improve issm api

Author: lostella

src/gluonts/model/deepstate/_estimator.py

@@ -269,7 +269,7 @@ def create_transformation(self) -> Transformation:
                 ),
                 # Unnormalized seasonal features
                 AddTimeFeatures(
-                    time_features=CompositeISSM.seasonal_features(self.freq),
+                    time_features=self.issm.time_features(),
                     pred_length=self.prediction_length,
                     start_field=FieldName.START,
                     target_field=FieldName.TARGET,

src/gluonts/model/deepstate/issm.py

@@ -13,6 +13,8 @@
 
 from typing import List, Tuple
 
+import numpy as np
+import pandas as pd
 from pandas.tseries.frequencies import to_offset
 
 from gluonts.core.component import validated
@@ -89,14 +91,23 @@ def _make_2_block_diagonal(F, left: Tensor, right: Tensor) -> Tensor:
     return _block_diagonal
 
 
+class ZeroFeature(TimeFeature):
+    """
+    A feature that is identically zero.
+    """
+
+    def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:
+        return np.zeros(index.values.shape)
+
+
 class ISSM:
     r"""
     An abstract class for providing the basic structure of Innovation State Space Model (ISSM).
 
     The structure of ISSM is given by
 
         * dimension of the latent state
-        * transition and emission coefficents of the transition model
+        * transition and innovation coefficents of the transition model
         * emission coefficient of the observation model
 
     """
@@ -106,27 +117,30 @@ def __init__(self):
         pass
 
     def latent_dim(self) -> int:
-        raise NotImplemented()
+        raise NotImplementedError
 
     def output_dim(self) -> int:
-        raise NotImplemented()
+        raise NotImplementedError
+
+    def time_features(self) -> List[TimeFeature]:
+        raise NotImplementedError
 
-    def emission_coeff(self, seasonal_indicators: Tensor):
-        raise NotImplemented()
+    def emission_coeff(self, features: Tensor) -> Tensor:
+        raise NotImplementedError
 
-    def transition_coeff(self, seasonal_indicators: Tensor):
-        raise NotImplemented()
+    def transition_coeff(self, features: Tensor) -> Tensor:
+        raise NotImplementedError
 
-    def innovation_coeff(self, seasonal_indicators: Tensor):
-        raise NotImplemented()
+    def innovation_coeff(self, features: Tensor) -> Tensor:
+        raise NotImplementedError
 
     def get_issm_coeff(
-        self, seasonal_indicators: Tensor
+        self, features: Tensor
     ) -> Tuple[Tensor, Tensor, Tensor]:
         return (
-            self.emission_coeff(seasonal_indicators),
-            self.transition_coeff(seasonal_indicators),
-            self.innovation_coeff(seasonal_indicators),
+            self.emission_coeff(features),
+            self.transition_coeff(features),
+            self.innovation_coeff(features),
         )
 
 
@@ -137,52 +151,47 @@ def latent_dim(self) -> int:
     def output_dim(self) -> int:
         return 1
 
+    def time_features(self) -> List[TimeFeature]:
+        return [ZeroFeature()]
+
     def emission_coeff(
-        self, seasonal_indicators: Tensor  # (batch_size, time_length)
+        self, feature: Tensor  # (batch_size, time_length)
     ) -> Tensor:
-        F = getF(seasonal_indicators)
+        F = getF(feature)
 
         _emission_coeff = F.ones(shape=(1, 1, 1, self.latent_dim()))
 
         # get the right shape: (batch_size, seq_length, obs_dim, latent_dim)
         zeros = _broadcast_param(
-            F.zeros_like(
-                seasonal_indicators.slice_axis(
-                    axis=-1, begin=0, end=1
-                ).squeeze(axis=-1)
-            ),
+            feature.slice_axis(axis=-1, begin=0, end=1).squeeze(axis=-1),
             axes=[2, 3],
             sizes=[1, self.latent_dim()],
         )
 
         return _emission_coeff.broadcast_like(zeros)
 
     def transition_coeff(
-        self, seasonal_indicators: Tensor  # (batch_size, time_length)
+        self, feature: Tensor  # (batch_size, time_length)
     ) -> Tensor:
-        F = getF(seasonal_indicators)
+        F = getF(feature)
 
