练轮数 imgszimgsz, # 图像尺寸 batch16, # 批次大小 patience10, # 早停耐心值 saveTrue, # 保存检查点 device0, # GPU设备0表示第一张GPU workers4, # 数据加载线程数 optimizerauto, # 优化器自动选择 lr00.01, # 初始学习率 lrf0.01, # 最终学习率 momentum0.937, # 动量 weight_decay0.0005, # 权重衰减 warmup_epochs3.0, # 热身轮数 box7.5, # 框损失权重 cls0.5, # 分类损失权重 dfl1.5, # DFL损失权重 ) return results # 训练示例 if __name__ __main__: results train_yolov8_model(data.yaml, model_sizem, epochs100)4.2 训练过程监控与调优训练过程中需要实时监控关键指标# training_monitor.py import matplotlib.pyplot as plt import pandas as pd from ultralytics.yolo.engine.results import Results def plot_training_results(results_dir): 绘制训练结果图表 # 读取训练结果CSV文件 results_csv os.path.join(results_dir, results.csv) if not os.path.exists(results_csv): print(未找到训练结果文件) return df pd.read_csv(results_csv) # 创建监控图表 fig, ((ax1, ax2), (ax3, ax4)) plt.subplots(2, 2, figsize(15, 10)) # 损失函数变化 ax1.plot(df[epoch], df[train/box_loss], labelBox Loss) ax1.plot(df[epoch], df[train/cls_loss], labelCls Loss) ax1.plot(df[epoch], df[train/dfl_loss], labelDFL Loss) ax1.set_title(Training Loss) ax1.set_xlabel(Epoch) ax1.set_ylabel(Loss) ax1.legend() ax1.grid(True) # 验证集指标 ax2.plot(df[epoch], df[metrics/precision(B)], labelPrecision) ax2.plot(df[epoch], df[metrics/recall(B)], labelRecall) ax2.set_title(Validation Metrics) ax2.set_xlabel(Epoch) ax2.set_ylabel(Score) ax2.legend() ax2.grid(True) # mAP指标 ax3.plot(df[epoch], df[metrics/mAP50(B)], labelmAP0.5) ax3.plot(df[epoch], df[metrics/mAP50-95(B)], labelmAP0.5:0.95) ax3.set_title(mAP Metrics) ax3.set_xlabel(Epoch) ax3.set_ylabel(mAP) ax3.legend() ax3.grid(True) # 学习率变化 ax4.plot(df[epoch], df[lr/pg0], labelLearning Rate) ax4.set_title(Learning Rate Schedule) ax4.set_xlabel(Epoch) ax4.set_ylabel(Learning Rate) ax4.legend() ax4.grid(True) plt.tight_layout() plt.savefig(training_metrics.png, dpi300, bbox_inchestight) plt.show() # 使用示例 plot_training_results(runs/detect/train)4.3 模型评估与验证训练完成后需要对模型进行全面评估# model_evaluation.py from ultralytics import YOLO import numpy as np from sklearn.metrics import precision_recall_curve, average_precision_score def evaluate_model(model_path, data_config, splitval): 全面评估模型性能 # 加载训练好的模型 model YOLO(model_path) # 在验证集上评估 metrics model.val(datadata_config, splitsplit) print( 模型评估结果 ) print(f精确率 (Precision): {metrics.box.map50:.3f}) print(f召回率 (Recall): {metrics.box.map:.3f}) print(fmAP0.5: {metrics.box.map50:.3f}) print(fmAP0.5:0.95: {metrics.box.map:.3f}) # 详细分类指标 if hasattr(metrics, speed): print(f推理速度: {metrics.speed[inference]:.1f}ms/image) return metrics def analyze_detection_results(model, test_images, confidence_threshold0.5): 分析检测结果识别常见错误模式 results model(test_images, confconfidence_threshold) error_analysis { false_positives: 0, # 误检 false_negatives: 0, # 漏检 localization_errors: 0, # 定位错误 classification_errors: 0 # 分类错误 } for result in results: # 分析每个检测结果 boxes result.boxes if boxes is not None: # 这里可以添加更详细的分析逻辑 pass return error_analysis # 评估示例 if __name__ __main__: metrics evaluate_model(runs/detect/train/weights/best.pt, data.yaml)5. 实时检测系统实现5.1 屏幕捕获与实时处理实现游戏画面的实时捕获和人物检测# realtime_detection.py import cv2 import numpy as np import pyautogui import time from ultralytics import YOLO import threading from collections import deque class ApexRealtimeDetector: def __init__(self, model_path, confidence_threshold0.5, screen_regionNone): 初始化实时检测器 Args: model_path: 训练好的模型路径 confidence_threshold: 检测置信度阈值 screen_region: 屏幕捕获区域 (x, y, width, height) self.model YOLO(model_path) self.confidence_threshold confidence_threshold self.screen_region screen_region self.running False self.detection_results deque(maxlen30) # 保存最近30帧结果 self.fps 0 self.frame_count 0 