本文为365天深度学习训练营中的学习记录博客原作者K同学啊本周是学习深度学习的第15周。编译器使用的是vscode安装的是CPU版PyTorchtorch 2.12.0cpu。本周的学习内容跟J1周深度学习第J1周ResNet-50算法学习-CSDN博客相似数据集一样本次学习只给出不一样的部分。1 模型本周使用了ResNetV1和ResNetV2进行训练。ResNetV1部分class ResNetV1(nn.Module): def __init__(self, classes3): super(ResNetV1, self).__init__() self.conv1 nn.Sequential( nn.Conv2d(3, 64, 7, stride2, padding3, biasFalse, padding_modezeros), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(kernel_size3, stride2, padding1) ) self.conv2 nn.Sequential( ConvBlock(64, 3, [64, 64, 256], stride1), IdentityBlock(256, 3, [64, 64, 256]), IdentityBlock(256, 3, [64, 64, 256]) ) self.conv3 nn.Sequential( ConvBlock(256, 3, [128, 128, 512]), IdentityBlock(512, 3, [128, 128, 512]), IdentityBlock(512, 3, [128, 128, 512]), IdentityBlock(512, 3, [128, 128, 512]) ) self.conv4 nn.Sequential( ConvBlock(512, 3, [256, 256, 1024]), IdentityBlock(1024, 3, [256, 256, 1024]), IdentityBlock(1024, 3, [256, 256, 1024]), IdentityBlock(1024, 3, [256, 256, 1024]), IdentityBlock(1024, 3, [256, 256, 1024]), IdentityBlock(1024, 3, [256, 256, 1024]) ) self.conv5 nn.Sequential( ConvBlock(1024, 3, [512, 512, 2048]), IdentityBlock(2048, 3, [512, 512, 2048]), IdentityBlock(2048, 3, [512, 512, 2048]) ) self.pool nn.AvgPool2d(kernel_size7, stride7, padding0) self.fc nn.Linear(2048, classes) def forward(self, x): x self.conv1(x) x self.conv2(x) x self.conv3(x) x self.conv4(x) x self.conv5(x) x self.pool(x) x torch.flatten(x, start_dim1) x self.fc(x) return xResNetV2 的预激活残差块BN - ReLU - Conv相加后不再使用 ReLU。模型部分class PreActBlock(nn.Module): def __init__(self, in_channel, kernel_size, filters, stride1): super(PreActBlock, self).__init__() filters1, filters2, filters3 filters self.bn1 nn.BatchNorm2d(in_channel) self.conv1 nn.Conv2d(in_channel, filters1, 1, stridestride, biasFalse) self.bn2 nn.BatchNorm2d(filters1) self.conv2 nn.Conv2d(filters1, filters2, kernel_size, paddingautopad(kernel_size), biasFalse) self.bn3 nn.BatchNorm2d(filters2) self.conv3 nn.Conv2d(filters2, filters3, 1, biasFalse) self.relu nn.ReLU() self.shortcut None if stride ! 1 or in_channel ! filters3: self.shortcut nn.Conv2d(in_channel, filters3, 1, stridestride, biasFalse) def forward(self, x): preact self.relu(self.bn1(x)) shortcut x if self.shortcut is None else self.shortcut(preact) x1 self.conv1(preact) x1 self.conv2(self.relu(self.bn2(x1))) x1 self.conv3(self.relu(self.bn3(x1))) return x1 shortcut # ResNet-50 V2 class ResNetV2(nn.Module): def __init__(self, classes3): super(ResNetV2, self).__init__() self.conv1 nn.Sequential( nn.Conv2d(3, 64, 7, stride2, padding3, biasFalse), nn.MaxPool2d(kernel_size3, stride2, padding1) ) self.conv2 nn.Sequential( PreActBlock(64, 3, [64, 64, 256]), PreActBlock(256, 3, [64, 64, 256]), PreActBlock(256, 3, [64, 64, 256]) ) self.conv3 nn.Sequential( PreActBlock(256, 3, [128, 128, 512], stride2), PreActBlock(512, 3, [128, 128, 512]), PreActBlock(512, 3, [128, 128, 512]), PreActBlock(512, 3, [128, 128, 512]) ) self.conv4 nn.Sequential( PreActBlock(512, 