Using BCEWithLogitsLossFlat in Tensorflow for Unknown Labels During Inference

Question

I have a classification problem in which I have about 100 classes. I am using a CNN with SoftMax output, and categorical_cross_entropy as the loss function and this works. B