PyTorch中CrossEntropyLoss、BCELoss、BCEWithLogitsLoss的理解

cv2016_DL2024-06-03 13:49

import torch

predict =torch.Tensor( $\[0.5796,0.4403,0.9087$ , $-1.5673,-0.3150,1.6660$ ])

#predict =torch.Tensor( $\[0.5796,0.4403$ , $-1.5673,-0.3150$ ])

print(predict)

target =torch.tensor( $0,2$ )

target_bce =torch.Tensor( $\[1,0,0$ , $0,0,1$ ])

ce_loss=torch.nn.CrossEntropyLoss()

soft_max=torch.nn.Softmax(dim=-1)

sig_max=torch.nn.Sigmoid()

soft_out=soft_max(predict)

sig_out=sig_max(predict)

bce_loss=torch.nn.BCELoss()

bce_loss1=torch.nn.BCEWithLogitsLoss()

print(ce_loss(predict,target))

print(bce_loss(soft_out,target_bce))

print(bce_loss(sig_out,target_bce))

print(bce_loss1(predict,target_bce))

输出：

#predict:

tensor( $\[ 0.5796, 0.4403, 0.9087$ ,
$-1.5673, -0.3150, 1.6660$ ])

#print(ce_loss(predict,target)):

tensor(0.6725)

#print(bce_loss(soft_out,target_bce))
tensor(0.3950)

#print(bce_loss(sig_out,target_bce))

tensor(0.5900)

print(bce_loss1(predict,target_bce))
tensor(0.5900)

结论：

1.sigmoid激活+BCELoss等于BCEWithLogitsLoss

2.BCEWithLogitsLoss和CrossEntropyLoss不一样，但都可以不加激活

3.sigmoid激活+BCELoss和softmax激活+BCELoss有很大区别