Logits are unnormalized outputs of a neural network. In a classification network with softmax as the last layer, logits are the inputs to the softmax layer, but since some libraries compute the softmax as part of the loss function, the network doesn't need a softmax layer, so the outputs of the network are the logits.