self.start_time time.time() def capture_screen(self): 捕获屏幕图像 if self.screen_region: screenshot pyautogui.screenshot(regionself.screen_region) else: screenshot pyautogui.screenshot() frame np.array(screenshot) frame cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) return frame def detect_players(self, frame): 检测游戏人物 results self.model(frame, confself.confidence_threshold, verboseFalse) return results[0] if results else None def draw_detections(self, frame, results): 在图像上绘制检测结果 if results and results.boxes is not None: boxes results.boxes for box in boxes: # 获取框坐标 x1, y1, x2, y2 box.xyxy[0].cpu().numpy() confidence box.conf[0].cpu().numpy() class_id int(box.cls[0].cpu().numpy()) # 绘制边界框 cv2.rectangle(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2) # 添加标签 label f{results.names[class_id]}: {confidence:.2f} cv2.putText(frame, label, (int(x1), int(y1)-10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2) # 显示FPS cv2.putText(frame, fFPS: {self.fps:.1f}, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2) return frame def calculate_fps(self): 计算实时FPS self.frame_count 1 if self.frame_count 30: end_time time.time() self.fps self.frame_count / (end_time - self.start_time) self.frame_count 0 self.start_time end_time def run_detection(self): 主检测循环 self.running True print(开始实时检测...) while self.running: try: # 捕获屏幕 frame self.capture_screen() # 人物检测 results self.detect_players(frame) # 绘制结果 if results: frame self.draw_detections(frame, results) self.detection_results.append(results) # 计算FPS self.calculate_fps() # 显示结果 cv2.imshow(Apex Player Detection, frame) # 退出检测 if cv2.waitKey(1) 0xFF ord(q): break except Exception as e: print(f检测错误: {e}) break cv2.destroyAllWindows() self.running False def start(self): 启动检测线程 detection_thread threading.Thread(targetself.run_detection) detection_thread.daemon True detection_thread.start() def stop(self): 停止检测 self.running False # 使用示例 if __name__ __main__: detector ApexRealtimeDetector( model_pathruns/detect/train/weights/best.pt, confidence_threshold0.6, screen_region(0, 0, 1920, 1080) # 根据实际游戏窗口调整 ) detector.start() # 保持主线程运行 try: while detector.running: time.sleep(1) except KeyboardInterrupt: detector.stop()5.2 性能优化技巧针对游戏实时检测的特殊需求提供多种优化方案# performance_optimization.py import torch from ultralytics import YOLO class OptimizedDetector: def __init__(self, model_path, optimization_levelbalanced): 优化版检测器 Args: optimization_level: speed速度优先, balanced平衡, accuracy精度优先 self.model YOLO(model_path) self.optimization_level optimization_level self.setup_optimization() def setup_optimization(self): 根据优化级别配置模型 if self.optimization_level speed: # 速度优先配置 self.inference_size 320 # 较小输入尺寸 self.confidence_threshold 0.7 # 较高置信度阈值 self.iou_threshold 0.4 # 较低IOU阈值 self.half_precision True # 半精度推理 elif self.optimization_level balanced: # 平衡配置 self.inference_size 640 self.confidence_threshold 0.5 self.iou_threshold 0.5 self.half_precision False else: # accuracy # 精度优先配置 self.inference_size 1280 self.confidence_threshold 0.3 self.iou_threshold 0.6 self.half_precision False # 启用GPU加速 if torch.cuda.is_available(): self.model.to(cuda) if self.half_precision: self.model.model.half() def warmup_model(self, warmup_iters100): 模型预热避免首次推理延迟 print(正在进行模型预热...) dummy_input torch.randn(1, 3, self.inference_size, self.inference_size) if torch.cuda.is_available(): dummy_input dummy_input.cuda() if self.half_precision: dummy_input dummy_input.half() for _ in range(warmup_iters): with torch.no_grad(): _ self.model.model(dummy_input) print(模型预热完成) def optimized_detect(self, image): 优化后的检测方法 results self.model( image, imgszself.inference_size, confself.confidence_threshold, iouself.iou_threshold, verboseFalse ) return results[0] if results else None # 使用示例 optimized_detector OptimizedDetector(best.pt, optimization_levelspeed) optimized_detector.warmup_model()6. 