Differences with torch.nn.SmoothL1Loss

torch.nn.SmoothL1Loss

class torch.nn.SmoothL1Loss(size_average=None, reduce=None, reduction='mean', beta=1.0)(input, target) -> Tensor

For more information, see torch.nn.SmoothL1Loss.

mindspore.nn.SmoothL1Loss

class mindspore.nn.SmoothL1Loss(beta=1.0, reduction='none')(logits, labels) -> Tensor

For more information, see mindspore.nn.SmoothL1Loss.

Differences

PyTorch: SmoothL1Loss loss function. If the element-wise absolute error between the predicted and target values is less than a set threshold beta, use a squared term, otherwise with an absolute error term.

MindSpore: There are no functional differences except for two parameters that have been deprecated in PyTorch.

Categories	Subcategories	PyTorch	MindSpore	Difference
Parameters	Parameter 1	size_average	-	Deprecated, replaced by reduction. MindSpore does not have this parameter
	Parameter 2	reduce	-	Deprecated, replaced by reduction. MindSpore does not have this parameter
	Parameter 3	reduction	reduction	Same function, different parameter names
	Parameter 4	beta	beta	-
Input	Input 1	input	logits	Same function, different parameter names
	Input 2	target	labels	Same function, different parameter names

Code Example

The two APIs achieve the same function and have the same usage.

# PyTorch
import torch
import torch.nn as nn

beta = 1
loss = nn.SmoothL1Loss(reduction="none", beta=beta)
logits = torch.FloatTensor([1, 2, 3])
labels = torch.FloatTensor([1, 2, 2])
output = loss(logits, labels)
print(output.numpy())
# [0.  0.  0.5]

# MindSpore
import mindspore
import numpy as np
from mindspore import Tensor

loss = mindspore.nn.SmoothL1Loss()
logits = Tensor(np.array([1, 2, 3]), mindspore.float32)
labels = Tensor(np.array([1, 2, 2]), mindspore.float32)
output = loss(logits, labels)
print(output)
# [0.  0.  0.5]