上篇介绍了InceptionV3论文，指路：经典神经网络论文超详细解读（四）——InceptionV2-V3学习笔记（翻译＋精读＋代码复现）
本篇我们来复现一下InceptionV3代码
InceptionV1回顾：GoogLeNet InceptionV1代码复现＋超详细注释（PyTorch）
第一步：定义基础卷积模块
卷积模块较上次多了BN层
BatchNorm2d()函数：
作用：卷积层之后总会添加BatchNorm2d进行数据的归一化处理，这使得数据在进行Relu之前不会因为数据过大而导致网络性能的不稳定
举例：
BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
1.num_features：一般输入参数的shape为batch_size*num_features*height*width，即为其中特征的数量，即为输入BN层的通道数；
2.eps：分母中添加的一个值，目的是为了计算的稳定性，默认为：1e-5,避免分母为0；
3.momentum：一个用于运行过程中均值和方差的一个估计参数（可以理解是一个稳定系数，类似于SGD中的momentum的系数）；
4.affine：当设为true时，会给定可以学习的系数矩阵gamma和beta
class BasicConv2d(nn.Module): def __init__(self, in_channels, out_channels, **kwargs): super(BasicConv2d, self).__init__() self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs) self.bn = nn.BatchNorm2d(out_channels, eps=0.001) def forward(self, x): x = self.conv(x) x = self.bn(x) return F.relu(x, inplace=True)
第二步：定义Inceptionv3模块
PyTorch提供的有六种基本的Inception模块，分别是InceptionA——InceptionE。
InceptionA
得到输入大小不变，通道数为224+pool_features的特征图。
假如输入为(35, 35, 192)的数据：
第一个branch：
经过branch1x1为带有64个1*1的卷积核，所以生成第一张特征图(35, 35, 64)；
第二个branch：
首先经过branch5x5_1为带有48个1*1的卷积核，所以第二张特征图(35, 35, 48)，
然后经过branch5x5_2为带有64个5*5大小且填充为2的卷积核，特征图大小依旧不变，因此第二张特征图最终为(35, 35, 64)；
第三个branch：
首先经过branch3x3dbl_1为带有64个1*1的卷积核，所以第三张特征图(35, 35, 64)，
然后经过branch3x3dbl_2为带有96个3*3大小且填充为1的卷积核，特征图大小依旧不变，因此进一步生成第三张特征图(35, 35, 96)，
最后经过branch3x3dbl_3为带有96个3*3大小且填充为1的卷积核，特征图大小和通道数不变，因此第三张特征图最终为(35, 35, 96)；
第四个branch：
首先经过avg_pool2d，其中池化核3*3，步长为1，填充为1，所以第四张特征图大小不变，通道数不变，第四张特征图为(35, 35, 192)，
然后经过branch_pool为带有pool_features个的1*1卷积，因此第四张特征图最终为(35, 35, pool_features)；
最后将四张特征图进行拼接，最终得到(35，35，64+64+96+pool_features)的特征图。
'''---InceptionA---''' class InceptionA(nn.Module): def __init__(self, in_channels, pool_features, conv_block=None): super(InceptionA, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 64, kernel_size=1) self.branch5x5_1 = conv_block(in_channels, 48, kernel_size=1) self.branch5x5_2 = conv_block(48, 64, kernel_size=5, padding=2) self.branch3x3dbl_1 = conv_block(in_channels, 64, kernel_size=1) self.branch3x3dbl_2 = conv_block(64, 96, kernel_size=3, padding=1) self.branch3x3dbl_3 = conv_block(96, 96, kernel_size=3, padding=1) self.branch_pool = conv_block(in_channels, pool_features, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch5x5 = self.branch5x5_1(x) branch5x5 = self.branch5x5_2(branch5x5) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1)
InceptionB
得到输入大小减半，通道数为480的特征图，
假如输入为(35, 35, 288)的数据：
第一个branch：
经过branch1x1为带有384个3*3大小且步长2的卷积核，(35-3+2*0)/2+1=17所以生成第一张特征图(17, 17, 384)；
第二个branch：
首先经过branch3x3dbl_1为带有64个1*1的卷积核，特征图大小不变，即(35, 35, 64)；
然后经过branch3x3dbl_2为带有96个3*3大小填充1的卷积核，特征图大小不变，即(35, 35, 96)，
再经过branch3x3dbl_3为带有96个3*3大小步长2的卷积核，(35-3+2*0)/2+1=17，即第二张特征图为(17, 17, 96)；
第三个branch：
经过max_pool2d，池化核大小3*3，步长为2，所以是二倍最大值下采样，通道数保持不变，第三张特征图为(17, 17, 288)；
最后将三张特征图进行拼接，最终得到(17(即Hin/2)，17(即Win/2)，384+96+288(Cin)=768)的特征图。
'''---InceptionB---''' class InceptionB(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionB, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch3x3 = conv_block(in_channels, 384, kernel_size=3, stride=2) self.branch3x3dbl_1 = conv_block(in_channels, 64, kernel_size=1) self.branch3x3dbl_2 = conv_block(64, 96, kernel_size=3, padding=1) self.branch3x3dbl_3 = conv_block(96, 96, kernel_size=3, stride=2) def _forward(self, x): branch3x3 = self.branch3x3(x) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = F.max_pool2d(x, kernel_size=3, stride=2) outputs = [branch3x3, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1)
InceptionC
得到输入大小不变，通道数为768的特征图。
假如输入为(17,17, 768)的数据：
第一个branch：
首先经过branch1x1为带有192个1*1的卷积核，所以生成第一张特征图(17,17, 192)；
第二个branch：
首先经过branch7x7_1为带有c7个1*1的卷积核，所以第二张特征图(17,17, c7)，
然后经过branch7x7_2为带有c7个1*7大小且填充为0*3的卷积核，特征图大小不变，进一步生成第二张特征图(17,17, c7)，
然后经过branch7x7_3为带有192个7*1大小且填充为3*0的卷积核，特征图大小不变，进一步生成第二张特征图(17,17, 192)，因此第二张特征图最终为(17,17, 192)；
第三个branch：
首先经过branch7x7dbl_1为带有c7个1*1的卷积核，所以第三张特征图(17,17, c7)，
然后经过branch7x7dbl_2为带有c7个7*1大小且填充为3*0的卷积核，特征图大小不变，进一步生成第三张特征图(17,17, c7)，
然后经过branch7x7dbl_3为带有c7个1*7大小且填充为0*3的卷积核，特征图大小不变，进一步生成第三张特征图(17,17, c7)，
然后经过branch7x7dbl_4为带有c7个7*1大小且填充为3*0的卷积核，特征图大小不变，进一步生成第三张特征图(17,17, c7)，