3, [256, 256, 1024], stride2), PreActBlock(1024, 3, [256, 256, 1024]), PreActBlock(1024, 3, [256, 256, 1024]), PreActBlock(1024, 3, [256, 256, 1024]), PreActBlock(1024, 3, [256, 256, 1024]), PreActBlock(1024, 3, [256, 256, 1024]), ) self.conv5 nn.Sequential( PreActBlock(1024, 3, [512, 512, 2048], stride2), PreActBlock(2048, 3, [512, 512, 2048]), PreActBlock(2048, 3, [512, 512, 2048]) ) self.bn nn.BatchNorm2d(2048) self.relu nn.ReLU() self.pool nn.AvgPool2d(kernel_size7, stride7, padding0) self.fc nn.Linear(2048, classes) def forward(self, x): x self.conv1(x) x self.conv2(x) x self.conv3(x) x self.conv4(x) x self.conv5(x) x self.relu(self.bn(x)) x self.pool(x) x torch.flatten(x, start_dim1) x self.fc(x) return x2 训练过程train和test步骤一样绘图部分for model_name, model_class in [(ResNetV1, ResNetV1), (ResNetV2, ResNetV2)]: print(f\nTraining {model_name}) model model_class(classeslen(total_data.classes)).to(device) optimizer torch.optim.AdamW(model.parameters(),lr 1e-4) loss_fn nn.CrossEntropyLoss() epochs 10 train_loss [] train_acc [] test_loss [] test_acc [] best_acc -1 for epoch in range(epochs): model.train() epoch_train_acc,epoch_train_loss train(train_dl,model,loss_fn,optimizer) model.eval() epoch_test_acc,epoch_test_loss test(test_dl,model,loss_fn) if epoch_test_acc best_acc: best_acc epoch_test_acc best_model copy.deepcopy(model) train_acc.append(epoch_train_acc) train_loss.append(epoch_train_loss) test_acc.append(epoch_test_acc) test_loss.append(epoch_test_loss) lr optimizer.state_dict()[param_groups][0][lr] template (Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}) print(template.format(epoch1,epoch_train_acc*100,epoch_train_loss, epoch_test_acc*100,epoch_test_loss,lr)) PATH f./{model_name}_best_model.pth torch.save(best_model.state_dict(), PATH) print(Done) # Plot training and validation metrics current_time datetime.now() # 获取当前时间 epochs_range range(epochs) plt.figure(figsize(12, 3)) plt.subplot(1, 2, 1) plt.plot(epochs_range, train_acc, labelTraining Accuracy) plt.plot(epochs_range, test_acc, labelTest Accuracy) plt.legend(loclower right) plt.title(f{model_name} Accuracy) plt.xlabel(current_time) # 打卡需带出时间戳否则代码截图无效 plt.subplot(1, 2, 2) plt.plot(epochs_range, train_loss, labelTraining Loss) plt.plot(epochs_range, test_loss, labelTest Loss) plt.legend(locupper right) plt.title(f{model_name} Loss) plt.tight_layout() plt.savefig(f./{model_name}_ACC_LOSS.png) best_model.load_state_dict(torch.load(PATH,map_locationdevice)) epoch_test_acc,epoch_test_loss test(test_dl,best_model,loss_fn) print(epoch_test_acc) print(epoch_test_loss) plt.show()3 训练结果进行ResNetV1训练进行ResNetV2训练4 个人总结ResNetV1 和 ResNetV2 均采用 ResNet-50 结构主要区别在于残差块中归一化与激活操作的位置。V1 采用“卷积→BN→ReLU”的顺序并在残差相加后再次进行 ReLU 激活V2 采用“BN→ReLU→卷积”的预激活方式相加后直接输出使信息和梯度能够更顺畅地传播。不过结构上的改进并不保证 V2 在所有任务中都取得更高准确率。从本次 10 轮训练曲线来看两个模型的测试准确率均接近 90%训练损失总体下降未出现明显的持续过拟合趋势。V1 最终训练准确率约为 89.5%测试损失约为 0.26V2 最终训练准确率约为 87%测试损失约为 0.30前期波动较大、后期趋于稳定。因此本次实验中两者分类准确率接近V1 的最终损失略低但由于类别数量不均衡仍需结合各类别召回率和混淆矩阵进一步判断骨折识别效果。