Resnet18训练CIFAR10 准确率95% |
您所在的位置：网站首页 › cifar10 resnet9 › Resnet18训练CIFAR10 准确率95% |
|
准确率 95.31% 几个关键点： 1、改模型：原始的resnet18首层使用的7x7的卷积核，CIFAR10图片太小不适合，要改成3x3的，步长和padding都要一并改成1。因为图太小，最大池化层也同样没用，删掉。最后一个全连接层输出改成10。 2、图片增强不要太多，只要训练集和验证集结果没有出现10%以上的差距都算不上过拟合。 3、学习率从0.1开始，10个epoch跑完loss值没有下降的话衰减50% 4、损失函数用CrossEntropyLoss 5、优化器用SGD 改模型代码： # 定义模型 model_ft = torchvision.models.resnet18(pretrained=False) # 修改模型 model_ft.conv1 = nn.Conv2d(3, 64, 3, stride=1, padding=1, bias=False) # 首层改成3x3卷积核 model_ft.maxpool = nn.MaxPool2d(1, 1, 0) # 图像太小 本来就没什么特征 所以这里通过1x1的池化核让池化层失效 num_ftrs = model_ft.fc.in_features # 获取（fc）层的输入的特征数 model_ft.fc = nn.Linear(num_ftrs, 10)这里的最大池化层实在是想不出什么好办法直接删掉，只能用这个办法让其失效 如果不想用原始的模型也能自己写个 下面是我是随便写的一个 # 实现Resnet18 """ ResNet18 是由17个卷积层和1个全连接层组成 下采样层的1x1卷积不算 池化不算 激活不算 主要思想是一个基础层 然后反复的重复这个基础层 """ import torch from torch import nn # 基础块 from torch.nn import Conv2d, BatchNorm2d, ReLU, MaxPool2d, AdaptiveAvgPool2d, Linear class BasicBlock(nn.Module): def __init__(self, in_features, out_features) -> None: super().__init__() self.in_features = in_features self.out_features = out_features stride = 1 _features = out_features if self.in_features != self.out_features: # 在输入通道和输出通道不相等的情况下计算通道是否为2倍差值 if self.out_features / self.in_features == 2.0: stride = 2 # 在输出特征是输入特征的2倍的情况下 要想参数不翻倍 步长就必须翻倍 else: raise ValueError("输出特征数最多为输入特征数的2倍！") self.conv1 = Conv2d(in_features, _features, kernel_size=3, stride=stride, padding=1, bias=False) self.bn1 = BatchNorm2d(_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) self.relu = ReLU(inplace=True) self.conv2 = Conv2d(_features, _features, kernel_size=3, stride=1, padding=1, bias=False) self.bn2 = BatchNorm2d(_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) # 下采样 self.downsample = None if self.in_features == self.out_features else nn.Sequential( Conv2d(in_features, out_features, kernel_size=1, stride=2, bias=False), BatchNorm2d(out_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) 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) # 输入输出的特征数不同时使用下采样层 if self.in_features != self.out_features: identity = self.downsample(x) # 残差求和 out += identity out = self.relu(out) return out class ResNet18(nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) self.relu = ReLU(inplace=True) # self.maxpool = MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.layer1 = nn.Sequential( BasicBlock(64, 64), BasicBlock(64, 64) ) self.layer2 = nn.Sequential( BasicBlock(64, 128), BasicBlock(128, 128) ) self.layer3 = nn.Sequential( BasicBlock(128, 256), BasicBlock(256, 256) ) self.layer4 = nn.Sequential( BasicBlock(256, 512), BasicBlock(512, 512) ) self.avgpool = AdaptiveAvgPool2d(output_size=(1, 1)) self.fc = Linear(in_features=512, out_features=10, bias=True) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) # x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) # |
| CopyRight 2018-2019 办公设备维修网 版权所有 豫ICP备15022753号-3 |