Faster R-CNN PyTorch 1.2 自定义数据集训练VOC格式20类mAP 80.36%实战指南1. 环境配置与项目准备在开始训练之前我们需要确保开发环境配置正确。以下是推荐的配置步骤# 创建Python虚拟环境 conda create -n fasterrcnn python3.7 conda activate fasterrcnn # 安装PyTorch 1.2及相关依赖 pip install torch1.2.0 torchvision0.4.0 pip install opencv-python pillow matplotlib tqdm项目结构建议如下faster-rcnn-pytorch/ ├── data/ │ ├── VOCdevkit/ │ │ └── VOC2007/ │ │ ├── Annotations/ │ │ ├── JPEGImages/ │ │ └── ImageSets/ ├── model_data/ │ ├── voc_classes.txt │ └── pretrained_weights/ ├── utils/ ├── train.py ├── predict.py └── get_map.py提示建议使用Git克隆官方仓库以获取完整代码结构git clone https://github.com/bubbliiiing/faster-rcnn-pytorch2. 数据集准备与VOC格式转换Faster R-CNN通常使用PASCAL VOC格式的数据集。以下是自定义数据集转换的关键步骤目录结构规范JPEGImages/存放所有训练图片(.jpg)Annotations/存放XML格式标注文件ImageSets/Main/包含train.txt, val.txt等划分文件标注文件示例annotation filename000001.jpg/filename size width800/width height600/height depth3/depth /size object namecat/name bndbox xmin100/xmin ymin200/ymin xmax300/xmax ymax400/ymax /bndbox /object /annotation自动生成训练集/验证集划分# voc_annotation.py关键参数配置 classes_path model_data/voc_classes.txt # 你的类别定义文件 trainval_percent 0.9 # 训练验证集比例 train_percent 0.9 # 训练集比例3. 模型架构关键参数解析Faster R-CNN的核心组件需要特别关注以下配置参数参数名称推荐值作用说明anchors_size[8,16,32]控制先验框的基准大小backboneresnet50特征提取网络选择input_shape[600,600]输入图像尺寸Freeze_Epoch50冻结训练轮数UnFreeze_Epoch100解冻训练总轮数Freeze_batch_size4冻结阶段batch sizeUnfreeze_batch_size2解冻阶段batch size特征提取网络配置示例class ResNet(nn.Module): def __init__(self, block, layers, num_classes1000): self.inplanes 64 super(ResNet, self).__init__() # 初始卷积层 self.conv1 nn.Conv2d(3, 64, kernel_size7, stride2, padding3, biasFalse) self.bn1 nn.BatchNorm2d(64) self.relu nn.ReLU(inplaceTrue) self.maxpool nn.MaxPool2d(kernel_size3, stride2, padding0, ceil_modeTrue) # 四个残差块 self.layer1 self._make_layer(block, 64, layers[0]) self.layer2 self._make_layer(block, 128, layers[1], stride2) self.layer3 self._make_layer(block, 256, layers[2], stride2) self.layer4 self._make_layer(block, 512, layers[3], stride2) # ROI头部网络 self.roi_pool RoIPool((7,7), spatial_scale1/16.0)4. 训练流程优化技巧4.1 两阶段训练策略冻结训练阶段# train.py中的关键配置 Freeze_Train True Init_Epoch 0 Freeze_Epoch 50 Freeze_batch_size 4 Freeze_lr 1e-4解冻训练阶段UnFreeze_Epoch 100 Unfreeze_batch_size 2 Unfreeze_lr 1e-54.2 学习率调整策略推荐使用余弦退火学习率def get_lr(optimizer): for param_group in optimizer.param_groups: return param_group[lr] lr_scheduler optim.lr_scheduler.CosineAnnealingLR( optimizer, T_max5, eta_min1e-5 )4.3 数据增强方案# 训练集数据增强 train_transform transforms.Compose([ transforms.Resize((600,600)), transforms.ColorJitter( brightness0.2, contrast0.2, saturation0.2, hue0.1), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize( mean[0.485, 0.456, 0.406], std[0.229, 0.224, 0.225]) ]) # 验证集只需基础转换 val_transform transforms.Compose([ transforms.Resize((600,600)), transforms.ToTensor(), transforms.Normalize( mean[0.485, 0.456, 0.406], std[0.229, 0.224, 0.225]) ])5. 模型评估与结果分析5.1 mAP计算流程使用get_map.py计算mAP的步骤生成预测结果python get_dr_txt.py --model-path logs/ep100-loss0.02.pth计算mAPpython get_map.py --dr-path results/detection-results/ \ --gt-path data/VOCdevkit/VOC2007/Annotations/5.2 关键评估指标指标名称本模型结果参考基准mAP0.580.36%77.5% (VGG16)mAP0.5:0.9556.2%53.4% (VGG16)推理速度8 FPS5 FPS (VGG16)5.3 常见问题排查低召回率调整RPN的NMS阈值默认0.7高误检率增加正样本的IoU阈值默认0.5训练震荡减小学习率或增大batch size6. 模型部署与推理优化6.1 预测脚本关键参数# predict.py配置示例 _defaults { model_path: logs/ep100-loss0.02.pth, classes_path: model_data/voc_classes.txt, confidence: 0.5, # 置信度阈值 nms_iou: 0.3, # NMS IoU阈值 backbone: resnet50 }6.2 TorchScript导出# 模型导出为TorchScript model.eval() example torch.rand(1, 3, 600, 600).to(device) traced_script torch.jit.trace(model, example) traced_script.save(fasterrcnn_res50.pt)6.3 性能优化技巧TensorRT加速trtexec --onnxfasterrcnn.onnx \ --saveEnginefasterrcnn.engine \ --fp16量化压缩model torch.quantization.quantize_dynamic( model, {nn.Linear, nn.Conv2d}, dtypetorch.qint8 )7. 进阶调优方向自定义backbonefrom torchvision.models import mobilenet_v3_large backbone mobilenet_v3_large(pretrainedTrue).features backbone.out_channels 960 # 必须设置输出通道数改进RPN结构class CustomRPN(nn.Module): def __init__(self, in_channels): super().__init__() self.conv nn.Conv2d(in_channels, 512, 3, padding1) self.cls_logits nn.Conv2d(512, 3*2, 1) # 3 anchors * 2 classes self.bbox_pred nn.Conv2d(512, 3*4, 1) # 3 anchors * 4 coords def forward(self, x): logits [] bbox_reg [] for feature in x: t F.relu(self.conv(feature)) logits.append(self.cls_logits(t)) bbox_reg.append(self.bbox_pred(t)) return logits, bbox_reg损失函数改进def focal_loss(pred, target, alpha0.25, gamma2.0): BCE_loss F.binary_cross_entropy_with_logits(pred, target, reductionnone) pt torch.exp(-BCE_loss) loss alpha * (1-pt)**gamma * BCE_loss return loss.mean()在实际项目中我们发现当使用ResNet50作为backbone时将anchors_size调整为[4,8,16]对小物体检测效果提升约5%。而采用余弦退火学习率策略相比固定学习率最终mAP可提升2-3个百分点。