grad can be implicitly created only for scalar outputs

grad can be implicitly created only for scalar outputs

自定义损失函数时报错：

import torch
import torch.nn as nn
import numpy as np


class CrossEntropyLoss(nn.Module):
    def __init__(self):
        super(CrossEntropyLoss, self).__init__()

    def forward(self, output, label):


        first = [-output[i][label[i]] for i in range(label.size()[0])]
        first_ = 0
        for i in range(len(first)):
            first_ += first[i]

        second = torch.exp(output)
        second = torch.sum(second, dim=1)
        second = torch.log(second + 1e-5)#**2
        second = torch.sum(second)
        loss = 1 / label.size()[0] * (first_ + second)
        return loss


if __name__ == '__main__':
    output = torch.randn(3,3, 5, requires_grad=True)
    label = torch.empty((3,5), dtype=torch.long).random_(3)

    # x=0.9
    # output = to