Function Differences with torch.nn.functional.binary_cross_entropy_with_logits

torch.nn.functional.binary_cross_entropy_with_logits

torch.nn.functional.binary_cross_entropy_with_logits(input, target, weight=None, size_average=None, reduce=None, reduction='mean', pos_weight=None)

For more information, see torch.nn.functional.binary_cross_entropy_with_logits.

mindspore.ops.binary_cross_entropy_with_logits

mindspore.ops.binary_cross_entropy_with_logits(logits, label, weight, pos_weight, reduction='mean')

For more information, see mindspore.ops.binary_cross_entropy_with_logits.

Differences

PyTorch: Compute the binary cross-entropy loss value between the target and predicted values.

MindSpore: MindSpore API basically implements the same function as PyTorch, but not set default value of weight and pos_weight.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	input	logits	Same function, different parameter names
	Parameter 2	target	label	Same function, different parameter names
	Parameter 3	weight	weight	default value is not set
	Parameter 4	size_average	-	PyTorch deprecated parameters, functionally replaced by the reduction parameter
	Parameter 5	reduce	-	PyTorch deprecated parameters, functionally replaced by the reduction parameter
	Parameter 6	reduction	reduction	Same function, same default values.
	Parameter 7	pos_weight	pos_weight	default value is not set

Code Example 1

import numpy as np
import mindspore
from mindspore import Tensor, ops

logits = Tensor(np.array([[-0.8, 1.2, 0.7], [-0.1, -0.4, 0.7]]), mindspore.float32)
label = Tensor(np.array([[0.3, 0.8, 1.2], [-0.6, 0.1, 2.2]]), mindspore.float32)
weight = Tensor(np.array([1.0, 1.0, 1.0]), mindspore.float32)
pos_weight = Tensor(np.array([1.0, 1.0, 1.0]), mindspore.float32)
output = ops.binary_cross_entropy_with_logits(logits, label, weight, pos_weight)
print(output)
# 0.34636116

import torch

logits = torch.tensor(np.array([[-0.8, 1.2, 0.7], [-0.1, -0.4, 0.7]]))
label = torch.tensor(np.array([[0.3, 0.8, 1.2], [-0.6, 0.1, 2.2]]))
output = torch.nn.functional.binary_cross_entropy_with_logits(logits, label)
print(output)
# tensor(0.3464, dtype=torch.float64)