Option to run as sequence-to-one and sequence-to-sequence about lstm HOT 5 OPEN

@jmframe,

Are you able to provide more details on this one? Specially, are you suggesting to modify the forward pass via neuralhydrology's customlstm as defined here?

Thank you in advance, cheers!

from lstm.

I think the two options are 1) change the way the forward pass is made, yes, but not through NH, do that through the BMI, and 2) change the way the model is trained with NH to match the way we have implemented it. I think option 1 is much more reasonable, for the short term. In the long term, the thing to do is develop a BMI directly in the NH code, so there is no potential for conflict between training and forward ngen predictions.

from lstm.

@jmframe I'm digging up old issues today. What are the advantages to resetting the state space each time step? It seems unlikely we'd want to multiply the current execution time by the sequence length unless some dramatic performance increases (i.e. better streamflow prediction) resulted. At some point, given time resources, we could retrain LSTM to eliminate the state resetting per your second recommendation.

from lstm.

Well. The advantage of resetting the state space and passing in the full sequence is that is how the model is trained, and that is what the weights of the model are trained to respond to. But, I did quite a bit of experimenting, and didn't find that it was much different, results-wise. The issue of computation and time constraints is an import one. It would be good to evaluate the performance of the large runs to see what this means in terms of extra costs.

There is a funded CIROH project to do all this, which was supposed to start in 2022, but you know the story...

from lstm.

That's good to know @jmframe. I'll leave this issue open.

from lstm.