Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "bdb78d22036604f5ecb43d0c3b10bc0d05729113"
Commit e3f7f7b3 authored by chenzk's avatar chenzk
Browse files

v1.0

parents
Pipeline #956 failed with stages
in 0 seconds
Hide whitespace changes
Inline Side-by-side
Showing with 8263 additions and 0 deletions
nbs/nbdev.yml 0 → 100644
View file @ e3f7f7b3
project:
output-dir: _docs
website:
title: "neuralforecast"
site-url: "https://Nixtla.github.io/neuralforecast/"
description: "Time series forecasting suite using deep learning models"
repo-branch: main
repo-url: "https://github.com/Nixtla/neuralforecast/"
nbs/styles.css 0 → 100644
View file @ e3f7f7b3
.cell-output pre {
margin-left: 0.8rem;
margin-top: 0;
background: none;
border-left: 2px solid lightsalmon;
border-top-left-radius: 0;
border-top-right-radius: 0;
}
.cell-output .sourceCode {
background: none;
margin-top: 0;
}
.cell > .sourceCode {
margin-bottom: 0;
}
\ No newline at end of file
nbs/tsdataset.ipynb 0 → 100644
View file @ e3f7f7b3
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "524620c1",
"metadata": {},
"outputs": [],
"source": [
"#| default_exp tsdataset"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "15392f6f",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"%load_ext autoreload\n",
"%autoreload 2"
]
},
{
"cell_type": "markdown",
"id": "12fa25a4",
"metadata": {},
"source": [
"# PyTorch Dataset/Loader\n",
"> Torch Dataset for Time Series\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "2508f7a9-1433-4ad8-8f2f-0078c6ed6c3c",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"from fastcore.test import test_eq\n",
"from nbdev.showdoc import show_doc\n",
"from neuralforecast.utils import generate_series"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "44065066-e72a-431f-938f-1528adef9fe8",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"import warnings\n",
"from collections.abc import Mapping\n",
"from typing import List, Optional, Union\n",
"\n",
"import numpy as np\n",
"import pandas as pd\n",
"import pytorch_lightning as pl\n",
"import torch\n",
"import utilsforecast.processing as ufp\n",
"from torch.utils.data import Dataset, DataLoader\n",
"from utilsforecast.compat import DataFrame, pl_Series"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "323a7a6e-38c3-496d-8f1e-cad05f643d41",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class TimeSeriesLoader(DataLoader):\n",
" \"\"\"TimeSeriesLoader DataLoader.\n",
" [Source code](https://github.com/Nixtla/neuralforecast1/blob/main/neuralforecast/tsdataset.py).\n",
"\n",
" Small change to PyTorch's Data loader. \n",
" Combines a dataset and a sampler, and provides an iterable over the given dataset.\n",
"\n",
" The class `~torch.utils.data.DataLoader` supports both map-style and\n",
" iterable-style datasets with single- or multi-process loading, customizing\n",
" loading order and optional automatic batching (collation) and memory pinning. \n",
" \n",
" **Parameters:**<br>\n",
" `batch_size`: (int, optional): how many samples per batch to load (default: 1).<br>\n",
" `shuffle`: (bool, optional): set to `True` to have the data reshuffled at every epoch (default: `False`).<br>\n",
" `sampler`: (Sampler or Iterable, optional): defines the strategy to draw samples from the dataset.<br>\n",
" Can be any `Iterable` with `__len__` implemented. If specified, `shuffle` must not be specified.<br>\n",
" \"\"\"\n",
" def __init__(self, dataset, **kwargs):\n",
" if 'collate_fn' in kwargs:\n",
" kwargs.pop('collate_fn')\n",
" kwargs_ = {**kwargs, **dict(collate_fn=self._collate_fn)}\n",
" DataLoader.__init__(self, dataset=dataset, **kwargs_)\n",
" \n",
" def _collate_fn(self, batch):\n",
" elem = batch[0]\n",
" elem_type = type(elem)\n",
"\n",
" if isinstance(elem, torch.Tensor):\n",
" out = None\n",
" if torch.utils.data.get_worker_info() is not None:\n",
" # If we're in a background process, concatenate directly into a\n",
" # shared memory tensor to avoid an extra copy\n",
" numel = sum(x.numel() for x in batch)\n",
" storage = elem.storage()._new_shared(numel, device=elem.device)\n",
" out = elem.new(storage).resize_(len(batch), *list(elem.size()))\n",
" return torch.stack(batch, 0, out=out)\n",
"\n",
" elif isinstance(elem, Mapping):\n",
" if elem['static'] is None:\n",
" return dict(temporal=self.collate_fn([d['temporal'] for d in batch]),\n",
" temporal_cols = elem['temporal_cols'],\n",
" y_idx=elem['y_idx'])\n",
" \n",
" return dict(static=self.collate_fn([d['static'] for d in batch]),\n",
" static_cols = elem['static_cols'],\n",
" temporal=self.collate_fn([d['temporal'] for d in batch]),\n",
" temporal_cols = elem['temporal_cols'],\n",
" y_idx=elem['y_idx'])\n",
"\n",
" raise TypeError(f'Unknown {elem_type}')"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "93e94050-0290-43ad-9a73-c4626bba9541",
"metadata": {},
"outputs": [],
"source": [
"show_doc(TimeSeriesLoader)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "05687429-c139-44c0-adb9-097c616908cc",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class TimeSeriesDataset(Dataset):\n",
"\n",
" def __init__(self,\n",
" temporal,\n",
" temporal_cols,\n",
" indptr,\n",
" max_size: int,\n",
" min_size: int,\n",
" y_idx: int,\n",
" static=None,\n",
" static_cols=None,\n",
" sorted=False,\n",
" ):\n",
" super().__init__()\n",
" self.temporal = self._as_torch_copy(temporal)\n",
" self.temporal_cols = pd.Index(list(temporal_cols))\n",
"\n",
" if static is not None:\n",
" self.static = self._as_torch_copy(static)\n",
" self.static_cols = static_cols\n",
" else:\n",
" self.static = static\n",
" self.static_cols = static_cols\n",
"\n",
" self.indptr = indptr\n",
" self.n_groups = self.indptr.size - 1\n",
" self.max_size = max_size\n",
" self.min_size = min_size\n",
" self.y_idx = y_idx\n",
"\n",
" # Upadated flag. To protect consistency, dataset can only be updated once\n",
" self.updated = False\n",
" self.sorted = sorted\n",
"\n",
" def __getitem__(self, idx):\n",
" if isinstance(idx, int):\n",
" # Parse temporal data and pad its left\n",
" temporal = torch.zeros(size=(len(self.temporal_cols), self.max_size),\n",
" dtype=torch.float32)\n",
" ts = self.temporal[self.indptr[idx] : self.indptr[idx + 1], :]\n",
" temporal[:len(self.temporal_cols), -len(ts):] = ts.permute(1, 0)\n",
"\n",
" # Add static data if available\n",
" static = None if self.static is None else self.static[idx,:]\n",
"\n",
" item = dict(temporal=temporal, temporal_cols=self.temporal_cols,\n",
" static=static, static_cols=self.static_cols,\n",
" y_idx=self.y_idx)\n",
"\n",
" return item\n",
" raise ValueError(f'idx must be int, got {type(idx)}')\n",
"\n",
" def __len__(self):\n",
" return self.n_groups\n",
"\n",
" def __repr__(self):\n",
" return f'TimeSeriesDataset(n_data={self.temporal.shape[0]:,}, n_groups={self.n_groups:,})'\n",
"\n",
" def __eq__(self, other):\n",
" if not hasattr(other, 'data') or not hasattr(other, 'indptr'):\n",
" return False\n",
" return np.allclose(self.data, other.data) and np.array_equal(self.indptr, other.indptr)\n",
"\n",
" def _as_torch_copy(\n",
" self,\n",
" x: Union[np.ndarray, torch.Tensor],\n",
" dtype: torch.dtype = torch.float32,\n",
" ) -> torch.Tensor:\n",
" if isinstance(x, np.ndarray):\n",
" x = torch.from_numpy(x)\n",
" return x.to(dtype, copy=False).clone()\n",
"\n",
" def align(self, df: DataFrame, id_col: str, time_col: str, target_col: str) -> 'TimeSeriesDataset':\n",
" # Protect consistency\n",
" df = ufp.copy_if_pandas(df, deep=False)\n",
"\n",
" # Add Nones to missing columns (without available_mask)\n",
" temporal_cols = self.temporal_cols.copy()\n",
" for col in temporal_cols:\n",
" if col not in df.columns:\n",
" df = ufp.assign_columns(df, col, np.nan)\n",
" if col == 'available_mask':\n",
" df = ufp.assign_columns(df, col, 1.0)\n",
" \n",
" # Sort columns to match self.temporal_cols (without available_mask)\n",
" df = df[ [id_col, time_col] + temporal_cols.tolist() ]\n",
"\n",
" # Process future_df\n",
" dataset, *_ = TimeSeriesDataset.from_df(\n",
" df=df,\n",
" sort_df=self.sorted,\n",
" id_col=id_col,\n",
" time_col=time_col,\n",
" target_col=target_col,\n",
" )\n",
" return dataset\n",
"\n",
" def append(self, futr_dataset: 'TimeSeriesDataset') -> 'TimeSeriesDataset':\n",
" \"\"\"Add future observations to the dataset. Returns a copy\"\"\"\n",
" if self.indptr.size != futr_dataset.indptr.size:\n",
" raise ValueError('Cannot append `futr_dataset` with different number of groups.')\n",
" # Define and fill new temporal with updated information\n",
" len_temporal, col_temporal = self.temporal.shape\n",
" len_futr = futr_dataset.temporal.shape[0]\n",
" new_temporal = torch.empty(size=(len_temporal + len_futr, col_temporal))\n",
" new_sizes = np.diff(self.indptr) + np.diff(futr_dataset.indptr)\n",
" new_indptr = np.append(0, new_sizes.cumsum()).astype(np.int32)\n",
" new_max_size = np.max(new_sizes)\n",
"\n",
" for i in range(self.n_groups):\n",
" curr_slice = slice(self.indptr[i], self.indptr[i + 1])\n",
" curr_size = curr_slice.stop - curr_slice.start\n",
" futr_slice = slice(futr_dataset.indptr[i], futr_dataset.indptr[i + 1])\n",
" new_temporal[new_indptr[i] : new_indptr[i] + curr_size] = self.temporal[curr_slice]\n",
" new_temporal[new_indptr[i] + curr_size : new_indptr[i + 1]] = futr_dataset.temporal[futr_slice]\n",
" \n",
" # Define new dataset\n",
" updated_dataset = TimeSeriesDataset(temporal=new_temporal,\n",
" temporal_cols=self.temporal_cols.copy(),\n",
" indptr=new_indptr,\n",
" max_size=new_max_size,\n",
" min_size=self.min_size,\n",
" static=self.static,\n",
" y_idx=self.y_idx,\n",
" static_cols=self.static_cols,\n",
" sorted=self.sorted)\n",
"\n",
" return updated_dataset\n",
"\n",
" @staticmethod\n",
" def update_dataset(dataset, futr_df, id_col='unique_id', time_col='ds', target_col='y'):\n",
" futr_dataset = dataset.align(\n",
" futr_df, id_col=id_col, time_col=time_col, target_col=target_col\n",
" )\n",
" return dataset.append(futr_dataset)\n",
" \n",
" @staticmethod\n",
" def trim_dataset(dataset, left_trim: int = 0, right_trim: int = 0):\n",
" \"\"\"\n",
" Trim temporal information from a dataset.\n",
" Returns temporal indexes [t+left:t-right] for all series.\n",
" \"\"\"\n",
" if dataset.min_size <= left_trim + right_trim:\n",
" raise Exception(f'left_trim + right_trim ({left_trim} + {right_trim}) \\\n",
" must be lower than the shorter time series ({dataset.min_size})')\n",
"\n",
" # Define and fill new temporal with trimmed information \n",
" len_temporal, col_temporal = dataset.temporal.shape\n",
" total_trim = (left_trim + right_trim) * dataset.n_groups\n",
" new_temporal = torch.zeros(size=(len_temporal-total_trim, col_temporal))\n",
" new_indptr = [0]\n",
"\n",
" acum = 0\n",
" for i in range(dataset.n_groups):\n",
" series_length = dataset.indptr[i + 1] - dataset.indptr[i]\n",
" new_length = series_length - left_trim - right_trim\n",
" new_temporal[acum:(acum+new_length), :] = dataset.temporal[dataset.indptr[i]+left_trim : \\\n",
" dataset.indptr[i + 1]-right_trim, :]\n",
" acum += new_length\n",
" new_indptr.append(acum)\n",
"\n",
" new_max_size = dataset.max_size-left_trim-right_trim\n",
" new_min_size = dataset.min_size-left_trim-right_trim\n",
" \n",
" # Define new dataset\n",
" updated_dataset = TimeSeriesDataset(temporal=new_temporal,\n",
" temporal_cols= dataset.temporal_cols.copy(),\n",
" indptr=np.array(new_indptr, dtype=np.int32),\n",
" max_size=new_max_size,\n",
" min_size=new_min_size,\n",
" y_idx=dataset.y_idx,\n",
" static=dataset.static,\n",
" static_cols=dataset.static_cols,\n",
" sorted=dataset.sorted)\n",
"\n",
" return updated_dataset\n",
"\n",
" @staticmethod\n",
" def from_df(df, static_df=None, sort_df=False, id_col='unique_id', time_col='ds', target_col='y'):\n",
" # TODO: protect on equality of static_df + df indexes\n",
" if isinstance(df, pd.DataFrame) and df.index.name == id_col:\n",
" warnings.warn(\n",
" \"Passing the id as index is deprecated, please provide it as a column instead.\",\n",
" FutureWarning,\n",
" )\n",
" df = df.reset_index(id_col)\n",
" # Define indexes if not given\n",
" if static_df is not None:\n",
" if isinstance(static_df, pd.DataFrame) and static_df.index.name == id_col:\n",
" warnings.warn(\n",
" \"Passing the id as index is deprecated, please provide it as a column instead.\",\n",
" FutureWarning,\n",
" )\n",
" if sort_df:\n",
" static_df = ufp.sort(static_df, by=id_col)\n",
"\n",
" ids, times, data, indptr, sort_idxs = ufp.process_df(df, id_col, time_col, target_col)\n",
" # processor sets y as the first column\n",
" temporal_cols = pd.Index(\n",
" [target_col] + [c for c in df.columns if c not in (id_col, time_col, target_col)]\n",
" )\n",
" temporal = data.astype(np.float32, copy=False)\n",
" indices = ids\n",
" if isinstance(df, pd.DataFrame):\n",
" dates = pd.Index(times, name=time_col)\n",
" else:\n",
" dates = pl_Series(time_col, times)\n",
" sizes = np.diff(indptr)\n",
" max_size = max(sizes)\n",
" min_size = min(sizes)\n",
"\n",
" # Add Available mask efficiently (without adding column to df)\n",
" if 'available_mask' not in df.columns:\n",
" available_mask = np.ones((len(temporal),1), dtype=np.float32)\n",
" temporal = np.append(temporal, available_mask, axis=1)\n",
" temporal_cols = temporal_cols.append(pd.Index(['available_mask']))\n",
"\n",
" # Static features\n",
" if static_df is not None:\n",
" static_cols = [col for col in static_df.columns if col != id_col]\n",
" static = ufp.to_numpy(static_df[static_cols])\n",
" static_cols = pd.Index(static_cols)\n",
" else:\n",
" static = None\n",
" static_cols = None\n",
"\n",
" dataset = TimeSeriesDataset(\n",
" temporal=temporal,\n",
" temporal_cols=temporal_cols,\n",
" static=static,\n",
" static_cols=static_cols,\n",
" indptr=indptr,\n",
" max_size=max_size,\n",
" min_size=min_size,\n",
" sorted=sort_df,\n",
" y_idx=0,\n",
" )\n",
" ds = df[time_col].to_numpy()\n",
" if sort_idxs is not None:\n",
" ds = ds[sort_idxs]\n",
" return dataset, indices, dates, ds"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "61a818bf-28d2-4561-8036-475f6fe78d0a",
"metadata": {},
"outputs": [],
"source": [
"show_doc(TimeSeriesDataset)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "52c07552-b6fa-4d10-8792-71743dcdfd1d",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# Testing sort_df=True functionality\n",
"temporal_df = generate_series(n_series=1000, \n",
" n_temporal_features=0, equal_ends=False)\n",
"sorted_temporal_df = temporal_df.sort_values(['unique_id', 'ds'])\n",
"unsorted_temporal_df = sorted_temporal_df.sample(frac=1.0)\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=unsorted_temporal_df,\n",
" sort_df=True)\n",
"\n",
"np.testing.assert_allclose(dataset.temporal[:,:-1], \n",
" sorted_temporal_df.drop(columns=['unique_id', 'ds']).values)\n",
"test_eq(indices, pd.Series(sorted_temporal_df['unique_id'].unique()))\n",
"test_eq(dates, temporal_df.groupby('unique_id')['ds'].max().values)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "24e51cf3",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class _FilesDataset:\n",
" def __init__(\n",
" self,\n",
" files: List[str],\n",
" temporal_cols: List[str],\n",
" static_cols: Optional[List[str]],\n",
" id_col: str,\n",
" time_col: str,\n",
" target_col: str,\n",
" min_size: int,\n",
" ):\n",
" self.files = files\n",
" self.temporal_cols = pd.Index(temporal_cols)\n",
" self.static_cols = pd.Index(static_cols) if static_cols is not None else None\n",
" self.id_col = id_col\n",
" self.time_col = time_col\n",
" self.target_col = target_col\n",
" self.min_size = min_size"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c4dae43c-4d11-4bbc-a431-ac33b004859a",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class TimeSeriesDataModule(pl.LightningDataModule):\n",
" \n",
" def __init__(\n",
" self, \n",
" dataset: TimeSeriesDataset,\n",
" batch_size=32, \n",
" valid_batch_size=1024,\n",
" num_workers=0,\n",
" drop_last=False,\n",
" shuffle_train=True,\n",
" ):\n",
" super().__init__()\n",
" self.dataset = dataset\n",
" self.batch_size = batch_size\n",
" self.valid_batch_size = valid_batch_size\n",
" self.num_workers = num_workers\n",
" self.drop_last = drop_last\n",
" self.shuffle_train = shuffle_train\n",
" \n",
" def train_dataloader(self):\n",
" loader = TimeSeriesLoader(\n",
" self.dataset,\n",
" batch_size=self.batch_size, \n",
" num_workers=self.num_workers,\n",
" shuffle=self.shuffle_train,\n",
" drop_last=self.drop_last\n",
" )\n",
" return loader\n",
" \n",
" def val_dataloader(self):\n",
" loader = TimeSeriesLoader(\n",
" self.dataset, \n",
" batch_size=self.valid_batch_size, \n",
" num_workers=self.num_workers,\n",
" shuffle=False,\n",
" drop_last=self.drop_last\n",
" )\n",
" return loader\n",
" \n",
" def predict_dataloader(self):\n",
" loader = TimeSeriesLoader(\n",
" self.dataset,\n",
" batch_size=self.valid_batch_size, \n",
" num_workers=self.num_workers,\n",
" shuffle=False\n",
" )\n",
" return loader"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "8535a15f-b5cf-4ca1-bfa2-e53a9e8c3bc0",
"metadata": {},
"outputs": [],
"source": [
"show_doc(TimeSeriesDataModule)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "b534d29d-eecc-43ba-8468-c23305fa24a2",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"batch_size = 128\n",
"data = TimeSeriesDataModule(dataset=dataset, \n",
" batch_size=batch_size, drop_last=True)\n",
"for batch in data.train_dataloader():\n",
" test_eq(batch['temporal'].shape, (batch_size, 2, 500))\n",
" test_eq(batch['temporal_cols'], ['y', 'available_mask'])"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "4481272a-ea3a-4b63-8f14-9445d8f41338",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"batch_size = 128\n",
"n_static_features = 2\n",
"n_temporal_features = 4\n",
"temporal_df, static_df = generate_series(n_series=1000,\n",
" n_static_features=n_static_features,\n",
" n_temporal_features=n_temporal_features, \n",
" equal_ends=False)\n",
"\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=temporal_df,\n",
" static_df=static_df,\n",
" sort_df=True)\n",
"data = TimeSeriesDataModule(dataset=dataset,\n",
" batch_size=batch_size, drop_last=True)\n",
"\n",
"for batch in data.train_dataloader():\n",
" test_eq(batch['temporal'].shape, (batch_size, n_temporal_features + 2, 500))\n",
" test_eq(batch['temporal_cols'],\n",
" ['y'] + [f'temporal_{i}' for i in range(n_temporal_features)] + ['available_mask'])\n",
" \n",
" test_eq(batch['static'].shape, (batch_size, n_static_features))\n",
" test_eq(batch['static_cols'], [f'static_{i}' for i in range(n_static_features)])"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "252b59f6",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# Testing sort_df=True functionality\n",
"temporal_df = generate_series(n_series=2,\n",
" n_temporal_features=2, equal_ends=True)\n",
"temporal_df = temporal_df.groupby('unique_id').tail(10)\n",
"temporal_df = temporal_df.reset_index()\n",
"temporal_full_df = temporal_df.sort_values(['unique_id', 'ds']).reset_index(drop=True)\n",
"temporal_full_df.loc[temporal_full_df.ds > '2001-05-11', ['y', 'temporal_0']] = None\n",
"\n",
"split1_df = temporal_full_df.loc[temporal_full_df.ds <= '2001-05-11']\n",
"split2_df = temporal_full_df.loc[temporal_full_df.ds > '2001-05-11']"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "6eab7367",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# Testing available mask\n",
"temporal_df_w_mask = temporal_df.copy()\n",
"temporal_df_w_mask['available_mask'] = 1\n",
"\n",
"# Mask with all 1's\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=temporal_df_w_mask,\n",
" sort_df=True)\n",
"mask_average = dataset.temporal[:, -1].mean()\n",
"np.testing.assert_almost_equal(mask_average, 1.0000)\n",
"\n",
"# Add 0's to available mask\n",
"temporal_df_w_mask.loc[temporal_df_w_mask.ds > '2001-05-11', 'available_mask'] = 0\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=temporal_df_w_mask,\n",
" sort_df=True)\n",
"mask_average = dataset.temporal[:, -1].mean()\n",
"np.testing.assert_almost_equal(mask_average, 0.7000)\n",
"\n",
"# Available mask not in last column\n",
"temporal_df_w_mask = temporal_df_w_mask[['unique_id','ds','y','available_mask', 'temporal_0','temporal_1']]\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=temporal_df_w_mask,\n",
" sort_df=True)\n",
"mask_average = dataset.temporal[:, 1].mean()\n",
"np.testing.assert_almost_equal(mask_average, 0.7000)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a0d23f1a",
"metadata": {},
"outputs": [],
"source": [
"# To test correct future_df wrangling of the `update_df` method\n",
"# We are checking that we are able to recover the AirPassengers dataset\n",
"# using the dataframe or splitting it into parts and initializing."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "39f999c2",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# FULL DATASET\n",
"dataset_full, indices_full, dates_full, ds_full = TimeSeriesDataset.from_df(df=temporal_full_df,\n",
" sort_df=False)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "30f927e2",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# SPLIT_1 DATASET\n",
"dataset_1, indices_1, dates_1, ds_1 = TimeSeriesDataset.from_df(df=split1_df,\n",
" sort_df=False)\n",
"dataset_1 = dataset_1.update_dataset(dataset_1, split2_df)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "468a6879",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"np.testing.assert_almost_equal(dataset_full.temporal.numpy(), dataset_1.temporal.numpy())\n",
"test_eq(dataset_full.max_size, dataset_1.max_size)\n",
"test_eq(dataset_full.indptr, dataset_1.indptr)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "556f852c",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"\n",
"# Testing trim_dataset functionality\n",
"n_static_features = 0\n",
"n_temporal_features = 2\n",
"temporal_df = generate_series(n_series=100,\n",
" min_length=50,\n",
" max_length=100,\n",
" n_static_features=n_static_features,\n",
" n_temporal_features=n_temporal_features, \n",
" equal_ends=False)\n",
"dataset, indices, dates, ds = TimeSeriesDataset.from_df(df=temporal_df,\n",
" static_df=static_df,\n",
" sort_df=True)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "db7b1a51",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"left_trim = 10\n",
"right_trim = 20\n",
"dataset_trimmed = dataset.trim_dataset(dataset, left_trim=left_trim, right_trim=right_trim)\n",
"\n",
"np.testing.assert_almost_equal(dataset.temporal[dataset.indptr[50]+left_trim:dataset.indptr[51]-right_trim].numpy(),\n",
" dataset_trimmed.temporal[dataset_trimmed.indptr[50]:dataset_trimmed.indptr[51]].numpy())"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "624a3fbb-cb78-4440-a645-54699fd82660",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"#| polars\n",
"import polars"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a1bdd479-b4c7-4a40-93eb-2b7c9b969a80",
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"#| polars\n",
"temporal_df2 = temporal_df.copy()\n",
"for col in ('unique_id', 'temporal_0', 'temporal_1'):\n",
" temporal_df2[col] = temporal_df2[col].cat.codes\n",
"temporal_pl = polars.from_pandas(temporal_df2).sample(fraction=1.0)\n",
"static_pl = polars.from_pandas(static_df.assign(unique_id=lambda df: df['unique_id'].astype('int64')))\n",
"dataset_pl, indices_pl, dates_pl, ds_pl = TimeSeriesDataset.from_df(df=temporal_pl, static_df=static_df, sort_df=True)\n",
"for attr in ('static_cols', 'temporal_cols', 'min_size', 'max_size', 'n_groups'):\n",
" test_eq(getattr(dataset, attr), getattr(dataset_pl, attr))\n",
"torch.testing.assert_allclose(dataset.temporal, dataset_pl.temporal)\n",
"torch.testing.assert_allclose(dataset.static, dataset_pl.static)\n",
"pd.testing.assert_series_equal(indices.astype('int64'), indices_pl.to_pandas().astype('int64'))\n",
"pd.testing.assert_index_equal(dates, pd.Index(dates_pl, name='ds'))\n",
"np.testing.assert_array_equal(ds, ds_pl)\n",
"np.testing.assert_array_equal(dataset.indptr, dataset_pl.indptr)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "959ea63c",
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class _DistributedTimeSeriesDataModule(TimeSeriesDataModule):\n",
" def __init__(\n",
" self,\n",
" dataset: _FilesDataset,\n",
" batch_size=32,\n",
" valid_batch_size=1024,\n",
" num_workers=0,\n",
" drop_last=False,\n",
" shuffle_train=True,\n",
" ):\n",
" super(TimeSeriesDataModule, self).__init__()\n",
" self.files_ds = dataset\n",
" self.batch_size = batch_size\n",
" self.valid_batch_size = valid_batch_size\n",
" self.num_workers = num_workers\n",
" self.drop_last = drop_last\n",
" self.shuffle_train = shuffle_train\n",
"\n",
" def setup(self, stage):\n",
" import torch.distributed as dist\n",
"\n",
" df = pd.read_parquet(self.files_ds.files[dist.get_rank()])\n",
" if self.files_ds.static_cols is not None:\n",
" static_df = (\n",
" df[[self.files_ds.id_col] + self.files_ds.static_cols.tolist()]\n",
" .groupby(self.files_ds.id_col, observed=True)\n",
" .head(1)\n",
" )\n",
" df = df.drop(columns=self.files_ds.static_cols)\n",
" else:\n",
" static_df = None\n",
" self.dataset, *_ = TimeSeriesDataset.from_df(\n",
" df=df,\n",
" static_df=static_df,\n",
" sort_df=True,\n",
" id_col=self.files_ds.id_col,\n",
" time_col=self.files_ds.time_col,\n",
" target_col=self.files_ds.target_col,\n",
" )"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "python3",
"language": "python",
"name": "python3"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
nbs/utils.ipynb 0 → 100644
View file @ e3f7f7b3
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| default_exp utils"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"%load_ext autoreload\n",
"%autoreload 2"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Example Data\n",
"\n",
"> The `core.NeuralForecast` class allows you to efficiently fit multiple `NeuralForecast` models for large sets of time series. It operates with pandas DataFrame `df` that identifies individual series and datestamps with the `unique_id` and `ds` columns, and the `y` column denotes the target time series variable. To assist development, we declare useful datasets that we use throughout all `NeuralForecast`'s unit tests.<br><br>"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"import random\n",
"from itertools import chain\n",
"from typing import List\n",
"\n",
"import numpy as np\n",
"import pandas as pd"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"import matplotlib.pyplot as plt\n",
"\n",
"from nbdev.showdoc import add_docs, show_doc"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 1. Synthetic Panel Data"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"def generate_series(n_series: int,\n",
" freq: str = 'D',\n",
" min_length: int = 50,\n",
" max_length: int = 500,\n",
" n_temporal_features: int = 0,\n",
" n_static_features: int = 0,\n",
" equal_ends: bool = False,\n",
" seed: int = 0) -> pd.DataFrame:\n",
" \"\"\"Generate Synthetic Panel Series.\n",
"\n",
" Generates `n_series` of frequency `freq` of different lengths in the interval [`min_length`, `max_length`].\n",
" If `n_temporal_features > 0`, then each serie gets temporal features with random values.\n",
" If `n_static_features > 0`, then a static dataframe is returned along the temporal dataframe.\n",
" If `equal_ends == True` then all series end at the same date.\n",
"\n",
" **Parameters:**<br>\n",
" `n_series`: int, number of series for synthetic panel.<br>\n",
" `min_length`: int, minimal length of synthetic panel's series.<br>\n",
" `max_length`: int, minimal length of synthetic panel's series.<br>\n",
" `n_temporal_features`: int, default=0, number of temporal exogenous variables for synthetic panel's series.<br>\n",
" `n_static_features`: int, default=0, number of static exogenous variables for synthetic panel's series.<br>\n",
" `equal_ends`: bool, if True, series finish in the same date stamp `ds`.<br>\n",
" `freq`: str, frequency of the data, [panda's available frequencies](https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases).<br>\n",
"\n",
" **Returns:**<br>\n",
" `freq`: pandas.DataFrame, synthetic panel with columns [`unique_id`, `ds`, `y`] and exogenous.\n",
" \"\"\"\n",
" seasonalities = {'D': 7, 'M': 12}\n",
" season = seasonalities[freq]\n",
"\n",
" rng = np.random.RandomState(seed)\n",
" series_lengths = rng.randint(min_length, max_length + 1, n_series)\n",
" total_length = series_lengths.sum()\n",
"\n",
" dates = pd.date_range('2000-01-01', periods=max_length, freq=freq).values\n",
" uids = [\n",
" np.repeat(i, serie_length) for i, serie_length in enumerate(series_lengths)\n",
" ]\n",
" if equal_ends:\n",
" ds = [dates[-serie_length:] for serie_length in series_lengths]\n",
" else:\n",
" ds = [dates[:serie_length] for serie_length in series_lengths]\n",
"\n",
" y = np.arange(total_length) % season + rng.rand(total_length) * 0.5\n",
" temporal_df = pd.DataFrame(dict(unique_id=chain.from_iterable(uids),\n",
" ds=chain.from_iterable(ds),\n",
" y=y))\n",
"\n",
" random.seed(seed)\n",
" for i in range(n_temporal_features):\n",
" random.seed(seed)\n",
" temporal_values = [\n",
" [random.randint(0, 100)] * serie_length for serie_length in series_lengths\n",
" ]\n",
" temporal_df[f'temporal_{i}'] = np.hstack(temporal_values)\n",
" temporal_df[f'temporal_{i}'] = temporal_df[f'temporal_{i}'].astype('category')\n",
" if i == 0:\n",
" temporal_df['y'] = temporal_df['y'] * \\\n",
" (1 + temporal_df[f'temporal_{i}'].cat.codes)\n",
"\n",
" temporal_df['unique_id'] = temporal_df['unique_id'].astype('category')\n",
