Hi guys,
I would like to implement a model using as input features cumulative sums of classes resulting from sampled sequences classification.
Let’s imagine we want to build signals (hourly or daily) which would be the sums of classified sequences of text, into a fix number of classes. And we want to feed a larger model with this multivariate feature. So in forward method we should iterate over sequences to classify it and build resulting sums of classes for each time steps.
What would be the most efficient way to implement this ? Ideally, I would like the last linear layer from the seq2class model to be trainable.