feat(auto): support input_size (#451)

Nixtla · Nov 20, 2024 · 151f9e8 · 151f9e8
1 parent 59f6cc5
commit 151f9e8
Show file tree

Hide file tree

Showing 4 changed files with 65 additions and 8 deletions.
diff --git a/mlforecast/auto.py b/mlforecast/auto.py
@@ -445,6 +445,7 @@ def fit(
         h: int,
         num_samples: int,
         step_size: Optional[int] = None,
+        input_size: Optional[int] = None,
         refit: Union[bool, int] = False,
         loss: Optional[Callable[[DataFrame, DataFrame], float]] = None,
         id_col: str = "unique_id",
@@ -470,6 +471,8 @@ def fit(
             Number of trials to run
         step_size : int, optional (default=None)
             Step size between each cross validation window. If None it will be equal to `h`.
+        input_size : int, optional (default=None)
+            Maximum training samples per serie in each window. If None, will use an expanding window.
         refit : bool or int (default=False)
             Retrain model for each cross validation window.
             If False, the models are trained at the beginning and then used to predict each window.
@@ -550,6 +553,7 @@ def config_fn(trial: optuna.Trial) -> Dict[str, Any]:
                 n_windows=n_windows,
                 h=h,
                 step_size=step_size,
+                input_size=input_size,
                 refit=refit,
                 id_col=id_col,
                 time_col=time_col,

diff --git a/mlforecast/optimization.py b/mlforecast/optimization.py
@@ -30,6 +30,7 @@ def mlforecast_objective(
     n_windows: int,
     h: int,
     step_size: Optional[int] = None,
+    input_size: Optional[int] = None,
     refit: Union[bool, int] = False,
     id_col: str = "unique_id",
     time_col: str = "ds",
@@ -56,6 +57,8 @@ def mlforecast_objective(
         Forecast horizon.
     step_size : int, optional (default=None)
         Step size between each cross validation window. If None it will be equal to `h`.
+    input_size : int, optional (default=None)
+        Maximum training samples per serie in each window. If None, will use an expanding window.
     refit : bool or int (default=False)
         Retrain model for each cross validation window.
         If False, the models are trained at the beginning and then used to predict each window.
@@ -86,6 +89,7 @@ def objective(trial: optuna.Trial) -> float:
             time_col=time_col,
             freq=freq,
             step_size=step_size,
+            input_size=input_size,
         )
         model_copy = clone(model)
         model_params = config["model_params"]

