use scalers from coreforecast (#873)

* use scalers from coreforecast * update base class * update docstrnig * remove base scaler
Nixtla · Jan 25, 2024 · f3ff57b · f3ff57b
1 parent 8a05d89
commit f3ff57b
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 29 deletions.
diff --git a/environment.yml b/environment.yml
@@ -3,6 +3,7 @@ channels:
   - conda-forge
   - pytorch
 dependencies:
+  - coreforecast>=0.0.3
   - hyperopt
   - jupyterlab
   - matplotlib
@@ -19,4 +20,4 @@ dependencies:
     - black
     - polars
     - "ray[tune]>=2.2.0"
-    - utilsforecast>=0.0.24
+    - utilsforecast>=0.0.25
diff --git a/nbs/core.ipynb b/nbs/core.ipynb
@@ -65,15 +65,13 @@
     "import pandas as pd\n",
     "import torch\n",
     "import utilsforecast.processing as ufp\n",
-    "from utilsforecast.compat import DataFrame, Series, pl_DataFrame, pl_Series\n",
-    "from utilsforecast.grouped_array import GroupedArray\n",
-    "from utilsforecast.target_transforms import (\n",
-    "    BaseTargetTransform,\n",
-    "    LocalBoxCox,\n",
-    "    LocalMinMaxScaler,            \n",
-    "    LocalRobustScaler,            \n",
+    "from coreforecast.grouped_array import GroupedArray\n",
+    "from coreforecast.scalers import (\n",
+    "    LocalMinMaxScaler,\n",
+    "    LocalRobustScaler,\n",
     "    LocalStandardScaler,\n",
     ")\n",
+    "from utilsforecast.compat import DataFrame, Series, pl_DataFrame, pl_Series\n",
     "from utilsforecast.validation import validate_freq\n",
     "\n",
     "from neuralforecast.tsdataset import TimeSeriesDataset\n",
@@ -233,7 +231,6 @@
     "    'robust': lambda: LocalRobustScaler(scale='mad'),\n",
     "    'robust-iqr': lambda: LocalRobustScaler(scale='iqr'),\n",
     "    'minmax': LocalMinMaxScaler,\n",
-    "    'boxcox': LocalBoxCox,\n",
     "}"
    ]
   },
@@ -286,7 +283,7 @@
     "            Frequency of the data. Must be a valid pandas or polars offset alias, or an integer.\n",
     "        local_scaler_type : str, optional (default=None)\n",
     "            Scaler to apply per-serie to all features before fitting, which is inverted after predicting.\n",
-    "            Can be 'standard', 'robust', 'robust-iqr', 'minmax' or 'boxcox'\n",
+    "            Can be 'standard', 'robust', 'robust-iqr' or 'minmax'\n",
     "        \n",
     "        Returns\n",
     "        -------\n",
@@ -301,7 +298,7 @@
     "        if local_scaler_type is not None and local_scaler_type not in _type2scaler:\n",
     "            raise ValueError(f'scaler_type must be one of {_type2scaler.keys()}')\n",
     "        self.local_scaler_type = local_scaler_type\n",
-    "        self.scalers_: Dict[str, BaseTargetTransform]\n",
+    "        self.scalers_: Dict\n",
     "\n",
     "        # Flags and attributes\n",
     "        self._fitted = False\n",
@@ -314,10 +311,10 @@
     "        for i, col in enumerate(dataset.temporal_cols):\n",
     "            if col == 'available_mask':\n",
     "                continue\n",
-    "            self.scalers_[col] = _type2scaler[self.local_scaler_type]()                \n",
-    "            ga = GroupedArray(dataset.temporal[:, i].numpy(), dataset.indptr)\n",
-    "            dataset.temporal[:, i] = torch.from_numpy(self.scalers_[col].fit_transform(ga))\n",
-    "    \n",
+    "            ga = GroupedArray(dataset.temporal[:, i].numpy(), dataset.indptr)                \n",
+    "            self.scalers_[col] = _type2scaler[self.local_scaler_type]().fit(ga)\n",
+    "            dataset.temporal[:, i] = torch.from_numpy(self.scalers_[col].transform(ga))\n",
+    "\n",
     "    def _scalers_transform(self, dataset: TimeSeriesDataset) -> None:\n",
     "        if not self.scalers_:\n",
     "            return None\n",
@@ -1466,7 +1463,7 @@
     ")\n",
     "\n",
     "# CV\n",
-    "nf = NeuralForecast(models=models, freq='M', local_scaler_type='robust-iqr')\n",
+    "nf = NeuralForecast(models=models, freq='M', local_scaler_type='robust')\n",
     "cv_res = nf.cross_validation(AirPassengersPanel)\n",
     "# check that the forecasts are similar to the original values (originals are restored directly from the df)\n",
     "np.testing.assert_allclose(\n",

