P26网络中的NiN

本质是NiN块，用1*1的卷积网络取代全连接层

import torch
from torch import nn
from d2l import torch as d2l

#显卡
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.backends.cudnn.is_available())
print(torch.cuda_version)
print(torch.backends.cudnn.version())

1
True
True
11.3
8200

网络架构

def nin_block(in_channels, out_channels, kernel_size, strides, padding):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, kernel_size, strides, padding),
        # 自定义的kernel和padding和stride都用在第一个卷积层上面
        nn.ReLU(),
        nn.Conv2d(out_channels, out_channels, kernel_size=1), nn.ReLU(),
        # ReLU激活后，kernel=1，那stride和padding默认也等于1，这不会改变通道数，所以输入和输出通道数相同
        nn.Conv2d(out_channels, out_channels, kernel_size=1),
        # 再加一层和上面一样的卷积
        nn.ReLU()
        # 最后再加一层激活函数
    )

NiN模型

net = nn.Sequential(
    nin_block(1, 96, kernel_size=11, strides=4, padding=0),
    #输入是1因为是灰度图，96是AlexNet里的96，这里是把AlexNet的卷积层搬过来，后面加了1*1的卷积层
    nn.MaxPool2d(3, stride=2),
    #高宽减半
    nin_block(96, 256, kernel_size=5, strides=1, padding=2),
    #通道增加到256
    nn.MaxPool2d(3, stride=2),
    nin_block(256, 384, kernel_size=3, strides=1, padding=1),
    #这里没在后面加2个卷积层
    nn.MaxPool2d(3, stride=2),
    nn.Dropout(0.5),
    nin_block(384, 10, kernel_size=3, strides=1, padding=1),
    #把通道数降到10，因为Fashion-MNIST的类别是10，根据自己的需求改变输出的通道数
    nn.AdaptiveAvgPool2d((1, 1)),
    #全局平均池化层，1*1是高宽都会变成1
    nn.Flatten()
    #把最后两个维度flatten，直接可以用去softmax
)

查看每个块的输出形状

X = torch.rand(size=(1,1,224,224))
for layer in net:
    X = layer(X) #之前漏掉了这一行，导致所有层都是224*224
    #print(layer.__class__.__name__,'output shape:\t', X.shape)

Sequential output shape:     torch.Size([1, 96, 54, 54])
输入是224*224，第一个人NIN block后，直接把高宽变成了54*54
MaxPool2d output shape:     torch.Size([1, 96, 26, 26])
Sequential output shape:    torch.Size([1, 256, 26, 26])
第二个NIN BLOCK的输出是26*26
MaxPool2d output shape:     torch.Size([1, 256, 12, 12])
用maxpool不改变通道数，但是高宽减半
Sequential output shape:    torch.Size([1, 384, 12, 12])
MaxPool2d output shape:     torch.Size([1, 384, 5, 5])
Dropout output shape:   torch.Size([1, 384, 5, 5])
Sequential output shape:    torch.Size([1, 10, 5, 5])
这里把输出变成了我需要的类别数，从384变成10
AdaptiveAvgPool2d output shape:     torch.Size([1, 10, 1, 1])
Flatten output shape:   torch.Size([1, 10])

训练模型

#训练模型

lr,num_epochs,batch_size = 0.1,10,128
train_iter,test_iter = d2l.load_data_fashion_mnist(batch_size,resize=224)
d2l.train_ch6(net,train_iter,test_iter,num_epochs,lr,d2l.try_gpu())

loss 0.349, train acc 0.869, test acc 0.869
1772.7 examples/sec on cuda:0

用了GPU后果然快乐很多