然后经过branch7x7dbl_5为带有192个1*7大小且填充为0*3的卷积核，特征图大小不变，因此第二张特征图最终为(17,17, 192)；
第四个branch：
首先经过avg_pool2d，其中池化核3*3，步长为1，填充为1，所以第四张特征图大小不变，通道数不变，第四张特征图为(17,17, 768)，
然后经过branch_pool为带有192个的1*1卷积，因此第四张特征图最终为(17,17, 192)；
最后将四张特征图进行拼接，最终得到(17, 17, 192+192+192+192=768)的特征图。
'''---InceptionC---''' class InceptionC(nn.Module): def __init__(self, in_channels, channels_7x7, conv_block=None): super(InceptionC, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 192, kernel_size=1) c7 = channels_7x7 self.branch7x7_1 = conv_block(in_channels, c7, kernel_size=1) self.branch7x7_2 = conv_block(c7, c7, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7_3 = conv_block(c7, 192, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_1 = conv_block(in_channels, c7, kernel_size=1) self.branch7x7dbl_2 = conv_block(c7, c7, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_3 = conv_block(c7, c7, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7dbl_4 = conv_block(c7, c7, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_5 = conv_block(c7, 192, kernel_size=(1, 7), padding=(0, 3)) self.branch_pool = conv_block(in_channels, 192, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch7x7 = self.branch7x7_1(x) branch7x7 = self.branch7x7_2(branch7x7) branch7x7 = self.branch7x7_3(branch7x7) branch7x7dbl = self.branch7x7dbl_1(x) branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch7x7, branch7x7dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1)
InceptionD
得到输入大小减半，通道数512的特征图，
假如输入为(17, 17, 768)的数据：
第一个branch：
首先经过branch3x3_1为带有192个1*1的卷积核，所以生成第一张特征图(17, 17, 192)；
然后经过branch3x3_2为带有320个3*3大小步长为2的卷积核，(17-3+2*0)/2+1=8，最终第一张特征图(8, 8, 320)；
第二个branch：
首先经过branch7x7x3_1为带有192个1*1的卷积核，特征图大小不变，即(17, 17, 192)；
然后经过branch7x7x3_2为带有192个1*7大小且填充为0*3的卷积核，特征图大小不变，进一步生成第三张特征图(17,17, 192)；
再经过branch7x7x3_3为带有192个7*1大小且填充为3*0的卷积核，特征图大小不变，进一步生成第三张特征图(17,17, 192)；
最后经过branch7x7x3_4为带有192个3*3大小步长为2的卷积核，最终第一张特征图(8, 8, 192)；
第三个branch：
首先经过max_pool2d，池化核大小3*3，步长为2，所以是二倍最大值下采样，通道数保持不变，第三张特征图为(8, 8, 768)；
最后将三张特征图进行拼接，最终得到(8(即Hin/2)，8(即Win/2)，320+192+768(Cin)=1280)的特征图。
'''---InceptionD---''' class InceptionD(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionD, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch3x3_1 = conv_block(in_channels, 192, kernel_size=1) self.branch3x3_2 = conv_block(192, 320, kernel_size=3, stride=2) self.branch7x7x3_1 = conv_block(in_channels, 192, kernel_size=1) self.branch7x7x3_2 = conv_block(192, 192, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7x3_3 = conv_block(192, 192, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7x3_4 = conv_block(192, 192, kernel_size=3, stride=2) def _forward(self, x): branch3x3 = self.branch3x3_1(x) branch3x3 = self.branch3x3_2(branch3x3) branch7x7x3 = self.branch7x7x3_1(x) branch7x7x3 = self.branch7x7x3_2(branch7x7x3) branch7x7x3 = self.branch7x7x3_3(branch7x7x3) branch7x7x3 = self.branch7x7x3_4(branch7x7x3) branch_pool = F.max_pool2d(x, kernel_size=3, stride=2) outputs = [branch3x3, branch7x7x3, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1)
InceptionE
最终得到输入大小不变，通道数为2048的特征图。
假如输入为(8,8, 1280)的数据：
第一个branch：
首先经过branch1x1为带有320个1*1的卷积核，所以生成第一张特征图(8, 8, 320)；
第二个branch：
首先经过branch3x3_1为带有384个1*1的卷积核，所以第二张特征图(8, 8, 384)，
经过分支branch3x3_2a为带有384个1*3大小且填充为0*1的卷积核，特征图大小不变，进一步生成特征图(8,8, 384)，
经过分支branch3x3_2b为带有192个3*1大小且填充为1*0的卷积核，特征图大小不变，进一步生成特征图(8,8, 384)，
因此第二张特征图最终为两个分支拼接(8,8, 384+384=768)；
第三个branch：
首先经过branch3x3dbl_1为带有448个1*1的卷积核，所以第三张特征图(8,8, 448)，
然后经过branch3x3dbl_2为带有384个3*3大小且填充为1的卷积核，特征图大小不变，进一步生成第三张特征图(8,8, 384)，
经过分支branch3x3dbl_3a为带有384个1*3大小且填充为0*1的卷积核，特征图大小不变，进一步生成特征图(8,8, 384)，
经过分支branch3x3dbl_3b为带有384个3*1大小且填充为1*0的卷积核，特征图大小不变，进一步生成特征图(8,8, 384)，
因此第三张特征图最终为两个分支拼接(8,8, 384+384=768)；
第四个branch：
首先经过avg_pool2d，其中池化核3*3，步长为1，填充为1，所以第四张特征图大小不变，通道数不变，第四张特征图为(8,8, 1280)，
然后经过branch_pool为带有192个的1*1卷积，因此第四张特征图最终为(8,8, 192)；
最后将四张特征图进行拼接，最终得到(8, 8, 320+768+768+192=2048)的特征图。
'''---InceptionE---''' class InceptionE(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionE, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 320, kernel_size=1) self.branch3x3_1 = conv_block(in_channels, 384, kernel_size=1) self.branch3x3_2a = conv_block(384, 384, kernel_size=(1, 3), padding=(0, 1)) self.branch3x3_2b = conv_block(384, 384, kernel_size=(3, 1), padding=(1, 0)) self.branch3x3dbl_1 = conv_block(in_channels, 448, kernel_size=1) self.branch3x3dbl_2 = conv_block(448, 384, kernel_size=3, padding=1) self.branch3x3dbl_3a = conv_block(384, 384, kernel_size=(1, 3), padding=(0, 1)) self.branch3x3dbl_3b = conv_block(384, 384, kernel_size=(3, 1), padding=(1, 0)) self.branch_pool = conv_block(in_channels, 192, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch3x3 = self.branch3x3_1(x) branch3x3 = [ self.branch3x3_2a(branch3x3), self.branch3x3_2b(branch3x3), ] branch3x3 = torch.cat(branch3x3, 1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = [ self.branch3x3dbl_3a(branch3x3dbl), self.branch3x3dbl_3b(branch3x3dbl), ] branch3x3dbl = torch.cat(branch3x3dbl, 1) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1)
第三步：定义辅助分类器InceptionAux
class InceptionAux(nn.Module): def __init__(self, in_channels, num_classes, conv_block=None): super(InceptionAux, self).__init__() if conv_block is None: conv_block = BasicConv2d self.conv0 = conv_block(in_channels, 128, kernel_size=1) self.conv1 = conv_block(128, 768, kernel_size=5) self.conv1.stddev = 0.01 self.fc = nn.Linear(768, num_classes) self.fc.stddev = 0.001 def forward(self, x): # N x 768 x 17 x 17 x = F.avg_pool2d(x, kernel_size=5, stride=3) # N x 768 x 5 x 5 x = self.conv0(x) # N x 128 x 5 x 5 x = self.conv1(x) # N x 768 x 1 x 1 # Adaptive average pooling x = F.adaptive_avg_pool2d(x, (1, 1)) # N x 768 x 1 x 1 x = torch.flatten(x, 1) # N x 768 x = self.fc(x) # N x 1000 return x
第四步：搭建GoogLeNet网络
'''-----------------------搭建GoogLeNet网络--------------------------''' class GoogLeNet(nn.Module): def __init__(self, num_classes=1000, aux_logits=True, transform_input=False, inception_blocks=None): super(GoogLeNet, self).__init__() if inception_blocks is None: inception_blocks = [ BasicConv2d, InceptionA, InceptionB, InceptionC, InceptionD, InceptionE, InceptionAux ] assert len(inception_blocks) == 7 conv_block = inception_blocks[0] inception_a = inception_blocks[1] inception_b = inception_blocks[2] inception_c = inception_blocks[3] inception_d = inception_blocks[4] inception_e = inception_blocks[5] inception_aux = inception_blocks[6] self.aux_logits = aux_logits self.transform_input = transform_input self.Conv2d_1a_3x3 = conv_block(3, 32, kernel_size=3, stride=2) self.Conv2d_2a_3x3 = conv_block(32, 32, kernel_size=3) self.Conv2d_2b_3x3 = conv_block(32, 64, kernel_size=3, padding=1) self.Conv2d_3b_1x1 = conv_block(64, 80, kernel_size=1) self.Conv2d_4a_3x3 = conv_block(80, 192, kernel_size=3) self.Mixed_5b = inception_a(192, pool_features=32) self.Mixed_5c = inception_a(256, pool_features=64) self.Mixed_5d = inception_a(288, pool_features=64) self.Mixed_6a = inception_b(288) self.Mixed_6b = inception_c(768, channels_7x7=128) self.Mixed_6c = inception_c(768, channels_7x7=160) self.Mixed_6d = inception_c(768, channels_7x7=160) self.Mixed_6e = inception_c(768, channels_7x7=192) if aux_logits: self.AuxLogits = inception_aux(768, num_classes) self.Mixed_7a = inception_d(768) self.Mixed_7b = inception_e(1280) self.Mixed_7c = inception_e(2048) self.fc = nn.Linear(2048, num_classes) '''输入(229,229,3)的数据，首先归一化输入，经过5个卷积，2个最大池化层。''' def _forward(self, x): # N x 3 x 299 x 299 x = self.Conv2d_1a_3x3(x) # N x 32 x 149 x 149 x = self.Conv2d_2a_3x3(x) # N x 32 x 147 x 147 x = self.Conv2d_2b_3x3(x) # N x 64 x 147 x 147 x = F.max_pool2d(x, kernel_size=3, stride=2) # N x 64 x 73 x 73 x = self.Conv2d_3b_1x1(x) # N x 80 x 73 x 73 x = self.Conv2d_4a_3x3(x) # N x 192 x 71 x 71 x = F.max_pool2d(x, kernel_size=3, stride=2) '''然后经过3个InceptionA结构， 1个InceptionB，3个InceptionC，1个InceptionD，2个InceptionE， 其中InceptionA，辅助分类器AuxLogits以经过最后一个InceptionC的输出为输入。''' # 35 x 35 x 192 x = self.Mixed_5b(x) # InceptionA(192, pool_features=32) # 35 x 35 x 256 x = self.Mixed_5c(x) # InceptionA(256, pool_features=64) # 35 x 35 x 288 x = self.Mixed_5d(x) # InceptionA(288, pool_features=64) # 35 x 35 x 288 x = self.Mixed_6a(x) # InceptionB(288) # 17 x 17 x 768 x = self.Mixed_6b(x) # InceptionC(768, channels_7x7=128) # 17 x 17 x 768 x = self.Mixed_6c(x) # InceptionC(768, channels_7x7=160) # 17 x 17 x 768 x = self.Mixed_6d(x) # InceptionC(768, channels_7x7=160) # 17 x 17 x 768 x = self.Mixed_6e(x) # InceptionC(768, channels_7x7=192) # 17 x 17 x 768 if self.training and self.aux_logits: aux = self.AuxLogits(x) # InceptionAux(768, num_classes) # 17 x 17 x 768 x = self.Mixed_7a(x) # InceptionD(768) # 8 x 8 x 1280 x = self.Mixed_7b(x) # InceptionE(1280) # 8 x 8 x 2048 x = self.Mixed_7c(x) # InceptionE(2048) '''进入分类部分。 