From cb6bdc7eb86d7455b11bc8ec83c043c992d528d9 Mon Sep 17 00:00:00 2001 From: yukun-hh Date: Thu, 16 Apr 2026 14:12:25 +0800 Subject: [PATCH] =?UTF-8?q?=E6=A8=A1=E5=9E=8B=E6=8F=90=E5=8D=87=E5=88=B050?= =?UTF-8?q?=E5=B1=82?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- Model.py | 184 +++++++++++++++++++++++++++++++++---------------------- Train.py | 7 +-- 2 files changed, 113 insertions(+), 78 deletions(-) diff --git a/Model.py b/Model.py index 313b2d0..a28aa78 100644 --- a/Model.py +++ b/Model.py @@ -1,10 +1,8 @@ """ -这个文件是模型的定义文件，请不要擅自修改，如有疑问微信群里反馈 -单独运行本文件将会输出模型结构 -目前的话是一个36层的模型，模型总量应该是在80M左右 如果到时候还是欠拟合的话再考虑去做更深的结构 +模型定义文件 - 使用瓶颈结构 (Bottleneck) 的深度残差网络 +目标：约50层，参数量约80M author : yukun-hh date : 2026-4-10 - """ import torch from torch import nn @@ -12,97 +10,135 @@ from torch.nn import functional as F from torchsummary import summary -# 残差块 -class Resblock(nn.Module): - def __init__(self, input_channels, output_channels, use_1x1conv=False, strides=1): +class Bottleneck(nn.Module): + """ + 瓶颈残差块：1x1(降维) -> 3x3 -> 1x1(升维) + 若需要下采样或通道变化，则在跳跃连接中使用1x1卷积 + """ + expansion = 4 # 输出通道是中间通道的4倍 + + def __init__(self, in_channels, mid_channels, stride=1, downsample=None): """ - :param input_channels: 进入残差块时的原通道 - :param output_channels: 输出时的通道数 - :param use_1x1conv: 如果输入和输出通道不相等时，需要用一个1x1的卷积层对原来的输入进行一个通道提升 - :param strides: 默认1，如果大于1起到缩小张量的作用 + :param in_channels: 输入通道数 + :param mid_channels: 中间层通道数（1x1降维后的通道数） + :param stride: 步长，用于下采样 + :param downsample: 下采样模块（当stride≠1或通道变化时使用） """ super().__init__() - self.conv1 = nn.Conv2d(input_channels, output_channels, kernel_size=3, padding=1, stride=strides) - self.conv2 = nn.Conv2d(output_channels, output_channels, kernel_size=3, padding=1, stride=1) - if use_1x1conv: - self.conv3 = nn.Conv2d(input_channels, output_channels, kernel_size=1, stride=strides) - else: - self.conv3 = None - self.bn1 = nn.BatchNorm2d(output_channels) - self.bn2 = nn.BatchNorm2d(output_channels) + self.conv1 = nn.Conv2d(in_channels, mid_channels, kernel_size=1, bias=False) + self.bn1 = nn.BatchNorm2d(mid_channels) + self.conv2 = nn.Conv2d(mid_channels, mid_channels, kernel_size=3, stride=stride, padding=1, bias=False) + self.bn2 = nn.BatchNorm2d(mid_channels) + self.conv3 = nn.Conv2d(mid_channels, mid_channels * self.expansion, kernel_size=1, bias=False) + self.bn3 = nn.BatchNorm2d(mid_channels * self.expansion) + self.relu = nn.ReLU(inplace=True) + self.downsample = downsample - def forward(self, X): - Y = F.relu(self.bn1(self.conv1(X))) - Y = self.bn2(self.conv2(Y)) - if self.conv3 is not None: - X = self.conv3(X) - Y += X - return F.relu(Y) + def forward(self, x): + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample is not None: + identity = self.downsample(x) + + out += identity + out = self.relu(out) + return out class Net(nn.Module): """ - 模型的主要结构就在这里了，到时也好该和调用 - 现在必须实现的方法： - 目前还是以图片缩放到256＊256构建残差块 + 基于 Bottleneck 的 ResNet 风格模型 + 各阶段配置仿照 ResNet-50，适当调整宽度以达到约80M参数 """ - def __init__(self): + def __init__(self, num_classes=4): super().__init__() - # 定义残差块的辅助方法 - def resnet_block(input_channels, num_channels, num_residuals, first_block=False): - """ - :param input_channels: 输入维度 - :param num_channels: 输出维度 - :param num_residuals: 单个残差层的残差块数 - :param first_block: 第一块不用下采样 特殊控制 - :return: list[nn.Module] - """ - blk = [] - for i in range(num_residuals): - if i == 0 and not first_block: - blk.append(Resblock(input_channels, num_channels, use_1x1conv=True, strides=2)) - else: - blk.append(Resblock(num_channels, num_channels)) - return blk + # 第一阶段：7x7卷积 + 最大池化 + self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) + self.bn1 = nn.BatchNorm2d(64) + self.relu = nn.ReLU(inplace=True) + self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) - # 构建网络各层 - self.b1 = nn.Sequential( - nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3), - nn.BatchNorm2d(64), - nn.ReLU(), - nn.MaxPool2d(kernel_size=3, stride=2, padding=1) - ) - """ - 7×7 卷积层，输出通道 64，步长 2，填充 3 - (3×256×256)->(64×128×128) - 批归一化 relu层 - 最大池化 - (64×128×128)->(64×64×64) - """ - self.b2 = nn.Sequential(*resnet_block(64, 64, num_residuals=3, first_block=True)) - self.b3 = nn.Sequential(*resnet_block(64, 128, num_residuals=4)) - self.b4 = nn.Sequential(*resnet_block(128, 256, num_residuals=6)) - self.b5 = nn.Sequential(*resnet_block(256, 512, num_residuals=3)) + # 残差阶段定义 + # 每个阶段的参数：[块数, 中间通道数, 步长] + # 为了达到80M参数，我们略微加宽网络（相比标准ResNet-50） + layers_config = [ + (3, 64, 1), # stage2: 3个瓶颈块，输出通道 64*4=256 + (4, 128, 2), # stage3: 4个瓶颈块，输出通道 128*4=512 + (14, 256, 2), # stage4: 14个瓶颈块，输出通道 256*4=1024（加深至此阶段） + (3, 512, 2) # stage5: 3个瓶颈块，输出通道 512*4=2048 + ] + self.in_channels = 64 + self.stage2 = self._make_layer(layers_config[0]) + self.stage3 = self._make_layer(layers_config[1]) + self.stage4 = self._make_layer(layers_config[2]) + self.stage5 = self._make_layer(layers_config[3]) + + # 全局池化与分类层 self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) - self.flatten = nn.Flatten() - self.fc = nn.Linear(512, 4) + self.fc = nn.Linear(2048, num_classes) + + def _make_layer(self, config): + """ + 构建一个残差阶段 + :param config: (块数, 中间通道数, 第一阶段步长) + :return: nn.Sequential + """ + num_blocks, mid_channels, stride = config + downsample = None + layers = [] + + # 第一个块可能需要下采样和通道匹配 + if stride != 1 or self.in_channels != mid_channels * Bottleneck.expansion: + downsample = nn.Sequential( + nn.Conv2d(self.in_channels, mid_channels * Bottleneck.expansion, + kernel_size=1, stride=stride, bias=False), + nn.BatchNorm2d(mid_channels * Bottleneck.expansion), + ) + + layers.append( + Bottleneck(self.in_channels, mid_channels, stride, downsample) + ) + self.in_channels = mid_channels * Bottleneck.expansion + + # 后续块 + for _ in range(1, num_blocks): + layers.append( + Bottleneck(self.in_channels, mid_channels) + ) + + return nn.Sequential(*layers) def forward(self, x): - x = self.b1(x) - x = self.b2(x) - x = self.b3(x) - x = self.b4(x) - x = self.b5(x) + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + x = self.maxpool(x) + + x = self.stage2(x) + x = self.stage3(x) + x = self.stage4(x) + x = self.stage5(x) + x = self.avgpool(x) - x = self.flatten(x) + x = torch.flatten(x, 1) x = self.fc(x) return x if __name__ == '__main__': - model = Net() - # 使用 torchsummary 查看模型结构 + model = Net(num_classes=4) summary(model, input_size=(3, 256, 256)) \ No newline at end of file diff --git a/Train.py b/Train.py index 6a71ae2..8b75e7a 100644 --- a/Train.py +++ b/Train.py @@ -80,7 +80,6 @@ def train(model, train_loader, val_loader, epochs=50, lr=0.001, device='cuda'): # 1. 定义损失函数和优化器 criterion = nn.CrossEntropyLoss() # 多分类用交叉熵 - # 优化器选择（推荐 Adam 或 SGD） # 或者使用 SGD + 动量 optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9, weight_decay=1e-4) @@ -142,15 +141,15 @@ if __name__ == '__main__': # 假设你的 dataloader 已经写好了 train_loader, val_loader, class_names = create_dataloaders( data_root='../trash_division_data/ultimate_4_class/', # 与trash-division同级文件夹 - batch_size=32, # 根据你的显存调整 + batch_size=16, # 根据你的显存调整 image_size=256, # 与你模型输入一致 - num_workers=4, # Windows 可能需设为 0 + num_workers=8, # Windows 可能需设为 0 augment=True # 训练时使用数据增强 ) # 1. 创建模型 device = torch.device('cuda' if torch.cuda.is_available() else 'xpu' if torch.xpu.is_available() else 'cpu') - model = Net() # 根据你的 Net 类调整 + model = Net(num_classes=4) # 根据你的 Net 类调整 model = model.to(device) # 打印模型信息