How to measure the feature importance for the DeepAR model?

[andychu00]

I just trained a DeepAR model with both static and dynamic features, however, once I jump into the evaluation step, I found it hard to obtain the feature importance or error contribution. I tried SHAP packages, but it doesn't support the GluonTS dataframe. Therefore, I would like to know if there is any way to measure the feature importance so that I can drop the feature that may not useful for the forecast. Thank you!