" temporal_df['unique_id'] = temporal_df['unique_id'].cat.as_ordered()\n",
"\n",
" if n_static_features > 0:\n",
" static_features = np.random.uniform(low=0.0, high=1.0, \n",
" size=(n_series, n_static_features))\n",
" static_df = pd.DataFrame.from_records(static_features, \n",
" columns = [f'static_{i}'for i in range(n_static_features)])\n",
" \n",
" static_df['unique_id'] = np.arange(n_series)\n",
" static_df['unique_id'] = static_df['unique_id'].astype('category')\n",
" static_df['unique_id'] = static_df['unique_id'].cat.as_ordered()\n",
"\n",
" return temporal_df, static_df\n",
"\n",
" return temporal_df"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"show_doc(generate_series, title_level=3)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"synthetic_panel = generate_series(n_series=2)\n",
"synthetic_panel.groupby('unique_id').head(4)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"temporal_df, static_df = generate_series(n_series=1000, n_static_features=2,\n",
" n_temporal_features=4, equal_ends=False)\n",
"static_df.head(2)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 2. AirPassengers Data\n",
"\n",
"The classic Box & Jenkins airline data. Monthly totals of international airline passengers, 1949 to 1960.\n",
"\n",
"It has been used as a reference on several forecasting libraries, since it is a series that shows clear trends and seasonalities it offers a nice opportunity to quickly showcase a model's predictions performance."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"AirPassengers = np.array([112., 118., 132., 129., 121., 135., 148., 148., 136., 119., 104.,\n",
" 118., 115., 126., 141., 135., 125., 149., 170., 170., 158., 133.,\n",
" 114., 140., 145., 150., 178., 163., 172., 178., 199., 199., 184.,\n",
" 162., 146., 166., 171., 180., 193., 181., 183., 218., 230., 242.,\n",
" 209., 191., 172., 194., 196., 196., 236., 235., 229., 243., 264.,\n",
" 272., 237., 211., 180., 201., 204., 188., 235., 227., 234., 264.,\n",
" 302., 293., 259., 229., 203., 229., 242., 233., 267., 269., 270.,\n",
" 315., 364., 347., 312., 274., 237., 278., 284., 277., 317., 313.,\n",
" 318., 374., 413., 405., 355., 306., 271., 306., 315., 301., 356.,\n",
" 348., 355., 422., 465., 467., 404., 347., 305., 336., 340., 318.,\n",
" 362., 348., 363., 435., 491., 505., 404., 359., 310., 337., 360.,\n",
" 342., 406., 396., 420., 472., 548., 559., 463., 407., 362., 405.,\n",
" 417., 391., 419., 461., 472., 535., 622., 606., 508., 461., 390.,\n",
" 432.], dtype=np.float32)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"AirPassengersDF = pd.DataFrame({'unique_id': np.ones(len(AirPassengers)),\n",
" 'ds': pd.date_range(start='1949-01-01',\n",
" periods=len(AirPassengers), freq=pd.offsets.MonthEnd()),\n",
" 'y': AirPassengers})"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"AirPassengersDF.head(12)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#We are going to plot the ARIMA predictions, and the prediction intervals.\n",
"fig, ax = plt.subplots(1, 1, figsize = (20, 7))\n",
"plot_df = AirPassengersDF.set_index('ds')\n",
"\n",
"plot_df[['y']].plot(ax=ax, linewidth=2)\n",
"ax.set_title('AirPassengers Forecast', fontsize=22)\n",
"ax.set_ylabel('Monthly Passengers', fontsize=20)\n",
"ax.set_xlabel('Timestamp [t]', fontsize=20)\n",
"ax.legend(prop={'size': 15})\n",
"ax.grid()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import numpy as np\n",
"import pandas as pd"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"n_static_features = 3\n",
"n_series = 5\n",
"\n",
"static_features = np.random.uniform(low=0.0, high=1.0, \n",
" size=(n_series, n_static_features))\n",
"static_df = pd.DataFrame.from_records(static_features, \n",
" columns = [f'static_{i}'for i in range(n_static_features)])\n",
"static_df['unique_id'] = np.arange(n_series)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"static_df"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 3. Panel AirPassengers Data\n",
"\n",
"Extension to classic Box & Jenkins airline data. Monthly totals of international airline passengers, 1949 to 1960.\n",
"\n",
"It includes two series with static, temporal and future exogenous variables, that can help to explore the performance of models like `NBEATSx` and `TFT`."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"\n",
"# Declare Panel Data\n",
"unique_id = np.concatenate([['Airline1']*len(AirPassengers), ['Airline2']*len(AirPassengers)])\n",
"ds = np.tile(\n",
" pd.date_range(\n",
" start='1949-01-01', periods=len(AirPassengers), freq=pd.offsets.MonthEnd()\n",
" ).to_numpy(), \n",
" 2,\n",
")\n",
"y = np.concatenate([AirPassengers, AirPassengers+300])\n",
"\n",
"AirPassengersPanel = pd.DataFrame({'unique_id': unique_id, 'ds': ds, 'y': y})\n",
"\n",
"# For future exogenous variables\n",
"# Declare SeasonalNaive12 and fill first 12 values with y\n",
"snaive = AirPassengersPanel.groupby('unique_id')['y'].shift(periods=12).reset_index(drop=True)\n",
"AirPassengersPanel['trend'] = range(len(AirPassengersPanel))\n",
"AirPassengersPanel['y_[lag12]'] = snaive.fillna(AirPassengersPanel['y'])\n",
"\n",
"# Declare Static Data\n",
"unique_id = np.array(['Airline1', 'Airline2'])\n",
"airline1_dummy = [0, 1]\n",
"airline2_dummy = [1, 0]\n",
"AirPassengersStatic = pd.DataFrame({'unique_id': unique_id,\n",
" 'airline1': airline1_dummy,\n",
" 'airline2': airline2_dummy})\n",
"\n",
"AirPassengersPanel.groupby('unique_id').tail(4)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"fig, ax = plt.subplots(1, 1, figsize = (20, 7))\n",
"plot_df = AirPassengersPanel.set_index('ds')\n",
"\n",
"plot_df.groupby('unique_id')['y'].plot(legend=True)\n",
"ax.set_title('AirPassengers Panel Data', fontsize=22)\n",
"ax.set_ylabel('Monthly Passengers', fontsize=20)\n",
"ax.set_xlabel('Timestamp [t]', fontsize=20)\n",
"ax.legend(title='unique_id', prop={'size': 15})\n",
"ax.grid()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"fig, ax = plt.subplots(1, 1, figsize = (20, 7))\n",
"plot_df = AirPassengersPanel[AirPassengersPanel.unique_id=='Airline1'].set_index('ds')\n",
"\n",
"plot_df[['y', 'trend', 'y_[lag12]']].plot(ax=ax, linewidth=2)\n",
"ax.set_title('Box-Cox AirPassengers Data', fontsize=22)\n",
"ax.set_ylabel('Monthly Passengers', fontsize=20)\n",
"ax.set_xlabel('Timestamp [t]', fontsize=20)\n",
"ax.legend(prop={'size': 15})\n",
"ax.grid()"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# 4. Time Features"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"We have developed a utility that generates normalized calendar features for use as absolute positional embeddings in Transformer-based models. These embeddings capture seasonal patterns in time series data and can be easily incorporated into the model architecture. Additionally, the features can be used as exogenous variables in other models to inform them of calendar patterns in the data."
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"**References**<br>\n",
"- [Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, Wancai Zhang. \"Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting\"](https://arxiv.org/abs/2012.07436)<br>"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"class TimeFeature:\n",
" def __init__(self):\n",
" pass\n",
"\n",
" def __call__(self, index: pd.DatetimeIndex):\n",
" return print('Overwrite with corresponding feature')\n",
"\n",
" def __repr__(self):\n",
" return self.__class__.__name__ + \"()\"\n",
"\n",
"class SecondOfMinute(TimeFeature):\n",
" \"\"\"Minute of hour encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return index.second / 59.0 - 0.5\n",
"\n",
"class MinuteOfHour(TimeFeature):\n",
" \"\"\"Minute of hour encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return index.minute / 59.0 - 0.5\n",
"\n",
"class HourOfDay(TimeFeature):\n",
" \"\"\"Hour of day encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return index.hour / 23.0 - 0.5\n",
"\n",
"class DayOfWeek(TimeFeature):\n",
" \"\"\"Hour of day encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return index.dayofweek / 6.0 - 0.5\n",
"\n",
"class DayOfMonth(TimeFeature):\n",
" \"\"\"Day of month encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return (index.day - 1) / 30.0 - 0.5\n",
"\n",
"class DayOfYear(TimeFeature):\n",
" \"\"\"Day of year encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return (index.dayofyear - 1) / 365.0 - 0.5\n",
"\n",
"class MonthOfYear(TimeFeature):\n",
" \"\"\"Month of year encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return (index.month - 1) / 11.0 - 0.5\n",
"\n",
"class WeekOfYear(TimeFeature):\n",
" \"\"\"Week of year encoded as value between [-0.5, 0.5]\"\"\"\n",
" def __call__(self, index: pd.DatetimeIndex) -> np.ndarray:\n",
" return (index.week - 1) / 52.0 - 0.5\n",
"\n",
"def time_features_from_frequency_str(freq_str: str) -> List[TimeFeature]:\n",
" \"\"\"\n",
" Returns a list of time features that will be appropriate for the given frequency string.\n",
" Parameters\n",
" ----------\n",
" freq_str\n",
" Frequency string of the form [multiple][granularity] such as \"12H\", \"5min\", \"1D\" etc.\n",
" \"\"\"\n",
"\n",
" if freq_str not in ['Q', 'M', 'MS', 'W', 'D', 'B', 'H', 'T', 'S']:\n",
" raise Exception('Frequency not supported')\n",
" \n",
" if freq_str in ['Q','M', 'MS']:\n",
" return [cls() for cls in [MonthOfYear]]\n",
" elif freq_str == 'W':\n",
" return [cls() for cls in [DayOfMonth, WeekOfYear]]\n",
" elif freq_str in ['D','B']:\n",
" return [cls() for cls in [DayOfWeek, DayOfMonth, DayOfYear]]\n",
" elif freq_str == 'H':\n",
" return [cls() for cls in [HourOfDay, DayOfWeek, DayOfMonth, DayOfYear]]\n",
" elif freq_str == 'T':\n",
" return [cls() for cls in [MinuteOfHour, HourOfDay, DayOfWeek, DayOfMonth, DayOfYear]]\n",
" else:\n",
" return [cls() for cls in [SecondOfMinute, MinuteOfHour, HourOfDay, DayOfWeek, DayOfMonth, DayOfYear]]\n",
"\n",
"def augment_calendar_df(df, freq='H'):\n",
" \"\"\"\n",
" > * Q - [month]\n",
" > * M - [month]\n",
" > * W - [Day of month, week of year]\n",
" > * D - [Day of week, day of month, day of year]\n",
" > * B - [Day of week, day of month, day of year]\n",
" > * H - [Hour of day, day of week, day of month, day of year]\n",
" > * T - [Minute of hour*, hour of day, day of week, day of month, day of year]\n",
" > * S - [Second of minute, minute of hour, hour of day, day of week, day of month, day of year]\n",
" *minute returns a number from 0-3 corresponding to the 15 minute period it falls into.\n",
" \"\"\"\n",
" df = df.copy()\n",
"\n",
" freq_map = {\n",
" 'Q':['month'],\n",
" 'M':['month'],\n",
" 'MS':['month'],\n",
" 'W':['monthday', 'yearweek'],\n",
" 'D':['weekday','monthday','yearday'],\n",
" 'B':['weekday','monthday','yearday'],\n",
" 'H':['dayhour','weekday','monthday','yearday'],\n",
" 'T':['hourminute','dayhour','weekday','monthday','yearday'],\n",
" 'S':['minutesecond','hourminute','dayhour','weekday','monthday','yearday']\n",
" }\n",
"\n",
" ds_col = pd.to_datetime(df.ds.values)\n",
" ds_data = np.vstack([feat(ds_col) for feat in time_features_from_frequency_str(freq)]).transpose(1,0)\n",
" ds_data = pd.DataFrame(ds_data, columns=freq_map[freq])\n",
" \n",
" return pd.concat([df, ds_data], axis=1), freq_map[freq]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"AirPassengerPanelCalendar, calendar_cols = augment_calendar_df(df=AirPassengersPanel, freq='M')\n",
"AirPassengerPanelCalendar.head()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"plot_df = AirPassengerPanelCalendar[AirPassengerPanelCalendar.unique_id=='Airline1'].set_index('ds')\n",
"plt.plot(plot_df['month'])\n",
"plt.grid()\n",
"plt.xlabel('Datestamp')\n",
"plt.ylabel('Normalized Month')\n",
"plt.show()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| export\n",
"def get_indexer_raise_missing(idx: pd.Index, vals: List[str]) -> List[int]:\n",
" idxs = idx.get_indexer(vals)\n",
" missing = [v for i, v in zip(idxs, vals) if i == -1]\n",
" if missing:\n",
" raise ValueError(f'The following values are missing from the index: {missing}')\n",
" return idxs"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "python3",
"language": "python",
"name": "python3"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
neuralforecast/__init__.py 0 → 100644
View file @ e3f7f7b3
__version__ = "1.7.1"
__all__ = ['NeuralForecast']
from .core import NeuralForecast
from .common._base_model import DistributedConfig # noqa: F401
neuralforecast/_modidx.py 0 → 100644
View file @ e3f7f7b3
# Autogenerated by nbdev
d = { 'settings': { 'branch': 'main',
'doc_baseurl': '/neuralforecast/',
'doc_host': 'https://Nixtla.github.io',
'git_url': 'https://github.com/Nixtla/neuralforecast/',
'lib_path': 'neuralforecast'},
'syms': { 'neuralforecast.auto': { 'neuralforecast.auto.AutoAutoformer': ('models.html#autoautoformer', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoAutoformer.__init__': ( 'models.html#autoautoformer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoAutoformer.get_default_config': ( 'models.html#autoautoformer.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoBiTCN': ('models.html#autobitcn', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoBiTCN.__init__': ('models.html#autobitcn.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoBiTCN.get_default_config': ( 'models.html#autobitcn.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDLinear': ('models.html#autodlinear', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDLinear.__init__': ( 'models.html#autodlinear.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDLinear.get_default_config': ( 'models.html#autodlinear.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDeepAR': ('models.html#autodeepar', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDeepAR.__init__': ( 'models.html#autodeepar.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDeepAR.get_default_config': ( 'models.html#autodeepar.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDilatedRNN': ('models.html#autodilatedrnn', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDilatedRNN.__init__': ( 'models.html#autodilatedrnn.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoDilatedRNN.get_default_config': ( 'models.html#autodilatedrnn.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoFEDformer': ('models.html#autofedformer', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoFEDformer.__init__': ( 'models.html#autofedformer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoFEDformer.get_default_config': ( 'models.html#autofedformer.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoGRU': ('models.html#autogru', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoGRU.__init__': ('models.html#autogru.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoGRU.get_default_config': ( 'models.html#autogru.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoHINT': ('models.html#autohint', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoHINT.__init__': ('models.html#autohint.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoHINT._fit_model': ( 'models.html#autohint._fit_model',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoHINT.get_default_config': ( 'models.html#autohint.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoInformer': ('models.html#autoinformer', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoInformer.__init__': ( 'models.html#autoinformer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoInformer.get_default_config': ( 'models.html#autoinformer.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoLSTM': ('models.html#autolstm', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoLSTM.__init__': ('models.html#autolstm.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoLSTM.get_default_config': ( 'models.html#autolstm.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLP': ('models.html#automlp', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLP.__init__': ('models.html#automlp.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLP.get_default_config': ( 'models.html#automlp.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLPMultivariate': ( 'models.html#automlpmultivariate',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLPMultivariate.__init__': ( 'models.html#automlpmultivariate.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoMLPMultivariate.get_default_config': ( 'models.html#automlpmultivariate.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATS': ('models.html#autonbeats', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATS.__init__': ( 'models.html#autonbeats.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATS.get_default_config': ( 'models.html#autonbeats.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATSx': ('models.html#autonbeatsx', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATSx.__init__': ( 'models.html#autonbeatsx.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNBEATSx.get_default_config': ( 'models.html#autonbeatsx.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNHITS': ('models.html#autonhits', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNHITS.__init__': ('models.html#autonhits.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNHITS.get_default_config': ( 'models.html#autonhits.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNLinear': ('models.html#autonlinear', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNLinear.__init__': ( 'models.html#autonlinear.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoNLinear.get_default_config': ( 'models.html#autonlinear.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoPatchTST': ('models.html#autopatchtst', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoPatchTST.__init__': ( 'models.html#autopatchtst.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoPatchTST.get_default_config': ( 'models.html#autopatchtst.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoRNN': ('models.html#autornn', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoRNN.__init__': ('models.html#autornn.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoRNN.get_default_config': ( 'models.html#autornn.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoStemGNN': ('models.html#autostemgnn', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoStemGNN.__init__': ( 'models.html#autostemgnn.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoStemGNN.get_default_config': ( 'models.html#autostemgnn.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTCN': ('models.html#autotcn', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTCN.__init__': ('models.html#autotcn.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTCN.get_default_config': ( 'models.html#autotcn.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTFT': ('models.html#autotft', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTFT.__init__': ('models.html#autotft.__init__', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTFT.get_default_config': ( 'models.html#autotft.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixer': ('models.html#autotsmixer', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixer.__init__': ( 'models.html#autotsmixer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixer.get_default_config': ( 'models.html#autotsmixer.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixerx': ('models.html#autotsmixerx', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixerx.__init__': ( 'models.html#autotsmixerx.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTSMixerx.get_default_config': ( 'models.html#autotsmixerx.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTimesNet': ('models.html#autotimesnet', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTimesNet.__init__': ( 'models.html#autotimesnet.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoTimesNet.get_default_config': ( 'models.html#autotimesnet.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoVanillaTransformer': ( 'models.html#autovanillatransformer',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoVanillaTransformer.__init__': ( 'models.html#autovanillatransformer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoVanillaTransformer.get_default_config': ( 'models.html#autovanillatransformer.get_default_config',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoiTransformer': ('models.html#autoitransformer', 'neuralforecast/auto.py'),
'neuralforecast.auto.AutoiTransformer.__init__': ( 'models.html#autoitransformer.__init__',
'neuralforecast/auto.py'),
'neuralforecast.auto.AutoiTransformer.get_default_config': ( 'models.html#autoitransformer.get_default_config',
'neuralforecast/auto.py')},
'neuralforecast.compat': {},
'neuralforecast.core': { 'neuralforecast.core.NeuralForecast': ('core.html#neuralforecast', 'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.__init__': ( 'core.html#neuralforecast.__init__',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._check_nan': ( 'core.html#neuralforecast._check_nan',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._get_model_names': ( 'core.html#neuralforecast._get_model_names',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._get_needed_futr_exog': ( 'core.html#neuralforecast._get_needed_futr_exog',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._no_refit_cross_validation': ( 'core.html#neuralforecast._no_refit_cross_validation',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._prepare_fit': ( 'core.html#neuralforecast._prepare_fit',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._reset_models': ( 'core.html#neuralforecast._reset_models',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._scalers_fit_transform': ( 'core.html#neuralforecast._scalers_fit_transform',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._scalers_target_inverse_transform': ( 'core.html#neuralforecast._scalers_target_inverse_transform',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast._scalers_transform': ( 'core.html#neuralforecast._scalers_transform',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.cross_validation': ( 'core.html#neuralforecast.cross_validation',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.fit': ('core.html#neuralforecast.fit', 'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.get_missing_future': ( 'core.html#neuralforecast.get_missing_future',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.load': ('core.html#neuralforecast.load', 'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.make_future_dataframe': ( 'core.html#neuralforecast.make_future_dataframe',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.predict': ( 'core.html#neuralforecast.predict',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.predict_insample': ( 'core.html#neuralforecast.predict_insample',
'neuralforecast/core.py'),
'neuralforecast.core.NeuralForecast.save': ('core.html#neuralforecast.save', 'neuralforecast/core.py'),
'neuralforecast.core._id_as_idx': ('core.html#_id_as_idx', 'neuralforecast/core.py'),
'neuralforecast.core._insample_times': ('core.html#_insample_times', 'neuralforecast/core.py'),
'neuralforecast.core._warn_id_as_idx': ('core.html#_warn_id_as_idx', 'neuralforecast/core.py')},
'neuralforecast.losses.numpy': { 'neuralforecast.losses.numpy._divide_no_nan': ( 'losses.numpy.html#_divide_no_nan',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy._metric_protections': ( 'losses.numpy.html#_metric_protections',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.mae': ('losses.numpy.html#mae', 'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.mape': ( 'losses.numpy.html#mape',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.mase': ( 'losses.numpy.html#mase',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.mqloss': ( 'losses.numpy.html#mqloss',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.mse': ('losses.numpy.html#mse', 'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.quantile_loss': ( 'losses.numpy.html#quantile_loss',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.rmae': ( 'losses.numpy.html#rmae',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.rmse': ( 'losses.numpy.html#rmse',
'neuralforecast/losses/numpy.py'),
'neuralforecast.losses.numpy.smape': ( 'losses.numpy.html#smape',
'neuralforecast/losses/numpy.py')},
'neuralforecast.losses.pytorch': { 'neuralforecast.losses.pytorch.Accuracy': ( 'losses.pytorch.html#accuracy',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Accuracy.__call__': ( 'losses.pytorch.html#accuracy.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Accuracy.__init__': ( 'losses.pytorch.html#accuracy.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Accuracy.domain_map': ( 'losses.pytorch.html#accuracy.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.BasePointLoss': ( 'losses.pytorch.html#basepointloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.BasePointLoss.__init__': ( 'losses.pytorch.html#basepointloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.BasePointLoss._compute_weights': ( 'losses.pytorch.html#basepointloss._compute_weights',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.BasePointLoss.domain_map': ( 'losses.pytorch.html#basepointloss.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.DistributionLoss': ( 'losses.pytorch.html#distributionloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.DistributionLoss.__call__': ( 'losses.pytorch.html#distributionloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.DistributionLoss.__init__': ( 'losses.pytorch.html#distributionloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.DistributionLoss.get_distribution': ( 'losses.pytorch.html#distributionloss.get_distribution',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.DistributionLoss.sample': ( 'losses.pytorch.html#distributionloss.sample',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM': ( 'losses.pytorch.html#gmm',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.__call__': ( 'losses.pytorch.html#gmm.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.__init__': ( 'losses.pytorch.html#gmm.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.domain_map': ( 'losses.pytorch.html#gmm.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.neglog_likelihood': ( 'losses.pytorch.html#gmm.neglog_likelihood',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.sample': ( 'losses.pytorch.html#gmm.sample',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.GMM.scale_decouple': ( 'losses.pytorch.html#gmm.scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberLoss': ( 'losses.pytorch.html#huberloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberLoss.__call__': ( 'losses.pytorch.html#huberloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberLoss.__init__': ( 'losses.pytorch.html#huberloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberMQLoss': ( 'losses.pytorch.html#hubermqloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberMQLoss.__call__': ( 'losses.pytorch.html#hubermqloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberMQLoss.__init__': ( 'losses.pytorch.html#hubermqloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberMQLoss._compute_weights': ( 'losses.pytorch.html#hubermqloss._compute_weights',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberMQLoss.domain_map': ( 'losses.pytorch.html#hubermqloss.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberQLoss': ( 'losses.pytorch.html#huberqloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberQLoss.__call__': ( 'losses.pytorch.html#huberqloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.HuberQLoss.__init__': ( 'losses.pytorch.html#huberqloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAE': ( 'losses.pytorch.html#mae',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAE.__call__': ( 'losses.pytorch.html#mae.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAE.__init__': ( 'losses.pytorch.html#mae.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAPE': ( 'losses.pytorch.html#mape',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAPE.__call__': ( 'losses.pytorch.html#mape.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MAPE.__init__': ( 'losses.pytorch.html#mape.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MASE': ( 'losses.pytorch.html#mase',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MASE.__call__': ( 'losses.pytorch.html#mase.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MASE.__init__': ( 'losses.pytorch.html#mase.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MQLoss': ( 'losses.pytorch.html#mqloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MQLoss.__call__': ( 'losses.pytorch.html#mqloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MQLoss.__init__': ( 'losses.pytorch.html#mqloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MQLoss._compute_weights': ( 'losses.pytorch.html#mqloss._compute_weights',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MQLoss.domain_map': ( 'losses.pytorch.html#mqloss.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MSE': ( 'losses.pytorch.html#mse',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MSE.__call__': ( 'losses.pytorch.html#mse.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.MSE.__init__': ( 'losses.pytorch.html#mse.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM': ( 'losses.pytorch.html#nbmm',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.__call__': ( 'losses.pytorch.html#nbmm.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.