chexpert | before add weights to each model , my best auc | Machine Learning library

It looks like you have adapted the training correctly for the binary classification, but the prediction wasn't, as you are still trying it as if it were a multi-class prediction.

The output of your model (varOutput) has the size (batch_size, 1), since there is only one class. The maximum across that dimension will always be 0, since that is the only class available, there is no separate class for 1.

This single class represents both cases (0 and 1), so you can consider it is a the probability of it being positive (1). To get the distinct value of either 0 or 1, you simply use a threshold of 0.5, so everything below that receives the class 0 and above that 1. This can be easily done with torch.round.

But you also have another problem, you're applying the sigmoid function twice in a row, once in the classifier nn.Sigmoid() and then afterwards again torch.sigmoid(varOutput). That is problematic, because sigmoid(0) = 0.5, hence all your probabilities are over 0.5.

The output of your model are already the probabilities, the only thing left is to round them: