比较与torch.nn.NLLLoss的功能差异

torch.nn.NLLLoss

torch.nn.NLLLoss(
    weight=None,
    size_average=None,
    ignore_index=-100,
    reduce=None,
    reduction='mean'
)(input, target)

更多内容详见torch.nn.NLLLoss。

mindspore.nn.NLLLoss

class mindspore.nn.NLLLoss(
    weight=None,
    ignore_index=-100,
    reduction='mean'
)(logits, labels)

更多内容详见mindspore.nn.NLLLoss。

差异对比

PyTorch：计算预测值和目标值之间的负对数似然损失。

MindSpore：除两个在PyTorch已弃用的参数不同外，功能上无差异。

分类	子类	PyTorch	MindSpore	差异
参数	参数1	weight	weight	指定各类别的权重
	参数2	size_average	-	已弃用，被reduction取代，MindSpore无此参数
	参数3	ignore_index	ignore_index	指定labels中需要忽略的值(一般为填充值)，使其不对梯度产生影响
	参数4	reduce	-	已弃用，被reduction取代，MindSpore无此参数
	参数5	reduction	reduction	指定应用于输出结果的计算方式
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

代码示例

import numpy as np

data = np.random.randn(2, 2, 3, 3)

# In MindSpore
import mindspore as ms

loss = ms.nn.NLLLoss(ignore_index=-110, reduction="none")
input = ms.Tensor(data, dtype=ms.float32)
target = ms.ops.zeros((2, 3, 3), dtype=ms.int32)
output = loss(input, target)
print(output)
# Out:
# [[[ 0.7047795   0.8196785  -0.7913506 ]
#   [ 0.22157642 -0.18818447 -0.65975004]
#   [ 1.7223285  -0.9269855   0.46461168]]

#  [[ 0.21305805 -2.213903    0.36110482]
#   [-0.1900587  -0.56938815  0.12274747]
#   [ 1.149195   -0.8739661  -1.7944012 ]]]


# In PyTorch
import torch

loss = torch.nn.NLLLoss(ignore_index=-110, reduction="none")
input = torch.tensor(data, dtype=torch.float32)
target = torch.zeros((2, 3, 3), dtype=torch.long)
output = loss(input, target)
print(output)
# Out：
# tensor([[[ 0.7048,  0.8197, -0.7914],
#          [ 0.2216, -0.1882, -0.6598],
#          [ 1.7223, -0.9270,  0.4646]],

#         [[ 0.2131, -2.2139,  0.3611],
#          [-0.1901, -0.5694,  0.1227],
#          [ 1.1492, -0.8740, -1.7944]]])