Does “cnn_learner” class creates different NN architecture depending if it is single label or multi label data set? I mean does it change in the case last layer from Softmax to Sigmoids?
The softmax/sigmoid is in the loss function, but yes, the function chooses a loss function by looking at what the targets are, so if you assembled your data through fastai, it will pick the right one for you.
You can always check with
learn.loss_func it didn’t make a mistake.