__init__': ( 'losses.pytorch.html#nbmm.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.domain_map': ( 'losses.pytorch.html#nbmm.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.neglog_likelihood': ( 'losses.pytorch.html#nbmm.neglog_likelihood',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.sample': ( 'losses.pytorch.html#nbmm.sample',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.NBMM.scale_decouple': ( 'losses.pytorch.html#nbmm.scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM': ( 'losses.pytorch.html#pmm',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.__call__': ( 'losses.pytorch.html#pmm.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.__init__': ( 'losses.pytorch.html#pmm.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.domain_map': ( 'losses.pytorch.html#pmm.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.neglog_likelihood': ( 'losses.pytorch.html#pmm.neglog_likelihood',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.sample': ( 'losses.pytorch.html#pmm.sample',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.PMM.scale_decouple': ( 'losses.pytorch.html#pmm.scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.QuantileLoss': ( 'losses.pytorch.html#quantileloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.QuantileLoss.__call__': ( 'losses.pytorch.html#quantileloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.QuantileLoss.__init__': ( 'losses.pytorch.html#quantileloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.RMSE': ( 'losses.pytorch.html#rmse',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.RMSE.__call__': ( 'losses.pytorch.html#rmse.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.RMSE.__init__': ( 'losses.pytorch.html#rmse.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.SMAPE': ( 'losses.pytorch.html#smape',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.SMAPE.__call__': ( 'losses.pytorch.html#smape.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.SMAPE.__init__': ( 'losses.pytorch.html#smape.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.TukeyLoss': ( 'losses.pytorch.html#tukeyloss',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.TukeyLoss.__call__': ( 'losses.pytorch.html#tukeyloss.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.TukeyLoss.__init__': ( 'losses.pytorch.html#tukeyloss.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.TukeyLoss.domain_map': ( 'losses.pytorch.html#tukeyloss.domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.TukeyLoss.masked_mean': ( 'losses.pytorch.html#tukeyloss.masked_mean',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie': ( 'losses.pytorch.html#tweedie',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie.__init__': ( 'losses.pytorch.html#tweedie.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie.log_prob': ( 'losses.pytorch.html#tweedie.log_prob',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie.mean': ( 'losses.pytorch.html#tweedie.mean',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie.sample': ( 'losses.pytorch.html#tweedie.sample',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.Tweedie.variance': ( 'losses.pytorch.html#tweedie.variance',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch._divide_no_nan': ( 'losses.pytorch.html#_divide_no_nan',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch._weighted_mean': ( 'losses.pytorch.html#_weighted_mean',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.bernoulli_domain_map': ( 'losses.pytorch.html#bernoulli_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.bernoulli_scale_decouple': ( 'losses.pytorch.html#bernoulli_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.est_alpha': ( 'losses.pytorch.html#est_alpha',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.est_beta': ( 'losses.pytorch.html#est_beta',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.est_lambda': ( 'losses.pytorch.html#est_lambda',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.level_to_outputs': ( 'losses.pytorch.html#level_to_outputs',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.nbinomial_domain_map': ( 'losses.pytorch.html#nbinomial_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.nbinomial_scale_decouple': ( 'losses.pytorch.html#nbinomial_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.normal_domain_map': ( 'losses.pytorch.html#normal_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.normal_scale_decouple': ( 'losses.pytorch.html#normal_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.poisson_domain_map': ( 'losses.pytorch.html#poisson_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.poisson_scale_decouple': ( 'losses.pytorch.html#poisson_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.quantiles_to_outputs': ( 'losses.pytorch.html#quantiles_to_outputs',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.relMSE': ( 'losses.pytorch.html#relmse',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.relMSE.__call__': ( 'losses.pytorch.html#relmse.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.relMSE.__init__': ( 'losses.pytorch.html#relmse.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.sCRPS': ( 'losses.pytorch.html#scrps',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.sCRPS.__call__': ( 'losses.pytorch.html#scrps.__call__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.sCRPS.__init__': ( 'losses.pytorch.html#scrps.__init__',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.student_domain_map': ( 'losses.pytorch.html#student_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.student_scale_decouple': ( 'losses.pytorch.html#student_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.tweedie_domain_map': ( 'losses.pytorch.html#tweedie_domain_map',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.tweedie_scale_decouple': ( 'losses.pytorch.html#tweedie_scale_decouple',
'neuralforecast/losses/pytorch.py'),
'neuralforecast.losses.pytorch.weighted_average': ( 'losses.pytorch.html#weighted_average',
'neuralforecast/losses/pytorch.py')},
'neuralforecast.models.autoformer': { 'neuralforecast.models.autoformer.AutoCorrelation': ( 'models.autoformer.html#autocorrelation',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelation.__init__': ( 'models.autoformer.html#autocorrelation.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelation.forward': ( 'models.autoformer.html#autocorrelation.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelation.time_delay_agg_full': ( 'models.autoformer.html#autocorrelation.time_delay_agg_full',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelation.time_delay_agg_inference': ( 'models.autoformer.html#autocorrelation.time_delay_agg_inference',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelation.time_delay_agg_training': ( 'models.autoformer.html#autocorrelation.time_delay_agg_training',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelationLayer': ( 'models.autoformer.html#autocorrelationlayer',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelationLayer.__init__': ( 'models.autoformer.html#autocorrelationlayer.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.AutoCorrelationLayer.forward': ( 'models.autoformer.html#autocorrelationlayer.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Autoformer': ( 'models.autoformer.html#autoformer',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Autoformer.__init__': ( 'models.autoformer.html#autoformer.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Autoformer.forward': ( 'models.autoformer.html#autoformer.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Decoder': ( 'models.autoformer.html#decoder',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Decoder.__init__': ( 'models.autoformer.html#decoder.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Decoder.forward': ( 'models.autoformer.html#decoder.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.DecoderLayer': ( 'models.autoformer.html#decoderlayer',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.DecoderLayer.__init__': ( 'models.autoformer.html#decoderlayer.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.DecoderLayer.forward': ( 'models.autoformer.html#decoderlayer.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Encoder': ( 'models.autoformer.html#encoder',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Encoder.__init__': ( 'models.autoformer.html#encoder.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.Encoder.forward': ( 'models.autoformer.html#encoder.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.EncoderLayer': ( 'models.autoformer.html#encoderlayer',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.EncoderLayer.__init__': ( 'models.autoformer.html#encoderlayer.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.EncoderLayer.forward': ( 'models.autoformer.html#encoderlayer.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.LayerNorm': ( 'models.autoformer.html#layernorm',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.LayerNorm.__init__': ( 'models.autoformer.html#layernorm.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.LayerNorm.forward': ( 'models.autoformer.html#layernorm.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.MovingAvg': ( 'models.autoformer.html#movingavg',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.MovingAvg.__init__': ( 'models.autoformer.html#movingavg.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.MovingAvg.forward': ( 'models.autoformer.html#movingavg.forward',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.SeriesDecomp': ( 'models.autoformer.html#seriesdecomp',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.SeriesDecomp.__init__': ( 'models.autoformer.html#seriesdecomp.__init__',
'neuralforecast/models/autoformer.py'),
'neuralforecast.models.autoformer.SeriesDecomp.forward': ( 'models.autoformer.html#seriesdecomp.forward',
'neuralforecast/models/autoformer.py')},
'neuralforecast.models.bitcn': { 'neuralforecast.models.bitcn.BiTCN': ( 'models.bitcn.html#bitcn',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.BiTCN.__init__': ( 'models.bitcn.html#bitcn.__init__',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.BiTCN.forward': ( 'models.bitcn.html#bitcn.forward',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.CustomConv1d': ( 'models.bitcn.html#customconv1d',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.CustomConv1d.__init__': ( 'models.bitcn.html#customconv1d.__init__',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.CustomConv1d.forward': ( 'models.bitcn.html#customconv1d.forward',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.TCNCell': ( 'models.bitcn.html#tcncell',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.TCNCell.__init__': ( 'models.bitcn.html#tcncell.__init__',
'neuralforecast/models/bitcn.py'),
'neuralforecast.models.bitcn.TCNCell.forward': ( 'models.bitcn.html#tcncell.forward',
'neuralforecast/models/bitcn.py')},
'neuralforecast.models.deepar': { 'neuralforecast.models.deepar.Decoder': ( 'models.deepar.html#decoder',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.Decoder.__init__': ( 'models.deepar.html#decoder.__init__',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.Decoder.forward': ( 'models.deepar.html#decoder.forward',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR': ( 'models.deepar.html#deepar',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.__init__': ( 'models.deepar.html#deepar.__init__',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.forward': ( 'models.deepar.html#deepar.forward',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.predict_step': ( 'models.deepar.html#deepar.predict_step',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.train_forward': ( 'models.deepar.html#deepar.train_forward',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.training_step': ( 'models.deepar.html#deepar.training_step',
'neuralforecast/models/deepar.py'),
'neuralforecast.models.deepar.DeepAR.validation_step': ( 'models.deepar.html#deepar.validation_step',
'neuralforecast/models/deepar.py')},
'neuralforecast.models.dilated_rnn': { 'neuralforecast.models.dilated_rnn.AttentiveLSTMLayer': ( 'models.dilated_rnn.html#attentivelstmlayer',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.AttentiveLSTMLayer.__init__': ( 'models.dilated_rnn.html#attentivelstmlayer.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.AttentiveLSTMLayer.forward': ( 'models.dilated_rnn.html#attentivelstmlayer.forward',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN': ( 'models.dilated_rnn.html#drnn',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN.__init__': ( 'models.dilated_rnn.html#drnn.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN._apply_cell': ( 'models.dilated_rnn.html#drnn._apply_cell',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN._pad_inputs': ( 'models.dilated_rnn.html#drnn._pad_inputs',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN._prepare_inputs': ( 'models.dilated_rnn.html#drnn._prepare_inputs',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN._split_outputs': ( 'models.dilated_rnn.html#drnn._split_outputs',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN._unpad_outputs': ( 'models.dilated_rnn.html#drnn._unpad_outputs',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN.drnn_layer': ( 'models.dilated_rnn.html#drnn.drnn_layer',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DRNN.forward': ( 'models.dilated_rnn.html#drnn.forward',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DilatedRNN': ( 'models.dilated_rnn.html#dilatedrnn',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DilatedRNN.__init__': ( 'models.dilated_rnn.html#dilatedrnn.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.DilatedRNN.forward': ( 'models.dilated_rnn.html#dilatedrnn.forward',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.LSTMCell': ( 'models.dilated_rnn.html#lstmcell',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.LSTMCell.__init__': ( 'models.dilated_rnn.html#lstmcell.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.LSTMCell.forward': ( 'models.dilated_rnn.html#lstmcell.forward',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMCell': ( 'models.dilated_rnn.html#reslstmcell',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMCell.__init__': ( 'models.dilated_rnn.html#reslstmcell.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMCell.forward': ( 'models.dilated_rnn.html#reslstmcell.forward',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMLayer': ( 'models.dilated_rnn.html#reslstmlayer',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMLayer.__init__': ( 'models.dilated_rnn.html#reslstmlayer.__init__',
'neuralforecast/models/dilated_rnn.py'),
'neuralforecast.models.dilated_rnn.ResLSTMLayer.forward': ( 'models.dilated_rnn.html#reslstmlayer.forward',
'neuralforecast/models/dilated_rnn.py')},
'neuralforecast.models.dlinear': { 'neuralforecast.models.dlinear.DLinear': ( 'models.dlinear.html#dlinear',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.DLinear.__init__': ( 'models.dlinear.html#dlinear.__init__',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.DLinear.forward': ( 'models.dlinear.html#dlinear.forward',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.MovingAvg': ( 'models.dlinear.html#movingavg',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.MovingAvg.__init__': ( 'models.dlinear.html#movingavg.__init__',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.MovingAvg.forward': ( 'models.dlinear.html#movingavg.forward',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.SeriesDecomp': ( 'models.dlinear.html#seriesdecomp',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.SeriesDecomp.__init__': ( 'models.dlinear.html#seriesdecomp.__init__',
'neuralforecast/models/dlinear.py'),
'neuralforecast.models.dlinear.SeriesDecomp.forward': ( 'models.dlinear.html#seriesdecomp.forward',
'neuralforecast/models/dlinear.py')},
'neuralforecast.models.fedformer': { 'neuralforecast.models.fedformer.AutoCorrelationLayer': ( 'models.fedformer.html#autocorrelationlayer',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.AutoCorrelationLayer.__init__': ( 'models.fedformer.html#autocorrelationlayer.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.AutoCorrelationLayer.forward': ( 'models.fedformer.html#autocorrelationlayer.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Decoder': ( 'models.fedformer.html#decoder',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Decoder.__init__': ( 'models.fedformer.html#decoder.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Decoder.forward': ( 'models.fedformer.html#decoder.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.DecoderLayer': ( 'models.fedformer.html#decoderlayer',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.DecoderLayer.__init__': ( 'models.fedformer.html#decoderlayer.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.DecoderLayer.forward': ( 'models.fedformer.html#decoderlayer.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Encoder': ( 'models.fedformer.html#encoder',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Encoder.__init__': ( 'models.fedformer.html#encoder.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.Encoder.forward': ( 'models.fedformer.html#encoder.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.EncoderLayer': ( 'models.fedformer.html#encoderlayer',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.EncoderLayer.__init__': ( 'models.fedformer.html#encoderlayer.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.EncoderLayer.forward': ( 'models.fedformer.html#encoderlayer.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FEDformer': ( 'models.fedformer.html#fedformer',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FEDformer.__init__': ( 'models.fedformer.html#fedformer.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FEDformer.forward': ( 'models.fedformer.html#fedformer.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierBlock': ( 'models.fedformer.html#fourierblock',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierBlock.__init__': ( 'models.fedformer.html#fourierblock.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierBlock.compl_mul1d': ( 'models.fedformer.html#fourierblock.compl_mul1d',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierBlock.forward': ( 'models.fedformer.html#fourierblock.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierCrossAttention': ( 'models.fedformer.html#fouriercrossattention',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierCrossAttention.__init__': ( 'models.fedformer.html#fouriercrossattention.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierCrossAttention.compl_mul1d': ( 'models.fedformer.html#fouriercrossattention.compl_mul1d',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.FourierCrossAttention.forward': ( 'models.fedformer.html#fouriercrossattention.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.LayerNorm': ( 'models.fedformer.html#layernorm',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.LayerNorm.__init__': ( 'models.fedformer.html#layernorm.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.LayerNorm.forward': ( 'models.fedformer.html#layernorm.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.MovingAvg': ( 'models.fedformer.html#movingavg',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.MovingAvg.__init__': ( 'models.fedformer.html#movingavg.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.MovingAvg.forward': ( 'models.fedformer.html#movingavg.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.SeriesDecomp': ( 'models.fedformer.html#seriesdecomp',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.SeriesDecomp.__init__': ( 'models.fedformer.html#seriesdecomp.__init__',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.SeriesDecomp.forward': ( 'models.fedformer.html#seriesdecomp.forward',
'neuralforecast/models/fedformer.py'),
'neuralforecast.models.fedformer.get_frequency_modes': ( 'models.fedformer.html#get_frequency_modes',
'neuralforecast/models/fedformer.py')},
'neuralforecast.models.gru': { 'neuralforecast.models.gru.GRU': ('models.gru.html#gru', 'neuralforecast/models/gru.py'),
'neuralforecast.models.gru.GRU.__init__': ( 'models.gru.html#gru.__init__',
'neuralforecast/models/gru.py'),
'neuralforecast.models.gru.GRU.forward': ( 'models.gru.html#gru.forward',
'neuralforecast/models/gru.py')},
'neuralforecast.models.hint': { 'neuralforecast.models.hint.HINT': ('models.hint.html#hint', 'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.__init__': ( 'models.hint.html#hint.__init__',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.__repr__': ( 'models.hint.html#hint.__repr__',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.fit': ( 'models.hint.html#hint.fit',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.get_test_size': ( 'models.hint.html#hint.get_test_size',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.predict': ( 'models.hint.html#hint.predict',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.save': ( 'models.hint.html#hint.save',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.HINT.set_test_size': ( 'models.hint.html#hint.set_test_size',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.get_bottomup_P': ( 'models.hint.html#get_bottomup_p',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.get_identity_P': ( 'models.hint.html#get_identity_p',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.get_mintrace_ols_P': ( 'models.hint.html#get_mintrace_ols_p',
'neuralforecast/models/hint.py'),
'neuralforecast.models.hint.get_mintrace_wls_P': ( 'models.hint.html#get_mintrace_wls_p',
'neuralforecast/models/hint.py')},
'neuralforecast.models.informer': { 'neuralforecast.models.informer.ConvLayer': ( 'models.informer.html#convlayer',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ConvLayer.__init__': ( 'models.informer.html#convlayer.__init__',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ConvLayer.forward': ( 'models.informer.html#convlayer.forward',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.Informer': ( 'models.informer.html#informer',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.Informer.__init__': ( 'models.informer.html#informer.__init__',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.Informer.forward': ( 'models.informer.html#informer.forward',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention': ( 'models.informer.html#probattention',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention.__init__': ( 'models.informer.html#probattention.__init__',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention._get_initial_context': ( 'models.informer.html#probattention._get_initial_context',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention._prob_QK': ( 'models.informer.html#probattention._prob_qk',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention._update_context': ( 'models.informer.html#probattention._update_context',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbAttention.forward': ( 'models.informer.html#probattention.forward',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbMask': ( 'models.informer.html#probmask',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbMask.__init__': ( 'models.informer.html#probmask.__init__',
'neuralforecast/models/informer.py'),
'neuralforecast.models.informer.ProbMask.mask': ( 'models.informer.html#probmask.mask',
'neuralforecast/models/informer.py')},
'neuralforecast.models.itransformer': { 'neuralforecast.models.itransformer.DataEmbedding_inverted': ( 'models.itransformer.html#dataembedding_inverted',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.DataEmbedding_inverted.__init__': ( 'models.itransformer.html#dataembedding_inverted.__init__',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.DataEmbedding_inverted.forward': ( 'models.itransformer.html#dataembedding_inverted.forward',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.FullAttention': ( 'models.itransformer.html#fullattention',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.FullAttention.__init__': ( 'models.itransformer.html#fullattention.__init__',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.FullAttention.forward': ( 'models.itransformer.html#fullattention.forward',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.TriangularCausalMask': ( 'models.itransformer.html#triangularcausalmask',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.TriangularCausalMask.__init__': ( 'models.itransformer.html#triangularcausalmask.__init__',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.TriangularCausalMask.mask': ( 'models.itransformer.html#triangularcausalmask.mask',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.iTransformer': ( 'models.itransformer.html#itransformer',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.iTransformer.__init__': ( 'models.itransformer.html#itransformer.__init__',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.iTransformer.forecast': ( 'models.itransformer.html#itransformer.forecast',
'neuralforecast/models/itransformer.py'),
'neuralforecast.models.itransformer.iTransformer.forward': ( 'models.itransformer.html#itransformer.forward',
'neuralforecast/models/itransformer.py')},
'neuralforecast.models.lstm': { 'neuralforecast.models.lstm.LSTM': ('models.lstm.html#lstm', 'neuralforecast/models/lstm.py'),
'neuralforecast.models.lstm.LSTM.__init__': ( 'models.lstm.html#lstm.__init__',
'neuralforecast/models/lstm.py'),
'neuralforecast.models.lstm.LSTM.forward': ( 'models.lstm.html#lstm.forward',
'neuralforecast/models/lstm.py')},
'neuralforecast.models.mlp': { 'neuralforecast.models.mlp.MLP': ('models.mlp.html#mlp', 'neuralforecast/models/mlp.py'),
'neuralforecast.models.mlp.MLP.__init__': ( 'models.mlp.html#mlp.__init__',
'neuralforecast/models/mlp.py'),
'neuralforecast.models.mlp.MLP.forward': ( 'models.mlp.html#mlp.forward',
'neuralforecast/models/mlp.py')},
'neuralforecast.models.mlpmultivariate': { 'neuralforecast.models.mlpmultivariate.MLPMultivariate': ( 'models.mlpmultivariate.html#mlpmultivariate',
'neuralforecast/models/mlpmultivariate.py'),
'neuralforecast.models.mlpmultivariate.MLPMultivariate.__init__': ( 'models.mlpmultivariate.html#mlpmultivariate.__init__',
'neuralforecast/models/mlpmultivariate.py'),
'neuralforecast.models.mlpmultivariate.MLPMultivariate.forward': ( 'models.mlpmultivariate.html#mlpmultivariate.forward',
'neuralforecast/models/mlpmultivariate.py')},
'neuralforecast.models.nbeats': { 'neuralforecast.models.nbeats.IdentityBasis': ( 'models.nbeats.html#identitybasis',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.IdentityBasis.__init__': ( 'models.nbeats.html#identitybasis.__init__',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.IdentityBasis.forward': ( 'models.nbeats.html#identitybasis.forward',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATS': ( 'models.nbeats.html#nbeats',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATS.__init__': ( 'models.nbeats.html#nbeats.__init__',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATS.create_stack': ( 'models.nbeats.html#nbeats.create_stack',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATS.forward': ( 'models.nbeats.html#nbeats.forward',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATSBlock': ( 'models.nbeats.html#nbeatsblock',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATSBlock.__init__': ( 'models.nbeats.html#nbeatsblock.__init__',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.NBEATSBlock.forward': ( 'models.nbeats.html#nbeatsblock.forward',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.SeasonalityBasis': ( 'models.nbeats.html#seasonalitybasis',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.SeasonalityBasis.__init__': ( 'models.nbeats.html#seasonalitybasis.__init__',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.SeasonalityBasis.forward': ( 'models.nbeats.html#seasonalitybasis.forward',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.TrendBasis': ( 'models.nbeats.html#trendbasis',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.TrendBasis.__init__': ( 'models.nbeats.html#trendbasis.__init__',
'neuralforecast/models/nbeats.py'),
'neuralforecast.models.nbeats.TrendBasis.forward': ( 'models.nbeats.html#trendbasis.forward',
'neuralforecast/models/nbeats.py')},
