比较与torch.nn.CrossEntropyLoss的差异

torch.nn.CrossEntropyLoss

class torch.nn.CrossEntropyLoss(
    weight=None,
    size_average=None,
    ignore_index=-100,
    reduce=None,
    reduction='mean'
)(input, target) -> Tensor

更多内容详见torch.nn.CrossEntropyLoss。

mindspore.nn.CrossEntropyLoss

class mindspore.nn.CrossEntropyLoss(
    weight=None,
    ignore_index=-100,
    reduction='mean',
    label_smoothing=0.0
)(logits, labels) -> Tensor

更多内容详见mindspore.nn.CrossEntropyLoss。

差异对比

PyTorch：计算预测值和目标值之间的交叉熵损失。

MindSpore：MindSpore此API实现功能与PyTorch基本一致，而且目标值支持两种不同的数据形式：类别索引和类别概率。

分类	子类	PyTorch	MindSpore	差异
参数	参数1	weight	weight	-
	参数2	size_average	-	PyTorch已弃用该参数，功能由reduction接替
	参数3	ignore_index	ignore_index	-
	参数4	reduce	-	PyTorch已弃用该参数，功能由reduction接替
	参数5	reduction	reduction	-
	参数6	-	label_smoothing	标签平滑值，用于计算Loss时防止模型过拟合的正则化手段。取值范围为[0.0, 1.0]。默认值：0.0
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

PyTorch和MindSpore都支持目标值为类别索引的情况。

# PyTorch
import torch
import numpy as np

input_torch = np.array([[1.624, -0.611, -0.528, -1.072, 0.865], [-2.301, 1.744, -0.761, 0.319, -0.249], [1.462, -2.060, -0.322, -0.384, 1.133]])
target_torch = np.array([1, 0, 4])
loss = torch.nn.CrossEntropyLoss()
input_torch = torch.tensor(input_torch, requires_grad=True)
target_torch = torch.tensor(target_torch, dtype=torch.long)
output = loss(input_torch, target_torch)
print(round(float(output.detach().numpy()), 3))
# 2.764

# MindSpore
import mindspore
import numpy as np

input_ms = np.array([[1.624, -0.611, -0.528, -1.072, 0.865], [-2.301, 1.744, -0.761, 0.319, -0.249], [1.462, -2.060, -0.322, -0.384, 1.133]])
target_ms = np.array([1, 0, 4])
input_ms = mindspore.Tensor(input_ms, mindspore.float32)
target_ms = mindspore.Tensor(target_ms, mindspore.int32)
loss = mindspore.nn.CrossEntropyLoss()
output = loss(input_ms, target_ms)
print(round(float(output), 3))
# 2.764