经过平均池化层+dropout+打平+全连接层输出''' x = F.adaptive_avg_pool2d(x, (1, 1)) # N x 2048 x 1 x 1 x = F.dropout(x, training=self.training) # N x 2048 x 1 x 1 x = torch.flatten(x, 1)#Flatten（）就是将2D的特征图压扁为1D的特征向量，是展平操作，进入全连接层之前使用,类才能写进nn.Sequential # N x 2048 x = self.fc(x) # N x 1000 (num_classes) return x, aux def forward(self, x): x, aux = self._forward(x) return x, aux
第五步*：网络结构参数初始化
'''-----------------------网络结构参数初始化--------------------------''' # 目的：使网络更好收敛，准确率更高 def _initialize_weights(self): # 将各种初始化方法定义为一个initialize_weights()的函数并在模型初始后进行使用。 # 遍历网络中的每一层 for m in self.modules(): # isinstance(object, type)，如果指定的对象拥有指定的类型，则isinstance()函数返回True '''如果是卷积层Conv2d''' if isinstance(m, nn.Conv2d): # Kaiming正态分布方式的权重初始化 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') '''判断是否有偏置：''' # 如果偏置不是0，将偏置置成0，对偏置进行初始化 if m.bias is not None: # torch.nn.init.constant_(tensor, val)，初始化整个矩阵为常数val nn.init.constant_(m.bias, 0) '''如果是全连接层''' elif isinstance(m, nn.Linear): # init.normal_(tensor, mean=0.0, std=1.0)，使用从正态分布中提取的值填充输入张量 # 参数：tensor：一个n维Tensor，mean：正态分布的平均值，std：正态分布的标准差 nn.init.normal_(m.weight, 0, 0.01) nn.init.constant_(m.bias, 0)
实现效果
完整代码
from __future__ import division import torch import torch.nn as nn import torch.nn.functional as F '''-------------------------第一步：定义基础卷积模块-------------------------------''' class BasicConv2d(nn.Module): def __init__(self, in_channels, out_channels, **kwargs): super(BasicConv2d, self).__init__() self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs) self.bn = nn.BatchNorm2d(out_channels, eps=0.001) def forward(self, x): x = self.conv(x) x = self.bn(x) return F.relu(x, inplace=True) '''-----------------第二步：定义Inceptionv3模块---------------------''' '''---InceptionA---''' class InceptionA(nn.Module): def __init__(self, in_channels, pool_features, conv_block=None): super(InceptionA, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 64, kernel_size=1) self.branch5x5_1 = conv_block(in_channels, 48, kernel_size=1) self.branch5x5_2 = conv_block(48, 64, kernel_size=5, padding=2) self.branch3x3dbl_1 = conv_block(in_channels, 64, kernel_size=1) self.branch3x3dbl_2 = conv_block(64, 96, kernel_size=3, padding=1) self.branch3x3dbl_3 = conv_block(96, 96, kernel_size=3, padding=1) self.branch_pool = conv_block(in_channels, pool_features, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch5x5 = self.branch5x5_1(x) branch5x5 = self.branch5x5_2(branch5x5) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1) '''---InceptionB---''' class InceptionB(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionB, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch3x3 = conv_block(in_channels, 384, kernel_size=3, stride=2) self.branch3x3dbl_1 = conv_block(in_channels, 64, kernel_size=1) self.branch3x3dbl_2 = conv_block(64, 96, kernel_size=3, padding=1) self.branch3x3dbl_3 = conv_block(96, 96, kernel_size=3, stride=2) def _forward(self, x): branch3x3 = self.branch3x3(x) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) branch_pool = F.max_pool2d(x, kernel_size=3, stride=2) outputs = [branch3x3, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1) '''---InceptionC---''' class InceptionC(nn.Module): def __init__(self, in_channels, channels_7x7, conv_block=None): super(InceptionC, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 192, kernel_size=1) c7 = channels_7x7 self.branch7x7_1 = conv_block(in_channels, c7, kernel_size=1) self.branch7x7_2 = conv_block(c7, c7, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7_3 = conv_block(c7, 192, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_1 = conv_block(in_channels, c7, kernel_size=1) self.branch7x7dbl_2 = conv_block(c7, c7, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_3 = conv_block(c7, c7, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7dbl_4 = conv_block(c7, c7, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7dbl_5 = conv_block(c7, 192, kernel_size=(1, 7), padding=(0, 3)) self.branch_pool = conv_block(in_channels, 192, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch7x7 = self.branch7x7_1(x) branch7x7 = self.branch7x7_2(branch7x7) branch7x7 = self.branch7x7_3(branch7x7) branch7x7dbl = self.branch7x7dbl_1(x) branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch7x7, branch7x7dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1) '''---InceptionD---''' class InceptionD(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionD, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch3x3_1 = conv_block(in_channels, 192, kernel_size=1) self.branch3x3_2 = conv_block(192, 320, kernel_size=3, stride=2) self.branch7x7x3_1 = conv_block(in_channels, 192, kernel_size=1) self.branch7x7x3_2 = conv_block(192, 192, kernel_size=(1, 7), padding=(0, 3)) self.branch7x7x3_3 = conv_block(192, 192, kernel_size=(7, 1), padding=(3, 0)) self.branch7x7x3_4 = conv_block(192, 192, kernel_size=3, stride=2) def _forward(self, x): branch3x3 = self.branch3x3_1(x) branch3x3 = self.branch3x3_2(branch3x3) branch7x7x3 = self.branch7x7x3_1(x) branch7x7x3 = self.branch7x7x3_2(branch7x7x3) branch7x7x3 = self.branch7x7x3_3(branch7x7x3) branch7x7x3 = self.branch7x7x3_4(branch7x7x3) branch_pool = F.max_pool2d(x, kernel_size=3, stride=2) outputs = [branch3x3, branch7x7x3, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1) '''---InceptionE---''' class InceptionE(nn.Module): def __init__(self, in_channels, conv_block=None): super(InceptionE, self).__init__() if conv_block is None: conv_block = BasicConv2d self.branch1x1 = conv_block(in_channels, 320, kernel_size=1) self.branch3x3_1 = conv_block(in_channels, 384, kernel_size=1) self.branch3x3_2a = conv_block(384, 384, kernel_size=(1, 3), padding=(0, 1)) self.branch3x3_2b = conv_block(384, 384, kernel_size=(3, 1), padding=(1, 0)) self.branch3x3dbl_1 = conv_block(in_channels, 448, kernel_size=1) self.branch3x3dbl_2 = conv_block(448, 384, kernel_size=3, padding=1) self.branch3x3dbl_3a = conv_block(384, 384, kernel_size=(1, 3), padding=(0, 1)) self.branch3x3dbl_3b = conv_block(384, 384, kernel_size=(3, 1), padding=(1, 0)) self.branch_pool = conv_block(in_channels, 192, kernel_size=1) def _forward(self, x): branch1x1 = self.branch1x1(x) branch3x3 = self.branch3x3_1(x) branch3x3 = [ self.branch3x3_2a(branch3x3), self.branch3x3_2b(branch3x3), ] branch3x3 = torch.cat(branch3x3, 1) branch3x3dbl = self.branch3x3dbl_1(x) branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) branch3x3dbl = [ self.branch3x3dbl_3a(branch3x3dbl), self.branch3x3dbl_3b(branch3x3dbl), ] branch3x3dbl = torch.cat(branch3x3dbl, 1) branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) branch_pool = self.branch_pool(branch_pool) outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] return outputs def forward(self, x): outputs = self._forward(x) return torch.cat(outputs, 1) '''-------------------第三步：定义辅助分类器InceptionAux-----------------------''' class InceptionAux(nn.Module): def __init__(self, in_channels, num_classes, conv_block=None): super(InceptionAux, self).__init__() if conv_block is None: conv_block = BasicConv2d self.conv0 = conv_block(in_channels, 128, kernel_size=1) self.conv1 = conv_block(128, 768, kernel_size=5) self.conv1.stddev = 0.01 self.fc = nn.Linear(768, num_classes) self.fc.stddev = 0.001 def forward(self, x): # N x 768 x 17 x 17 x = F.avg_pool2d(x, kernel_size=5, stride=3) # N x 768 x 5 x 5 x = self.conv0(x) # N x 128 x 5 x 5 x = self.conv1(x) # N x 768 x 1 x 1 # Adaptive average pooling x = F.adaptive_avg_pool2d(x, (1, 1)) # N x 768 x 1 x 1 x = torch.flatten(x, 1) # N x 768 x = self.fc(x) # N x 1000 return x '''-----------------------第四步：搭建GoogLeNet网络--------------------------''' class GoogLeNet(nn.Module): def __init__(self, num_classes=1000, aux_logits=True, transform_input=False, inception_blocks=None): super(GoogLeNet, self).