         _transition_coeff = (
             F.eye(self.latent_dim()).expand_dims(axis=0).expand_dims(axis=0)
         )
 
         # get the right shape: (batch_size, seq_length, latent_dim, latent_dim)
         zeros = _broadcast_param(
-            F.zeros_like(
-                seasonal_indicators.slice_axis(
-                    axis=-1, begin=0, end=1
-                ).squeeze(axis=-1)
-            ),
+            feature.slice_axis(axis=-1, begin=0, end=1).squeeze(axis=-1),
             axes=[2, 3],
             sizes=[self.latent_dim(), self.latent_dim()],
         )
 
         return _transition_coeff.broadcast_like(zeros)
 
     def innovation_coeff(
-        self, seasonal_indicators: Tensor  # (batch_size, time_length)
+        self, feature: Tensor  # (batch_size, time_length)
     ) -> Tensor:
-        return self.emission_coeff(seasonal_indicators).squeeze(axis=2)
+        return self.emission_coeff(feature).squeeze(axis=2)
 
 
 class LevelTrendISSM(LevelISSM):
@@ -192,10 +201,13 @@ def latent_dim(self) -> int:
     def output_dim(self) -> int:
         return 1
 
+    def time_features(self) -> List[TimeFeature]:
+        return [ZeroFeature()]
+
     def transition_coeff(
-        self, seasonal_indicators: Tensor  # (batch_size, time_length)
+        self, feature: Tensor  # (batch_size, time_length)
     ) -> Tensor:
-        F = getF(seasonal_indicators)
+        F = getF(feature)
 
         _transition_coeff = (
             (F.diag(F.ones(shape=(2,)), k=0) + F.diag(F.ones(shape=(1,)), k=1))
@@ -205,11 +217,7 @@ def transition_coeff(
 
         # get the right shape: (batch_size, seq_length, latent_dim, latent_dim)
         zeros = _broadcast_param(
-            F.zeros_like(
-                seasonal_indicators.slice_axis(
-                    axis=-1, begin=0, end=1
-                ).squeeze(axis=-1)
-            ),
+            feature.slice_axis(axis=-1, begin=0, end=1).squeeze(axis=-1),
             axes=[2, 3],
             sizes=[self.latent_dim(), self.latent_dim()],
         )
@@ -223,26 +231,28 @@ class SeasonalityISSM(LevelISSM):
     """
 
     @validated()
-    def __init__(self, num_seasons: int) -> None:
+    def __init__(self, num_seasons: int, time_feature: TimeFeature) -> None:
         super(SeasonalityISSM, self).__init__()
         self.num_seasons = num_seasons
+        self.time_feature = time_feature
 
     def latent_dim(self) -> int:
         return self.num_seasons
 
     def output_dim(self) -> int:
         return 1
 
-    def emission_coeff(self, seasonal_indicators: Tensor) -> Tensor:
-        F = getF(seasonal_indicators)
-        return F.one_hot(seasonal_indicators, depth=self.latent_dim())
+    def time_features(self) -> List[TimeFeature]:
+        return [self.time_feature]
 
-    def innovation_coeff(self, seasonal_indicators: Tensor) -> Tensor:
-        F = getF(seasonal_indicators)
-        # seasonal_indicators = F.modulo(seasonal_indicators - 1, self.latent_dim)
-        return F.one_hot(seasonal_indicators, depth=self.latent_dim()).squeeze(
-            axis=2
-        )
+    def emission_coeff(self, feature: Tensor) -> Tensor:
+        F = getF(feature)
+        return F.one_hot(feature, depth=self.latent_dim())
+
+    def innovation_coeff(self, feature: Tensor) -> Tensor:
+        F = getF(feature)
+        # feature = F.modulo(feature - 1, self.latent_dim)
+        return F.one_hot(feature, depth=self.latent_dim()).squeeze(axis=2)
 
 
 class CompositeISSM(ISSM):
@@ -269,6 +279,12 @@ def latent_dim(self) -> int:
     def output_dim(self) -> int:
         return self.nonseasonal_issm.output_dim()
 
