Temporal Model Analyzer - Enhancement

[edpclau]

Hi!

I've been using TemporalModelAnalyzer to explain my RNNs and I'm running into the issue that my model outputs and targets are 2D arrays, rather than the 1D arrays required by TemporalModelAnalyzer. That is, my model makes a prediction for each timepoint in my sequence. Is there a chance that support for 2D targets and outputs could be added?

[cloudbopper]

Thanks for using our work! The method currently only supports many-to-one models, not many-to-many models, though it's one of our highest priority items for future work. I don't yet have an ETA on when that might be added, however. You might explore other explanation methods that are designed for sequential models. An approach I'm aware of that does local explanations of sequence-to-sequence models is Feature Importance in Time (FIT) - https://github.com/sanatonek/time_series_explainability