'neuralforecast.models.nbeatsx': { 'neuralforecast.models.nbeatsx.ExogenousBasis': ( 'models.nbeatsx.html#exogenousbasis',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.ExogenousBasis.__init__': ( 'models.nbeatsx.html#exogenousbasis.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.ExogenousBasis.forward': ( 'models.nbeatsx.html#exogenousbasis.forward',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.IdentityBasis': ( 'models.nbeatsx.html#identitybasis',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.IdentityBasis.__init__': ( 'models.nbeatsx.html#identitybasis.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.IdentityBasis.forward': ( 'models.nbeatsx.html#identitybasis.forward',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSBlock': ( 'models.nbeatsx.html#nbeatsblock',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSBlock.__init__': ( 'models.nbeatsx.html#nbeatsblock.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSBlock.forward': ( 'models.nbeatsx.html#nbeatsblock.forward',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSx': ( 'models.nbeatsx.html#nbeatsx',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSx.__init__': ( 'models.nbeatsx.html#nbeatsx.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSx.create_stack': ( 'models.nbeatsx.html#nbeatsx.create_stack',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.NBEATSx.forward': ( 'models.nbeatsx.html#nbeatsx.forward',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.SeasonalityBasis': ( 'models.nbeatsx.html#seasonalitybasis',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.SeasonalityBasis.__init__': ( 'models.nbeatsx.html#seasonalitybasis.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.SeasonalityBasis.forward': ( 'models.nbeatsx.html#seasonalitybasis.forward',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.TrendBasis': ( 'models.nbeatsx.html#trendbasis',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.TrendBasis.__init__': ( 'models.nbeatsx.html#trendbasis.__init__',
'neuralforecast/models/nbeatsx.py'),
'neuralforecast.models.nbeatsx.TrendBasis.forward': ( 'models.nbeatsx.html#trendbasis.forward',
'neuralforecast/models/nbeatsx.py')},
'neuralforecast.models.nhits': { 'neuralforecast.models.nhits.NHITS': ( 'models.nhits.html#nhits',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITS.__init__': ( 'models.nhits.html#nhits.__init__',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITS.create_stack': ( 'models.nhits.html#nhits.create_stack',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITS.forward': ( 'models.nhits.html#nhits.forward',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITSBlock': ( 'models.nhits.html#nhitsblock',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITSBlock.__init__': ( 'models.nhits.html#nhitsblock.__init__',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits.NHITSBlock.forward': ( 'models.nhits.html#nhitsblock.forward',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits._IdentityBasis': ( 'models.nhits.html#_identitybasis',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits._IdentityBasis.__init__': ( 'models.nhits.html#_identitybasis.__init__',
'neuralforecast/models/nhits.py'),
'neuralforecast.models.nhits._IdentityBasis.forward': ( 'models.nhits.html#_identitybasis.forward',
'neuralforecast/models/nhits.py')},
'neuralforecast.models.nlinear': { 'neuralforecast.models.nlinear.NLinear': ( 'models.nlinear.html#nlinear',
'neuralforecast/models/nlinear.py'),
'neuralforecast.models.nlinear.NLinear.__init__': ( 'models.nlinear.html#nlinear.__init__',
'neuralforecast/models/nlinear.py'),
'neuralforecast.models.nlinear.NLinear.forward': ( 'models.nlinear.html#nlinear.forward',
'neuralforecast/models/nlinear.py')},
'neuralforecast.models.patchtst': { 'neuralforecast.models.patchtst.Coord1dPosEncoding': ( 'models.patchtst.html#coord1dposencoding',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Coord2dPosEncoding': ( 'models.patchtst.html#coord2dposencoding',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Flatten_Head': ( 'models.patchtst.html#flatten_head',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Flatten_Head.__init__': ( 'models.patchtst.html#flatten_head.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Flatten_Head.forward': ( 'models.patchtst.html#flatten_head.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST': ( 'models.patchtst.html#patchtst',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST.__init__': ( 'models.patchtst.html#patchtst.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST.forward': ( 'models.patchtst.html#patchtst.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST_backbone': ( 'models.patchtst.html#patchtst_backbone',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST_backbone.__init__': ( 'models.patchtst.html#patchtst_backbone.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST_backbone.create_pretrain_head': ( 'models.patchtst.html#patchtst_backbone.create_pretrain_head',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PatchTST_backbone.forward': ( 'models.patchtst.html#patchtst_backbone.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.PositionalEncoding': ( 'models.patchtst.html#positionalencoding',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN': ( 'models.patchtst.html#revin',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN.__init__': ( 'models.patchtst.html#revin.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN._denormalize': ( 'models.patchtst.html#revin._denormalize',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN._get_statistics': ( 'models.patchtst.html#revin._get_statistics',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN._init_params': ( 'models.patchtst.html#revin._init_params',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN._normalize': ( 'models.patchtst.html#revin._normalize',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.RevIN.forward': ( 'models.patchtst.html#revin.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoder': ( 'models.patchtst.html#tstencoder',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoder.__init__': ( 'models.patchtst.html#tstencoder.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoder.forward': ( 'models.patchtst.html#tstencoder.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoderLayer': ( 'models.patchtst.html#tstencoderlayer',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoderLayer.__init__': ( 'models.patchtst.html#tstencoderlayer.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTEncoderLayer.forward': ( 'models.patchtst.html#tstencoderlayer.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTiEncoder': ( 'models.patchtst.html#tstiencoder',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTiEncoder.__init__': ( 'models.patchtst.html#tstiencoder.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.TSTiEncoder.forward': ( 'models.patchtst.html#tstiencoder.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Transpose': ( 'models.patchtst.html#transpose',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Transpose.__init__': ( 'models.patchtst.html#transpose.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.Transpose.forward': ( 'models.patchtst.html#transpose.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._MultiheadAttention': ( 'models.patchtst.html#_multiheadattention',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._MultiheadAttention.__init__': ( 'models.patchtst.html#_multiheadattention.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._MultiheadAttention.forward': ( 'models.patchtst.html#_multiheadattention.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._ScaledDotProductAttention': ( 'models.patchtst.html#_scaleddotproductattention',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._ScaledDotProductAttention.__init__': ( 'models.patchtst.html#_scaleddotproductattention.__init__',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst._ScaledDotProductAttention.forward': ( 'models.patchtst.html#_scaleddotproductattention.forward',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.get_activation_fn': ( 'models.patchtst.html#get_activation_fn',
'neuralforecast/models/patchtst.py'),
'neuralforecast.models.patchtst.positional_encoding': ( 'models.patchtst.html#positional_encoding',
'neuralforecast/models/patchtst.py')},
'neuralforecast.models.rnn': { 'neuralforecast.models.rnn.RNN': ('models.rnn.html#rnn', 'neuralforecast/models/rnn.py'),
'neuralforecast.models.rnn.RNN.__init__': ( 'models.rnn.html#rnn.__init__',
'neuralforecast/models/rnn.py'),
'neuralforecast.models.rnn.RNN.forward': ( 'models.rnn.html#rnn.forward',
'neuralforecast/models/rnn.py')},
'neuralforecast.models.stemgnn': { 'neuralforecast.models.stemgnn.GLU': ( 'models.stemgnn.html#glu',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.GLU.__init__': ( 'models.stemgnn.html#glu.__init__',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.GLU.forward': ( 'models.stemgnn.html#glu.forward',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN': ( 'models.stemgnn.html#stemgnn',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.__init__': ( 'models.stemgnn.html#stemgnn.__init__',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.cheb_polynomial': ( 'models.stemgnn.html#stemgnn.cheb_polynomial',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.forward': ( 'models.stemgnn.html#stemgnn.forward',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.get_laplacian': ( 'models.stemgnn.html#stemgnn.get_laplacian',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.graph_fft': ( 'models.stemgnn.html#stemgnn.graph_fft',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.latent_correlation_layer': ( 'models.stemgnn.html#stemgnn.latent_correlation_layer',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StemGNN.self_graph_attention': ( 'models.stemgnn.html#stemgnn.self_graph_attention',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StockBlockLayer': ( 'models.stemgnn.html#stockblocklayer',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StockBlockLayer.__init__': ( 'models.stemgnn.html#stockblocklayer.__init__',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StockBlockLayer.forward': ( 'models.stemgnn.html#stockblocklayer.forward',
'neuralforecast/models/stemgnn.py'),
'neuralforecast.models.stemgnn.StockBlockLayer.spe_seq_cell': ( 'models.stemgnn.html#stockblocklayer.spe_seq_cell',
'neuralforecast/models/stemgnn.py')},
'neuralforecast.models.tcn': { 'neuralforecast.models.tcn.TCN': ('models.tcn.html#tcn', 'neuralforecast/models/tcn.py'),
'neuralforecast.models.tcn.TCN.__init__': ( 'models.tcn.html#tcn.__init__',
'neuralforecast/models/tcn.py'),
'neuralforecast.models.tcn.TCN.forward': ( 'models.tcn.html#tcn.forward',
'neuralforecast/models/tcn.py')},
'neuralforecast.models.tft': { 'neuralforecast.models.tft.GLU': ('models.tft.html#glu', 'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.GLU.__init__': ( 'models.tft.html#glu.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.GLU.forward': ( 'models.tft.html#glu.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.GRN': ('models.tft.html#grn', 'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.GRN.__init__': ( 'models.tft.html#grn.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.GRN.forward': ( 'models.tft.html#grn.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.InterpretableMultiHeadAttention': ( 'models.tft.html#interpretablemultiheadattention',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.InterpretableMultiHeadAttention.__init__': ( 'models.tft.html#interpretablemultiheadattention.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.InterpretableMultiHeadAttention.forward': ( 'models.tft.html#interpretablemultiheadattention.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.MaybeLayerNorm': ( 'models.tft.html#maybelayernorm',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.MaybeLayerNorm.__init__': ( 'models.tft.html#maybelayernorm.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.MaybeLayerNorm.forward': ( 'models.tft.html#maybelayernorm.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.StaticCovariateEncoder': ( 'models.tft.html#staticcovariateencoder',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.StaticCovariateEncoder.__init__': ( 'models.tft.html#staticcovariateencoder.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.StaticCovariateEncoder.forward': ( 'models.tft.html#staticcovariateencoder.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFT': ('models.tft.html#tft', 'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFT.__init__': ( 'models.tft.html#tft.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFT.forward': ( 'models.tft.html#tft.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFTEmbedding': ( 'models.tft.html#tftembedding',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFTEmbedding.__init__': ( 'models.tft.html#tftembedding.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFTEmbedding._apply_embedding': ( 'models.tft.html#tftembedding._apply_embedding',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TFTEmbedding.forward': ( 'models.tft.html#tftembedding.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalCovariateEncoder': ( 'models.tft.html#temporalcovariateencoder',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalCovariateEncoder.__init__': ( 'models.tft.html#temporalcovariateencoder.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalCovariateEncoder.forward': ( 'models.tft.html#temporalcovariateencoder.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalFusionDecoder': ( 'models.tft.html#temporalfusiondecoder',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalFusionDecoder.__init__': ( 'models.tft.html#temporalfusiondecoder.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.TemporalFusionDecoder.forward': ( 'models.tft.html#temporalfusiondecoder.forward',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.VariableSelectionNetwork': ( 'models.tft.html#variableselectionnetwork',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.VariableSelectionNetwork.__init__': ( 'models.tft.html#variableselectionnetwork.__init__',
'neuralforecast/models/tft.py'),
'neuralforecast.models.tft.VariableSelectionNetwork.forward': ( 'models.tft.html#variableselectionnetwork.forward',
'neuralforecast/models/tft.py')},
'neuralforecast.models.timellm': { 'neuralforecast.models.timellm.FlattenHead': ( 'models.timellm.html#flattenhead',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.FlattenHead.__init__': ( 'models.timellm.html#flattenhead.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.FlattenHead.forward': ( 'models.timellm.html#flattenhead.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize': ( 'models.timellm.html#normalize',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize.__init__': ( 'models.timellm.html#normalize.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize._denormalize': ( 'models.timellm.html#normalize._denormalize',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize._get_statistics': ( 'models.timellm.html#normalize._get_statistics',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize._init_params': ( 'models.timellm.html#normalize._init_params',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize._normalize': ( 'models.timellm.html#normalize._normalize',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.Normalize.forward': ( 'models.timellm.html#normalize.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.PatchEmbedding': ( 'models.timellm.html#patchembedding',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.PatchEmbedding.__init__': ( 'models.timellm.html#patchembedding.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.PatchEmbedding.forward': ( 'models.timellm.html#patchembedding.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReplicationPad1d': ( 'models.timellm.html#replicationpad1d',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReplicationPad1d.__init__': ( 'models.timellm.html#replicationpad1d.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReplicationPad1d.forward': ( 'models.timellm.html#replicationpad1d.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReprogrammingLayer': ( 'models.timellm.html#reprogramminglayer',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReprogrammingLayer.__init__': ( 'models.timellm.html#reprogramminglayer.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReprogrammingLayer.forward': ( 'models.timellm.html#reprogramminglayer.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.ReprogrammingLayer.reprogramming': ( 'models.timellm.html#reprogramminglayer.reprogramming',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TimeLLM': ( 'models.timellm.html#timellm',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TimeLLM.__init__': ( 'models.timellm.html#timellm.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TimeLLM.calcute_lags': ( 'models.timellm.html#timellm.calcute_lags',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TimeLLM.forecast': ( 'models.timellm.html#timellm.forecast',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TimeLLM.forward': ( 'models.timellm.html#timellm.forward',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TokenEmbedding': ( 'models.timellm.html#tokenembedding',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TokenEmbedding.__init__': ( 'models.timellm.html#tokenembedding.__init__',
'neuralforecast/models/timellm.py'),
'neuralforecast.models.timellm.TokenEmbedding.forward': ( 'models.timellm.html#tokenembedding.forward',
'neuralforecast/models/timellm.py')},
'neuralforecast.models.timesnet': { 'neuralforecast.models.timesnet.FFT_for_Period': ( 'models.timesnet.html#fft_for_period',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.Inception_Block_V1': ( 'models.timesnet.html#inception_block_v1',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.Inception_Block_V1.__init__': ( 'models.timesnet.html#inception_block_v1.__init__',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.Inception_Block_V1._initialize_weights': ( 'models.timesnet.html#inception_block_v1._initialize_weights',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.Inception_Block_V1.forward': ( 'models.timesnet.html#inception_block_v1.forward',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesBlock': ( 'models.timesnet.html#timesblock',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesBlock.__init__': ( 'models.timesnet.html#timesblock.__init__',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesBlock.forward': ( 'models.timesnet.html#timesblock.forward',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesNet': ( 'models.timesnet.html#timesnet',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesNet.__init__': ( 'models.timesnet.html#timesnet.__init__',
'neuralforecast/models/timesnet.py'),
'neuralforecast.models.timesnet.TimesNet.forward': ( 'models.timesnet.html#timesnet.forward',
'neuralforecast/models/timesnet.py')},
'neuralforecast.models.tsmixer': { 'neuralforecast.models.tsmixer.FeatureMixing': ( 'models.tsmixer.html#featuremixing',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.FeatureMixing.__init__': ( 'models.tsmixer.html#featuremixing.__init__',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.FeatureMixing.forward': ( 'models.tsmixer.html#featuremixing.forward',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.MixingLayer': ( 'models.tsmixer.html#mixinglayer',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.MixingLayer.__init__': ( 'models.tsmixer.html#mixinglayer.__init__',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.MixingLayer.forward': ( 'models.tsmixer.html#mixinglayer.forward',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.ReversibleInstanceNorm1d': ( 'models.tsmixer.html#reversibleinstancenorm1d',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.ReversibleInstanceNorm1d.__init__': ( 'models.tsmixer.html#reversibleinstancenorm1d.__init__',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.ReversibleInstanceNorm1d.forward': ( 'models.tsmixer.html#reversibleinstancenorm1d.forward',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.ReversibleInstanceNorm1d.reverse': ( 'models.tsmixer.html#reversibleinstancenorm1d.reverse',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TSMixer': ( 'models.tsmixer.html#tsmixer',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TSMixer.__init__': ( 'models.tsmixer.html#tsmixer.__init__',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TSMixer.forward': ( 'models.tsmixer.html#tsmixer.forward',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TemporalMixing': ( 'models.tsmixer.html#temporalmixing',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TemporalMixing.__init__': ( 'models.tsmixer.html#temporalmixing.__init__',
'neuralforecast/models/tsmixer.py'),
'neuralforecast.models.tsmixer.TemporalMixing.forward': ( 'models.tsmixer.html#temporalmixing.forward',
'neuralforecast/models/tsmixer.py')},
'neuralforecast.models.tsmixerx': { 'neuralforecast.models.tsmixerx.FeatureMixing': ( 'models.tsmixerx.html#featuremixing',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.FeatureMixing.__init__': ( 'models.tsmixerx.html#featuremixing.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.FeatureMixing.forward': ( 'models.tsmixerx.html#featuremixing.forward',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayer': ( 'models.tsmixerx.html#mixinglayer',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayer.__init__': ( 'models.tsmixerx.html#mixinglayer.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayer.forward': ( 'models.tsmixerx.html#mixinglayer.forward',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayerWithStaticExogenous': ( 'models.tsmixerx.html#mixinglayerwithstaticexogenous',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayerWithStaticExogenous.__init__': ( 'models.tsmixerx.html#mixinglayerwithstaticexogenous.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.MixingLayerWithStaticExogenous.forward': ( 'models.tsmixerx.html#mixinglayerwithstaticexogenous.forward',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.ReversibleInstanceNorm1d': ( 'models.tsmixerx.html#reversibleinstancenorm1d',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.ReversibleInstanceNorm1d.__init__': ( 'models.tsmixerx.html#reversibleinstancenorm1d.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.ReversibleInstanceNorm1d.forward': ( 'models.tsmixerx.html#reversibleinstancenorm1d.forward',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.ReversibleInstanceNorm1d.reverse': ( 'models.tsmixerx.html#reversibleinstancenorm1d.reverse',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TSMixerx': ( 'models.tsmixerx.html#tsmixerx',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TSMixerx.__init__': ( 'models.tsmixerx.html#tsmixerx.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TSMixerx.forward': ( 'models.tsmixerx.html#tsmixerx.forward',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TemporalMixing': ( 'models.tsmixerx.html#temporalmixing',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TemporalMixing.__init__': ( 'models.tsmixerx.html#temporalmixing.__init__',
'neuralforecast/models/tsmixerx.py'),
'neuralforecast.models.tsmixerx.TemporalMixing.forward': ( 'models.tsmixerx.html#temporalmixing.forward',
'neuralforecast/models/tsmixerx.py')},
'neuralforecast.models.vanillatransformer': { 'neuralforecast.models.vanillatransformer.FullAttention': ( 'models.vanillatransformer.html#fullattention',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.FullAttention.__init__': ( 'models.vanillatransformer.html#fullattention.__init__',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.FullAttention.forward': ( 'models.vanillatransformer.html#fullattention.forward',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.TriangularCausalMask': ( 'models.vanillatransformer.html#triangularcausalmask',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.TriangularCausalMask.__init__': ( 'models.vanillatransformer.html#triangularcausalmask.__init__',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.TriangularCausalMask.mask': ( 'models.vanillatransformer.html#triangularcausalmask.mask',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.VanillaTransformer': ( 'models.vanillatransformer.html#vanillatransformer',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.VanillaTransformer.__init__': ( 'models.vanillatransformer.html#vanillatransformer.__init__',
'neuralforecast/models/vanillatransformer.py'),
'neuralforecast.models.vanillatransformer.VanillaTransformer.forward': ( 'models.vanillatransformer.html#vanillatransformer.forward',
'neuralforecast/models/vanillatransformer.py')},
'neuralforecast.tsdataset': { 'neuralforecast.tsdataset.TimeSeriesDataModule': ( 'tsdataset.html#timeseriesdatamodule',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataModule.__init__': ( 'tsdataset.html#timeseriesdatamodule.__init__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataModule.predict_dataloader': ( 'tsdataset.html#timeseriesdatamodule.predict_dataloader',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataModule.train_dataloader': ( 'tsdataset.html#timeseriesdatamodule.train_dataloader',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataModule.val_dataloader': ( 'tsdataset.html#timeseriesdatamodule.val_dataloader',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset': ( 'tsdataset.html#timeseriesdataset',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.__eq__': ( 'tsdataset.html#timeseriesdataset.__eq__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.__getitem__': ( 'tsdataset.html#timeseriesdataset.__getitem__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.__init__': ( 'tsdataset.html#timeseriesdataset.__init__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.__len__': ( 'tsdataset.html#timeseriesdataset.__len__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.__repr__': ( 'tsdataset.html#timeseriesdataset.__repr__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset._as_torch_copy': ( 'tsdataset.html#timeseriesdataset._as_torch_copy',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.align': ( 'tsdataset.html#timeseriesdataset.align',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.append': ( 'tsdataset.html#timeseriesdataset.append',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.from_df': ( 'tsdataset.html#timeseriesdataset.from_df',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.trim_dataset': ( 'tsdataset.html#timeseriesdataset.trim_dataset',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesDataset.update_dataset': ( 'tsdataset.html#timeseriesdataset.update_dataset',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesLoader': ( 'tsdataset.html#timeseriesloader',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesLoader.__init__': ( 'tsdataset.html#timeseriesloader.__init__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset.TimeSeriesLoader._collate_fn': ( 'tsdataset.html#timeseriesloader._collate_fn',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset._DistributedTimeSeriesDataModule': ( 'tsdataset.html#_distributedtimeseriesdatamodule',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset._DistributedTimeSeriesDataModule.__init__': ( 'tsdataset.html#_distributedtimeseriesdatamodule.__init__',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset._DistributedTimeSeriesDataModule.setup': ( 'tsdataset.html#_distributedtimeseriesdatamodule.setup',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset._FilesDataset': ( 'tsdataset.html#_filesdataset',
'neuralforecast/tsdataset.py'),
'neuralforecast.tsdataset._FilesDataset.__init__': ( 'tsdataset.html#_filesdataset.__init__',
'neuralforecast/tsdataset.py')},
'neuralforecast.utils': { 'neuralforecast.utils.DayOfMonth': ('utils.html#dayofmonth', 'neuralforecast/utils.py'),
'neuralforecast.utils.DayOfMonth.__call__': ( 'utils.html#dayofmonth.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.DayOfWeek': ('utils.html#dayofweek', 'neuralforecast/utils.py'),
'neuralforecast.utils.DayOfWeek.__call__': ( 'utils.html#dayofweek.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.DayOfYear': ('utils.html#dayofyear', 'neuralforecast/utils.py'),
'neuralforecast.utils.DayOfYear.__call__': ( 'utils.html#dayofyear.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.HourOfDay': ('utils.html#hourofday', 'neuralforecast/utils.py'),
'neuralforecast.utils.HourOfDay.__call__': ( 'utils.html#hourofday.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.MinuteOfHour': ('utils.html#minuteofhour', 'neuralforecast/utils.py'),
'neuralforecast.utils.MinuteOfHour.__call__': ( 'utils.html#minuteofhour.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.MonthOfYear': ('utils.html#monthofyear', 'neuralforecast/utils.py'),
'neuralforecast.utils.MonthOfYear.__call__': ( 'utils.html#monthofyear.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.SecondOfMinute': ('utils.html#secondofminute', 'neuralforecast/utils.py'),
'neuralforecast.utils.SecondOfMinute.__call__': ( 'utils.html#secondofminute.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.TimeFeature': ('utils.html#timefeature', 'neuralforecast/utils.py'),
'neuralforecast.utils.TimeFeature.__call__': ( 'utils.html#timefeature.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.TimeFeature.__init__': ( 'utils.html#timefeature.__init__',
'neuralforecast/utils.py'),
'neuralforecast.utils.TimeFeature.__repr__': ( 'utils.html#timefeature.__repr__',
'neuralforecast/utils.py'),
'neuralforecast.utils.WeekOfYear': ('utils.html#weekofyear', 'neuralforecast/utils.py'),
'neuralforecast.utils.WeekOfYear.__call__': ( 'utils.html#weekofyear.__call__',
'neuralforecast/utils.py'),
'neuralforecast.utils.augment_calendar_df': ( 'utils.html#augment_calendar_df',
'neuralforecast/utils.py'),
'neuralforecast.utils.generate_series': ('utils.html#generate_series', 'neuralforecast/utils.py'),
'neuralforecast.utils.get_indexer_raise_missing': ( 'utils.html#get_indexer_raise_missing',
'neuralforecast/utils.py'),
'neuralforecast.utils.time_features_from_frequency_str': ( 'utils.html#time_features_from_frequency_str',
'neuralforecast/utils.py')}}}
neuralforecast/auto.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/models.ipynb.
