Prediction using Tensors to avoid detachment

[damandak]

Hi,

I'm working on a explainable AI project using N-BEATSx and N-HiTS models with exogenous variables, using gradient-based perturbations. For this, I need the predict functions to receive pytorch tensors, but the predict functions of N-BEATSx and N-HiTS models accept Pandas DataFrames as input. This breaks the computational graph when converting from tensors. This loss of gradient information makes gradient-based perturbations impossible.

I need a way to run predictions directly on PyTorch tensors to maintain the computational graph and enable gradient computation. Is there an existing method or a recommended approach to achieve this within the library?

Any guidance or suggestions would be greatly appreciated.

Thank you!

[jmoralez]

Hey. The dataframes interface is a convenience, if you want to you can directly use the forward methods (NBEATSx, NHITS) which take dicts of tensors and return a tensor.

If you need utilities for going from a dataframe to batches you can use TimeSeriesDataset.from_df and the data module.

[damandak]

Thank you! Using the forward methods, those work great!