diff --git a/nbs/auto.ipynb b/nbs/auto.ipynb
@@ -523,6 +523,7 @@
     "        h: int,\n",
     "        num_samples: int,\n",
     "        step_size: Optional[int] = None,\n",
+    "        input_size: Optional[int] = None,\n",
     "        refit: Union[bool, int] = False,\n",
     "        loss: Optional[Callable[[DataFrame, DataFrame], float]] = None,\n",
     "        id_col: str = 'unique_id',\n",
@@ -548,6 +549,8 @@
     "            Number of trials to run\n",
     "        step_size : int, optional (default=None)\n",
     "            Step size between each cross validation window. If None it will be equal to `h`.\n",
+    "        input_size : int, optional (default=None)\n",
+    "            Maximum training samples per serie in each window. If None, will use an expanding window.\n",
     "        refit : bool or int (default=False)\n",
     "            Retrain model for each cross validation window.\n",
     "            If False, the models are trained at the beginning and then used to predict each window.\n",
@@ -625,6 +628,7 @@
     "                n_windows=n_windows,\n",
     "                h=h,\n",
     "                step_size=step_size,\n",
+    "                input_size=input_size,\n",
     "                refit=refit,\n",
     "                id_col=id_col,\n",
     "                time_col=time_col,\n",
@@ -818,6 +822,7 @@
        ">                          (df:Union[pandas.core.frame.DataFrame,polars.datafram\n",
        ">                          e.frame.DataFrame], n_windows:int, h:int,\n",
        ">                          num_samples:int, step_size:Optional[int]=None,\n",
+       ">                          input_size:Optional[int]=None,\n",
        ">                          refit:Union[bool,int]=False, loss:Optional[Callable[[\n",
        ">                          Union[pandas.core.frame.DataFrame,polars.dataframe.fr\n",
        ">                          ame.DataFrame],Union[pandas.core.frame.DataFrame,pola\n",
@@ -839,6 +844,7 @@
        "| h | int |  | Forecast horizon. |\n",
        "| num_samples | int |  | Number of trials to run |\n",
        "| step_size | Optional | None | Step size between each cross validation window. If None it will be equal to `h`. |\n",
+       "| input_size | Optional | None | Maximum training samples per serie in each window. If None, will use an expanding window. |\n",
        "| refit | Union | False | Retrain model for each cross validation window.<br>If False, the models are trained at the beginning and then used to predict each window.<br>If positive int, the models are retrained every `refit` windows. |\n",
        "| loss | Optional | None | Function that takes the validation and train dataframes and produces a float.<br>If `None` will use the average SMAPE across series. |\n",
        "| id_col | str | unique_id | Column that identifies each serie. |\n",
@@ -861,6 +867,7 @@
        ">                          (df:Union[pandas.core.frame.DataFrame,polars.datafram\n",
        ">                          e.frame.DataFrame], n_windows:int, h:int,\n",
        ">                          num_samples:int, step_size:Optional[int]=None,\n",
+       ">                          input_size:Optional[int]=None,\n",
        ">                          refit:Union[bool,int]=False, loss:Optional[Callable[[\n",
        ">                          Union[pandas.core.frame.DataFrame,polars.dataframe.fr\n",
        ">                          ame.DataFrame],Union[pandas.core.frame.DataFrame,pola\n",
@@ -882,6 +889,7 @@
        "| h | int |  | Forecast horizon. |\n",
        "| num_samples | int |  | Number of trials to run |\n",
        "| step_size | Optional | None | Step size between each cross validation window. If None it will be equal to `h`. |\n",
+       "| input_size | Optional | None | Maximum training samples per serie in each window. If None, will use an expanding window. |\n",
        "| refit | Union | False | Retrain model for each cross validation window.<br>If False, the models are trained at the beginning and then used to predict each window.<br>If positive int, the models are retrained every `refit` windows. |\n",
        "| loss | Optional | None | Function that takes the validation and train dataframes and produces a float.<br>If `None` will use the average SMAPE across series. |\n",
        "| id_col | str | unique_id | Column that identifies each serie. |\n",
@@ -914,7 +922,7 @@
       "text/markdown": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L574){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L592){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.predict\n",
        "\n",
@@ -934,7 +942,7 @@
       "text/plain": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L574){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L592){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.predict\n",
        "\n",
@@ -972,7 +980,7 @@
       "text/markdown": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L606){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L624){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.save\n",
        "\n",
@@ -988,7 +996,7 @@
       "text/plain": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L606){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L624){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.save\n",
        "\n",
@@ -1022,7 +1030,7 @@
       "text/markdown": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L616){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L634){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.forecast_fitted_values\n",
        "\n",
@@ -1040,7 +1048,7 @@
       "text/plain": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L616){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/auto.py#L634){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "### AutoMLForecast.forecast_fitted_values\n",
        "\n",
@@ -1072,6 +1080,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "import time\n",
+    "\n",
     "import pandas as pd\n",
     "from datasetsforecast.m4 import M4, M4Evaluation, M4Info\n",
     "from sklearn.linear_model import Ridge\n",
@@ -1788,6 +1798,33 @@
     "    preds2.rename(columns={'id': 'unique_id', 'time': 'ds'}),\n",
     ")"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "caf14191-370e-41b9-a322-2c9b0a7a5f6e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#| hide\n",
+    "# using input_size\n",
+    "fit_kwargs = dict(\n",
+    "    n_windows=3,\n",
+    "    h=h,\n",
+    "    num_samples=5,\n",
+    "    optimize_kwargs={'timeout': 60},    \n",
+    ")\n",
+    "\n",
+    "start = time.perf_counter()\n",
+    "auto_mlf.fit(df=train, **fit_kwargs)\n",
+    "no_limit = time.perf_counter() - start\n",
+    "\n",
+    "start = time.perf_counter()\n",
+    "auto_mlf.fit(df=train, input_size=50, **fit_kwargs)\n",
+    "with_limit = time.perf_counter() - start\n",
+    "\n",
+    "assert with_limit < no_limit"
+   ]
   }
  ],
  "metadata": {

