含并行连结的网络（GoogLeNet）

中文版 pytorch
22

如果只要让代码能运行，最小的输入应该可以是1x1吧。

23

Q1.1 和 Q1.2

class Inception2(nn.Module):

def __init__(self, in_channels, c1, c2, c3, c4, **kwargs):
    super(Inception2, self).__init__(**kwargs)
    # 线路1，单1x1卷积层
    self.p1_1 = nn.Conv2d(in_channels, c1, kernel_size=1)
    # 线路2，1x1卷积层后接3x3卷积层
    self.p2_1 = nn.Conv2d(in_channels, c2[0], kernel_size=1)
    self.p2_2 = nn.Conv2d(c2[0], c2[1], kernel_size=3, padding=1)
    # 线路3，1x1卷积层后接5x5卷积层
    self.p3_1 = nn.Conv2d(in_channels, c3[0], kernel_size=1)
    self.p3_2 = nn.Conv2d(c3[0], c3[1], kernel_size=5, padding=2)
    # 线路4，3x3最大汇聚层后接1x1卷积层
    self.p4_1 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
    self.p4_2 = nn.Conv2d(in_channels, c4, kernel_size=1)
    
    self.bn1 = nn.BatchNorm2d(c1)
    self.bn2 = nn.BatchNorm2d(c2[1])
    self.bn3 = nn.BatchNorm2d(c3[1])
    self.bn4 = nn.BatchNorm2d(c4)
    

def forward(self, x):
    p1 = F.relu(self.bn1(self.p1_1(x)))
    p2 = F.relu(self.bn2(self.p2_2(F.relu(self.p2_1(x)))))
    p3 = F.relu(self.bn3(self.p3_2(F.relu(self.p3_1(x)))))
    p4 = F.relu(self.bn4(self.p4_2(self.p4_1(x))))

    # 在通道维度上连结输出
    return torch.cat((p1, p2, p3, p4), dim=1)

b31 = nn.Sequential(Inception2(192, 64, (96, 64), (16, 64), 64),
Inception2(256, 128, (128, 128), (32, 96), 128),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1))

net2 = nn.Sequential(b1, b2, b31, b4, b5, nn.Linear(1024, 10))

lr, num_epochs, batch_size = 0.1, 10, 128
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=96)
d2l.train_ch6(net2, train_iter, test_iter, num_epochs, lr, d2l.try_gpu())

loss 0.177, train acc 0.935, test acc 0.909
1024.5 examples/sec on cuda:0
1713977554545

24

你好，会不会是梯度消失？还是过拟合了呢？