图形用户界面开发6.1 使用PyQt5创建检测界面开发用户友好的图形界面# gui_interface.py import sys import cv2 from PyQt5.QtWidgets import (QApplication, QMainWindow, QVBoxLayout, QHBoxLayout, QPushButton, QLabel, QSlider, QGroupBox, QTextEdit, QFileDialog, QWidget) from PyQt5.QtCore import QTimer, Qt, pyqtSignal, QThread from PyQt5.QtGui import QImage, QPixmap import numpy as np from realtime_detection import ApexRealtimeDetector class DetectionThread(QThread): 检测线程避免界面卡顿 frame_ready pyqtSignal(np.ndarray) def __init__(self, detector): super().__init__() self.detector detector self.running False def run(self): self.running True while self.running: frame self.detector.capture_screen() results self.detector.detect_players(frame) if results: frame self.detector.draw_detections(frame, results) self.frame_ready.emit(frame) QThread.msleep(33) # 约30FPS def stop(self): self.running False class MainWindow(QMainWindow): def __init__(self): super().__init__() self.detector None self.detection_thread None self.init_ui() def init_ui(self): 初始化用户界面 self.setWindowTitle(Apex游戏人物检测系统) self.setGeometry(100, 100, 1200, 800) # 中央部件 central_widget QWidget() self.setCentralWidget(central_widget) # 主布局 main_layout QHBoxLayout() central_widget.setLayout(main_layout) # 左侧控制面板 control_panel self.create_control_panel() main_layout.addWidget(control_panel, 1) # 右侧视频显示 video_panel self.create_video_panel() main_layout.addWidget(video_panel, 3) def create_control_panel(self): 创建控制面板 panel QGroupBox(控制面板) layout QVBoxLayout() # 模型加载按钮 self.load_model_btn QPushButton(加载模型) self.load_model_btn.clicked.connect(self.load_model) layout.addWidget(self.load_model_btn) # 开始/停止检测按钮 self.start_btn QPushButton(开始检测) self.start_btn.clicked.connect(self.toggle_detection) self.start_btn.setEnabled(False) layout.addWidget(self.start_btn) # 置信度滑块 confidence_layout QHBoxLayout() confidence_layout.addWidget(QLabel(置信度阈值:)) self.confidence_slider QSlider(Qt.Horizontal) self.confidence_slider.setRange(30, 90) # 0.3-0.9 self.confidence_slider.setValue(50) # 默认0.5 self.confidence_slider.valueChanged.connect(self.update_confidence) confidence_layout.addWidget(self.confidence_slider) self.confidence_label QLabel(0.50) confidence_layout.addWidget(self.confidence_label) layout.addLayout(confidence_layout) # 统计信息 self.stats_text QTextEdit() self.stats_text.setMaximumHeight(200) layout.addWidget(QLabel(检测统计:)) layout.addWidget(self.stats_text) panel.setLayout(layout) return panel def create_video_panel(self): 创建视频显示面板 panel QGroupBox(实时检测画面) layout QVBoxLayout() self.video_label QLabel() self.video_label.setAlignment(Qt.AlignCenter) self.video_label.setMinimumSize(640, 480) self.video_label.setText(等待开始检测...) layout.addWidget(self.video_label) panel.setLayout(layout) return panel def load_model(self): 加载模型文件 model_path, _ QFileDialog.getOpenFileName( self, 选择YOLOv8模型文件, , 模型文件 (*.pt) ) if model_path: try: self.detector ApexRealtimeDetector(model_path) self.start_btn.setEnabled(True) self.stats_text.append(f模型加载成功: {model_path}) except Exception as e: self.stats_text.append(f模型加载失败: {str(e)}) def toggle_detection(self): 切换检测状态 if self.detection_thread and self.detection_thread.isRunning(): self.stop_detection() else: self.start_detection() def start_detection(self): 开始检测 if self.detector: self.detection_thread DetectionThread(self.detector) self.detection_thread.frame_ready.connect(self.update_video_frame) self.detection_thread.start() self.start_btn.setText(停止检测) self.stats_text.append(检测已启动) def