# %% auto 0
__all__ = ['AutoRNN', 'AutoLSTM', 'AutoGRU', 'AutoTCN', 'AutoDeepAR', 'AutoDilatedRNN', 'AutoBiTCN', 'AutoMLP', 'AutoNBEATS',
'AutoNBEATSx', 'AutoNHITS', 'AutoDLinear', 'AutoNLinear', 'AutoTFT', 'AutoVanillaTransformer',
'AutoInformer', 'AutoAutoformer', 'AutoFEDformer', 'AutoPatchTST', 'AutoiTransformer', 'AutoTimesNet',
'AutoStemGNN', 'AutoHINT', 'AutoTSMixer', 'AutoTSMixerx', 'AutoMLPMultivariate']
# %% ../nbs/models.ipynb 2
from os import cpu_count
import torch
from ray import tune
from ray.tune.search.basic_variant import BasicVariantGenerator
from .common._base_auto import BaseAuto
from .common._base_auto import MockTrial
from .models.rnn import RNN
from .models.gru import GRU
from .models.tcn import TCN
from .models.lstm import LSTM
from .models.deepar import DeepAR
from .models.dilated_rnn import DilatedRNN
from .models.bitcn import BiTCN
from .models.mlp import MLP
from .models.nbeats import NBEATS
from .models.nbeatsx import NBEATSx
from .models.nhits import NHITS
from .models.dlinear import DLinear
from .models.nlinear import NLinear
from .models.tft import TFT
from .models.vanillatransformer import VanillaTransformer
from .models.informer import Informer
from .models.autoformer import Autoformer
from .models.fedformer import FEDformer
from .models.patchtst import PatchTST
from .models.timesnet import TimesNet
from .models.itransformer import iTransformer
from .models.stemgnn import StemGNN
from .models.hint import HINT
from .models.tsmixer import TSMixer
from .models.tsmixerx import TSMixerx
from .models.mlpmultivariate import MLPMultivariate
from .losses.pytorch import MAE, MQLoss, DistributionLoss
# %% ../nbs/models.ipynb 13
class AutoRNN(BaseAuto):
default_config = {
"input_size_multiplier": [-1, 4, 16, 64],
"inference_input_size_multiplier": [-1],
"h": None,
"encoder_hidden_size": tune.choice([50, 100, 200, 300]),
"encoder_n_layers": tune.randint(1, 4),
"context_size": tune.choice([5, 10, 50]),
"decoder_hidden_size": tune.choice([64, 128, 256, 512]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([16, 32]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
"""Auto RNN
**Parameters:**<br>
"""
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoRNN, self).__init__(
cls_model=RNN,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["inference_input_size"] = tune.choice(
[h * x for x in config["inference_input_size_multiplier"]]
)
del config["input_size_multiplier"], config["inference_input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 17
class AutoLSTM(BaseAuto):
default_config = {
"input_size_multiplier": [-1, 4, 16, 64],
"inference_input_size_multiplier": [-1],
"h": None,
"encoder_hidden_size": tune.choice([50, 100, 200, 300]),
"encoder_n_layers": tune.randint(1, 4),
"context_size": tune.choice([5, 10, 50]),
"decoder_hidden_size": tune.choice([64, 128, 256, 512]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([16, 32]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoLSTM, self).__init__(
cls_model=LSTM,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["inference_input_size"] = tune.choice(
[h * x for x in config["inference_input_size_multiplier"]]
)
del config["input_size_multiplier"], config["inference_input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 21
class AutoGRU(BaseAuto):
default_config = {
"input_size_multiplier": [-1, 4, 16, 64],
"inference_input_size_multiplier": [-1],
"h": None,
"encoder_hidden_size": tune.choice([50, 100, 200, 300]),
"encoder_n_layers": tune.randint(1, 4),
"context_size": tune.choice([5, 10, 50]),
"decoder_hidden_size": tune.choice([64, 128, 256, 512]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([16, 32]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoGRU, self).__init__(
cls_model=GRU,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["inference_input_size"] = tune.choice(
[h * x for x in config["inference_input_size_multiplier"]]
)
del config["input_size_multiplier"], config["inference_input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 25
class AutoTCN(BaseAuto):
default_config = {
"input_size_multiplier": [-1, 4, 16, 64],
"inference_input_size_multiplier": [-1],
"h": None,
"encoder_hidden_size": tune.choice([50, 100, 200, 300]),
"context_size": tune.choice([5, 10, 50]),
"decoder_hidden_size": tune.choice([64, 128]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([16, 32]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoTCN, self).__init__(
cls_model=TCN,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["inference_input_size"] = tune.choice(
[h * x for x in config["inference_input_size_multiplier"]]
)
del config["input_size_multiplier"], config["inference_input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 29
class AutoDeepAR(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"lstm_hidden_size": tune.choice([32, 64, 128, 256]),
"lstm_n_layers": tune.randint(1, 4),
"lstm_dropout": tune.uniform(0.0, 0.5),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice(["robust", "minmax1"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=DistributionLoss(
distribution="StudentT", level=[80, 90], return_params=False
),
valid_loss=MQLoss(level=[80, 90]),
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoDeepAR, self).__init__(
cls_model=DeepAR,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 33
class AutoDilatedRNN(BaseAuto):
default_config = {
"input_size_multiplier": [-1, 4, 16, 64],
"inference_input_size_multiplier": [-1],
"h": None,
"cell_type": tune.choice(["LSTM", "GRU"]),
"encoder_hidden_size": tune.choice([50, 100, 200, 300]),
"dilations": tune.choice([[[1, 2], [4, 8]], [[1, 2, 4, 8]]]),
"context_size": tune.choice([5, 10, 50]),
"decoder_hidden_size": tune.choice([64, 128, 256, 512]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([16, 32]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoDilatedRNN, self).__init__(
cls_model=DilatedRNN,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["inference_input_size"] = tune.choice(
[h * x for x in config["inference_input_size_multiplier"]]
)
del config["input_size_multiplier"], config["inference_input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 37
class AutoBiTCN(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([16, 32]),
"dropout": tune.uniform(0.0, 0.99),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoBiTCN, self).__init__(
cls_model=BiTCN,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 42
class AutoMLP(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([256, 512, 1024]),
"num_layers": tune.randint(2, 6),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoMLP, self).__init__(
cls_model=MLP,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 46
class AutoNBEATS(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoNBEATS, self).__init__(
cls_model=NBEATS,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 50
class AutoNBEATSx(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoNBEATSx, self).__init__(
cls_model=NBEATSx,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 54
class AutoNHITS(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"n_pool_kernel_size": tune.choice(
[[2, 2, 1], 3 * [1], 3 * [2], 3 * [4], [8, 4, 1], [16, 8, 1]]
),
"n_freq_downsample": tune.choice(
[
[168, 24, 1],
[24, 12, 1],
[180, 60, 1],
[60, 8, 1],
[40, 20, 1],
[1, 1, 1],
]
),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.quniform(lower=500, upper=1500, q=100),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(lower=1, upper=20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoNHITS, self).__init__(
cls_model=NHITS,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 58
class AutoDLinear(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"moving_avg_window": tune.choice([11, 25, 51]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.quniform(lower=500, upper=1500, q=100),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(lower=1, upper=20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoDLinear, self).__init__(
cls_model=DLinear,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 62
class AutoNLinear(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.quniform(lower=500, upper=1500, q=100),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(lower=1, upper=20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoNLinear, self).__init__(
cls_model=NLinear,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 67
class AutoTFT(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([64, 128, 256]),
"n_head": tune.choice([4, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoTFT, self).__init__(
cls_model=TFT,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 71
class AutoVanillaTransformer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([64, 128, 256]),
"n_head": tune.choice([4, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoVanillaTransformer, self).__init__(
cls_model=VanillaTransformer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 75
class AutoInformer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([64, 128, 256]),
"n_head": tune.choice([4, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoInformer, self).__init__(
cls_model=Informer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 79
class AutoAutoformer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([64, 128, 256]),
"n_head": tune.choice([4, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoAutoformer, self).__init__(
cls_model=Autoformer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 83
class AutoFEDformer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([64, 128, 256]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoFEDformer, self).__init__(
cls_model=FEDformer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 87
class AutoPatchTST(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3],
"h": None,
"hidden_size": tune.choice([16, 128, 256]),
"n_heads": tune.choice([4, 16]),
"patch_len": tune.choice([16, 24]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"revin": tune.choice([False, True]),
"max_steps": tune.choice([500, 1000, 5000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"windows_batch_size": tune.choice([128, 256, 512, 1024]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoPatchTST, self).__init__(
cls_model=PatchTST,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 91
class AutoiTransformer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"n_series": None,
"hidden_size": tune.choice([64, 128, 256]),
"n_heads": tune.choice([4, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
n_series,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend, n_series=n_series)
# Always use n_series from parameters, raise exception with Optuna because we can't enforce it
if backend == "ray":
config["n_series"] = n_series
elif backend == "optuna":
mock_trial = MockTrial()
if (
"n_series" in config(mock_trial)
and config(mock_trial)["n_series"] != n_series
) or ("n_series" not in config(mock_trial)):
raise Exception(f"config needs 'n_series': {n_series}")
super(AutoiTransformer, self).__init__(
cls_model=iTransformer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
# Rolling windows with step_size=1 or step_size=h
# See `BaseWindows` and `BaseRNN`'s create_windows
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
# Always use n_series from parameters
config["n_series"] = n_series
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 96
class AutoTimesNet(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"hidden_size": tune.choice([32, 64, 128]),
"conv_hidden_size": tune.choice([32, 64, 128]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice(["robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128]),
"windows_batch_size": tune.choice([32, 64, 128, 256]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend)
super(AutoTimesNet, self).__init__(
cls_model=TimesNet,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series=None):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 101
class AutoStemGNN(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4],
"h": None,
"n_series": None,
"n_stacks": tune.choice([2]),
"multi_layer": tune.choice([3, 5, 7]),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
n_series,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend, n_series=n_series)
# Always use n_series from parameters, raise exception with Optuna because we can't enforce it
if backend == "ray":
config["n_series"] = n_series
elif backend == "optuna":
mock_trial = MockTrial()
if (
"n_series" in config(mock_trial)
and config(mock_trial)["n_series"] != n_series
) or ("n_series" not in config(mock_trial)):
raise Exception(f"config needs 'n_series': {n_series}")
super(AutoStemGNN, self).__init__(
cls_model=StemGNN,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
# Rolling windows with step_size=1 or step_size=h
# See `BaseWindows` and `BaseRNN`'s create_windows
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
# Always use n_series from parameters
config["n_series"] = n_series
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 105
class AutoHINT(BaseAuto):
def __init__(
self,
cls_model,
h,
loss,
valid_loss,
S,
config,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
refit_with_val=False,
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
super(AutoHINT, self).__init__(
cls_model=cls_model,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
if backend == "optuna":
raise Exception("Optuna is not supported for AutoHINT.")
# Validate presence of reconciliation strategy
# parameter in configuration space
if not ("reconciliation" in config.keys()):
raise Exception(
"config needs reconciliation, \
try tune.choice(['BottomUp', 'MinTraceOLS', 'MinTraceWLS'])"
)
self.S = S
def _fit_model(
self, cls_model, config, dataset, val_size, test_size, distributed_config=None
):
# Overwrite _fit_model for HINT two-stage instantiation
reconciliation = config.pop("reconciliation")
base_model = cls_model(**config)
model = HINT(
h=base_model.h, model=base_model, S=self.S, reconciliation=reconciliation
)
model.test_size = test_size
model = model.fit(
dataset,
val_size=val_size,
test_size=test_size,
distributed_config=distributed_config,
)
return model
@classmethod
def get_default_config(cls, h, backend, n_series=None):
raise Exception("AutoHINT has no default configuration.")
# %% ../nbs/models.ipynb 110
class AutoTSMixer(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4],
"h": None,
"n_series": None,
"n_block": tune.choice([1, 2, 4, 6, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-2),
"ff_dim": tune.choice([32, 64, 128]),
"scaler_type": tune.choice(["identity", "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"dropout": tune.uniform(0.0, 0.99),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
n_series,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend, n_series=n_series)
# Always use n_series from parameters, raise exception with Optuna because we can't enforce it
if backend == "ray":
config["n_series"] = n_series
elif backend == "optuna":
mock_trial = MockTrial()
if (
"n_series" in config(mock_trial)
and config(mock_trial)["n_series"] != n_series
) or ("n_series" not in config(mock_trial)):
raise Exception(f"config needs 'n_series': {n_series}")
super(AutoTSMixer, self).__init__(
cls_model=TSMixer,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
# Rolling windows with step_size=1 or step_size=h
# See `BaseWindows` and `BaseRNN`'s create_windows
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
# Always use n_series from parameters
config["n_series"] = n_series
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 114
class AutoTSMixerx(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4],
"h": None,
"n_series": None,
"n_block": tune.choice([1, 2, 4, 6, 8]),
"learning_rate": tune.loguniform(1e-4, 1e-2),
"ff_dim": tune.choice([32, 64, 128]),
"scaler_type": tune.choice(["identity", "robust", "standard"]),
"max_steps": tune.choice([500, 1000, 2000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"dropout": tune.uniform(0.0, 0.99),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
n_series,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend, n_series=n_series)
# Always use n_series from parameters, raise exception with Optuna because we can't enforce it
if backend == "ray":
config["n_series"] = n_series
elif backend == "optuna":
mock_trial = MockTrial()
if (
"n_series" in config(mock_trial)
and config(mock_trial)["n_series"] != n_series
) or ("n_series" not in config(mock_trial)):
raise Exception(f"config needs 'n_series': {n_series}")
super(AutoTSMixerx, self).__init__(
cls_model=TSMixerx,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
# Rolling windows with step_size=1 or step_size=h
# See `BaseWindows` and `BaseRNN`'s create_windows
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
# Always use n_series from parameters
config["n_series"] = n_series
config = cls._ray_config_to_optuna(config)
return config
# %% ../nbs/models.ipynb 118
class AutoMLPMultivariate(BaseAuto):
default_config = {
"input_size_multiplier": [1, 2, 3, 4, 5],
"h": None,
"n_series": None,
"hidden_size": tune.choice([256, 512, 1024]),
"num_layers": tune.randint(2, 6),
"learning_rate": tune.loguniform(1e-4, 1e-1),
"scaler_type": tune.choice([None, "robust", "standard"]),
"max_steps": tune.choice([500, 1000]),
"batch_size": tune.choice([32, 64, 128, 256]),
"loss": None,
"random_seed": tune.randint(1, 20),
}
def __init__(
self,
h,
n_series,
loss=MAE(),
valid_loss=None,
config=None,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
refit_with_val=False,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
# Define search space, input/output sizes
if config is None:
config = self.get_default_config(h=h, backend=backend, n_series=n_series)
# Always use n_series from parameters, raise exception with Optuna because we can't enforce it
if backend == "ray":
config["n_series"] = n_series
elif backend == "optuna":
mock_trial = MockTrial()
if (
"n_series" in config(mock_trial)
and config(mock_trial)["n_series"] != n_series
) or ("n_series" not in config(mock_trial)):
raise Exception(f"config needs 'n_series': {n_series}")
super(AutoMLPMultivariate, self).__init__(
cls_model=MLPMultivariate,
h=h,
loss=loss,
valid_loss=valid_loss,
config=config,
search_alg=search_alg,
num_samples=num_samples,
refit_with_val=refit_with_val,
cpus=cpus,
gpus=gpus,
verbose=verbose,
alias=alias,
backend=backend,
callbacks=callbacks,
)
@classmethod
def get_default_config(cls, h, backend, n_series):
config = cls.default_config.copy()
config["input_size"] = tune.choice(
[h * x for x in config["input_size_multiplier"]]
)
# Rolling windows with step_size=1 or step_size=h
# See `BaseWindows` and `BaseRNN`'s create_windows
config["step_size"] = tune.choice([1, h])
del config["input_size_multiplier"]
if backend == "optuna":
# Always use n_series from parameters
config["n_series"] = n_series
config = cls._ray_config_to_optuna(config)
return config
neuralforecast/common/_base_auto.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.base_auto.ipynb.
# %% auto 0
__all__ = ['BaseAuto']
# %% ../../nbs/common.base_auto.ipynb 5
from copy import deepcopy
from os import cpu_count
import torch
import pytorch_lightning as pl
from ray import air, tune
from ray.tune.integration.pytorch_lightning import TuneReportCallback
from ray.tune.search.basic_variant import BasicVariantGenerator
# %% ../../nbs/common.base_auto.ipynb 6
class MockTrial:
def suggest_int(*args, **kwargs):
return "int"
def suggest_categorical(*args, **kwargs):
return "categorical"
def suggest_uniform(*args, **kwargs):
return "uniform"
def suggest_loguniform(*args, **kwargs):
return "loguniform"
def suggest_float(*args, **kwargs):
if "log" in kwargs:
return "quantized_log"
elif "step" in kwargs:
return "quantized_loguniform"
return "float"
# %% ../../nbs/common.base_auto.ipynb 7
class BaseAuto(pl.LightningModule):
"""
Class for Automatic Hyperparameter Optimization, it builds on top of `ray` to
give access to a wide variety of hyperparameter optimization tools ranging
from classic grid search, to Bayesian optimization and HyperBand algorithm.
The validation loss to be optimized is defined by the `config['loss']` dictionary
value, the config also contains the rest of the hyperparameter search space.
It is important to note that the success of this hyperparameter optimization
heavily relies on a strong correlation between the validation and test periods.
Parameters
----------
cls_model : PyTorch/PyTorchLightning model
See `neuralforecast.models` [collection here](https://nixtla.github.io/neuralforecast/models.html).
h : int
Forecast horizon
loss : PyTorch module
Instantiated train loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).
valid_loss : PyTorch module
Instantiated valid loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).
config : dict or callable
Dictionary with ray.tune defined search space or function that takes an optuna trial and returns a configuration dict.
search_alg : ray.tune.search variant or optuna.sampler
For ray see https://docs.ray.io/en/latest/tune/api_docs/suggestion.html
For optuna see https://optuna.readthedocs.io/en/stable/reference/samplers/index.html.
num_samples : int
Number of hyperparameter optimization steps/samples.
cpus : int (default=os.cpu_count())
Number of cpus to use during optimization. Only used with ray tune.
gpus : int (default=torch.cuda.device_count())
Number of gpus to use during optimization, default all available. Only used with ray tune.
refit_with_val : bool
Refit of best model should preserve val_size.
verbose : bool
Track progress.
alias : str, optional (default=None)
Custom name of the model.
backend : str (default='ray')
Backend to use for searching the hyperparameter space, can be either 'ray' or 'optuna'.
callbacks : list of callable, optional (default=None)
List of functions to call during the optimization process.
ray reference: https://docs.ray.io/en/latest/tune/tutorials/tune-metrics.html
optuna reference: https://optuna.readthedocs.io/en/stable/tutorial/20_recipes/007_optuna_callback.html
"""
def __init__(
self,
cls_model,
h,
loss,
valid_loss,
config,
search_alg=BasicVariantGenerator(random_state=1),
num_samples=10,
cpus=cpu_count(),
gpus=torch.cuda.device_count(),
refit_with_val=False,
verbose=False,
alias=None,
backend="ray",
callbacks=None,
):
super(BaseAuto, self).__init__()
self.save_hyperparameters() # Allows instantiation from a checkpoint from class
if backend == "ray":
if not isinstance(config, dict):
raise ValueError(
"You have to provide a dict as `config` when using `backend='ray'`"
)
config_base = deepcopy(config)
elif backend == "optuna":
if not callable(config):
raise ValueError(
"You have to provide a function that takes a trial and returns a dict as `config` when using `backend='optuna'`"
)
# extract constant values from the config fn for validations
config_base = config(MockTrial())
else:
raise ValueError(
f"Unknown backend {backend}. The supported backends are 'ray' and 'optuna'."
)
if config_base.get("h", None) is not None:
raise Exception("Please use `h` init argument instead of `config['h']`.")
if config_base.get("loss", None) is not None:
raise Exception(
"Please use `loss` init argument instead of `config['loss']`."
)
if config_base.get("valid_loss", None) is not None:
raise Exception(
"Please use `valid_loss` init argument instead of `config['valid_loss']`."
