Differences with torch.nn.BCEWithLogitsLoss

torch.nn.BCEWithLogitsLoss

torch.nn.BCEWithLogitsLoss(
    weight=None,
    size_average=None,
    reduce=None,
    reduction='mean',
    pos_weight=None
)(input, target) -> Tensor

For more information, see torch.nn.BCEWithLogitsLoss.

mindspore.nn.BCEWithLogitsLoss

class mindspore.nn.BCEWithLogitsLoss(
    reduction='mean',
    weight=None,
    pos_weight=None
)(logits, labels) -> Tensor

For more information, see mindspore.nn.BCEWithLogitsLoss.

Differences

PyTorch: Combine the Sigmoid layer and BCELoss in one class to calculate the binary cross-entropy loss between the predicted and target values, making it numerically more stable than using Sigmoid followed by BCELoss separately.

MindSpore: MindSpore API basically implements the same function as PyTorch. Only the input parameter names are different.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	input	logits	input Tensor
	Input 2	target	labels	input Tensor
Parameters	Parameter 1	weight	weight	Same function, same parameter name
	Parameter 2	size_average	-	Same function. PyTorch has deprecated this parameter, while MindSpore does not have this parameter
	Parameter 3	reduce	-	Same function. PyTorch has deprecated this parameter, while MindSpore does not have this parameter
	Parameter 4	reduction	reduction	Same function, same parameter name
	Parameter 5	pos_weight	pos_weight	Same function, same parameter name

Code Example 1

The two APIs achieve the same function and have the same usage. The three parameters of PyTorch BCEWithLogitsLoss operator, weight, reduction, and pos_weight, are functionally identical to the corresponding three parameters of MindSpore BCEWithLogitsLoss operator, with the same parameter names and the same default values. By default, MindSpore can get the same results as PyTorch.

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

np.random.seed(1)
input = Tensor(np.random.rand(1, 2, 3).astype(np.float32))
print(input.numpy())
# [[[4.17021990e-01 7.20324516e-01 1.14374816e-04]
#   [3.02332580e-01 1.46755889e-01 9.23385918e-02]]]
target = Tensor(np.random.randint(2,size=(1, 2, 3)).astype(np.float32))
print(target.numpy())
# [[[0. 1. 1.]
#   [0. 0. 1.]]]
torch_BCEWithLogitsLoss = torch.nn.BCEWithLogitsLoss()
torch_output = torch_BCEWithLogitsLoss(input, target)
torch_output_np = torch_output.numpy()
print(torch_output_np)
# 0.7142954

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

np.random.seed(1)
logits = Tensor(np.random.rand(1, 2, 3).astype(np.float32))
print(logits.asnumpy())
# [[[4.17021990e-01 7.20324516e-01 1.14374816e-04]
#   [3.02332580e-01 1.46755889e-01 9.23385918e-02]]]
labels = Tensor(np.random.randint(2,size=(1, 2, 3)).astype(np.float32))
print(labels.asnumpy())
# [[[0. 1. 1.]
#   [0. 0. 1.]]]
ms_BCEWithLogitsLoss = mindspore.nn.BCEWithLogitsLoss()
ms_output = ms_BCEWithLogitsLoss(logits, labels)
ms_output_np = ms_output.asnumpy()
print(ms_output_np)
# 0.71429545