softmax_with_cross_entropy_loss.py 文件源码

def backward(self, sth):
        input, label = self.saved_tensors
        grad_fs = grad_label = None

        if self.needs_input_grad[0]:
            fs = torch.nn.Softmax()(
                torch.autograd.Variable(input, requires_grad=False)
            ).data

            # neg. one hot label
            y = input.new().resize_as_(input).zero_()
            for i, l in enumerate(label):
                y[i, l] = -1.

            fs.add_(y).mul_(1. / len(label))
            grad_fs = fs
        if self.needs_input_grad[1]:
            raise NotImplementedError()

        return grad_fs, grad_label