I am trying to use some open source neural network architecture written in pytorch. The original model cannot handle multi output, the situation where one example is associated with multiple labels. I wonder whether there is any easy way one can adapt it to be able to handle multi output. I can certainly define a class that inherits the model and then change the fit and predict method, but it seems quite complicated.
If you don’t have the (log_)softmax at the end (more often one has not these days), it should be trivial to use sigmoid + thresholding on the scores (the raw model outputs).