Pytorch 不支持单热向量? [英] Pytorch doesn't support one-hot vector?

问题描述

我对 Pytorch 如何处理 one-hot 向量感到非常困惑.在这个

然而，labels 不是单热矢量格式.我得到以下 size

print(labels.size())打印(输出.大小())输出>>>火炬.大小([4])输出>>>torch.Size([4, 10])

奇迹般地，我将 outputs 和 labels 传递给 criterion=CrossEntropyLoss()，完全没有错误.

loss = criteria(outputs, labels) # 怎么没有错误?

我的假设:

也许 pytorch 会自动将 labels 转换为 one-hot 矢量形式.因此，我尝试将标签转换为单热向量，然后再将其传递给损失函数.

def to_one_hot_vector(num_class, label):b = np.zeros((label.shape[0], num_class))b[np.arange(label.shape[0]), label] = 1返回 blabel_one_hot = to_one_hot_vector(10,labels)label_one_hot = torch.Tensor(labels_one_hot)label_one_hot = labels_one_hot.type(torch.LongTensor)loss =criteria(outputs, labels_one_hot) # 现在它给了我错误

但是，我收到以下错误

<块引用>

运行时错误:不支持多目标/opt/pytorch/pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15

那么，Pytorch 不支持单热向量?Pytorch 如何计算两个张量 outputs = [1,0,0],[0,0,1] 的 cross entropy 和labels = [0,2] ?目前对我来说完全没有意义.

解决方案

PyTorch 在其 CrossEntropyLoss那个

<块引用>

该标准期望一个类索引(0 到 C-1)作为 minibatch 大小的一维张量的每个值的目标

换句话说，它有你的 to_one_hot_vector 函数在概念上内置在 CEL 中，并且没有公开 one-hot API.请注意，与存储类标签相比，one-hot 向量的内存效率较低.

如果你得到 one-hot 向量并且需要使用类标签格式(例如与 CEL 兼容)，你可以使用 argmax，如下所示:

导入火炬标签 = torch.tensor([1, 2, 3, 5])one_hot = torch.zeros(4, 6)one_hot[torch.arange(4), 标签] = 1恢复 = torch.argmax(one_hot，dim=1)断言(标签==恢复).所有().项目()

I am very confused by how Pytorch deals with one-hot vectors. In this tutorial, the neural network will generate a one-hot vector as its output. As far as I understand, the schematic structure of the neural network in the tutorial should be like:

However, the labels are not in one-hot vector format. I get the following size

print(labels.size())
print(outputs.size())

output>>> torch.Size([4]) 
output>>> torch.Size([4, 10])

Miraculously, I they pass the outputs and labels to criterion=CrossEntropyLoss(), there's no error at all.

loss = criterion(outputs, labels) # How come it has no error?

My hypothesis:

Maybe pytorch automatically convert the labels to one-hot vector form. So, I try to convert labels to one-hot vector before passing it to the loss function.

def to_one_hot_vector(num_class, label):
    b = np.zeros((label.shape[0], num_class))
    b[np.arange(label.shape[0]), label] = 1

    return b

labels_one_hot = to_one_hot_vector(10,labels)
labels_one_hot = torch.Tensor(labels_one_hot)
labels_one_hot = labels_one_hot.type(torch.LongTensor)

loss = criterion(outputs, labels_one_hot) # Now it gives me error

However, I got the following error

RuntimeError: multi-target not supported at /opt/pytorch/pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15

So, one-hot vectors are not supported in Pytorch? How does Pytorch calculates the cross entropy for the two tensor outputs = [1,0,0],[0,0,1] and labels = [0,2] ? It doesn't make sense to me at all at the moment.

解决方案

PyTorch states in its documentation for CrossEntropyLoss that

This criterion expects a class index (0 to C-1) as the target for each value of a 1D tensor of size minibatch

In other words, it has your to_one_hot_vector function conceptually built in CEL and does not expose the one-hot API. Notice that one-hot vectors are memory inefficient compared to storing class labels.

If you are given one-hot vectors and need to go to class labels format (for instance to be compatible with CEL), you can use argmax like below:

import torch
 
labels = torch.tensor([1, 2, 3, 5])
one_hot = torch.zeros(4, 6)
one_hot[torch.arange(4), labels] = 1
 
reverted = torch.argmax(one_hot, dim=1)
assert (labels == reverted).all().item()

这篇关于Pytorch 不支持单热向量?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！