nn.BCEWithLogisticLoss() = F.sigmoid() + F.binary_cross_entropy()/nn.BCELoss()
import torch
import torch.nn as nn
import torch.nn.functional as F
"""
nn.BCEWithLogisticLoss() = F.sigmoid() + F.binary_cross_entropy()/nn.BCELoss()
"""
pred = torch.tensor([[-0.2],[0.2],[0.8]])
target = torch.FloatTensor([[0],[0],[1]])
bce_loss = F.binary_cross_entropy(F.sigmoid(pred),target)
print('bce_out',bce_loss)
bcew_loss = F.binary_cross_entropy_with_logits(pred,target)
print('bcewlogistic_out',bcew_loss )
bce_out tensor(0.5891)
bcewlogistic_out tensor(0.5891)
nn.CrossEntropyLoss = F.log_softmax() + F.nll_loss()/nn.NLLLoss()
import torch
import torch.nn as nn
import torch.nn.functional as F
"""
nn.CrossEntropyLoss = F.log_softmax() + F.nll_loss()/nn.NLLLoss()
"""
data = torch.randn(2, 5)
nn.Softmax()
log_soft = F.log_softmax(data, dim=1)
print('log_soft:', log_soft, '\n')
target = torch.tensor([1, 2])
EntropyLoss = nn.CrossEntropyLoss()
entropy_out = EntropyLoss(data, target)
NLLLoss = nn.NLLLoss()
nll_out = NLLLoss(log_soft, target)
print('entropy_out:', entropy_out)
print('nll_out:', nll_out)
log_soft: tensor([[-3.1368, -2.0374, -1.2469, -1.2996, -1.3236],
[-1.2834, -1.0692, -2.5731, -1.5507, -2.3944]])
entropy_out: tensor(2.3053)
nll_out: tensor(2.3053)