I am trying to develop a sequential model (using
LSTMs) that has to predict several features at each time-step. I originally wanted to use
NLLLoss as a loss function but I am so sure about how to handle in such a multi-feature case.
The input of the model corresponds to the different one-hot encoded features concatenated together.
In the doc, it is said that
NLLLoss can accept an input of the form
(N,C,d1,d2,...,dK), where I assume the different d’s are dimensions you have to create for each of the feature to predict ? I tried it but it didn’t work.
Am I using the wrong loss function or is there a proper way to use NLLLoss for thie specific case ?
I googled a lot, but am still lost …