动手学PyTorch（李沐）14 ----LeNet

LeNet

​ LeNet网络结构如上，两个卷积层，两个池化层，再加连个全连接层构成了LeNet。

QA:

1. 时序性数据同样可以用卷积网络。
2. 在数据压缩高宽减半的同时，可以把通道数翻倍，通道数增加可以使得能匹配的模式变多，所以同样一个像素表示的信息变大，通道变多相当于把信息分散到通道中。
3. 输出通道可以很大程度上认为是匹配的某一个局部特征。

---

卷积神经网络（LeNet）

LeNet-由两部分组成，卷积编码器和全连接层密集块

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

class Reshape(torch.nn.Module):
  def forward(self,x):
    return x.view(-1,1,28,28)

net = torch.nn.Sequential(
  Reshape(),
  #卷积
  nn.Conv2d(1, 6, kernel_size=5,padding=2), nn.Sigmoid(),
  #均值池化层2*2窗口
  nn.MaxPool2d(kernel_size=2, stride=2),
  nn.Conv2d(6, 16, kernel_size=5), nn.Sigmoid(),
  nn.MaxPool2d(kernel_size=2, stride=2),
  #卷积层出来是4D的东西，要把高宽拉成一个向量，Flatten（）就是保留一维，后面拉伸
  nn.Flatten(),
  # 16*5*5是该层输入，120是输出
  nn.Linear(16 * 5 * 5, 120), nn.Sigmoid(),
  nn.Linear(120, 84), nn.Sigmoid(),
  nn.Linear(84, 10))

检查模型

X = torch.rand(size=(1, 1, 28, 28), dtype=torch.float32)
for layer in net:
  X = layer(X)
  #layer.__class__指的是每一层， .__name__得到每一次的名字
  print(layer.__class__.__name__,'output shape: \t',X.shape)

LeNet在FashionMNIST数据集的表现

from d2l import torch as d2l
batch_size = 256
train_iter,test_iter = d2l.load_data_fashion_mnist(batch_size)

def evaluate_accuracy_gpu(net, data_iter, device=None): 
  """使⽤GPU计算模型在数据集上的精度"""
  if isinstance(net, nn.Module):
    net.eval() # 设置为评估模式
    if not device:
      # 如果没有传入device的话，就看网络存的地方
      device = next(iter(net.parameters())).device
    # 正确预测的数量，总预测的数量
  metric = d2l.Accumulator(2)
  with torch.no_grad():
    for X, y in data_iter:
      if isinstance(X, list):
        # BERT微调所需的（之后将介绍）
        X = [x.to(device) for x in X]
      else:
        X = X.to(device)
      y = y.to(device)
      metric.add(d2l.accuracy(net(X), y), y.numel())
  return metric[0] / metric[1]

为使用GPU,训练函数需要改动

def train_ch6(net,train_iter,test_iter,num_epochs,lr,device):
  "trian a model with a gpu(defined in chapter 6)"
  # 权重初始化,使得输入和输出的方差不太大
  def init_weights(m):
    if type(m) == nn.Linear or type(m) == nn.Conv2d:
      nn.init.xavier_uniform_(m.weight)
  # 每一层都调用函数，初始化网络权值
  net.apply(init_weights)
  print('training on',device)
  net.to(device)
  # 随机梯度下降
  optimizer = torch.optim.SGD(net.parameters(),lr=lr)
  # 交叉熵
  loss = nn.CrossEntropyLoss()
  # 动画效果
  animator   = d2l.Animator(xlabel='epoch',xlim=[1,num_epochs],
                            legend = ['train_loss','train_acc','test_acc'])
  timer,num_batches = d2l.Timer(),len(train_iter)
  for epoch in range(num_epochs):
    metric = d2l.Accumulator(3)
    net.train()
    for i,(x,y) in enumerate(train_iter):
      timer.start()
      optimizer.zero_grad()
      x,y = x.to(device),y.to(device)
      y_hat = net(x)
      # 计算损失
      l = loss(y_hat,y)
      # 计算梯度
      l.backward()
      # 迭代优化
      optimizer.step()
      metric.add(l*x.shape[0],d2l.accuracy(y_hat,y),x.shape[0])
      timer.stop()
      train_l = metric[0] / metric[2]
      train_acc = metric[1] / metric[2]
      if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
        animator.add(epoch + (i + 1) / num_batches,
         (train_l, train_acc, None))
    test_acc = evaluate_accuracy_gpu(net, test_iter)
    animator.add(epoch + 1, (None, None, test_acc))
  print(f'loss {train_l:.3f}, train acc {train_acc:.3f}, '
      f'test acc {test_acc:.3f}')
  print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec '
      f'on {str(device)}')

lr, num_epochs = 1.5, 15
train_ch6(net, train_iter, test_iter, num_epochs, lr, d2l.try_gpu())