python - 如何修复损失函数/softmax中的维度误差？

Question

我正在 PyTorch 中为 XOR 实现逻辑回归（我不希望它运行良好，这只是一个演示）。出于某种原因，我收到错误“IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)”。我不清楚这起源于哪里。错误指向训练期间的 log_softmax。

import torch.nn as nn
import torch.nn.functional as F

class LogisticRegression(nn.Module):
  # input_size: Dimensionality of input feature vector.
  # num_classes: The number of classes in the classification problem.
  def __init__(self, input_size, num_classes):
    # Always call the superclass (nn.Module) constructor first!
    super(LogisticRegression, self).__init__()
    # Set up the linear transform
    self.linear = nn.Linear(input_size, num_classes)

  # Forward's sole argument is the input.
  # input is of shape (batch_size, input_size)
  def forward(self, x):
    # Apply the linear transform.
    # out is of shape (batch_size, num_classes)
    out = self.linear(x)
    # Softmax the out tensor to get a log-probability distribution
    # over classes for each example.
    out_distribution = F.softmax(out, dim=-1)
    return out_distribution


# Binary classifiation
num_outputs = 1
num_input_features = 2

# Create the logistic regression model
logreg_clf = LogisticRegression(num_input_features, num_outputs)

print(logreg_clf)


lr_rate = 0.001

X = torch.Tensor([[0,0],[0,1], [1,0], [1,1]])
Y = torch.Tensor([0,1,1,0]).view(-1,1) #view is similar to numpy.reshape()

# Run the forward pass of the logistic regression model
sample_output = logreg_clf(X) #completely random at the moment
print(X)

loss_function = nn.CrossEntropyLoss() # computes softmax and then the cross entropy
optimizer = torch.optim.SGD(logreg_clf.parameters(), lr=lr_rate)



from torch.autograd import Variable
#training loop: 

epochs = 201 #how many times we go through the training set
steps = X.size(0) #steps = 4; we have 4 training examples

for i in range(epochs):
    for j in range(steps):
        #sample from the training set: 
        data_point = np.random.randint(X.size(0))
        x_var = Variable(X[data_point], requires_grad=False)
        y_var = Variable(Y[data_point], requires_grad=False)

        optimizer.zero_grad() # zero the gradient buffers
        y_hat = logreg_clf(x_var) #get the output from the model
        loss = loss_function.forward(y_hat, y_var) #calculate the loss
        loss.backward() #backprop
        optimizer.step() #does the update

    if i % 500 == 0:
        print ("Epoch: {0}, Loss: {1}, ".format(i, loss.data.numpy()))

标签： pythonpytorch

Answer 1

首先，您正在执行二进制分类任务。所以输出特征的数量应该是2；即，num_outputs = 1。

其次，正如在nn.CrossEntropyLoss()文档中声明的那样，该.forward方法接受两个张量，如下所示：

• Input: (N, C)哪里C是班级的数量（在你的情况下是2）。
• Target: (N)

N在上面的示例中是您传递给损失函数的训练示例的数量；为简单起见，您可以将其设置为 1（即，对每个实例进行前向传递，然后更新梯度）。

注意：此外，您不需要使用.Softmax()beforenn.CrossEntropyLoss()模块，因为此类已nn.LogSoftmax包含在其自身中。

我修改了您的代码，如下所示，这是您的代码段的一个工作示例：

import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import torch

class LogisticRegression(nn.Module):
  # input_size: Dimensionality of input feature vector.
  # num_classes: The number of classes in the classification problem.
  def __init__(self, input_size, num_classes):
    # Always call the superclass (nn.Module) constructor first!
    super(LogisticRegression, self).__init__()
    # Set up the linear transform
    self.linear = nn.Linear(input_size, num_classes)

  # Forward's sole argument is the input.
  # input is of shape (batch_size, input_size)
  def forward(self, x):
    # Apply the linear transform.
    # out is of shape (batch_size, num_classes)
    out = self.linear(x)
    # Softmax the out tensor to get a log-probability distribution
    # over classes for each example.
    return out


# Binary classifiation
num_outputs = 2
num_input_features = 2

# Create the logistic regression model
logreg_clf = LogisticRegression(num_input_features, num_outputs)

print(logreg_clf)


lr_rate = 0.001

X = torch.Tensor([[0,0],[0,1], [1,0], [1,1]])
Y = torch.Tensor([0,1,1,0]).view(-1,1) #view is similar to numpy.reshape()

# Run the forward pass of the logistic regression model
sample_output = logreg_clf(X) #completely random at the moment
print(X)

loss_function = nn.CrossEntropyLoss() # computes softmax and then the cross entropy
optimizer = torch.optim.SGD(logreg_clf.parameters(), lr=lr_rate)



from torch.autograd import Variable
#training loop:

epochs = 201 #how many times we go through the training set
steps = X.size(0) #steps = 4; we have 4 training examples

for i in range(epochs):
    for j in range(steps):
        #sample from the training set:
        data_point = np.random.randint(X.size(0))
        x_var = Variable(X[data_point], requires_grad=False).unsqueeze(0)
        y_var = Variable(Y[data_point], requires_grad=False).long()

        optimizer.zero_grad() # zero the gradient buffers
        y_hat = logreg_clf(x_var) #get the output from the model
        loss = loss_function(y_hat, y_var) #calculate the loss
        loss.backward() #backprop
        optimizer.step() #does the update

    if i % 500 == 0:
        print ("Epoch: {0}, Loss: {1}, ".format(i, loss.data.numpy()))

更新

要获得 0 或 1 的预测类标签：

pred = np.argmax(y_hat.detach().numpy, axis=0)

至于.detach()函数，numpy 期望张量/数组与计算图分离；即，张量不应该有require_grad=True，而分离方法会为你解决问题。