diff --git a/neuralforecast/core.py b/neuralforecast/core.py
@@ -15,15 +15,13 @@
 import pandas as pd
 import torch
 import utilsforecast.processing as ufp
-from utilsforecast.compat import DataFrame, Series, pl_DataFrame, pl_Series
-from utilsforecast.grouped_array import GroupedArray
-from utilsforecast.target_transforms import (
-    BaseTargetTransform,
-    LocalBoxCox,
+from coreforecast.grouped_array import GroupedArray
+from coreforecast.scalers import (
     LocalMinMaxScaler,
     LocalRobustScaler,
     LocalStandardScaler,
 )
+from utilsforecast.compat import DataFrame, Series, pl_DataFrame, pl_Series
 from utilsforecast.validation import validate_freq
 
 from .tsdataset import TimeSeriesDataset
@@ -147,7 +145,6 @@ def _insample_times(
     "robust": lambda: LocalRobustScaler(scale="mad"),
     "robust-iqr": lambda: LocalRobustScaler(scale="iqr"),
     "minmax": LocalMinMaxScaler,
-    "boxcox": LocalBoxCox,
 }
 
 # %% ../nbs/core.ipynb 9
@@ -186,7 +183,7 @@ def __init__(
             Frequency of the data. Must be a valid pandas or polars offset alias, or an integer.
         local_scaler_type : str, optional (default=None)
             Scaler to apply per-serie to all features before fitting, which is inverted after predicting.
-            Can be 'standard', 'robust', 'robust-iqr', 'minmax' or 'boxcox'
+            Can be 'standard', 'robust', 'robust-iqr' or 'minmax'
 
         Returns
         -------
@@ -203,7 +200,7 @@ def __init__(
         if local_scaler_type is not None and local_scaler_type not in _type2scaler:
             raise ValueError(f"scaler_type must be one of {_type2scaler.keys()}")
         self.local_scaler_type = local_scaler_type
-        self.scalers_: Dict[str, BaseTargetTransform]
+        self.scalers_: Dict
 
         # Flags and attributes
         self._fitted = False
@@ -216,11 +213,9 @@ def _scalers_fit_transform(self, dataset: TimeSeriesDataset) -> None:
         for i, col in enumerate(dataset.temporal_cols):
             if col == "available_mask":
                 continue
-            self.scalers_[col] = _type2scaler[self.local_scaler_type]()
             ga = GroupedArray(dataset.temporal[:, i].numpy(), dataset.indptr)
-            dataset.temporal[:, i] = torch.from_numpy(
-                self.scalers_[col].fit_transform(ga)
-            )
+            self.scalers_[col] = _type2scaler[self.local_scaler_type]().fit(ga)
+            dataset.temporal[:, i] = torch.from_numpy(self.scalers_[col].transform(ga))
 
     def _scalers_transform(self, dataset: TimeSeriesDataset) -> None:
         if not self.scalers_:

diff --git a/settings.ini b/settings.ini
@@ -15,7 +15,7 @@ language = English
 custom_sidebar = True
 license = apache2
 status = 2
-requirements = numpy>=1.21.6 pandas>=1.3.5 torch>=2.0.0 pytorch-lightning>=2.0.0 ray[tune]>=2.2.0 optuna utilsforecast>=0.0.24 numba
+requirements = coreforecast>=0.0.3 numpy>=1.21.6 pandas>=1.3.5 torch>=2.0.0 pytorch-lightning>=2.0.0 ray[tune]>=2.2.0 optuna utilsforecast>=0.0.25
 dev_requirements = nbdev black mypy ruff matplotlib hyperopt polars pyarrow
 nbs_path = nbs
 doc_path = _docs