__init__() if inception_blocks is None: inception_blocks = [ BasicConv2d, InceptionA, InceptionB, InceptionC, InceptionD, InceptionE, InceptionAux ] assert len(inception_blocks) == 7 conv_block = inception_blocks[0] inception_a = inception_blocks[1] inception_b = inception_blocks[2] inception_c = inception_blocks[3] inception_d = inception_blocks[4] inception_e = inception_blocks[5] inception_aux = inception_blocks[6] self.aux_logits = aux_logits self.transform_input = transform_input self.Conv2d_1a_3x3 = conv_block(3, 32, kernel_size=3, stride=2) self.Conv2d_2a_3x3 = conv_block(32, 32, kernel_size=3) self.Conv2d_2b_3x3 = conv_block(32, 64, kernel_size=3, padding=1) self.Conv2d_3b_1x1 = conv_block(64, 80, kernel_size=1) self.Conv2d_4a_3x3 = conv_block(80, 192, kernel_size=3) self.Mixed_5b = inception_a(192, pool_features=32) self.Mixed_5c = inception_a(256, pool_features=64) self.Mixed_5d = inception_a(288, pool_features=64) self.Mixed_6a = inception_b(288) self.Mixed_6b = inception_c(768, channels_7x7=128) self.Mixed_6c = inception_c(768, channels_7x7=160) self.Mixed_6d = inception_c(768, channels_7x7=160) self.Mixed_6e = inception_c(768, channels_7x7=192) if aux_logits: self.AuxLogits = inception_aux(768, num_classes) self.Mixed_7a = inception_d(768) self.Mixed_7b = inception_e(1280) self.Mixed_7c = inception_e(2048) self.fc = nn.Linear(2048, num_classes) '''输入(229,229,3)的数据，首先归一化输入，经过5个卷积，2个最大池化层。''' def _forward(self, x): # N x 3 x 299 x 299 x = self.Conv2d_1a_3x3(x) # N x 32 x 149 x 149 x = self.Conv2d_2a_3x3(x) # N x 32 x 147 x 147 x = self.Conv2d_2b_3x3(x) # N x 64 x 147 x 147 x = F.max_pool2d(x, kernel_size=3, stride=2) # N x 64 x 73 x 73 x = self.Conv2d_3b_1x1(x) # N x 80 x 73 x 73 x = self.Conv2d_4a_3x3(x) # N x 192 x 71 x 71 x = F.max_pool2d(x, kernel_size=3, stride=2) '''然后经过3个InceptionA结构， 1个InceptionB，3个InceptionC，1个InceptionD，2个InceptionE， 其中InceptionA，辅助分类器AuxLogits以经过最后一个InceptionC的输出为输入。''' # 35 x 35 x 192 x = self.Mixed_5b(x) # InceptionA(192, pool_features=32) # 35 x 35 x 256 x = self.Mixed_5c(x) # InceptionA(256, pool_features=64) # 35 x 35 x 288 x = self.Mixed_5d(x) # InceptionA(288, pool_features=64) # 35 x 35 x 288 x = self.Mixed_6a(x) # InceptionB(288) # 17 x 17 x 768 x = self.Mixed_6b(x) # InceptionC(768, channels_7x7=128) # 17 x 17 x 768 x = self.Mixed_6c(x) # InceptionC(768, channels_7x7=160) # 17 x 17 x 768 x = self.Mixed_6d(x) # InceptionC(768, channels_7x7=160) # 17 x 17 x 768 x = self.Mixed_6e(x) # InceptionC(768, channels_7x7=192) # 17 x 17 x 768 if self.training and self.aux_logits: aux = self.AuxLogits(x) # InceptionAux(768, num_classes) # 17 x 17 x 768 x = self.Mixed_7a(x) # InceptionD(768) # 8 x 8 x 1280 x = self.Mixed_7b(x) # InceptionE(1280) # 8 x 8 x 2048 x = self.Mixed_7c(x) # InceptionE(2048) '''进入分类部分。 经过平均池化层+dropout+打平+全连接层输出''' x = F.adaptive_avg_pool2d(x, (1, 1)) # N x 2048 x 1 x 1 x = F.dropout(x, training=self.training) # N x 2048 x 1 x 1 x = torch.flatten(x, 1)#Flatten（）就是将2D的特征图压扁为1D的特征向量，是展平操作，进入全连接层之前使用,类才能写进nn.Sequential # N x 2048 x = self.fc(x) # N x 1000 (num_classes) return x, aux def forward(self, x): x, aux = self._forward(x) return x, aux '''-----------------------第五步：网络结构参数初始化--------------------------''' # 目的：使网络更好收敛，准确率更高 def _initialize_weights(self): # 将各种初始化方法定义为一个initialize_weights()的函数并在模型初始后进行使用。 # 遍历网络中的每一层 for m in self.modules(): # isinstance(object, type)，如果指定的对象拥有指定的类型，则isinstance()函数返回True '''如果是卷积层Conv2d''' if isinstance(m, nn.Conv2d): # Kaiming正态分布方式的权重初始化 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') '''判断是否有偏置：''' # 如果偏置不是0，将偏置置成0，对偏置进行初始化 if m.bias is not None: # torch.nn.init.constant_(tensor, val)，初始化整个矩阵为常数val nn.init.constant_(m.bias, 0) '''如果是全连接层''' elif isinstance(m, nn.Linear): # init.normal_(tensor, mean=0.0, std=1.0)，使用从正态分布中提取的值填充输入张量 # 参数：tensor：一个n维Tensor，mean：正态分布的平均值，std：正态分布的标准差 nn.init.normal_(m.weight, 0, 0.01) nn.init.constant_(m.bias, 0) '''---------------------------------------显示网络结构-------------------------------''' if __name__ == '__main__': net = GoogLeNet(1000).cuda() from torchsummary import summary summary(net, (3, 299, 299))
