I have a dataset where I calculate one-hot encoded labels for the hugging face trainer. However I have to drop some labels before training, but I don't know which ones exactly. So the resulting label space looks something like this:
{[1,0,0,0], [0,0,1,0], [0,0,0,1]}
Note how [0,1,0,0] is not in the list.
Now, when evaluating the model, it returns an array of probabilities. Assuming I have hundreds of labels and don't know which ones have been dropped, how do I map the trainer's output (e.g. after softmax) to the correct labels?
