#675 add first draft for tft_explainer

CHANGELOG.md

@@ -7,6 +7,7 @@ Darts is still in an early development phase, and we cannot always guarantee bac
 - Fixed edge case in ShapExplainer for regression models where covariates series > target series [#1310](https://https://github.com/unit8co/darts/pull/1310) by [Rijk van der Meulen](https://github.com/rijkvandermeulen)
 - Removed `IPython` as a dependency. [#1331](https://github.com/unit8co/darts/pull/1331) by [Erik Hasse](https://github.com/erik-hasse)
 - New models: `DLinearModel` and `NLinearModel` as proposed in [this paper](https://arxiv.org/pdf/2205.13504.pdf). [#1139](https://github.com/unit8co/darts/pull/1139) by [Julien Herzen](https://github.com/hrzn) and [Greg DeVos](https://github.com/gdevos010).
+- Added new `TFTExplainer` class to implement the Explainable AI part described in [the paper](https://arxiv.org/abs/1912.09363) of the `TFT` model. [#1392](https://github.com/unit8co/darts/pull/1392) by [Sebastian Cattes](https://github.com/cattes).
 
 [Full Changelog](https://github.com/unit8co/darts/compare/0.22.0...master)
 

darts/explainability/__init__.py

@@ -5,3 +5,4 @@
 
 from darts.explainability.explainability_result import ExplainabilityResult
 from darts.explainability.shap_explainer import ShapExplainer
+from darts.explainability.tft_explainer import TFTExplainer

darts/explainability/explainability_result.py

@@ -58,9 +58,7 @@ def get_explanation(
 
         raise_if(
             component is None and len(self.available_components) > 1,
-            ValueError(
-                "The component parameter is required when the model has more than one component."
-            ),
+            "The component parameter is required when the model has more than one component.",
             logger,
         )
 