番外
上面实现的是torchvision中的Inception v3结构，和论文中不太一样。因此我上GitHub上找到了论文复现的代码。先放个链接：https://github.com/AlgorithmicIntelligence/GoogLeNetv3_Pytorch/tree/master/models
（论文中结构）
代码
import torch import torch.nn as nn from functools import partial # functools.partial()：减少某个函数的参数个数。 partial() 函数允许你给一个或多个参数设置固定的值，减少接下来被调用时的参数个数 '''-----------------------第一步：定义卷积模块-----------------------''' #基础卷积模块 class Conv2d(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, output=False): super(Conv2d, self).__init__() '''卷积层''' self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding) '''输出层''' self.output = output if self.output == False: '''bn层''' self.bn = nn.BatchNorm2d(out_channels) '''relu层''' self.relu = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) if self.output: return x else: x = self.bn(x) x = self.relu(x) return x class Separable_Conv2d(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0): super(Separable_Conv2d, self).__init__() self.conv_h = nn.Conv2d(in_channels, in_channels, (kernel_size, 1), stride=(stride, 1), padding=(padding, 0)) self.conv_w = nn.Conv2d(in_channels, out_channels, (1, kernel_size), stride=(1, stride), padding=(0, padding)) self.bn = nn.BatchNorm2d(out_channels) self.relu = nn.ReLU(inplace=True) def forward(self, x): x = self.conv_h(x) x = self.conv_w(x) x = self.bn(x) x = self.relu(x) return x class Concat_Separable_Conv2d(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0): super(Concat_Separable_Conv2d, self).__init__() self.conv_h = nn.Conv2d(in_channels, out_channels, (kernel_size, 1), stride=(stride, 1), padding=(padding, 0)) self.conv_w = nn.Conv2d(in_channels, out_channels, (1, kernel_size), stride=(1, stride), padding=(0, padding)) self.bn = nn.BatchNorm2d(out_channels * 2) self.relu = nn.ReLU(inplace=True) def forward(self, x): x_h = self.conv_h(x) x_w = self.conv_w(x) x = torch.cat([x_h, x_w], dim=1) x = self.bn(x) x = self.relu(x) return x #Flatten（）就是将2D的特征图压扁为1D的特征向量，是展平操作，进入全连接层之前使用,类才能写进nn.Sequential class Flatten(nn.Module): # 传入输入维度和输出维度 def __init__(self): # 调用父类构造函数 super(Flatten, self).__init__() # 实现forward函数 def forward(self, x): # 保存batch维度，后面的维度全部压平 return torch.flatten(x, 1) #Squeeze（）降维 class Squeeze(nn.Module): def __init__(self): super(Squeeze, self).__init__() def forward(self, x): return torch.squeeze(x) '''-----------------------搭建GoogLeNet网络--------------------------''' class GoogLeNet(nn.Module): def __init__(self, num_classes, mode='train'): super(GoogLeNet, self).__init__() self.num_classes = num_classes self.mode = mode self.layers = nn.Sequential( Conv2d(3, 32, 3, stride=2), Conv2d(32, 32, 3, stride=1), Conv2d(32, 64, 3, stride=1, padding=1), nn.MaxPool2d(kernel_size=3, stride=2), Conv2d(64, 80, kernel_size=3), Conv2d(80, 192, kernel_size=3, stride=2), Conv2d(192, 288, kernel_size=3, stride=1, padding=1), #输入：35*35*288。将5*5用两个3*3代替 Inceptionv3(288, 64, 48, 64, 64, 96, 64, mode='1'), # 3a Inceptionv3(288, 64, 48, 64, 64, 96, 64, mode='1'), # 3b Inceptionv3(288, 0, 128, 384, 64, 96, 0, stride=2, pool_type='MAX', mode='1'), # 3c #输入：17*17*768。 Inceptionv3(768, 192, 128, 192, 128, 192, 192, mode='2'), # 4a Inceptionv3(768, 192, 160, 192, 160, 192, 192, mode='2'), # 4b Inceptionv3(768, 192, 160, 192, 160, 192, 192, mode='2'), # 4c Inceptionv3(768, 192, 192, 192, 192, 192, 192, mode='2'), # 4d Inceptionv3(768, 0, 192, 320, 192, 192, 0, stride=2, pool_type='MAX', mode='2'), # 4e #8*8*1280 Inceptionv3(1280, 320, 384, 384, 448, 384, 192, mode='3'), # 5a Inceptionv3(2048, 320, 384, 384, 448, 384, 192, pool_type='MAX', mode='3'), # 5b nn.AvgPool2d(8, 1), Conv2d(2048, num_classes, kernel_size=1, output=True), Squeeze(), ) if mode == 'train': self.aux = InceptionAux(768, num_classes) def forward(self, x): for idx, layer in enumerate(self.layers): if (idx == 14 and self.mode == 'train'): aux = self.aux(x) x = layer(x) if self.mode == 'train': return x, aux else: return x '''-----------------------网络结构参数初始化--------------------------''' # 目的：使网络更好收敛，准确率更高 def _initialize_weights(self): # 将各种初始化方法定义为一个initialize_weights()的函数并在模型初始后进行使用。 # 遍历网络中的每一层 for m in self.modules(): # isinstance(object, type)，如果指定的对象拥有指定的类型，则isinstance()函数返回True '''如果是卷积层Conv2d''' if isinstance(m, nn.Conv2d): # Kaiming正态分布方式的权重初始化 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') '''判断是否有偏置：''' # 如果偏置不是0，将偏置置成0，对偏置进行初始化 if m.bias is not None: # torch.nn.init.constant_(tensor, val)，初始化整个矩阵为常数val nn.init.constant_(m.bias, 0) '''如果是全连接层''' elif isinstance(m, nn.Linear): # init.normal_(tensor, mean=0.0, std=1.0)，使用从正态分布中提取的值填充输入张量 # 参数：tensor：一个n维Tensor，mean：正态分布的平均值，std：正态分布的标准差 nn.init.normal_(m.weight, 0, 0.01) nn.init.constant_(m.bias, 0) '''---------------------Inceptionv3-------------------------------------''' ''' Inceptionv3由三个连续的Inception模块组组成 ''' class Inceptionv3(nn.Module): def __init__(self, input_channel, conv1_channel, conv3_reduce_channel, conv3_channel, conv3_double_reduce_channel, conv3_double_channel, pool_reduce_channel, stride=1, pool_type='AVG', mode='1'): super(Inceptionv3, self).__init__() self.stride = stride if stride == 2: padding_conv3 = 0 padding_conv7 = 2 else: padding_conv3 = 1 padding_conv7 = 3 if conv1_channel != 0: self.conv1 = Conv2d(input_channel, conv1_channel, kernel_size=1) else: self.conv1 = None self.conv3_reduce = Conv2d(input_channel, conv3_reduce_channel, kernel_size=1) #第一种Inception模式：输入的特征图尺寸为35x35x288,采用了论文中图5中的架构，将5x5以两个3x3代替。 if mode == '1': self.conv3 = Conv2d(conv3_reduce_channel, conv3_channel, kernel_size=3, stride=stride, padding=padding_conv3) self.conv3_double1 = Conv2d(conv3_double_reduce_channel, conv3_double_channel, kernel_size=3, padding=1) self.conv3_double2 = Conv2d(conv3_double_channel, conv3_double_channel, kernel_size=3, stride=stride, padding=padding_conv3) #第二种Inception模块：输入特征图尺寸为17x17x768，采用了论文中图6中nx1+1xn的不对称卷积结构 elif mode == '2': self.conv3 = Separable_Conv2d(conv3_reduce_channel, conv3_channel, kernel_size=7, stride=stride, padding=padding_conv7) self.conv3_double1 = Separable_Conv2d(conv3_double_reduce_channel, conv3_double_channel, kernel_size=7, padding=3) self.conv3_double2 = Separable_Conv2d(conv3_double_channel, conv3_double_channel, kernel_size=7, stride=stride, padding=padding_conv7) #第三种Inception模块：输入特征图尺寸为8x8x1280, 采用了论文图7中所示的并行模块的结构 elif mode == '3': self.conv3 = Concat_Separable_Conv2d(conv3_reduce_channel, conv3_channel, kernel_size=3, stride=stride, padding=1) self.conv3_double1 = Conv2d(conv3_double_reduce_channel, conv3_double_channel, kernel_size=3, padding=1) self.conv3_double2 = Concat_Separable_Conv2d(conv3_double_channel, conv3_double_channel, kernel_size=3, stride=stride, padding=1) self.conv3_double_reduce = Conv2d(input_channel, conv3_double_reduce_channel, kernel_size=1) if pool_type == 'MAX': self.pool = nn.MaxPool2d(kernel_size=3, stride=stride, padding=padding_conv3) elif pool_type == 'AVG': self.pool = nn.AvgPool2d(kernel_size=3, stride=stride, padding=padding_conv3) if pool_reduce_channel != 0: self.pool_reduce = Conv2d(input_channel, pool_reduce_channel, kernel_size=1) else: self.pool_reduce = None def forward(self, x): output_conv3 = self.conv3(self.conv3_reduce(x)) output_conv3_double = self.conv3_double2(self.conv3_double1(self.conv3_double_reduce(x))) if self.pool_reduce != None: output_pool = self.pool_reduce(self.pool(x)) else: output_pool = self.pool(x) if self.conv1 != None: output_conv1 = self.conv1(x) outputs = torch.cat([output_conv1, output_conv3, output_conv3_double, output_pool], dim=1) else: outputs = torch.cat([output_conv3, output_conv3_double, output_pool], dim=1) return outputs '''------------辅助分类器---------------------------''' class InceptionAux(nn.Module): def __init__(self, input_channel, num_classes): super(InceptionAux, self).__init__() self.layers = nn.Sequential( nn.AvgPool2d(5, 3), Conv2d(input_channel, 128, 1), Conv2d(128, 1024, kernel_size=5), Conv2d(1024, num_classes, kernel_size=1, output=True), Squeeze() ) def forward(self, x): x = self.layers(x) return x '''-------------------显示网络结构-------------------------------''' if __name__ == '__main__': net = GoogLeNet(1000).cuda() from torchsummary import summary summary(net, (3, 299, 299))
本篇到这里就结束了，有什么问题欢迎大家留言讨论呀！
