• Loss Function for Single Label Multiple Classes Classification (output is always one class)

    • softmax is the multi-category equivalent of sigmoid

    • softmax, and then the log likelihood (nll) of that - is called cross-entropy loss. In PyTorch, this is available as nn.CrossEntropyLoss

    • This will work only when the output is always a category(class).

      • If input image is something that does not belong to the Classifier’s “n” categories, the output will still be one of those categories which is wrong in real-life scenarios.

  • Loss Function for Multiple Label Multi Class Classification (with no classes for certain inputs). Input Image may not belong to any category that is being predicted.

    • Binary Cross Entropy is useful for such cases. It is roughly RMSE loss with log applied.

    • nn.BCEWithLogitsLoss is the Pytorch class instance