Question

我有一个网络在5D输入张量上执行3D卷积。我的网络的输出，如果大小（1,12,60,36,60）对应于（BatchSize，NumClasses，x-dim，y-dim，z-dim）。我需要计算一个体素方面的交叉熵损失。但是我继续犯错误。

尝试使用torch.nn.CrossEntropyLoss()计算交叉熵损失时，我不断收到以下错误消息：

RuntimeError: multi-target not supported at .../src/THCUNN/generic/ClassNLLCriterion.cu:16

这是我的代码的摘录：

import torch
import torch.nn as nn
from torch.autograd import Variable
criterion = torch.nn.CrossEntropyLoss()
images = Variable(torch.randn(1, 12, 60, 36, 60)).cuda()
labels = Variable(torch.zeros(1, 12, 60, 36, 60).random_(2)).long().cuda()
loss = criterion(images.view(1,-1), labels.view(1,-1))

当我为标签创建一个热张量时也会发生同样的情况：

nclasses = 12
labels = (np.random.randint(0,12,(1,60,36,60))) # Random labels with values between [0..11]
labels = (np.arange(nclasses) == labels[..., None] - 1).astype(int) # Converts labels to one_hot_tensor
a = np.transpose(labels,(0,4,3,2,1)) #  Reorder dimensions to match shape of "images" ([1, 12, 60, 36, 60])
b = Variable(torch.from_numpy(a)).cuda()
loss = criterion(images.view(1,-1), b.view(1,-1))

知道我做错了什么吗？有人能提供在5D输出张量上计算交叉熵的例子吗？

Answer 1

刚刚检查了一些实现（fcn）的2D语义分割，并尝试将其应用于3D语义分割。不能保证这是正确的，我不得不仔细检查......

 import torch
 import torch.nn.functional as F
 def cross_entropy3d(input, target, weight=None, size_average=True):
    # input: (n, c, h, w, z), target: (n, h, w, z)
    n, c, h, w , z = input.size()
    # log_p: (n, c, h, w, z)
    log_p = F.log_softmax(input, dim=1)
    # log_p: (n*h*w*z, c)
    log_p = log_p.permute(0, 4, 3, 2, 1).contiguous().view(-1, c) # make class dimension last dimension
    log_p = log_p[target.view(n, h, w, z, 1).repeat(1, 1, 1, 1, c) >= 0] # this looks wrong -> Should rather be a one-hot vector
    log_p = log_p.view(-1, c)
    # target: (n*h*w*z,)
    mask = target >= 0
    target = target[mask]
    loss = F.nll_loss(log_p, target.view(-1), weight=weight, size_average=False)
    if size_average:
        loss /= mask.data.sum()
    return loss
images = Variable(torch.randn(5, 3, 16, 16, 16))
labels = Variable(torch.LongTensor(5, 16, 16, 16).random_(3))
cross_entropy3d(images, labels, weight=None, size_average=True)

Answer 2

docs解释了这种行为（底线，看起来它实际上正在计算稀疏交叉熵损失，因此不需要输出的所有维度的目标，但只有所需的索引... ...他们具体说明：

Input: (N,C), where C = number of classes
Target: (N), where each value is 0 <= targets[i] <= C-1
Output: scalar. If reduce is False, then (N) instead.

我不确定您的用例，但您可能希望使用KL Divergence或Binary Cross Entropy Loss。两者都是在相同大小的输入和目标上定义的。

示例CrossEntropyLoss用于pytorch中的3D语义分段

2 个答案: