[ENH] Soft-DTW (#2160)

* started working on soft-dtw

* added tests for soft-dtw

* uncommented distances

* fix example

* fixed example

* fixed indent

* indent

* indent

aeon/distances/__init__.py

@@ -74,6 +74,10 @@
     "shift_scale_invariant_distance",
     "shift_scale_invariant_pairwise_distance",
     "shift_scale_invariant_best_shift",
+    "soft_dtw_distance",
+    "soft_dtw_pairwise_distance",
+    "soft_dtw_alignment_path",
+    "soft_dtw_cost_matrix",
 ]
 
 
@@ -151,6 +155,12 @@
     shift_scale_invariant_distance,
     shift_scale_invariant_pairwise_distance,
 )
+from aeon.distances._soft_dtw import (
+    soft_dtw_alignment_path,
+    soft_dtw_cost_matrix,
+    soft_dtw_distance,
+    soft_dtw_pairwise_distance,
+)
 from aeon.distances._squared import squared_distance, squared_pairwise_distance
 from aeon.distances._twe import (
     twe_alignment_path,

aeon/distances/_distance.py

@@ -62,6 +62,12 @@
     shift_scale_invariant_distance,
     shift_scale_invariant_pairwise_distance,
 )
+from aeon.distances._soft_dtw import (
+    soft_dtw_alignment_path,
+    soft_dtw_cost_matrix,
+    soft_dtw_distance,
+    soft_dtw_pairwise_distance,
+)
 from aeon.distances._squared import squared_distance, squared_pairwise_distance
 from aeon.distances._twe import (
     twe_alignment_path,
@@ -102,6 +108,7 @@ class DistanceKwargs(TypedDict, total=False):
     standardize: bool
     m: int
     max_shift: Optional[int]
+    gamma: float
 
 
 DistanceFunction = Callable[[np.ndarray, np.ndarray, Any], float]
@@ -256,6 +263,14 @@ def distance(
         return sbd_distance(x, y, kwargs.get("standardize", True))
     elif metric == "shift_scale":
         return shift_scale_invariant_distance(x, y, kwargs.get("max_shift", None))
+    elif metric == "soft_dtw":
+        return soft_dtw_distance(
+            x,
+            y,
+            gamma=kwargs.get("gamma", 1.0),
+            itakura_max_slope=kwargs.get("itakura_max_slope"),
+            window=kwargs.get("window"),
+        )
     else:
         if isinstance(metric, Callable):
             return metric(x, y, **kwargs)
@@ -438,6 +453,14 @@ def pairwise_distance(
         return shift_scale_invariant_pairwise_distance(
             x, y, kwargs.get("max_shift", None)
         )
+    elif metric == "soft_dtw":
+        return soft_dtw_pairwise_distance(
+            x,
+            y,
+            gamma=kwargs.get("gamma", 1.0),
+            itakura_max_slope=kwargs.get("itakura_max_slope"),
+            window=kwargs.get("window"),
+        )
     else:
         if isinstance(metric, Callable):
             if y is None and not symmetric:
@@ -630,6 +653,14 @@ def alignment_path(
             kwargs.get("itakura_max_slope"),
             kwargs.get("warp_penalty", 1.0),
         )
+    elif metric == "soft_dtw":
+        return soft_dtw_alignment_path(
+            x,
+            y,
+            gamma=kwargs.get("gamma", 1.0),
+            itakura_max_slope=kwargs.get("itakura_max_slope"),
+            window=kwargs.get("window"),
+        )
     else:
         raise ValueError("Metric must be one of the supported strings")
 
@@ -773,6 +804,14 @@ def cost_matrix(
             kwargs.get("itakura_max_slope"),
             kwargs.get("warp_penalty", 1.0),
         )
+    elif metric == "soft_dtw":
+        return soft_dtw_cost_matrix(
+            x,
+            y,
+            gamma=kwargs.get("gamma", 1.0),
+            itakura_max_slope=kwargs.get("itakura_max_slope"),
+            window=kwargs.get("window"),
+        )
     else:
         raise ValueError("Metric must be one of the supported strings")
 
@@ -826,6 +865,7 @@ def get_distance_function(metric: Union[str, DistanceFunction]) -> DistanceFunct
     'minkowski'     distances.minkowski_distance
     'sbd'           distances.sbd_distance
     'shift_scale'   distances.shift_scale_invariant_distance
+    'soft_dtw'      distances.soft_dtw_distance
     =============== ========================================
 
     Parameters
@@ -884,6 +924,7 @@ def get_pairwise_distance_function(
     'minkowski'     distances.minkowski_pairwise_distance
     'sbd'           distances.sbd_pairwise_distance
     'shift_scale'   distances.shift_scale_invariant_pairwise_distance
+    'soft_dtw'      distances.soft_dtw_pairwise_distance
     =============== ========================================
 
     Parameters
@@ -937,6 +978,7 @@ def get_alignment_path_function(metric: str) -> AlignmentPathFunction:
     'msm'           distances.msm_alignment_path
     'twe'           distances.twe_alignment_path
     'lcss'          distances.lcss_alignment_path
+    'soft_dtw'      distances.soft_dtw_alignment_path
     =============== ========================================
 
     Parameters
@@ -985,6 +1027,7 @@ def get_cost_matrix_function(metric: str) -> CostMatrixFunction:
     'msm'           distances.msm_cost_matrix
     'twe'           distances.twe_cost_matrix
     'lcss'          distances.lcss_cost_matrix
+    'soft_dtw'      distances.soft_dtw_cost_matrix
     =============== ========================================
 
     Parameters
@@ -1142,6 +1185,13 @@ def _resolve_key_from_distance(metric: Union[str, Callable], key: str) -> Any:
         "distance": shift_scale_invariant_distance,
         "pairwise_distance": shift_scale_invariant_pairwise_distance,
     },
+    {
+        "name": "soft_dtw",
+        "distance": soft_dtw_distance,
+        "pairwise_distance": soft_dtw_pairwise_distance,
+        "cost_matrix": soft_dtw_cost_matrix,
+        "alignment_path": soft_dtw_alignment_path,
+    },
 ]
 
 DISTANCES_DICT = {d["name"]: d for d in DISTANCES}