move tensor to cpu before numpy (#1949)

* move tensor to cpu before numpy

* update changelog

CHANGELOG.md

@@ -12,6 +12,7 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 
 **Fixed**
 - Fixed a bug in `TimeSeries.from_dataframe()` when using a pandas.DataFrame with `df.columns.name != None`. [#1938](https://github.com/unit8co/darts/pull/1938) by [Antoine Madrona](https://github.com/madtoinou).
+- Fixed a bug when using `TFTExplainer` with a `TFTModel` running on GPU. [#1949](https://github.com/unit8co/darts/pull/1949) by [Dennis Bader](https://github.com/dennisbader).
 
 
 ## [0.25.0](https://github.com/unit8co/darts/tree/0.25.0) (2023-08-04)

darts/explainability/tft_explainer.py

@@ -207,7 +207,7 @@ def explain(
         # get the weights and the attention head from the trained model for the prediction
         # aggregate over attention heads
         attention_heads = (
-            self.model.model._attn_out_weights.detach().numpy().sum(axis=-2)
+            self.model.model._attn_out_weights.detach().cpu().numpy().sum(axis=-2)
         )
         # get the variable importances (pd.DataFrame with rows corresponding to the number of input series)
         encoder_importance = self._encoder_importance
@@ -501,7 +501,8 @@ def _get_importance(
 
         # transform the encoder/decoder weights to percentages, rounded to n_decimals
         weights_percentage = (
-            weight.detach().numpy().mean(axis=1).squeeze(axis=1).round(n_decimals) * 100
+            weight.detach().cpu().numpy().mean(axis=1).squeeze(axis=1).round(n_decimals)
+            * 100
         )
         # create a dataframe with the variable names and the weights
         name_mapping = self._name_mapping