)
# This attribute helps to protect
# model and datasets interactions protections
if "early_stop_patience_steps" in config_base.keys():
self.early_stop_patience_steps = 1
else:
self.early_stop_patience_steps = -1
if callable(config):
# reset config_base here to save params to override in the config fn
config_base = {}
# Add losses to config and protect valid_loss default
config_base["h"] = h
config_base["loss"] = loss
if valid_loss is None:
valid_loss = loss
config_base["valid_loss"] = valid_loss
if isinstance(config, dict):
self.config = config_base
else:
def config_f(trial):
return {**config(trial), **config_base}
self.config = config_f
self.h = h
self.cls_model = cls_model
self.loss = loss
self.valid_loss = valid_loss
self.num_samples = num_samples
self.search_alg = search_alg
self.cpus = cpus
self.gpus = gpus
self.refit_with_val = refit_with_val
self.verbose = verbose
self.alias = alias
self.backend = backend
self.callbacks = callbacks
# Base Class attributes
self.SAMPLING_TYPE = cls_model.SAMPLING_TYPE
def __repr__(self):
return type(self).__name__ if self.alias is None else self.alias
def _train_tune(self, config_step, cls_model, dataset, val_size, test_size):
"""BaseAuto._train_tune
Internal function that instantiates a NF class model, then automatically
explores the validation loss (ptl/val_loss) on which the hyperparameter
exploration is based.
**Parameters:**<br>
`config_step`: Dict, initialization parameters of a NF model.<br>
`cls_model`: NeuralForecast model class, yet to be instantiated.<br>
`dataset`: NeuralForecast dataset, to fit the model.<br>
`val_size`: int, validation size for temporal cross-validation.<br>
`test_size`: int, test size for temporal cross-validation.<br>
"""
metrics = {"loss": "ptl/val_loss", "train_loss": "train_loss"}
callbacks = [TuneReportCallback(metrics, on="validation_end")]
if "callbacks" in config_step.keys():
callbacks.extend(config_step["callbacks"])
config_step = {**config_step, **{"callbacks": callbacks}}
# Protect dtypes from tune samplers
if "batch_size" in config_step.keys():
config_step["batch_size"] = int(config_step["batch_size"])
if "windows_batch_size" in config_step.keys():
config_step["windows_batch_size"] = int(config_step["windows_batch_size"])
# Tune session receives validation signal
# from the specialized PL TuneReportCallback
_ = self._fit_model(
cls_model=cls_model,
config=config_step,
dataset=dataset,
val_size=val_size,
test_size=test_size,
)
def _tune_model(
self,
cls_model,
dataset,
val_size,
test_size,
cpus,
gpus,
verbose,
num_samples,
search_alg,
config,
):
train_fn_with_parameters = tune.with_parameters(
self._train_tune,
cls_model=cls_model,
dataset=dataset,
val_size=val_size,
test_size=test_size,
)
# Device
if gpus > 0:
device_dict = {"gpu": gpus}
else:
device_dict = {"cpu": cpus}
# on Windows, prevent long trial directory names
import platform
trial_dirname_creator = (
(lambda trial: f"{trial.trainable_name}_{trial.trial_id}")
if platform.system() == "Windows"
else None
)
tuner = tune.Tuner(
tune.with_resources(train_fn_with_parameters, device_dict),
run_config=air.RunConfig(callbacks=self.callbacks, verbose=verbose),
tune_config=tune.TuneConfig(
metric="loss",
mode="min",
num_samples=num_samples,
search_alg=search_alg,
trial_dirname_creator=trial_dirname_creator,
),
param_space=config,
)
results = tuner.fit()
return results
@staticmethod
def _ray_config_to_optuna(ray_config):
def optuna_config(trial):
out = {}
for k, v in ray_config.items():
if hasattr(v, "sampler"):
sampler = v.sampler
if isinstance(
sampler, tune.search.sample.Integer.default_sampler_cls
):
v = trial.suggest_int(k, v.lower, v.upper)
elif isinstance(
sampler, tune.search.sample.Categorical.default_sampler_cls
):
v = trial.suggest_categorical(k, v.categories)
elif isinstance(sampler, tune.search.sample.Uniform):
v = trial.suggest_uniform(k, v.lower, v.upper)
elif isinstance(sampler, tune.search.sample.LogUniform):
v = trial.suggest_loguniform(k, v.lower, v.upper)
elif isinstance(sampler, tune.search.sample.Quantized):
if isinstance(
sampler.get_sampler(), tune.search.sample.Float._LogUniform
):
v = trial.suggest_float(k, v.lower, v.upper, log=True)
elif isinstance(
sampler.get_sampler(), tune.search.sample.Float._Uniform
):
v = trial.suggest_float(k, v.lower, v.upper, step=sampler.q)
else:
raise ValueError(f"Couldn't translate {type(v)} to optuna.")
out[k] = v
return out
return optuna_config
def _optuna_tune_model(
self,
cls_model,
dataset,
val_size,
test_size,
verbose,
num_samples,
search_alg,
config,
distributed_config,
):
import optuna
def objective(trial):
user_cfg = config(trial)
cfg = deepcopy(user_cfg)
model = self._fit_model(
cls_model=cls_model,
config=cfg,
dataset=dataset,
val_size=val_size,
test_size=test_size,
distributed_config=distributed_config,
)
trial.set_user_attr("ALL_PARAMS", user_cfg)
metrics = model.metrics
trial.set_user_attr(
"METRICS",
{
"loss": metrics["ptl/val_loss"],
"train_loss": metrics["train_loss"],
},
)
return trial.user_attrs["METRICS"]["loss"]
if isinstance(search_alg, optuna.samplers.BaseSampler):
sampler = search_alg
else:
sampler = None
study = optuna.create_study(sampler=sampler, direction="minimize")
study.optimize(
objective,
n_trials=num_samples,
show_progress_bar=verbose,
callbacks=self.callbacks,
)
return study
def _fit_model(
self, cls_model, config, dataset, val_size, test_size, distributed_config=None
):
model = cls_model(**config)
model = model.fit(
dataset,
val_size=val_size,
test_size=test_size,
distributed_config=distributed_config,
)
return model
def fit(
self,
dataset,
val_size=0,
test_size=0,
random_seed=None,
distributed_config=None,
):
"""BaseAuto.fit
Perform the hyperparameter optimization as specified by the BaseAuto configuration
dictionary `config`.
The optimization is performed on the `TimeSeriesDataset` using temporal cross validation with
the validation set that sequentially precedes the test set.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset` see details [here](https://nixtla.github.io/neuralforecast/tsdataset.html)<br>
`val_size`: int, size of temporal validation set (needs to be bigger than 0).<br>
`test_size`: int, size of temporal test set (default 0).<br>
`random_seed`: int=None, random_seed for hyperparameter exploration algorithms, not yet implemented.<br>
**Returns:**<br>
`self`: fitted instance of `BaseAuto` with best hyperparameters and results<br>.
"""
# we need val_size > 0 to perform
# hyperparameter selection.
search_alg = deepcopy(self.search_alg)
val_size = val_size if val_size > 0 else self.h
if self.backend == "ray":
if distributed_config is not None:
raise ValueError(
"distributed training is not supported for the ray backend."
)
results = self._tune_model(
cls_model=self.cls_model,
dataset=dataset,
val_size=val_size,
test_size=test_size,
cpus=self.cpus,
gpus=self.gpus,
verbose=self.verbose,
num_samples=self.num_samples,
search_alg=search_alg,
config=self.config,
)
best_config = results.get_best_result().config
else:
results = self._optuna_tune_model(
cls_model=self.cls_model,
dataset=dataset,
val_size=val_size,
test_size=test_size,
verbose=self.verbose,
num_samples=self.num_samples,
search_alg=search_alg,
config=self.config,
distributed_config=distributed_config,
)
best_config = results.best_trial.user_attrs["ALL_PARAMS"]
self.model = self._fit_model(
cls_model=self.cls_model,
config=best_config,
dataset=dataset,
val_size=val_size * (1 - self.refit_with_val),
test_size=test_size,
distributed_config=distributed_config,
)
self.results = results
# Added attributes for compatibility with NeuralForecast core
self.futr_exog_list = self.model.futr_exog_list
self.hist_exog_list = self.model.hist_exog_list
self.stat_exog_list = self.model.stat_exog_list
return self
def predict(self, dataset, step_size=1, **data_kwargs):
"""BaseAuto.predict
Predictions of the best performing model on validation.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset` see details [here](https://nixtla.github.io/neuralforecast/tsdataset.html)<br>
`step_size`: int, steps between sequential predictions, (default 1).<br>
`**data_kwarg`: additional parameters for the dataset module.<br>
`random_seed`: int=None, random_seed for hyperparameter exploration algorithms (not implemented).<br>
**Returns:**<br>
`y_hat`: numpy predictions of the `NeuralForecast` model.<br>
"""
return self.model.predict(dataset=dataset, step_size=step_size, **data_kwargs)
def set_test_size(self, test_size):
self.model.set_test_size(test_size)
def get_test_size(self):
return self.model.test_size
def save(self, path):
"""BaseAuto.save
Save the fitted model to disk.
**Parameters:**<br>
`path`: str, path to save the model.<br>
"""
self.model.save(path)
neuralforecast/common/_base_model.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.base_model.ipynb.
# %% auto 0
__all__ = ['DistributedConfig', 'BaseModel']
# %% ../../nbs/common.base_model.ipynb 2
import inspect
import random
import warnings
from contextlib import contextmanager
from copy import deepcopy
from dataclasses import dataclass
import fsspec
import numpy as np
import torch
import torch.nn as nn
import pytorch_lightning as pl
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from neuralforecast.tsdataset import (
TimeSeriesDataModule,
TimeSeriesDataset,
_DistributedTimeSeriesDataModule,
)
# %% ../../nbs/common.base_model.ipynb 3
@dataclass
class DistributedConfig:
partitions_path: str
num_nodes: int
devices: int
# %% ../../nbs/common.base_model.ipynb 4
@contextmanager
def _disable_torch_init():
"""Context manager used to disable pytorch's weight initialization.
This is especially useful when loading saved models, since when initializing
a model the weights are also initialized following some method
(e.g. kaiming uniform), and that time is wasted since we'll override them with
the saved weights."""
def noop(*args, **kwargs):
return
kaiming_uniform = nn.init.kaiming_uniform_
kaiming_normal = nn.init.kaiming_normal_
xavier_uniform = nn.init.xavier_uniform_
xavier_normal = nn.init.xavier_normal_
nn.init.kaiming_uniform_ = noop
nn.init.kaiming_normal_ = noop
nn.init.xavier_uniform_ = noop
nn.init.xavier_normal_ = noop
try:
yield
finally:
nn.init.kaiming_uniform_ = kaiming_uniform
nn.init.kaiming_normal_ = kaiming_normal
nn.init.xavier_uniform_ = xavier_uniform
nn.init.xavier_normal_ = xavier_normal
# %% ../../nbs/common.base_model.ipynb 5
class BaseModel(pl.LightningModule):
def __init__(
self,
random_seed,
loss,
valid_loss,
optimizer,
optimizer_kwargs,
futr_exog_list,
hist_exog_list,
stat_exog_list,
max_steps,
early_stop_patience_steps,
**trainer_kwargs,
):
super().__init__()
with warnings.catch_warnings(record=False):
warnings.filterwarnings("ignore")
# the following line issues a warning about the loss attribute being saved
# but we do want to save it
self.save_hyperparameters() # Allows instantiation from a checkpoint from class
self.random_seed = random_seed
pl.seed_everything(self.random_seed, workers=True)
# Loss
self.loss = loss
if valid_loss is None:
self.valid_loss = loss
else:
self.valid_loss = valid_loss
self.train_trajectories = []
self.valid_trajectories = []
# Optimization
if optimizer is not None and not issubclass(optimizer, torch.optim.Optimizer):
raise TypeError(
"optimizer is not a valid subclass of torch.optim.Optimizer"
)
self.optimizer = optimizer
self.optimizer_kwargs = optimizer_kwargs if optimizer_kwargs else {}
# Variables
self.futr_exog_list = list(futr_exog_list) if futr_exog_list is not None else []
self.hist_exog_list = list(hist_exog_list) if hist_exog_list is not None else []
self.stat_exog_list = list(stat_exog_list) if stat_exog_list is not None else []
## Trainer arguments ##
# Max steps, validation steps and check_val_every_n_epoch
trainer_kwargs = {**trainer_kwargs, "max_steps": max_steps}
if "max_epochs" in trainer_kwargs.keys():
raise Exception("max_epochs is deprecated, use max_steps instead.")
# Callbacks
if early_stop_patience_steps > 0:
if "callbacks" not in trainer_kwargs:
trainer_kwargs["callbacks"] = []
trainer_kwargs["callbacks"].append(
EarlyStopping(
monitor="ptl/val_loss", patience=early_stop_patience_steps
)
)
# Add GPU accelerator if available
if trainer_kwargs.get("accelerator", None) is None:
if torch.cuda.is_available():
trainer_kwargs["accelerator"] = "gpu"
if trainer_kwargs.get("devices", None) is None:
if torch.cuda.is_available():
trainer_kwargs["devices"] = -1
# Avoid saturating local memory, disabled fit model checkpoints
if trainer_kwargs.get("enable_checkpointing", None) is None:
trainer_kwargs["enable_checkpointing"] = False
self.trainer_kwargs = trainer_kwargs
def __repr__(self):
return type(self).__name__ if self.alias is None else self.alias
def _check_exog(self, dataset):
temporal_cols = set(dataset.temporal_cols.tolist())
static_cols = set(
dataset.static_cols.tolist() if dataset.static_cols is not None else []
)
missing_hist = set(self.hist_exog_list) - temporal_cols
missing_futr = set(self.futr_exog_list) - temporal_cols
missing_stat = set(self.stat_exog_list) - static_cols
if missing_hist:
raise Exception(
f"{missing_hist} historical exogenous variables not found in input dataset"
)
if missing_futr:
raise Exception(
f"{missing_futr} future exogenous variables not found in input dataset"
)
if missing_stat:
raise Exception(
f"{missing_stat} static exogenous variables not found in input dataset"
)
def _restart_seed(self, random_seed):
if random_seed is None:
random_seed = self.random_seed
torch.manual_seed(random_seed)
def _get_temporal_exogenous_cols(self, temporal_cols):
return list(
set(temporal_cols.tolist()) & set(self.hist_exog_list + self.futr_exog_list)
)
def _fit(
self,
dataset,
batch_size,
valid_batch_size=1024,
val_size=0,
test_size=0,
random_seed=None,
shuffle_train=True,
distributed_config=None,
):
self._check_exog(dataset)
self._restart_seed(random_seed)
self.val_size = val_size
self.test_size = test_size
is_local = isinstance(dataset, TimeSeriesDataset)
if is_local:
datamodule_constructor = TimeSeriesDataModule
else:
datamodule_constructor = _DistributedTimeSeriesDataModule
datamodule = datamodule_constructor(
dataset=dataset,
batch_size=batch_size,
valid_batch_size=valid_batch_size,
num_workers=self.num_workers_loader,
drop_last=self.drop_last_loader,
shuffle_train=shuffle_train,
)
if self.val_check_steps > self.max_steps:
warnings.warn(
"val_check_steps is greater than max_steps, "
"setting val_check_steps to max_steps."
)
val_check_interval = min(self.val_check_steps, self.max_steps)
self.trainer_kwargs["val_check_interval"] = int(val_check_interval)
self.trainer_kwargs["check_val_every_n_epoch"] = None
if is_local:
model = self
trainer = pl.Trainer(**model.trainer_kwargs)
trainer.fit(model, datamodule=datamodule)
model.metrics = trainer.callback_metrics
model.__dict__.pop("_trainer", None)
else:
assert distributed_config is not None
from pyspark.ml.torch.distributor import TorchDistributor
def train_fn(
model_cls,
model_params,
datamodule,
trainer_kwargs,
num_tasks,
num_proc_per_task,
val_size,
test_size,
):
import pytorch_lightning as pl
# we instantiate here to avoid pickling large tensors (weights)
model = model_cls(**model_params)
model.val_size = val_size
model.test_size = test_size
for arg in ("devices", "num_nodes"):
trainer_kwargs.pop(arg, None)
trainer = pl.Trainer(
strategy="ddp",
use_distributed_sampler=False, # to ensure our dataloaders are used as-is
num_nodes=num_tasks,
devices=num_proc_per_task,
**trainer_kwargs,
)
trainer.fit(model=model, datamodule=datamodule)
model.metrics = trainer.callback_metrics
model.__dict__.pop("_trainer", None)
return model
def is_gpu_accelerator(accelerator):
from pytorch_lightning.accelerators.cuda import CUDAAccelerator
return (
accelerator == "gpu"
or isinstance(accelerator, CUDAAccelerator)
or (accelerator == "auto" and CUDAAccelerator.is_available())
)
local_mode = distributed_config.num_nodes == 1
if local_mode:
num_tasks = 1
num_proc_per_task = distributed_config.devices
else:
num_tasks = distributed_config.devices * distributed_config.devices
num_proc_per_task = 1 # number of GPUs per task
num_proc = num_tasks * num_proc_per_task
use_gpu = is_gpu_accelerator(self.trainer_kwargs["accelerator"])
model = TorchDistributor(
num_processes=num_proc,
local_mode=local_mode,
use_gpu=use_gpu,
).run(
train_fn,
model_cls=type(self),
model_params=self.hparams,
datamodule=datamodule,
trainer_kwargs=self.trainer_kwargs,
num_tasks=num_tasks,
num_proc_per_task=num_proc_per_task,
val_size=val_size,
test_size=test_size,
)
return model
def on_fit_start(self):
torch.manual_seed(self.random_seed)
np.random.seed(self.random_seed)
random.seed(self.random_seed)
def configure_optimizers(self):
if self.optimizer:
optimizer_signature = inspect.signature(self.optimizer)
optimizer_kwargs = deepcopy(self.optimizer_kwargs)
if "lr" in optimizer_signature.parameters:
if "lr" in optimizer_kwargs:
warnings.warn(
"ignoring learning rate passed in optimizer_kwargs, using the model's learning rate"
)
optimizer_kwargs["lr"] = self.learning_rate
optimizer = self.optimizer(params=self.parameters(), **optimizer_kwargs)
else:
optimizer = torch.optim.Adam(self.parameters(), lr=self.learning_rate)
scheduler = {
"scheduler": torch.optim.lr_scheduler.StepLR(
optimizer=optimizer, step_size=self.lr_decay_steps, gamma=0.5
),
"frequency": 1,
"interval": "step",
}
return {"optimizer": optimizer, "lr_scheduler": scheduler}
def get_test_size(self):
return self.test_size
def set_test_size(self, test_size):
self.test_size = test_size
def on_validation_epoch_end(self):
if self.val_size == 0:
return
losses = torch.stack(self.validation_step_outputs)
avg_loss = losses.mean().item()
self.log(
"ptl/val_loss",
avg_loss,
batch_size=losses.size(0),
sync_dist=True,
)
self.valid_trajectories.append((self.global_step, avg_loss))
self.validation_step_outputs.clear() # free memory (compute `avg_loss` per epoch)
def save(self, path):
with fsspec.open(path, "wb") as f:
torch.save(
{"hyper_parameters": self.hparams, "state_dict": self.state_dict()},
f,
)
@classmethod
def load(cls, path, **kwargs):
with fsspec.open(path, "rb") as f:
content = torch.load(f, **kwargs)
with _disable_torch_init():
model = cls(**content["hyper_parameters"])
model.load_state_dict(content["state_dict"], strict=True, assign=True)
return model
neuralforecast/common/_base_multivariate.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.base_multivariate.ipynb.
# %% auto 0
__all__ = ['BaseMultivariate']
# %% ../../nbs/common.base_multivariate.ipynb 5
import numpy as np
import torch
import torch.nn as nn
import pytorch_lightning as pl
import neuralforecast.losses.pytorch as losses
from ._base_model import BaseModel
from ._scalers import TemporalNorm
from ..tsdataset import TimeSeriesDataModule
from ..utils import get_indexer_raise_missing
# %% ../../nbs/common.base_multivariate.ipynb 6
class BaseMultivariate(BaseModel):
"""Base Multivariate
Base class for all multivariate models. The forecasts for all time-series are produced simultaneously
within each window, which are randomly sampled during training.
This class implements the basic functionality for all windows-based models, including:
- PyTorch Lightning's methods training_step, validation_step, predict_step.<br>
- fit and predict methods used by NeuralForecast.core class.<br>
- sampling and wrangling methods to generate multivariate windows.
"""
def __init__(
self,
h,
input_size,
loss,
valid_loss,
learning_rate,
max_steps,
val_check_steps,
n_series,
batch_size,
step_size=1,
num_lr_decays=0,
early_stop_patience_steps=-1,
scaler_type="robust",
futr_exog_list=None,
hist_exog_list=None,
stat_exog_list=None,
num_workers_loader=0,
drop_last_loader=False,
random_seed=1,
alias=None,
optimizer=None,
optimizer_kwargs=None,
**trainer_kwargs,
):
super().__init__(
random_seed=random_seed,
loss=loss,
valid_loss=valid_loss,
optimizer=optimizer,
optimizer_kwargs=optimizer_kwargs,
futr_exog_list=futr_exog_list,
hist_exog_list=hist_exog_list,
stat_exog_list=stat_exog_list,
max_steps=max_steps,
early_stop_patience_steps=early_stop_patience_steps,
**trainer_kwargs,
)
# Padder to complete train windows,
# example y=[1,2,3,4,5] h=3 -> last y_output = [5,0,0]
self.h = h
self.input_size = input_size
self.n_series = n_series
self.padder = nn.ConstantPad1d(padding=(0, self.h), value=0)
# Multivariate models do not support these loss functions yet.
unsupported_losses = (
losses.sCRPS,
losses.MQLoss,
losses.DistributionLoss,
losses.PMM,
losses.GMM,
losses.HuberMQLoss,
losses.MASE,
losses.relMSE,
losses.NBMM,
)
if isinstance(self.loss, unsupported_losses):
raise Exception(f"{self.loss} is not supported in a Multivariate model.")
if isinstance(self.valid_loss, unsupported_losses):
raise Exception(
f"{self.valid_loss} is not supported in a Multivariate model."
