PyTorch之初级使用

• 使用流程
  ①. 数据准备; ②. 模型确立; ③. 损失函数确立; ④. 优化器确立; ⑤. 模型训练及保存
• 模块介绍
  Part Ⅰ: 数据准备
  torch.utils.data.Dataset 
  torch.utils.data.DataLoader 
  关于Dataset, 作为数据集, 需要实现基本的3个方法, 分别为: __init__、__len__、__getitem__. 示例如下,
  

  class TrainingDataset(Dataset):
  
      def __init__(self, X, Y_, transform=None, target_transform=None):
          self.__X = X
          self.__Y_ = Y_
          self.__transform = transform
          self.__target_transform = target_transform 
  
  
      def __len__(self):
          return len(self.__X)
  
  
      def __getitem__(self, idx):
          x = self.__X[idx]
          y_ = self.__Y_[idx]
          if self.__transform:
              x = self.__transform(x)
          if self.__target_transform:
              y_ = self.__target_transform(y_)
          return x, y_

  关于DataLoader, 作为数据集封装, 将数据集Dataset封装为可迭代对象. 示例如下,

  batch_size = 100
  trainingLoader = DataLoader(trainingData, batch_size=batch_size, shuffle=True)

  Part Ⅱ: 模型确立
  torch.nn 
  torch.nn.Module 
  网络模型由基类Module派生, 内部所有操作模块均由命名空间nn提供, 需要实现基本的2个方法, 分别为: __init__、forward. 其中, __init__方法定义操作, forward方法运用操作进行正向计算. 示例如下,

  class NeuralNetwork(nn.Module):
  
      def __init__(self):
          super(NeuralNetwork, self).__init__()
          self.__linear_tanh_stack = nn.Sequential(
              nn.Linear(3, 5),
              nn.Tanh(),
              nn.Linear(5, 3)
          )
  
  
      def forward(self, x):
          y = self.__linear_tanh_stack(x)
          return y
  
  
  model = NeuralNetwork()

  Part Ⅲ: 损失函数确立
  torch.nn 
  常见损失函数有: nn.MSELoss(回归任务)、nn.CrossEntropyLoss(多分类任务)等. 示例如下,

  loss_func = nn.MSELoss(reduction="sum")

  Part Ⅳ: 优化器确立
  torch.optim 
  常见的优化器有: optim.SGD、optim.Adam等. 示例如下,

  optimizer = optim.Adam(model.parameters(), lr=0.001)

  Part Ⅴ: 模型训练及保存
  有效整合数据、模型、损失函数及优化器. 注意, 模型参数之梯度默认累积, 每次参数优化需要显式清零. 示例如下,

  def train_loop(dataloader, model, loss_func, optimizer):
      for batchIdx, (X, Y_) in enumerate(dataloader):
          Y = model(X)
          loss = loss_func(Y, Y_)
  
          optimizer.zero_grad()
          loss.backward()
          optimizer.step()
  
          
  epoch = 50000
  for epochIdx in range(epoch):
      train_loop(trainingLoader, model, loss_func, optimizer)
  
  
  torch.save(model.state_dict(), "model_params.pth")

• 代码实现
  本文使用与Back Propagation - Python实现相同的网络架构及数据生成策略, 分别如下所示,
  

  $$
  \begin{equation*}
  \left\{
  \begin{split}
  x &= r + 2g + 3b \\
  y &= r^2 + 2g^2 + 3b^2 \\
  lv &= -3r - 4g - 5b
  \end{split}
  \right.
  \end{equation*} 
  $$
  具体实现如下,

  import numpy
  import torch
  from torch import nn
  from torch import optim
  from torch.utils.data import Dataset, DataLoader
  from matplotlib import pyplot as plt
  
  
  numpy.random.seed(1)
  torch.manual_seed(3)
  
  
  # 生成training数据
  def getData(n=100):
      rgbRange = (-1, 1)
      r = numpy.random.uniform(*rgbRange, (n, 1))
      g = numpy.random.uniform(*rgbRange, (n, 1))
      b = numpy.random.uniform(*rgbRange, (n, 1))
      x_ = r + 2 * g + 3 * b
      y_ = r ** 2 + 2 * g ** 2 + 3 * b ** 2
      lv_ = -3 * r - 4 * g - 5 * b
      RGB = numpy.hstack((r, g, b))
      XYLv_ = numpy.hstack((x_, y_, lv_))
      return RGB, XYLv_
  
  
  class TrainingDataset(Dataset):
  
      def __init__(self, X, Y_, transform=None, target_transform=None):
          self.__X = X
          self.__Y_ = Y_
          self.__transform = transform
          self.__target_transform = target_transform 
  
  
      def __len__(self):
          return len(self.__X)
  
  
      def __getitem__(self, idx):
          x = self.__X[idx]
          y_ = self.__Y_[idx]
          if self.__transform:
              x = self.__transform(x)
          if self.__target_transform:
              y_ = self.__target_transform(y_)
          return x, y_
  
  
  RGB, XYLv_ = getData(1000)
  trainingData = TrainingDataset(RGB, XYLv_, torch.Tensor, torch.Tensor)
  
  batch_size = 100
  trainingLoader = DataLoader(trainingData, batch_size=batch_size, shuffle=True)
  
  
  class NeuralNetwork(nn.Module):
  
      def __init__(self):
          super(NeuralNetwork, self).__init__()
          self.__linear_tanh_stack = nn.Sequential(
              nn.Linear(3, 5),
              nn.Tanh(),
              nn.Linear(5, 3)
          )
  
  
      def forward(self, x):
          y = self.__linear_tanh_stack(x)
          return y
  
  
  model = NeuralNetwork()
  loss_func = nn.MSELoss(reduction="sum")
  optimizer = optim.Adam(model.parameters(), lr=0.001)
  
  
  def train_loop(dataloader, model, loss_func, optimizer):
      JVal = 0
      for batchIdx, (X, Y_) in enumerate(dataloader):
          Y = model(X)
          loss = loss_func(Y, Y_)
  
          JVal += loss.item()
  
          optimizer.zero_grad()
          loss.backward()
          optimizer.step()
  
      JVal /= 2
      return JVal
  
  
  JPath = list()
  epoch = 50000
  for epochIdx in range(epoch):
      JVal = train_loop(trainingLoader, model, loss_func, optimizer)
      print("epoch: {:5d},   JVal = {:.5f}".format(epochIdx, JVal))
      JPath.append(JVal)
  
  
  torch.save(model.state_dict(), "model_params.pth")
  
  
  fig = plt.figure(figsize=(6, 4))
  ax1 = fig.add_subplot(1, 1, 1)
  
  ax1.plot(numpy.arange(len(JPath)), JPath, "k.", markersize=1)
  ax1.plot(0, JPath[0], "go", label="seed")
  ax1.plot(len(JPath)-1, JPath[-1], "r*", label="solution")
  
  ax1.legend()
  ax1.set(xlabel="$epoch$", ylabel="$JVal$", title="solution-JVal = {:.5f}".format(JPath[-1]))
  
  fig.tight_layout()
  fig.savefig("plot_fig.png", dpi=100)

• 结果展示
  可以看到, 在training data上总体loss随epoch增加逐渐降低.

• 使用建议
  ①. 分batch处理训练数据, 可以提升训练初始阶段模型参数收敛速度;
  ②. 常规优化器推荐Adam, 具备自动步长调节的能力.

• 参考文档
  ①. https://pytorch.org/tutorials/beginner/basics/intro.html