使用pytorch对一维数据实现简单多分类

Pytorch 是目前最好用的神经网络库之一，最近我写了一个pytorch的简单代码，在这里对其做一个全面的介绍。

在pytorch 中一些常用的功能都已经被封装成了模块，所以我们只需要继承并重写部分函数即可。首先介绍一下本文最终希望实现的目标，对本地的一维数据（1xn)的ndarry 进行一个多分类，数据集为mn的数据，标签为m1的数组。下面是结合代码记录一下踩坑过程。

继承Dataset类，可以看到我这里重写了三个函数，init 函数用于载入numpy数据并将其转化为相应的tensor，__gititem__函数用于定义训练时会返回的单个数据与标签，__len__表示数据数量m。

class MyDataSet(torch.utils.data.Dataset):
     def __init__(self, data, label):
         self.data = torch.from_numpy(data).float()
         self.label = torch.from_numpy(label)
         self.length = label.shape[0]
     def __getitem__(self, index):
         return self.data[index], self.label[index]
     def __len__(self):
         return self.length

通过继承nn.Module来自定义神经网络

　　其中__init__函数来自定义定义我们需要的网络参数，这里我们block1 的in_channels为1，输出参数可根据需要自己设定，但而且当前层的输出channel应该和下一层的输入channel相同，
注意：MaxPool1d的inchannel需要自己计算一下，当然如果你不想算，可以给个参数直接运行，看报错信息的提示
　　__forward__ 函数定义了网络的连接方式，注意此处应返回x。

class NeuralNetwork(nn.Module):
  def __init__(self):
      super(NeuralNetwork, self).__init__()
      self.block1 = nn.Sequential(
          nn.Conv1d(in_channels=filter_num[0], out_channels=filter_num[1], padding='same', stride=conv_stride_size,
                    kernel_size=conv_kernel_size, dtype=dtype),
          nn.BatchNorm1d(filter_num[1],dtype=dtype),
          nn.ReLU(),
          nn.Conv1d(filter_num[1], filter_num[1], kernel_size=conv_kernel_size, stride=conv_stride_size, padding='same',dtype=dtype),
          nn.BatchNorm1d(filter_num[1],dtype=dtype),
          nn.ReLU(),
          nn.MaxPool1d(kernel_size=pool_kernel_size, stride=pool_stride, padding=pool_padding),
          nn.Dropout(drop_float),
      )
      self.block2=nn.Sequential(
          nn.Conv1d(in_channels=filter_num[1],out_channels=filter_num[2], padding='same', stride=conv_stride_size,
                    kernel_size=conv_kernel_size, dtype=dtype),
          nn.BatchNorm1d(num_features=filter_num[2],dtype=dtype),
          nn.ReLU(),
          nn.Conv1d(filter_num[2],filter_num[2],kernel_size=conv_kernel_size,stride=conv_stride_size,
                    padding='same',dtype=dtype),
          nn.BatchNorm1d(filter_num[2],dtype=dtype),
          nn.ReLU(),
          nn.MaxPool1d(kernel_size=pool_kernel_size, stride=pool_stride,padding=pool_padding),
          nn.Dropout(drop_float),
      )
      self.block3=nn.Sequential(
          nn.Conv1d(in_channels=filter_num[2],out_channels=filter_num[3], padding='same', stride=conv_stride_size,
                    kernel_size=conv_kernel_size, dtype=dtype),
          nn.BatchNorm1d(num_features=filter_num[3],dtype=dtype),
          nn.ReLU(),
          nn.Conv1d(filter_num[3],filter_num[3],kernel_size=conv_kernel_size,stride=conv_stride_size,
                    padding='same',dtype=dtype),
          nn.BatchNorm1d(filter_num[3],dtype=dtype),
          nn.ReLU(),
          nn.MaxPool1d(kernel_size=pool_kernel_size, stride=pool_stride,padding=pool_padding),
          nn.Dropout(drop_float),
      )
      self.block4=nn.Sequential(
          nn.Conv1d(in_channels=filter_num[3],out_channels=filter_num[4], padding='same', stride=conv_stride_size,
                    kernel_size=conv_kernel_size, dtype=dtype),
          nn.BatchNorm1d(num_features=filter_num[4],dtype=dtype),
          nn.ReLU(),
          nn.Conv1d(filter_num[4],filter_num[4],kernel_size=conv_kernel_size,stride=conv_stride_size,
                    padding='same',dtype=dtype),
          nn.BatchNorm1d(filter_num[4],dtype=dtype),
          nn.ReLU(),
          nn.MaxPool1d(kernel_size=pool_kernel_size, stride=pool_stride,padding=pool_padding),
          nn.Dropout(drop_float),
      )
      self.connected = nn.Sequential(
          nn.Flatten(),
          nn.BatchNorm1d(num_features=1280 ,dtype=dtype),
          nn.ReLU(),
          nn.Dropout(0.7),
          nn.Flatten(),
          nn.BatchNorm1d(num_features=1280 ,dtype=dtype),
          nn.ReLU(),
          nn.Dropout(0.5),
          nn.Linear(1280, 180,dtype=dtype)
      )
  def forward(self, x):
      x = self.block1(x)
      x= self.block2(x)
      x= self.block3(x)
      x= self.block4(x)
      x = self.connected(x)
      return x

主程序。为了更好的说明，先放一下主程序。这里的程序是已经载入了数据的，data是mn 数组，label为m1数组。
实例化DataLoader的第一个参数是Dataset的实例，通过DataLoader，其功能是为下文训练和测试过程提供数据。

    print("data loading ...")
    train_data_set = MyDataSet(trainingDataLoadProcess.data, trainingDataLoadProcess.label)
    test_data_set = MyDataSet(testDataLoadProcess.data, testDataLoadProcess.label)
    train_dataloader = DataLoader(train_data_set, batch_size=batch_size, shuffle=shuffle)
    test_dataloader = DataLoader(test_data_set, batch_size=batch_size, shuffle=shuffle)
    print("model constructing...")
    model_test = NeuralNetwork()
    loss_fn = nn.CrossEntropyLoss()
    optimizer = torch.optim.Adamax(model_test.parameters(), lr=learning_rate, betas=(beta_1, beta_2), weight_decay=0.0)
    print(model_test)
    for t in range(epochs):
        print(f'Epoch{t + 1}\n-------------') 
        train_loop(train_dataloader, model_test, loss_fn, optimizer)
        test_loop(test_dataloader, model_test, loss_fn)
    print("Done!")

定义训练阶段，从DataLoader中取出数据，这里X，y分别为batch_sizen，batch_size1的数据。
首先要进行一个调整，将X调整为batch_size1n的float,设置float的转化过程放在Dataset的初始化函数里完成了
注意：如果没有这一步会报错

期望是long但得到了float的错误。(虽然我也不明白为啥错误不是期望float...）

y为1*batch的数组并转化成long（这里y的形式可能与损失函数有关）

def train_loop(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    for batch, (X, y) in enumerate(dataloader):
        X = X.unsqueeze(1)
        y=y.squeeze(1).long()
        pred = model(X)
        loss = loss_fn(pred, y)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

定义测试过程（同上）

def test_loop(dataloader, model, loss_fn):
    size = len(dataloader.dataset)
    num_batches = len(dataloader)
    test_loss, correct = 0, 0
    with torch.no_grad():
        for X, y in dataloader:
            X = X.unsqueeze(1)
            y=y.squeeze(1).long()
            pred = model(X)
            test_loss = loss_fn(pred, y).item()
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
    test_loss /= num_batches
    correct /= size