)
self.batch_size = batch_size
# Optimization
self.learning_rate = learning_rate
self.max_steps = max_steps
self.num_lr_decays = num_lr_decays
self.lr_decay_steps = (
max(max_steps // self.num_lr_decays, 1) if self.num_lr_decays > 0 else 10e7
)
self.early_stop_patience_steps = early_stop_patience_steps
self.val_check_steps = val_check_steps
self.step_size = step_size
# Scaler
self.scaler = TemporalNorm(
scaler_type=scaler_type, dim=2
) # Time dimension is in the second axis
# Fit arguments
self.val_size = 0
self.test_size = 0
# Model state
self.decompose_forecast = False
# DataModule arguments
self.num_workers_loader = num_workers_loader
self.drop_last_loader = drop_last_loader
# used by on_validation_epoch_end hook
self.validation_step_outputs = []
self.alias = alias
def _create_windows(self, batch, step):
# Parse common data
window_size = self.input_size + self.h
temporal_cols = batch["temporal_cols"]
temporal = batch["temporal"]
if step == "train":
if self.val_size + self.test_size > 0:
cutoff = -self.val_size - self.test_size
temporal = temporal[:, :, :cutoff]
temporal = self.padder(temporal)
windows = temporal.unfold(
dimension=-1, size=window_size, step=self.step_size
)
# [n_series, C, Ws, L+H] 0, 1, 2, 3
# Sample and Available conditions
available_idx = temporal_cols.get_loc("available_mask")
sample_condition = windows[:, available_idx, :, -self.h :]
sample_condition = torch.sum(sample_condition, axis=2) # Sum over time
sample_condition = torch.sum(
sample_condition, axis=0
) # Sum over time-series
available_condition = windows[:, available_idx, :, : -self.h]
available_condition = torch.sum(
available_condition, axis=2
) # Sum over time
available_condition = torch.sum(
available_condition, axis=0
) # Sum over time-series
final_condition = (sample_condition > 0) & (
available_condition > 0
) # Of shape [Ws]
windows = windows[:, :, final_condition, :]
# Get Static data
static = batch.get("static", None)
static_cols = batch.get("static_cols", None)
# Protection of empty windows
if final_condition.sum() == 0:
raise Exception("No windows available for training")
# Sample windows
n_windows = windows.shape[2]
if self.batch_size is not None:
w_idxs = np.random.choice(
n_windows,
size=self.batch_size,
replace=(n_windows < self.batch_size),
)
windows = windows[:, :, w_idxs, :]
windows = windows.permute(2, 1, 3, 0) # [Ws, C, L+H, n_series]
windows_batch = dict(
temporal=windows,
temporal_cols=temporal_cols,
static=static,
static_cols=static_cols,
)
return windows_batch
elif step in ["predict", "val"]:
if step == "predict":
predict_step_size = self.predict_step_size
cutoff = -self.input_size - self.test_size
temporal = batch["temporal"][:, :, cutoff:]
elif step == "val":
predict_step_size = self.step_size
cutoff = -self.input_size - self.val_size - self.test_size
if self.test_size > 0:
temporal = batch["temporal"][:, :, cutoff : -self.test_size]
else:
temporal = batch["temporal"][:, :, cutoff:]
if (
(step == "predict")
and (self.test_size == 0)
and (len(self.futr_exog_list) == 0)
):
temporal = self.padder(temporal)
windows = temporal.unfold(
dimension=-1, size=window_size, step=predict_step_size
)
# [n_series, C, Ws, L+H] -> [Ws, C, L+H, n_series]
windows = windows.permute(2, 1, 3, 0)
# Get Static data
static = batch.get("static", None)
static_cols = batch.get("static_cols", None)
windows_batch = dict(
temporal=windows,
temporal_cols=temporal_cols,
static=static,
static_cols=static_cols,
)
return windows_batch
else:
raise ValueError(f"Unknown step {step}")
def _normalization(self, windows, y_idx):
# windows are already filtered by train/validation/test
# from the `create_windows_method` nor leakage risk
temporal = windows["temporal"] # [Ws, C, L+H, n_series]
temporal_cols = windows["temporal_cols"].copy() # [Ws, C, L+H, n_series]
# To avoid leakage uses only the lags
temporal_data_cols = self._get_temporal_exogenous_cols(
temporal_cols=temporal_cols
)
temporal_idxs = get_indexer_raise_missing(temporal_cols, temporal_data_cols)
temporal_idxs = np.append(y_idx, temporal_idxs)
temporal_data = temporal[:, temporal_idxs, :, :]
temporal_mask = temporal[
:, temporal_cols.get_loc("available_mask"), :, :
].clone()
temporal_mask[:, -self.h :, :] = 0.0
# Normalize. self.scaler stores the shift and scale for inverse transform
temporal_mask = temporal_mask.unsqueeze(
1
) # Add channel dimension for scaler.transform.
temporal_data = self.scaler.transform(x=temporal_data, mask=temporal_mask)
# Replace values in windows dict
temporal[:, temporal_idxs, :, :] = temporal_data
windows["temporal"] = temporal
return windows
def _inv_normalization(self, y_hat, temporal_cols, y_idx):
# Receives window predictions [Ws, H, n_series]
# Broadcasts outputs and inverts normalization
# Add C dimension
# if y_hat.ndim == 2:
# remove_dimension = True
# y_hat = y_hat.unsqueeze(-1)
# else:
# remove_dimension = False
y_scale = self.scaler.x_scale[:, [y_idx], :].squeeze(1)
y_loc = self.scaler.x_shift[:, [y_idx], :].squeeze(1)
# y_scale = torch.repeat_interleave(y_scale, repeats=y_hat.shape[-1], dim=-1)
# y_loc = torch.repeat_interleave(y_loc, repeats=y_hat.shape[-1], dim=-1)
y_hat = self.scaler.inverse_transform(z=y_hat, x_scale=y_scale, x_shift=y_loc)
# if remove_dimension:
# y_hat = y_hat.squeeze(-1)
# y_loc = y_loc.squeeze(-1)
# y_scale = y_scale.squeeze(-1)
return y_hat, y_loc, y_scale
def _parse_windows(self, batch, windows):
# Temporal: [Ws, C, L+H, n_series]
# Filter insample lags from outsample horizon
mask_idx = batch["temporal_cols"].get_loc("available_mask")
y_idx = batch["y_idx"]
insample_y = windows["temporal"][:, y_idx, : -self.h, :]
insample_mask = windows["temporal"][:, mask_idx, : -self.h, :]
outsample_y = windows["temporal"][:, y_idx, -self.h :, :]
outsample_mask = windows["temporal"][:, mask_idx, -self.h :, :]
# Filter historic exogenous variables
if len(self.hist_exog_list):
hist_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.hist_exog_list
)
hist_exog = windows["temporal"][:, hist_exog_idx, : -self.h, :]
else:
hist_exog = None
# Filter future exogenous variables
if len(self.futr_exog_list):
futr_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.futr_exog_list
)
futr_exog = windows["temporal"][:, futr_exog_idx, :, :]
else:
futr_exog = None
# Filter static variables
if len(self.stat_exog_list):
static_idx = get_indexer_raise_missing(
windows["static_cols"], self.stat_exog_list
)
stat_exog = windows["static"][:, static_idx]
else:
stat_exog = None
return (
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
)
def training_step(self, batch, batch_idx):
# Create and normalize windows [batch_size, n_series, C, L+H]
windows = self._create_windows(batch, step="train")
y_idx = batch["y_idx"]
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
(
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [batch_size, L, n_series]
insample_mask=insample_mask, # [batch_size, L, n_series]
futr_exog=futr_exog, # [batch_size, n_feats, L+H, n_series]
hist_exog=hist_exog, # [batch_size, n_feats, L, n_series]
stat_exog=stat_exog,
) # [n_series, n_feats]
# Model Predictions
output = self(windows_batch)
if self.loss.is_distribution_output:
outsample_y, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
loss = self.loss(y=outsample_y, distr_args=distr_args, mask=outsample_mask)
else:
loss = self.loss(y=outsample_y, y_hat=output, mask=outsample_mask)
if torch.isnan(loss):
print("Model Parameters", self.hparams)
print("insample_y", torch.isnan(insample_y).sum())
print("outsample_y", torch.isnan(outsample_y).sum())
print("output", torch.isnan(output).sum())
raise Exception("Loss is NaN, training stopped.")
self.log(
"train_loss",
loss.item(),
batch_size=outsample_y.size(0),
prog_bar=True,
on_epoch=True,
)
self.train_trajectories.append((self.global_step, loss.item()))
return loss
def validation_step(self, batch, batch_idx):
if self.val_size == 0:
return np.nan
# Create and normalize windows [Ws, L+H, C]
windows = self._create_windows(batch, step="val")
y_idx = batch["y_idx"]
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
(
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [Ws, L]
insample_mask=insample_mask, # [Ws, L]
futr_exog=futr_exog, # [Ws, L+H]
hist_exog=hist_exog, # [Ws, L]
stat_exog=stat_exog,
) # [Ws, 1]
# Model Predictions
output = self(windows_batch)
if self.loss.is_distribution_output:
outsample_y, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
if str(type(self.valid_loss)) in [
"<class 'neuralforecast.losses.pytorch.sCRPS'>",
"<class 'neuralforecast.losses.pytorch.MQLoss'>",
]:
_, output = self.loss.sample(distr_args=distr_args)
# Validation Loss evaluation
if self.valid_loss.is_distribution_output:
valid_loss = self.valid_loss(
y=outsample_y, distr_args=distr_args, mask=outsample_mask
)
else:
valid_loss = self.valid_loss(
y=outsample_y, y_hat=output, mask=outsample_mask
)
if torch.isnan(valid_loss):
raise Exception("Loss is NaN, training stopped.")
self.log(
"valid_loss",
valid_loss.item(),
batch_size=outsample_y.size(0),
prog_bar=True,
on_epoch=True,
)
self.validation_step_outputs.append(valid_loss)
return valid_loss
def predict_step(self, batch, batch_idx):
# Create and normalize windows [Ws, L+H, C]
windows = self._create_windows(batch, step="predict")
y_idx = batch["y_idx"]
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
insample_y, insample_mask, _, _, hist_exog, futr_exog, stat_exog = (
self._parse_windows(batch, windows)
)
windows_batch = dict(
insample_y=insample_y, # [Ws, L]
insample_mask=insample_mask, # [Ws, L]
futr_exog=futr_exog, # [Ws, L+H]
hist_exog=hist_exog, # [Ws, L]
stat_exog=stat_exog,
) # [Ws, 1]
# Model Predictions
output = self(windows_batch)
if self.loss.is_distribution_output:
_, y_loc, y_scale = self._inv_normalization(
y_hat=output[0], temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
_, y_hat = self.loss.sample(distr_args=distr_args)
if self.loss.return_params:
distr_args = torch.stack(distr_args, dim=-1)
distr_args = torch.reshape(
distr_args, (len(windows["temporal"]), self.h, -1)
)
y_hat = torch.concat((y_hat, distr_args), axis=2)
else:
y_hat, _, _ = self._inv_normalization(
y_hat=output, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
return y_hat
def fit(
self,
dataset,
val_size=0,
test_size=0,
random_seed=None,
distributed_config=None,
):
"""Fit.
The `fit` method, optimizes the neural network's weights using the
initialization parameters (`learning_rate`, `windows_batch_size`, ...)
and the `loss` function as defined during the initialization.
Within `fit` we use a PyTorch Lightning `Trainer` that
inherits the initialization's `self.trainer_kwargs`, to customize
its inputs, see [PL's trainer arguments](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.trainer.trainer.Trainer.html?highlight=trainer).
The method is designed to be compatible with SKLearn-like classes
and in particular to be compatible with the StatsForecast library.
By default the `model` is not saving training checkpoints to protect
disk memory, to get them change `enable_checkpointing=True` in `__init__`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`val_size`: int, validation size for temporal cross-validation.<br>
`test_size`: int, test size for temporal cross-validation.<br>
"""
if distributed_config is not None:
raise ValueError(
"multivariate models cannot be trained using distributed data parallel."
)
return self._fit(
dataset=dataset,
batch_size=self.n_series,
valid_batch_size=self.n_series,
val_size=val_size,
test_size=test_size,
random_seed=random_seed,
shuffle_train=False,
distributed_config=None,
)
def predict(
self,
dataset,
test_size=None,
step_size=1,
random_seed=None,
**data_module_kwargs,
):
"""Predict.
Neural network prediction with PL's `Trainer` execution of `predict_step`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`test_size`: int=None, test size for temporal cross-validation.<br>
`step_size`: int=1, Step size between each window.<br>
`**data_module_kwargs`: PL's TimeSeriesDataModule args, see [documentation](https://pytorch-lightning.readthedocs.io/en/1.6.1/extensions/datamodules.html#using-a-datamodule).
"""
self._check_exog(dataset)
self._restart_seed(random_seed)
self.predict_step_size = step_size
self.decompose_forecast = False
datamodule = TimeSeriesDataModule(
dataset=dataset,
valid_batch_size=self.n_series,
batch_size=self.n_series,
**data_module_kwargs,
)
# Protect when case of multiple gpu. PL does not support return preds with multiple gpu.
pred_trainer_kwargs = self.trainer_kwargs.copy()
if (pred_trainer_kwargs.get("accelerator", None) == "gpu") and (
torch.cuda.device_count() > 1
):
pred_trainer_kwargs["devices"] = [0]
trainer = pl.Trainer(**pred_trainer_kwargs)
fcsts = trainer.predict(self, datamodule=datamodule)
fcsts = torch.vstack(fcsts).numpy()
fcsts = np.transpose(fcsts, (2, 0, 1))
fcsts = fcsts.flatten()
fcsts = fcsts.reshape(-1, len(self.loss.output_names))
return fcsts
def decompose(self, dataset, step_size=1, random_seed=None, **data_module_kwargs):
raise NotImplementedError("decompose")
neuralforecast/common/_base_recurrent.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.base_recurrent.ipynb.
# %% auto 0
__all__ = ['BaseRecurrent']
# %% ../../nbs/common.base_recurrent.ipynb 6
import numpy as np
import torch
import torch.nn as nn
import pytorch_lightning as pl
from ._base_model import BaseModel
from ._scalers import TemporalNorm
from ..tsdataset import TimeSeriesDataModule
from ..utils import get_indexer_raise_missing
# %% ../../nbs/common.base_recurrent.ipynb 7
class BaseRecurrent(BaseModel):
"""Base Recurrent
Base class for all recurrent-based models. The forecasts are produced sequentially between
windows.
This class implements the basic functionality for all windows-based models, including:
- PyTorch Lightning's methods training_step, validation_step, predict_step. <br>
- fit and predict methods used by NeuralForecast.core class. <br>
- sampling and wrangling methods to sequential windows. <br>
"""
def __init__(
self,
h,
input_size,
inference_input_size,
loss,
valid_loss,
learning_rate,
max_steps,
val_check_steps,
batch_size,
valid_batch_size,
scaler_type="robust",
num_lr_decays=0,
early_stop_patience_steps=-1,
futr_exog_list=None,
hist_exog_list=None,
stat_exog_list=None,
num_workers_loader=0,
drop_last_loader=False,
random_seed=1,
alias=None,
optimizer=None,
optimizer_kwargs=None,
**trainer_kwargs,
):
super().__init__(
random_seed=random_seed,
loss=loss,
valid_loss=valid_loss,
optimizer=optimizer,
optimizer_kwargs=optimizer_kwargs,
futr_exog_list=futr_exog_list,
hist_exog_list=hist_exog_list,
stat_exog_list=stat_exog_list,
max_steps=max_steps,
early_stop_patience_steps=early_stop_patience_steps,
**trainer_kwargs,
)
# Padder to complete train windows,
# example y=[1,2,3,4,5] h=3 -> last y_output = [5,0,0]
self.h = h
self.input_size = input_size
self.inference_input_size = inference_input_size
self.padder = nn.ConstantPad1d(padding=(0, self.h), value=0)
if (
str(type(self.loss))
== "<class 'neuralforecast.losses.pytorch.DistributionLoss'>"
and self.loss.distribution == "Bernoulli"
):
raise Exception(
"Temporal Classification not yet available for Recurrent-based models"
)
# Valid batch_size
self.batch_size = batch_size
if valid_batch_size is None:
self.valid_batch_size = batch_size
else:
self.valid_batch_size = valid_batch_size
# Optimization
self.learning_rate = learning_rate
self.max_steps = max_steps
self.num_lr_decays = num_lr_decays
self.lr_decay_steps = (
max(max_steps // self.num_lr_decays, 1) if self.num_lr_decays > 0 else 10e7
)
self.early_stop_patience_steps = early_stop_patience_steps
self.val_check_steps = val_check_steps
# Scaler
self.scaler = TemporalNorm(
scaler_type=scaler_type,
dim=-1, # Time dimension is -1.
num_features=1 + len(self.hist_exog_list) + len(self.futr_exog_list),
)
# Fit arguments
self.val_size = 0
self.test_size = 0
# DataModule arguments
self.num_workers_loader = num_workers_loader
self.drop_last_loader = drop_last_loader
# used by on_validation_epoch_end hook
self.validation_step_outputs = []
self.alias = alias
def _normalization(self, batch, val_size=0, test_size=0):
temporal = batch["temporal"] # B, C, T
temporal_cols = batch["temporal_cols"].copy()
y_idx = batch["y_idx"]
# Separate data and mask
temporal_data_cols = self._get_temporal_exogenous_cols(
temporal_cols=temporal_cols
)
temporal_idxs = get_indexer_raise_missing(temporal_cols, temporal_data_cols)
temporal_idxs = np.append(y_idx, temporal_idxs)
temporal_data = temporal[:, temporal_idxs, :]
temporal_mask = temporal[:, temporal_cols.get_loc("available_mask"), :].clone()
# Remove validation and test set to prevent leakeage
if val_size + test_size > 0:
cutoff = val_size + test_size
temporal_mask[:, -cutoff:] = 0
# Normalize. self.scaler stores the shift and scale for inverse transform
temporal_mask = temporal_mask.unsqueeze(
1
) # Add channel dimension for scaler.transform.
temporal_data = self.scaler.transform(x=temporal_data, mask=temporal_mask)
# Replace values in windows dict
temporal[:, temporal_idxs, :] = temporal_data
batch["temporal"] = temporal
return batch
def _inv_normalization(self, y_hat, temporal_cols, y_idx):
# Receives window predictions [B, seq_len, H, output]
# Broadcasts outputs and inverts normalization
# Get 'y' scale and shift, and add W dimension
y_loc = self.scaler.x_shift[:, [y_idx], 0].flatten() # [B,C,T] -> [B]
y_scale = self.scaler.x_scale[:, [y_idx], 0].flatten() # [B,C,T] -> [B]
# Expand scale and shift to y_hat dimensions
y_loc = y_loc.view(*y_loc.shape, *(1,) * (y_hat.ndim - 1)) # .expand(y_hat)
y_scale = y_scale.view(
*y_scale.shape, *(1,) * (y_hat.ndim - 1)
) # .expand(y_hat)
y_hat = self.scaler.inverse_transform(z=y_hat, x_scale=y_scale, x_shift=y_loc)
return y_hat, y_loc, y_scale
def _create_windows(self, batch, step):
temporal = batch["temporal"]
temporal_cols = batch["temporal_cols"]
if step == "train":
if self.val_size + self.test_size > 0:
cutoff = -self.val_size - self.test_size
temporal = temporal[:, :, :cutoff]
temporal = self.padder(temporal)
# Truncate batch to shorter time-series
av_condition = torch.nonzero(
torch.min(
temporal[:, temporal_cols.get_loc("available_mask")], axis=0
).values
)
min_time_stamp = int(av_condition.min())
available_ts = temporal.shape[-1] - min_time_stamp
if available_ts < 1 + self.h:
raise Exception(
"Time series too short for given input and output size. \n"
f"Available timestamps: {available_ts}"
)
temporal = temporal[:, :, min_time_stamp:]
if step == "val":
if self.test_size > 0:
temporal = temporal[:, :, : -self.test_size]
temporal = self.padder(temporal)
if step == "predict":
if (self.test_size == 0) and (len(self.futr_exog_list) == 0):
temporal = self.padder(temporal)
# Test size covers all data, pad left one timestep with zeros
if temporal.shape[-1] == self.test_size:
padder_left = nn.ConstantPad1d(padding=(1, 0), value=0)
temporal = padder_left(temporal)
# Parse batch
window_size = 1 + self.h # 1 for current t and h for future
windows = temporal.unfold(dimension=-1, size=window_size, step=1)
# Truncated backprogatation/inference (shorten sequence where RNNs unroll)
n_windows = windows.shape[2]
input_size = -1
if (step == "train") and (self.input_size > 0):
input_size = self.input_size
if (input_size > 0) and (n_windows > input_size):
max_sampleable_time = n_windows - self.input_size + 1
start = np.random.choice(max_sampleable_time)
windows = windows[:, :, start : (start + input_size), :]
if (step == "val") and (self.inference_input_size > 0):
cutoff = self.inference_input_size + self.val_size
windows = windows[:, :, -cutoff:, :]
if (step == "predict") and (self.inference_input_size > 0):
cutoff = self.inference_input_size + self.test_size
windows = windows[:, :, -cutoff:, :]
# [B, C, input_size, 1+H]
windows_batch = dict(
temporal=windows,
temporal_cols=temporal_cols,
static=batch.get("static", None),
static_cols=batch.get("static_cols", None),
)
return windows_batch
def _parse_windows(self, batch, windows):
# [B, C, seq_len, 1+H]
# Filter insample lags from outsample horizon
mask_idx = batch["temporal_cols"].get_loc("available_mask")
y_idx = batch["y_idx"]
insample_y = windows["temporal"][:, y_idx, :, : -self.h]
insample_mask = windows["temporal"][:, mask_idx, :, : -self.h]
outsample_y = windows["temporal"][:, y_idx, :, -self.h :].contiguous()
outsample_mask = windows["temporal"][:, mask_idx, :, -self.h :].contiguous()
# Filter historic exogenous variables
if len(self.hist_exog_list):
hist_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.hist_exog_list
)
hist_exog = windows["temporal"][:, hist_exog_idx, :, : -self.h]
else:
hist_exog = None
# Filter future exogenous variables
if len(self.futr_exog_list):
futr_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.futr_exog_list
)
futr_exog = windows["temporal"][:, futr_exog_idx, :, :]
else:
futr_exog = None
# Filter static variables
if len(self.stat_exog_list):
static_idx = get_indexer_raise_missing(
windows["static_cols"], self.stat_exog_list
)
stat_exog = windows["static"][:, static_idx]
else:
stat_exog = None
return (
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
)
def training_step(self, batch, batch_idx):
# Create and normalize windows [Ws, L+H, C]
batch = self._normalization(
batch, val_size=self.val_size, test_size=self.test_size
)
windows = self._create_windows(batch, step="train")
# Parse windows
(
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [B, seq_len, 1]
insample_mask=insample_mask, # [B, seq_len, 1]
futr_exog=futr_exog, # [B, F, seq_len, 1+H]
hist_exog=hist_exog, # [B, C, seq_len]
stat_exog=stat_exog,
) # [B, S]
# Model predictions
output = self(windows_batch) # tuple([B, seq_len, H, output])
if self.loss.is_distribution_output:
outsample_y, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y,
temporal_cols=batch["temporal_cols"],
y_idx=batch["y_idx"],
)
B = output[0].size()[0]
T = output[0].size()[1]
H = output[0].size()[2]
output = [arg.view(-1, *(arg.size()[2:])) for arg in output]
outsample_y = outsample_y.view(B * T, H)
outsample_mask = outsample_mask.view(B * T, H)
y_loc = y_loc.repeat_interleave(repeats=T, dim=0).squeeze(-1)
y_scale = y_scale.repeat_interleave(repeats=T, dim=0).squeeze(-1)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
loss = self.loss(y=outsample_y, distr_args=distr_args, mask=outsample_mask)
else:
loss = self.loss(y=outsample_y, y_hat=output, mask=outsample_mask)
if torch.isnan(loss):
print("Model Parameters", self.hparams)
print("insample_y", torch.isnan(insample_y).sum())
print("outsample_y", torch.isnan(outsample_y).sum())
print("output", torch.isnan(output).sum())
raise Exception("Loss is NaN, training stopped.")
self.log(
"train_loss",
loss.item(),
batch_size=outsample_y.size(0),
prog_bar=True,
on_epoch=True,
)
self.train_trajectories.append((self.global_step, loss.item()))
return loss
def validation_step(self, batch, batch_idx):
if self.val_size == 0:
return np.nan
# Create and normalize windows [Ws, L+H, C]
batch = self._normalization(
batch, val_size=self.val_size, test_size=self.test_size
)
windows = self._create_windows(batch, step="val")
y_idx = batch["y_idx"]
# Parse windows
(
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [B, seq_len, 1]
insample_mask=insample_mask, # [B, seq_len, 1]
futr_exog=futr_exog, # [B, F, seq_len, 1+H]
hist_exog=hist_exog, # [B, C, seq_len]
stat_exog=stat_exog,
) # [B, S]
# Remove train y_hat (+1 and -1 for padded last window with zeros)
# tuple([B, seq_len, H, output]) -> tuple([B, validation_size, H, output])
val_windows = (self.val_size) + 1
outsample_y = outsample_y[:, -val_windows:-1, :]
outsample_mask = outsample_mask[:, -val_windows:-1, :]
# Model predictions
output = self(windows_batch) # tuple([B, seq_len, H, output])
if self.loss.is_distribution_output:
output = [arg[:, -val_windows:-1] for arg in output]
outsample_y, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
B = output[0].size()[0]
T = output[0].size()[1]
H = output[0].size()[2]
output = [arg.reshape(-1, *(arg.size()[2:])) for arg in output]
outsample_y = outsample_y.reshape(B * T, H)
outsample_mask = outsample_mask.reshape(B * T, H)
y_loc = y_loc.repeat_interleave(repeats=T, dim=0).squeeze(-1)
y_scale = y_scale.repeat_interleave(repeats=T, dim=0).squeeze(-1)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
_, sample_mean, quants = self.loss.sample(distr_args=distr_args)
if str(type(self.valid_loss)) in [
"<class 'neuralforecast.losses.pytorch.sCRPS'>",
"<class 'neuralforecast.losses.pytorch.MQLoss'>",
]:
output = quants
elif str(type(self.valid_loss)) in [
"<class 'neuralforecast.losses.pytorch.relMSE'>"
]:
output = torch.unsqueeze(sample_mean, dim=-1) # [N,H,1] -> [N,H]
else:
output = output[:, -val_windows:-1, :]
# Validation Loss evaluation
if self.valid_loss.is_distribution_output:
valid_loss = self.valid_loss(
y=outsample_y, distr_args=distr_args, mask=outsample_mask
)
else:
outsample_y, _, _ = self._inv_normalization(
y_hat=outsample_y, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
output, _, _ = self._inv_normalization(
y_hat=output, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
valid_loss = self.valid_loss(
y=outsample_y, y_hat=output, mask=outsample_mask
)
if torch.isnan(valid_loss):
raise Exception("Loss is NaN, training stopped.")
self.log(
"valid_loss",
valid_loss.item(),
batch_size=outsample_y.size(0),
prog_bar=True,
on_epoch=True,
)
self.validation_step_outputs.append(valid_loss)
return valid_loss
def predict_step(self, batch, batch_idx):
# Create and normalize windows [Ws, L+H, C]
batch = self._normalization(batch, val_size=0, test_size=self.test_size)
windows = self._create_windows(batch, step="predict")
y_idx = batch["y_idx"]
# Parse windows
insample_y, insample_mask, _, _, hist_exog, futr_exog, stat_exog = (
self._parse_windows(batch, windows)
)
windows_batch = dict(
insample_y=insample_y, # [B, seq_len, 1]
insample_mask=insample_mask, # [B, seq_len, 1]
futr_exog=futr_exog, # [B, F, seq_len, 1+H]
hist_exog=hist_exog, # [B, C, seq_len]
stat_exog=stat_exog,
) # [B, S]
# Model Predictions
output = self(windows_batch) # tuple([B, seq_len, H], ...)
if self.loss.is_distribution_output:
_, y_loc, y_scale = self._inv_normalization(
y_hat=output[0], temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
B = output[0].size()[0]
T = output[0].size()[1]
H = output[0].size()[2]
output = [arg.reshape(-1, *(arg.size()[2:])) for arg in output]
y_loc = y_loc.repeat_interleave(repeats=T, dim=0).squeeze(-1)
y_scale = y_scale.repeat_interleave(repeats=T, dim=0).squeeze(-1)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
_, sample_mean, quants = self.loss.sample(distr_args=distr_args)
y_hat = torch.concat((sample_mean, quants), axis=2)
y_hat = y_hat.view(B, T, H, -1)
if self.loss.return_params:
distr_args = torch.stack(distr_args, dim=-1)
distr_args = torch.reshape(distr_args, (B, T, H, -1))
y_hat = torch.concat((y_hat, distr_args), axis=3)
else:
y_hat, _, _ = self._inv_normalization(
y_hat=output, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
return y_hat
def fit(
self,
dataset,
val_size=0,
test_size=0,
random_seed=None,
distributed_config=None,
):
"""Fit.
