pytorch で multi-labal classification に利用されそうなロスとその使い方

multi-label_classification_losses.py

# MultiLabelSoftMarginLoss only
ml_criterion = nn.MultiLabelSoftMarginLoss()
## torch.randn
data, labels = Variable(torch.randn([1, 5])), Variable(torch.randn([1, 5]))
print(data.data, labels.data)
print(ml_criterion(data, labels))
## fixed FloatTensor
data, labels = Variable(torch.FloatTensor([1, 50, 100, 50, 1])), Variable(torch.FloatTensor([0, 0, 1, 0, 0]))
print(data.data, labels.data)
print(ml_criterion(data, labels))

# Sigmoid + BCELoss
ml_criterion = nn.BCELoss()
sigmoid = nn.Sigmoid()
data, labels = Variable(torch.FloatTensor([1, 50, 100, 50, 1])), Variable(torch.FloatTensor([0, 0, 1, 0, 0]))
print(data.data, labels.data)
print(ml_criterion(sigmoid(data), labels))