#cross-entropy-loss

Question 1: Single choice

mediumAI-checkedPyTorch

A pretrained DenseNet gets a new classifier and loss:

model.classifier = nn.Sequential(nn.Linear(1024, 256), nn.ReLU(), nn.Dropout(0.2),
                                 nn.Linear(256, 2), nn.LogSoftmax(dim=1))
criterion = nn.NLLLoss()

Why is NLLLoss the right loss here rather than nn.CrossEntropyLoss?

The outputs are already log-probabilities; CrossEntropyLoss would normalize again1
CrossEntropyLoss only works for more than two classes2
NLLLoss applies the softmax that the model is missing3
CrossEntropyLoss needs one-hot encoded targets, while NLLLoss takes class indices4

Timed test