darts/explainability/tft_explainer.py

@@ -0,0 +1,222 @@
+from typing import Dict, List, Literal, Optional
+
+import matplotlib.pyplot as plt
+import pandas as pd
+from torch import Tensor
+
+from darts import TimeSeries
+from darts.explainability.explainability import (
+    ExplainabilityResult,
+    ForecastingModelExplainer,
+)
+from darts.models import TFTModel
+
+
+class TFTExplainer(ForecastingModelExplainer):
+    def __init__(
+        self,
+        model: TFTModel,
+    ):
+        """
+        Explainer class for the TFT model.
+
+        Parameters
+        ----------
+        model
+            The fitted TFT model to be explained.
+        """
+        super().__init__(model)
+
+        if not model._fit_called:
+            raise ValueError("The model needs to be trained before explaining it.")
+
+        self._model = model
+
+    @property
+    def encoder_importance(self):
+        return self._get_importance(
+            weight=self._model.model._encoder_sparse_weights,
+            names=self._model.model.encoder_variables,
+        )
+
+    @property
+    def decoder_importance(self):
+        return self._get_importance(
+            weight=self._model.model._decoder_sparse_weights,
+            names=self._model.model.decoder_variables,
+        )
+
+    def get_variable_selection_weight(self, plot=False) -> Dict[str, pd.DataFrame]:
+        """Returns the variable selection weight of the TFT model.
+
+        Parameters
+        ----------
+        plot
+            Whether to plot the variable selection weight.
+
+        Returns
+        -------
+        TimeSeries
+            The variable selection weight.
+
+        """
+
+        if plot:
+            # plot the encoder and decoder weights sorted descending
+            self.encoder_importance.plot(kind="bar", title="Encoder weights")
+            self.decoder_importance.plot(kind="bar", title="Decoder weights")
+
+        return {
+            "encoder_importance": self.encoder_importance,
+            "decoder_importance": self.decoder_importance,
+        }
+
+    def explain(self, **kwargs) -> ExplainabilityResult:
+        """Returns the explainability result of the TFT model.
+
+        The explainability result contains the attention heads of the TFT model.
+        The attention heads determine the contribution of time-varying inputs.
+
+        Parameters
+        ----------
+        kwargs
+            Arguments passed to the `predict` method of the TFT model.
+
+        Returns
+        -------
+        ExplainabilityResult
+            The explainability result containing the attention heads.
+
+        """
+        super().explain()
+        # without the predict call, the weights will still bet set to the last iteration of the forward() method
+        # of the _TFTModule class
+        if "n" not in kwargs:
+            kwargs["n"] = self._model.model.output_chunk_length
+
+        _ = self._model.predict(**kwargs)
+
+        # get the weights and the attention head from the trained model for the prediction
+        attention_heads = (
+            self._model.model._attn_out_weights.squeeze().sum(axis=1).detach()
+        )
+
+        # return the explainer result to be used in other methods
+        return ExplainabilityResult(
+            {
+                0: {
+                    "attention_heads": TimeSeries.from_dataframe(
+                        pd.DataFrame(attention_heads).transpose()
+                    ),
+                }
+            },
+        )
+
+    @staticmethod
+    def plot_attention_heads(
+        expl_result: ExplainabilityResult,
+        plot_type: Optional[Literal["all", "time", "heatmap"]] = "time",
+    ):
+        """Plots the attention heads of the TFT model."""
+        attention_heads = expl_result.get_explanation(
+            component="attention_heads",
+            horizon=0,
+        )
+        if plot_type == "all":
+            fig = plt.figure()
+            attention_heads.plot(
+                label="Attention Head",
+                plot_all_components=True,
+                figure=fig,
+            )
+            # move legend to the right side of the figure
+            plt.legend(bbox_to_anchor=(0.95, 1), loc="upper left")
+            plt.xlabel("Time steps in past")
+            plt.ylabel("Attention")
+        elif plot_type == "time":
+            fig = plt.figure()
+            attention_heads.mean(1).plot(label="Mean Attention Head", figure=fig)
+            plt.xlabel("Time steps in past")
+            plt.ylabel("Attention")
+        elif plot_type == "heatmap":
+            avg_attention = attention_heads.values().transpose()
+            fig = plt.figure()
+            plt.imshow(avg_attention, cmap="hot", interpolation="nearest", figure=fig)
+            plt.xlabel("Time steps in past")
+            plt.ylabel("Horizon")
+        else:
+            raise ValueError("`plot_type` must be either 'all', 'time' or 'heatmap'")
+
+    def _get_importance(
+        self,
+        weight: Tensor,
+        names: List[str],
+        n_decimals=3,
+    ) -> pd.DataFrame:
+        """Returns the encoder or decoder variable of the TFT model.
+
+        Parameters
+        ----------
+        weights
+            The weights of the encoder or decoder of the trained TFT model.
+        names
+            The encoder or decoder names saved in the TFT model class.
+        n_decimals
+            The number of decimals to round the importance to.
+
+        Returns
+        -------
+        pd.DataFrame
+            The importance of the variables.
+        """
+        # transform the encoder/decoder weights to percentages, rounded to n_decimals
+        weights_percentage = (
+            weight.mean(axis=1).detach().numpy().mean(axis=0).round(n_decimals) * 100
+        )
+
+        # create a dataframe with the variable names and the weights
+        name_mapping = self._name_mapping
+        importance = pd.DataFrame(
+            weights_percentage,
+            columns=[name_mapping[name] for name in names],
+        )
+
+        # return the importance sorted descending
+        return importance.transpose().sort_values(0, ascending=False).transpose()
+
+    @property
+    def _name_mapping(self) -> Dict[str, str]:
+        """Returns the feature name mapping of the TFT model.
+
+        Returns
+        -------
+        Dict[str, str]
+            The feature name mapping. For example
+            {
+                'past_covariate_0': 'heater',
+                'past_covariate_1': 'year',
+                'past_covariate_2': 'month',
+                'future_covariate_0': 'darts_enc_fc_cyc_month_sin',
+                'future_covariate_1': 'darts_enc_fc_cyc_month_cos',
+                'target_0': 'ice cream',
+             }
+
+        """
+        past_covariates_name_mapping = {
+            f"past_covariate_{i}": colname
+            for i, colname in enumerate(self._model.past_covariate_series.components)
+        }
+        future_covariates_name_mapping = {
+            f"future_covariate_{i}": colname
+            for i, colname in enumerate(self._model.future_covariate_series.components)
+        }
+        target_name_mapping = {
+            f"target_{i}": colname
+            for i, colname in enumerate(self._model.training_series.components)
+        }
+
+        return {
+            **past_covariates_name_mapping,
+            **future_covariates_name_mapping,
+            **target_name_mapping,
+        }

darts/models/forecasting/tft_model.py

@@ -325,6 +325,10 @@ def __init__(
 
         self.output_layer = nn.Linear(self.hidden_size, self.n_targets * self.loss_size)
 
+        self._encoder_sparse_weights = None
+        self._decoder_sparse_weights = None
+        self._attn_out_weights = None
+
     @property
     def reals(self) -> List[str]:
         """
@@ -632,6 +636,9 @@ def forward(
         out = out.view(
             batch_size, self.output_chunk_length, self.n_targets, self.loss_size
         )
+        self._encoder_sparse_weights = encoder_sparse_weights
+        self._decoder_sparse_weights = decoder_sparse_weights
+        self._attn_out_weights = attn_out_weights
 
         # TODO: (Darts) remember this in case we want to output interpretation
         # return self.to_network_output(