比较与torch.nn.GaussianNLLLoss的差异

torch.nn.GaussianNLLLoss

class torch.nn.GaussianNLLLoss(
    *,
    full=False,
    eps=1e-06,
    reduction='mean'
)(input, target, var) -> Tensor/Scalar

更多内容详见torch.nn.GaussianNLLLoss。

mindspore.nn.GaussianNLLLoss

class mindspore.nn.GaussianNLLLoss(
    *,
    full=False,
    eps=1e-06,
    reduction='mean'
)(logits, labels, var) -> Tensor/Scalar

更多内容详见mindspore.nn.GaussianNLLLoss。

差异对比

PyTorch：服从高斯分布的负对数似然损失。

MindSpore：与PyTorch实现同样的功能。如果var中存在小于0的数字，PyTorch会直接报错，而MindSpore则会计算max(var, eps) 之后，将结果传给log进行计算。

分类	子类	PyTorch	MindSpore	差异
参数	参数1	full	full	功能一致
	参数2	eps	eps	功能一致
	参数3	reduction	reduction	功能一致
输入	输入1	input	logits	功能一致，参数名不同
	输入2	target	labels	功能一致，参数名不同
	输入3	var	var	功能一致

代码示例

两API实现功能和使用方法基本相同，但PyTorch和MindSpore针对输入 var<0 的情况做了不同处理。

# PyTorch
import torch
from torch import nn
import numpy as np

arr1 = np.arange(8).reshape((4, 2))
arr2 = np.array([2, 3, 1, 4, 6, 4, 4, 9]).reshape((4, 2))
logits = torch.tensor(arr1, dtype=torch.float32)
labels = torch.tensor(arr2, dtype=torch.float32)
loss = nn.GaussianNLLLoss(reduction='mean')
var = torch.tensor(np.ones((4, 1)), dtype=torch.float32)
output = loss(logits, labels, var)
# tensor(1.4375)

# 如果var中有小于0的元素，PyTorch会直接报错
var[0] = -1
output2 = loss(logits, labels, var)
# ValueError: var has negative entry/entries

# MindSpore
import numpy as np
from mindspore import Tensor
import mindspore.nn as nn
from mindspore import dtype as mstype

arr1 = np.arange(8).reshape((4, 2))
arr2 = np.array([2, 3, 1, 4, 6, 4, 4, 9]).reshape((4, 2))
logits = Tensor(arr1, mstype.float32)
labels = Tensor(arr2, mstype.float32)
loss = nn.GaussianNLLLoss(reduction='mean')
var = Tensor(np.ones((4, 1)), mstype.float32)
output = loss(logits, labels, var)
print(output)
# 1.4374993

# 如果var中有小于0的元素，MindSpore会使用max(var, eps)的结果
var[0] = -1
output2 = loss(logits, labels, var)
print(output2)
# 499999.22