+    def time_features(self) -> List[TimeFeature]:
+        ans = self.nonseasonal_issm.time_features()
+        for issm in self.seasonal_issms:
+            ans.extend(issm.time_features())
+        return ans
+
     @classmethod
     def get_from_freq(cls, freq: str, add_trend: bool = DEFAULT_ADD_TREND):
         offset = to_offset(freq)
@@ -277,71 +293,63 @@ def get_from_freq(cls, freq: str, add_trend: bool = DEFAULT_ADD_TREND):
 
         if offset.name == "M":
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=12)  # month-of-year seasonality
+                SeasonalityISSM(  # month-of-year seasonality
+                    num_seasons=12, time_feature=MonthOfYear(normalized=False)
+                )
             ]
         elif offset.name == "W-SUN":
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=53)  # week-of-year seasonality
+                SeasonalityISSM(  # week-of-year seasonality
+                    num_seasons=53, time_feature=WeekOfYear(normalized=False)
+                )
             ]
         elif offset.name == "D":
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=7)
-            ]  # day-of-week seasonality
+                SeasonalityISSM(  # day-of-week seasonality
+                    num_seasons=7, time_feature=DayOfWeek(normalized=False)
+                )
+            ]
         elif offset.name == "B":  # TODO: check this case
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=7)
-            ]  # day-of-week seasonality
+                SeasonalityISSM(  # day-of-week seasonality
+                    num_seasons=7, time_feature=DayOfWeek(normalized=False)
+                )
+            ]
         elif offset.name == "H":
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=24),  # hour-of-day seasonality
-                SeasonalityISSM(num_seasons=7),  # day-of-week seasonality
+                SeasonalityISSM(  # hour-of-day seasonality
+                    num_seasons=24, time_feature=HourOfDay(normalized=False)
+                ),
+                SeasonalityISSM(  # day-of-week seasonality
+                    num_seasons=7, time_feature=DayOfWeek(normalized=False)
+                ),
             ]
         elif offset.name == "T":
             seasonal_issms = [
-                SeasonalityISSM(num_seasons=60),  # minute-of-hour seasonality
-                SeasonalityISSM(num_seasons=24),  # hour-of-day seasonality
+                SeasonalityISSM(  # minute-of-hour seasonality
+                    num_seasons=60, time_feature=MinuteOfHour(normalized=False)
+                ),
+                SeasonalityISSM(  # hour-of-day seasonality
+                    num_seasons=24, time_features=HourOfDay(normalized=False)
+                ),
             ]
         else:
             RuntimeError(f"Unsupported frequency {offset.name}")
 
         return cls(seasonal_issms=seasonal_issms, add_trend=add_trend)
 
-    @classmethod
-    def seasonal_features(cls, freq: str) -> List[TimeFeature]:
-        offset = to_offset(freq)
-        if offset.name == "M":
-            return [MonthOfYear(normalized=False)]
-        elif offset.name == "W-SUN":
-            return [WeekOfYear(normalized=False)]
-        elif offset.name == "D":
-            return [DayOfWeek(normalized=False)]
-        elif offset.name == "B":  # TODO: check this case
-            return [DayOfWeek(normalized=False)]
-        elif offset.name == "H":
-            return [HourOfDay(normalized=False), DayOfWeek(normalized=False)]
-        elif offset.name == "T":
-            return [
-                MinuteOfHour(normalized=False),
-                HourOfDay(normalized=False),
-            ]
-        else:
-            RuntimeError(f"Unsupported frequency {offset.name}")
-
-        return []
-
     def get_issm_coeff(
-        self, seasonal_indicators: Tensor  # (batch_size, time_length)
+        self, features: Tensor  # (batch_size, time_length)
     ) -> Tuple[Tensor, Tensor, Tensor]:
-        F = getF(seasonal_indicators)
+        F = getF(features)
         emission_coeff_ls, transition_coeff_ls, innovation_coeff_ls = zip(
-            self.nonseasonal_issm.get_issm_coeff(seasonal_indicators),
             *[
                 issm.get_issm_coeff(
-                    seasonal_indicators.slice_axis(
-                        axis=-1, begin=ix, end=ix + 1
-                    )
+                    features.slice_axis(axis=-1, begin=ix, end=ix + 1)
+                )
+                for ix, issm in enumerate(
+                    [self.nonseasonal_issm] + self.seasonal_issms
                 )
-                for ix, issm in enumerate(self.seasonal_issms)
             ],
         )