stop_detection(self): 停止检测 if self.detection_thread: self.detection_thread.stop() self.detection_thread.wait() self.start_btn.setText(开始检测) self.stats_text.append(检测已停止) def update_video_frame(self, frame): 更新视频帧显示 # 转换OpenCV图像为Qt图像 rgb_image cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) h, w, ch rgb_image.shape bytes_per_line ch * w qt_image QImage(rgb_image.data, w, h, bytes_per_line, QImage.Format_RGB888) pixmap QPixmap.fromImage(qt_image) # 缩放显示 scaled_pixmap pixmap.scaled( self.video_label.width(), self.video_label.height(), Qt.KeepAspectRatio, Qt.SmoothTransformation ) self.video_label.setPixmap(scaled_pixmap) def update_confidence(self, value): 更新置信度阈值 confidence value / 100.0 self.confidence_label.setText(f{confidence:.2f}) if self.detector: self.detector.confidence_threshold confidence def main(): app QApplication(sys.argv) window MainWindow() window.show() sys.exit(app.exec_()) if __name__ __main__: main()6.2 界面功能扩展为GUI添加更多实用功能# advanced_features.py import json import datetime from pathlib import Path class AdvancedDetectionSystem: def __init__(self, gui_window): self.gui gui_window self.detection_history [] self.recording False self.video_writer None def start_recording(self, output_path): 开始录制检测视频 fourcc cv2.VideoWriter_fourcc(*XVID) frame_size (1920, 1080) # 根据实际调整 self.video_writer cv2.VideoWriter(output_path, fourcc, 30.0, frame_size) self.recording True def stop_recording(self): 停止录制 if self.video_writer: self.video_writer.release() self.video_writer None self.recording False def save_detection_data(self, results, timestamp): 保存检测数据用于后续分析 detection_data { timestamp: timestamp.isoformat(), detections: [] } if results and results.boxes is not None: for box in results.boxes: detection { class: results.names[int(box.cls[0])], confidence: float(box.conf[0]), bbox: box.xyxy[0].cpu().numpy().tolist() } detection_data[detections].append(detection) self.detection_history.append(detection_data) # 定期保存到文件 if len(self.detection_history) 100: self.flush_detection_data() def flush_detection_data(self): 将检测数据写入文件 if self.detection_history: timestamp datetime.datetime.now().strftime(%Y%m%d_%H%M%S) filename fdetection_log_{timestamp}.json with open(filename, w) as f: json.dump(self.detection_history, f, indent2) self.detection_history.clear()7. 模型部署与优化7.1 模型导出与格式转换将训练好的模型转换为各种部署格式# model_export.py from ultralytics import YOLO import torch def export_model(model_path, export_formats[onnx, torchscript]): 导出模型为不同格式 model YOLO(model_path) for format in export_formats: try: if format onnx: # 导出为ONNX格式 model.export(formatonnx, dynamicTrue, simplifyTrue) print(ONNX导出成功) elif format torchscript: # 导出为TorchScript格式 model.export(formattorchscript) print(TorchScript导出成功) elif format tensorrt: # 导出为TensorRT格式需要GPU if torch.cuda.is_available(): model.export(formatengine, halfTrue) print(TensorRT导出成功) else: print(TensorRT导出需要GPU支持) except Exception as e: print(f{format}导出失败: {e}) # 导出示例 export_model(runs/detect/train/weights/best.pt, [onnx, torchscript])7.2 移动端部署考虑针对移动设备的优化方案# mobile_optimization.py def optimize_for_mobile(model_path, output_path): 为移动端优化模型 model YOLO(model_path) # 使用更小的输入尺寸 model.export( formatonnx, imgsz320, # 移动端使用较小尺寸 dynamicFalse, # 固定尺寸提高性能 simplifyTrue ) # 额外的移动端优化 optimized_model optimize_onnx_model(f{model_path[:-3]}_mobile.onnx) return optimized_model def optimize_onnx_model(onnx_path): 使用ONNX Runtime工具优化模型 import onnxruntime as ort from onnxruntime.transformers import optimizer # 基础优化 optimized_model optimizer.optimize_model(onnx_path) # 移动端特定优化 optimized_model.optimize_for_fixed_batch