diff --git a/nbs/optimization.ipynb b/nbs/optimization.ipynb
@@ -81,6 +81,7 @@
     "    n_windows: int,\n",
     "    h: int,\n",
     "    step_size: Optional[int] = None,\n",
+    "    input_size: Optional[int] = None,\n",
     "    refit: Union[bool, int] = False,\n",
     "    id_col: str = 'unique_id',\n",
     "    time_col: str = 'ds',\n",
@@ -107,6 +108,8 @@
     "        Forecast horizon.\n",
     "    step_size : int, optional (default=None)\n",
     "        Step size between each cross validation window. If None it will be equal to `h`.\n",
+    "    input_size : int, optional (default=None)\n",
+    "        Maximum training samples per serie in each window. If None, will use an expanding window.\n",
     "    refit : bool or int (default=False)\n",
     "        Retrain model for each cross validation window.\n",
     "        If False, the models are trained at the beginning and then used to predict each window.\n",
@@ -136,6 +139,7 @@
     "            time_col=time_col,\n",
     "            freq=freq,\n",
     "            step_size=step_size,\n",
+    "            input_size=input_size,\n",
     "        )\n",
     "        model_copy = clone(model)\n",
     "        model_params = config['model_params']\n",
@@ -232,11 +236,13 @@
        ">                            ial._trial.Trial],Dict[str,Any]], loss:Callable,\n",
        ">                            model:sklearn.base.BaseEstimator,\n",
        ">                            freq:Union[int,str], n_windows:int, h:int,\n",
+       ">                            step_size:Optional[int]=None,\n",
+       ">                            input_size:Optional[int]=None,\n",
        ">                            refit:Union[bool,int]=False,\n",
        ">                            id_col:str='unique_id', time_col:str='ds',\n",
        ">                            target_col:str='y')\n",
        "\n",
-       "optuna objective function for the MLForecast class\n",
+       "*optuna objective function for the MLForecast class*\n",
        "\n",
        "|    | **Type** | **Default** | **Details** |\n",
        "| -- | -------- | ----------- | ----------- |\n",
@@ -247,6 +253,8 @@
        "| freq | Union |  | pandas' or polars' offset alias or integer denoting the frequency of the series. |\n",
        "| n_windows | int |  | Number of windows to evaluate. |\n",
        "| h | int |  | Forecast horizon. |\n",
+       "| step_size | Optional | None | Step size between each cross validation window. If None it will be equal to `h`. |\n",
+       "| input_size | Optional | None |  |\n",
        "| refit | Union | False | Retrain model for each cross validation window.<br>If False, the models are trained at the beginning and then used to predict each window.<br>If positive int, the models are retrained every `refit` windows.         |\n",
        "| id_col | str | unique_id | Column that identifies each serie. |\n",
        "| time_col | str | ds | Column that identifies each timestep, its values can be timestamps or integers. |\n",
@@ -266,11 +274,13 @@
        ">                            ial._trial.Trial],Dict[str,Any]], loss:Callable,\n",
        ">                            model:sklearn.base.BaseEstimator,\n",
        ">                            freq:Union[int,str], n_windows:int, h:int,\n",
+       ">                            step_size:Optional[int]=None,\n",
+       ">                            input_size:Optional[int]=None,\n",
        ">                            refit:Union[bool,int]=False,\n",
        ">                            id_col:str='unique_id', time_col:str='ds',\n",
        ">                            target_col:str='y')\n",
        "\n",
-       "optuna objective function for the MLForecast class\n",
+       "*optuna objective function for the MLForecast class*\n",
        "\n",
        "|    | **Type** | **Default** | **Details** |\n",
        "| -- | -------- | ----------- | ----------- |\n",
@@ -281,6 +291,8 @@
        "| freq | Union |  | pandas' or polars' offset alias or integer denoting the frequency of the series. |\n",
        "| n_windows | int |  | Number of windows to evaluate. |\n",
        "| h | int |  | Forecast horizon. |\n",
+       "| step_size | Optional | None | Step size between each cross validation window. If None it will be equal to `h`. |\n",
+       "| input_size | Optional | None |  |\n",
        "| refit | Union | False | Retrain model for each cross validation window.<br>If False, the models are trained at the beginning and then used to predict each window.<br>If positive int, the models are retrained every `refit` windows.         |\n",
        "| id_col | str | unique_id | Column that identifies each serie. |\n",
        "| time_col | str | ds | Column that identifies each timestep, its values can be timestamps or integers. |\n",