The `fit` method, optimizes the neural network's weights using the
initialization parameters (`learning_rate`, `batch_size`, ...)
and the `loss` function as defined during the initialization.
Within `fit` we use a PyTorch Lightning `Trainer` that
inherits the initialization's `self.trainer_kwargs`, to customize
its inputs, see [PL's trainer arguments](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.trainer.trainer.Trainer.html?highlight=trainer).
The method is designed to be compatible with SKLearn-like classes
and in particular to be compatible with the StatsForecast library.
By default the `model` is not saving training checkpoints to protect
disk memory, to get them change `enable_checkpointing=True` in `__init__`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`val_size`: int, validation size for temporal cross-validation.<br>
`test_size`: int, test size for temporal cross-validation.<br>
`random_seed`: int=None, random_seed for pytorch initializer and numpy generators, overwrites model.__init__'s.<br>
"""
return self._fit(
dataset=dataset,
batch_size=self.batch_size,
valid_batch_size=self.valid_batch_size,
val_size=val_size,
test_size=test_size,
random_seed=random_seed,
distributed_config=distributed_config,
)
def predict(self, dataset, step_size=1, random_seed=None, **data_module_kwargs):
"""Predict.
Neural network prediction with PL's `Trainer` execution of `predict_step`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`step_size`: int=1, Step size between each window.<br>
`random_seed`: int=None, random_seed for pytorch initializer and numpy generators, overwrites model.__init__'s.<br>
`**data_module_kwargs`: PL's TimeSeriesDataModule args, see [documentation](https://pytorch-lightning.readthedocs.io/en/1.6.1/extensions/datamodules.html#using-a-datamodule).
"""
self._check_exog(dataset)
self._restart_seed(random_seed)
if step_size > 1:
raise Exception("Recurrent models do not support step_size > 1")
# fcsts (window, batch, h)
# Protect when case of multiple gpu. PL does not support return preds with multiple gpu.
pred_trainer_kwargs = self.trainer_kwargs.copy()
if (pred_trainer_kwargs.get("accelerator", None) == "gpu") and (
torch.cuda.device_count() > 1
):
pred_trainer_kwargs["devices"] = [0]
trainer = pl.Trainer(**pred_trainer_kwargs)
datamodule = TimeSeriesDataModule(
dataset=dataset,
valid_batch_size=self.valid_batch_size,
num_workers=self.num_workers_loader,
**data_module_kwargs,
)
fcsts = trainer.predict(self, datamodule=datamodule)
if self.test_size > 0:
# Remove warmup windows (from train and validation)
# [N,T,H,output], avoid indexing last dim for univariate output compatibility
fcsts = torch.vstack(
[fcst[:, -(1 + self.test_size - self.h) :, :] for fcst in fcsts]
)
fcsts = fcsts.numpy().flatten()
fcsts = fcsts.reshape(-1, len(self.loss.output_names))
else:
fcsts = torch.vstack([fcst[:, -1:, :] for fcst in fcsts]).numpy().flatten()
fcsts = fcsts.reshape(-1, len(self.loss.output_names))
return fcsts
neuralforecast/common/_base_windows.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.base_windows.ipynb.
# %% auto 0
__all__ = ['BaseWindows']
# %% ../../nbs/common.base_windows.ipynb 5
import numpy as np
import torch
import torch.nn as nn
import pytorch_lightning as pl
from ._base_model import BaseModel
from ._scalers import TemporalNorm
from ..tsdataset import TimeSeriesDataModule
from ..utils import get_indexer_raise_missing
# %% ../../nbs/common.base_windows.ipynb 6
class BaseWindows(BaseModel):
"""Base Windows
Base class for all windows-based models. The forecasts are produced separately
for each window, which are randomly sampled during training.
This class implements the basic functionality for all windows-based models, including:
- PyTorch Lightning's methods training_step, validation_step, predict_step.<br>
- fit and predict methods used by NeuralForecast.core class.<br>
- sampling and wrangling methods to generate windows.
"""
def __init__(
self,
h,
input_size,
loss,
valid_loss,
learning_rate,
max_steps,
val_check_steps,
batch_size,
valid_batch_size,
windows_batch_size,
inference_windows_batch_size,
start_padding_enabled,
step_size=1,
num_lr_decays=0,
early_stop_patience_steps=-1,
scaler_type="identity",
futr_exog_list=None,
hist_exog_list=None,
stat_exog_list=None,
exclude_insample_y=False,
num_workers_loader=0,
drop_last_loader=False,
random_seed=1,
alias=None,
optimizer=None,
optimizer_kwargs=None,
**trainer_kwargs,
):
super().__init__(
random_seed=random_seed,
loss=loss,
valid_loss=valid_loss,
optimizer=optimizer,
optimizer_kwargs=optimizer_kwargs,
futr_exog_list=futr_exog_list,
hist_exog_list=hist_exog_list,
stat_exog_list=stat_exog_list,
max_steps=max_steps,
early_stop_patience_steps=early_stop_patience_steps,
**trainer_kwargs,
)
# Padder to complete train windows,
# example y=[1,2,3,4,5] h=3 -> last y_output = [5,0,0]
self.h = h
self.input_size = input_size
self.windows_batch_size = windows_batch_size
self.start_padding_enabled = start_padding_enabled
if start_padding_enabled:
self.padder_train = nn.ConstantPad1d(
padding=(self.input_size - 1, self.h), value=0
)
else:
self.padder_train = nn.ConstantPad1d(padding=(0, self.h), value=0)
# Batch sizes
self.batch_size = batch_size
if valid_batch_size is None:
self.valid_batch_size = batch_size
else:
self.valid_batch_size = valid_batch_size
if inference_windows_batch_size is None:
self.inference_windows_batch_size = windows_batch_size
else:
self.inference_windows_batch_size = inference_windows_batch_size
# Optimization
self.learning_rate = learning_rate
self.max_steps = max_steps
self.num_lr_decays = num_lr_decays
self.lr_decay_steps = (
max(max_steps // self.num_lr_decays, 1) if self.num_lr_decays > 0 else 10e7
)
self.early_stop_patience_steps = early_stop_patience_steps
self.val_check_steps = val_check_steps
self.windows_batch_size = windows_batch_size
self.step_size = step_size
self.exclude_insample_y = exclude_insample_y
# Scaler
self.scaler = TemporalNorm(
scaler_type=scaler_type,
dim=1, # Time dimension is 1.
num_features=1 + len(self.hist_exog_list) + len(self.futr_exog_list),
)
# Fit arguments
self.val_size = 0
self.test_size = 0
# Model state
self.decompose_forecast = False
# DataModule arguments
self.num_workers_loader = num_workers_loader
self.drop_last_loader = drop_last_loader
# used by on_validation_epoch_end hook
self.validation_step_outputs = []
self.alias = alias
def _create_windows(self, batch, step, w_idxs=None):
# Parse common data
window_size = self.input_size + self.h
temporal_cols = batch["temporal_cols"]
temporal = batch["temporal"]
if step == "train":
if self.val_size + self.test_size > 0:
cutoff = -self.val_size - self.test_size
temporal = temporal[:, :, :cutoff]
temporal = self.padder_train(temporal)
if temporal.shape[-1] < window_size:
raise Exception(
"Time series is too short for training, consider setting a smaller input size or set start_padding_enabled=True"
)
windows = temporal.unfold(
dimension=-1, size=window_size, step=self.step_size
)
# [B, C, Ws, L+H] 0, 1, 2, 3
# -> [B * Ws, L+H, C] 0, 2, 3, 1
windows_per_serie = windows.shape[2]
windows = windows.permute(0, 2, 3, 1).contiguous()
windows = windows.reshape(-1, window_size, len(temporal_cols))
# Sample and Available conditions
available_idx = temporal_cols.get_loc("available_mask")
available_condition = windows[:, : self.input_size, available_idx]
available_condition = torch.sum(available_condition, axis=1)
final_condition = available_condition > 0
if self.h > 0:
sample_condition = windows[:, self.input_size :, available_idx]
sample_condition = torch.sum(sample_condition, axis=1)
final_condition = (sample_condition > 0) & (available_condition > 0)
windows = windows[final_condition]
# Parse Static data to match windows
# [B, S_in] -> [B, Ws, S_in] -> [B*Ws, S_in]
static = batch.get("static", None)
static_cols = batch.get("static_cols", None)
if static is not None:
static = torch.repeat_interleave(
static, repeats=windows_per_serie, dim=0
)
static = static[final_condition]
# Protection of empty windows
if final_condition.sum() == 0:
raise Exception("No windows available for training")
# Sample windows
n_windows = len(windows)
if self.windows_batch_size is not None:
w_idxs = np.random.choice(
n_windows,
size=self.windows_batch_size,
replace=(n_windows < self.windows_batch_size),
)
windows = windows[w_idxs]
if static is not None:
static = static[w_idxs]
# think about interaction available * sample mask
# [B, C, Ws, L+H]
windows_batch = dict(
temporal=windows,
temporal_cols=temporal_cols,
static=static,
static_cols=static_cols,
)
return windows_batch
elif step in ["predict", "val"]:
if step == "predict":
initial_input = temporal.shape[-1] - self.test_size
if (
initial_input <= self.input_size
): # There is not enough data to predict first timestamp
padder_left = nn.ConstantPad1d(
padding=(self.input_size - initial_input, 0), value=0
)
temporal = padder_left(temporal)
predict_step_size = self.predict_step_size
cutoff = -self.input_size - self.test_size
temporal = temporal[:, :, cutoff:]
elif step == "val":
predict_step_size = self.step_size
cutoff = -self.input_size - self.val_size - self.test_size
if self.test_size > 0:
temporal = batch["temporal"][:, :, cutoff : -self.test_size]
else:
temporal = batch["temporal"][:, :, cutoff:]
if temporal.shape[-1] < window_size:
initial_input = temporal.shape[-1] - self.val_size
padder_left = nn.ConstantPad1d(
padding=(self.input_size - initial_input, 0), value=0
)
temporal = padder_left(temporal)
if (
(step == "predict")
and (self.test_size == 0)
and (len(self.futr_exog_list) == 0)
):
padder_right = nn.ConstantPad1d(padding=(0, self.h), value=0)
temporal = padder_right(temporal)
windows = temporal.unfold(
dimension=-1, size=window_size, step=predict_step_size
)
# [batch, channels, windows, window_size] 0, 1, 2, 3
# -> [batch * windows, window_size, channels] 0, 2, 3, 1
windows_per_serie = windows.shape[2]
windows = windows.permute(0, 2, 3, 1).contiguous()
windows = windows.reshape(-1, window_size, len(temporal_cols))
static = batch.get("static", None)
static_cols = batch.get("static_cols", None)
if static is not None:
static = torch.repeat_interleave(
static, repeats=windows_per_serie, dim=0
)
# Sample windows for batched prediction
if w_idxs is not None:
windows = windows[w_idxs]
if static is not None:
static = static[w_idxs]
windows_batch = dict(
temporal=windows,
temporal_cols=temporal_cols,
static=static,
static_cols=static_cols,
)
return windows_batch
else:
raise ValueError(f"Unknown step {step}")
def _normalization(self, windows, y_idx):
# windows are already filtered by train/validation/test
# from the `create_windows_method` nor leakage risk
temporal = windows["temporal"] # B, L+H, C
temporal_cols = windows["temporal_cols"].copy() # B, L+H, C
# To avoid leakage uses only the lags
# temporal_data_cols = temporal_cols.drop('available_mask').tolist()
temporal_data_cols = self._get_temporal_exogenous_cols(
temporal_cols=temporal_cols
)
temporal_idxs = get_indexer_raise_missing(temporal_cols, temporal_data_cols)
temporal_idxs = np.append(y_idx, temporal_idxs)
temporal_data = temporal[:, :, temporal_idxs]
temporal_mask = temporal[:, :, temporal_cols.get_loc("available_mask")].clone()
if self.h > 0:
temporal_mask[:, -self.h :] = 0.0
# Normalize. self.scaler stores the shift and scale for inverse transform
temporal_mask = temporal_mask.unsqueeze(
-1
) # Add channel dimension for scaler.transform.
temporal_data = self.scaler.transform(x=temporal_data, mask=temporal_mask)
# Replace values in windows dict
temporal[:, :, temporal_idxs] = temporal_data
windows["temporal"] = temporal
return windows
def _inv_normalization(self, y_hat, temporal_cols, y_idx):
# Receives window predictions [B, H, output]
# Broadcasts outputs and inverts normalization
# Add C dimension
if y_hat.ndim == 2:
remove_dimension = True
y_hat = y_hat.unsqueeze(-1)
else:
remove_dimension = False
y_scale = self.scaler.x_scale[:, :, [y_idx]]
y_loc = self.scaler.x_shift[:, :, [y_idx]]
y_scale = torch.repeat_interleave(y_scale, repeats=y_hat.shape[-1], dim=-1).to(
y_hat.device
)
y_loc = torch.repeat_interleave(y_loc, repeats=y_hat.shape[-1], dim=-1).to(
y_hat.device
)
y_hat = self.scaler.inverse_transform(z=y_hat, x_scale=y_scale, x_shift=y_loc)
y_loc = y_loc.to(y_hat.device)
y_scale = y_scale.to(y_hat.device)
if remove_dimension:
y_hat = y_hat.squeeze(-1)
y_loc = y_loc.squeeze(-1)
y_scale = y_scale.squeeze(-1)
return y_hat, y_loc, y_scale
def _parse_windows(self, batch, windows):
# Filter insample lags from outsample horizon
y_idx = batch["y_idx"]
mask_idx = batch["temporal_cols"].get_loc("available_mask")
insample_y = windows["temporal"][:, : self.input_size, y_idx]
insample_mask = windows["temporal"][:, : self.input_size, mask_idx]
# Declare additional information
outsample_y = None
outsample_mask = None
hist_exog = None
futr_exog = None
stat_exog = None
if self.h > 0:
outsample_y = windows["temporal"][:, self.input_size :, y_idx]
outsample_mask = windows["temporal"][:, self.input_size :, mask_idx]
if len(self.hist_exog_list):
hist_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.hist_exog_list
)
hist_exog = windows["temporal"][:, : self.input_size, hist_exog_idx]
if len(self.futr_exog_list):
futr_exog_idx = get_indexer_raise_missing(
windows["temporal_cols"], self.futr_exog_list
)
futr_exog = windows["temporal"][:, :, futr_exog_idx]
if len(self.stat_exog_list):
static_idx = get_indexer_raise_missing(
windows["static_cols"], self.stat_exog_list
)
stat_exog = windows["static"][:, static_idx]
# TODO: think a better way of removing insample_y features
if self.exclude_insample_y:
insample_y = insample_y * 0
return (
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
)
def training_step(self, batch, batch_idx):
# Create and normalize windows [Ws, L+H, C]
windows = self._create_windows(batch, step="train")
y_idx = batch["y_idx"]
original_outsample_y = torch.clone(windows["temporal"][:, -self.h :, y_idx])
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
(
insample_y,
insample_mask,
outsample_y,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [Ws, L]
insample_mask=insample_mask, # [Ws, L]
futr_exog=futr_exog, # [Ws, L+H]
hist_exog=hist_exog, # [Ws, L]
stat_exog=stat_exog,
) # [Ws, 1]
# Model Predictions
output = self(windows_batch)
if self.loss.is_distribution_output:
_, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y, temporal_cols=batch["temporal_cols"], y_idx=y_idx
)
outsample_y = original_outsample_y
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
loss = self.loss(y=outsample_y, distr_args=distr_args, mask=outsample_mask)
else:
loss = self.loss(y=outsample_y, y_hat=output, mask=outsample_mask)
if torch.isnan(loss):
print("Model Parameters", self.hparams)
print("insample_y", torch.isnan(insample_y).sum())
print("outsample_y", torch.isnan(outsample_y).sum())
print("output", torch.isnan(output).sum())
raise Exception("Loss is NaN, training stopped.")
self.log(
"train_loss",
loss.item(),
batch_size=outsample_y.size(0),
prog_bar=True,
on_epoch=True,
)
self.train_trajectories.append((self.global_step, loss.item()))
return loss
def _compute_valid_loss(
self, outsample_y, output, outsample_mask, temporal_cols, y_idx
):
if self.loss.is_distribution_output:
_, y_loc, y_scale = self._inv_normalization(
y_hat=outsample_y, temporal_cols=temporal_cols, y_idx=y_idx
)
distr_args = self.loss.scale_decouple(
output=output, loc=y_loc, scale=y_scale
)
_, sample_mean, quants = self.loss.sample(distr_args=distr_args)
if str(type(self.valid_loss)) in [
"<class 'neuralforecast.losses.pytorch.sCRPS'>",
"<class 'neuralforecast.losses.pytorch.MQLoss'>",
]:
output = quants
elif str(type(self.valid_loss)) in [
"<class 'neuralforecast.losses.pytorch.relMSE'>"
]:
output = torch.unsqueeze(sample_mean, dim=-1) # [N,H,1] -> [N,H]
# Validation Loss evaluation
if self.valid_loss.is_distribution_output:
valid_loss = self.valid_loss(
y=outsample_y, distr_args=distr_args, mask=outsample_mask
)
else:
output, _, _ = self._inv_normalization(
y_hat=output, temporal_cols=temporal_cols, y_idx=y_idx
)
valid_loss = self.valid_loss(
y=outsample_y, y_hat=output, mask=outsample_mask
)
return valid_loss
def validation_step(self, batch, batch_idx):
if self.val_size == 0:
return np.nan
# TODO: Hack to compute number of windows
windows = self._create_windows(batch, step="val")
n_windows = len(windows["temporal"])
y_idx = batch["y_idx"]
# Number of windows in batch
windows_batch_size = self.inference_windows_batch_size
if windows_batch_size < 0:
windows_batch_size = n_windows
n_batches = int(np.ceil(n_windows / windows_batch_size))
valid_losses = []
batch_sizes = []
for i in range(n_batches):
# Create and normalize windows [Ws, L+H, C]
w_idxs = np.arange(
i * windows_batch_size, min((i + 1) * windows_batch_size, n_windows)
)
windows = self._create_windows(batch, step="val", w_idxs=w_idxs)
original_outsample_y = torch.clone(windows["temporal"][:, -self.h :, y_idx])
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
(
insample_y,
insample_mask,
_,
outsample_mask,
hist_exog,
futr_exog,
stat_exog,
) = self._parse_windows(batch, windows)
windows_batch = dict(
insample_y=insample_y, # [Ws, L]
insample_mask=insample_mask, # [Ws, L]
futr_exog=futr_exog, # [Ws, L+H]
hist_exog=hist_exog, # [Ws, L]
stat_exog=stat_exog,
) # [Ws, 1]
# Model Predictions
output_batch = self(windows_batch)
valid_loss_batch = self._compute_valid_loss(
outsample_y=original_outsample_y,
output=output_batch,
outsample_mask=outsample_mask,
temporal_cols=batch["temporal_cols"],
y_idx=batch["y_idx"],
)
valid_losses.append(valid_loss_batch)
batch_sizes.append(len(output_batch))
valid_loss = torch.stack(valid_losses)
batch_sizes = torch.tensor(batch_sizes, device=valid_loss.device)
batch_size = torch.sum(batch_sizes)
valid_loss = torch.sum(valid_loss * batch_sizes) / batch_size
if torch.isnan(valid_loss):
raise Exception("Loss is NaN, training stopped.")
self.log(
"valid_loss",
valid_loss.item(),
batch_size=batch_size,
prog_bar=True,
on_epoch=True,
)
self.validation_step_outputs.append(valid_loss)
return valid_loss
def predict_step(self, batch, batch_idx):
# TODO: Hack to compute number of windows
windows = self._create_windows(batch, step="predict")
n_windows = len(windows["temporal"])
y_idx = batch["y_idx"]
# Number of windows in batch
windows_batch_size = self.inference_windows_batch_size
if windows_batch_size < 0:
windows_batch_size = n_windows
n_batches = int(np.ceil(n_windows / windows_batch_size))
y_hats = []
for i in range(n_batches):
# Create and normalize windows [Ws, L+H, C]
w_idxs = np.arange(
i * windows_batch_size, min((i + 1) * windows_batch_size, n_windows)
)
windows = self._create_windows(batch, step="predict", w_idxs=w_idxs)
windows = self._normalization(windows=windows, y_idx=y_idx)
# Parse windows
insample_y, insample_mask, _, _, hist_exog, futr_exog, stat_exog = (
self._parse_windows(batch, windows)
)
windows_batch = dict(
insample_y=insample_y, # [Ws, L]
insample_mask=insample_mask, # [Ws, L]
futr_exog=futr_exog, # [Ws, L+H]
hist_exog=hist_exog, # [Ws, L]
stat_exog=stat_exog,
) # [Ws, 1]
# Model Predictions
output_batch = self(windows_batch)
# Inverse normalization and sampling
if self.loss.is_distribution_output:
_, y_loc, y_scale = self._inv_normalization(
y_hat=output_batch[0],
temporal_cols=batch["temporal_cols"],
y_idx=y_idx,
)
distr_args = self.loss.scale_decouple(
output=output_batch, loc=y_loc, scale=y_scale
)
_, sample_mean, quants = self.loss.sample(distr_args=distr_args)
y_hat = torch.concat((sample_mean, quants), axis=2)
if self.loss.return_params:
distr_args = torch.stack(distr_args, dim=-1)
distr_args = torch.reshape(
distr_args, (len(windows["temporal"]), self.h, -1)
)
y_hat = torch.concat((y_hat, distr_args), axis=2)
else:
y_hat, _, _ = self._inv_normalization(
y_hat=output_batch,
temporal_cols=batch["temporal_cols"],
y_idx=y_idx,
)
y_hats.append(y_hat)
y_hat = torch.cat(y_hats, dim=0)
return y_hat
def fit(
self,
dataset,
val_size=0,
test_size=0,
random_seed=None,
distributed_config=None,
):
"""Fit.
The `fit` method, optimizes the neural network's weights using the
initialization parameters (`learning_rate`, `windows_batch_size`, ...)
and the `loss` function as defined during the initialization.
Within `fit` we use a PyTorch Lightning `Trainer` that
inherits the initialization's `self.trainer_kwargs`, to customize
its inputs, see [PL's trainer arguments](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.trainer.trainer.Trainer.html?highlight=trainer).
The method is designed to be compatible with SKLearn-like classes
and in particular to be compatible with the StatsForecast library.
By default the `model` is not saving training checkpoints to protect
disk memory, to get them change `enable_checkpointing=True` in `__init__`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`val_size`: int, validation size for temporal cross-validation.<br>
`random_seed`: int=None, random_seed for pytorch initializer and numpy generators, overwrites model.__init__'s.<br>
`test_size`: int, test size for temporal cross-validation.<br>
"""
return self._fit(
dataset=dataset,
batch_size=self.batch_size,
valid_batch_size=self.valid_batch_size,
val_size=val_size,
test_size=test_size,
random_seed=random_seed,
distributed_config=distributed_config,
)
def predict(
self,
dataset,
test_size=None,
step_size=1,
random_seed=None,
**data_module_kwargs,
):
"""Predict.
Neural network prediction with PL's `Trainer` execution of `predict_step`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`test_size`: int=None, test size for temporal cross-validation.<br>
`step_size`: int=1, Step size between each window.<br>
`random_seed`: int=None, random_seed for pytorch initializer and numpy generators, overwrites model.__init__'s.<br>
`**data_module_kwargs`: PL's TimeSeriesDataModule args, see [documentation](https://pytorch-lightning.readthedocs.io/en/1.6.1/extensions/datamodules.html#using-a-datamodule).
"""
self._check_exog(dataset)
self._restart_seed(random_seed)
self.predict_step_size = step_size
self.decompose_forecast = False
datamodule = TimeSeriesDataModule(
dataset=dataset,
valid_batch_size=self.valid_batch_size,
**data_module_kwargs,
)
# Protect when case of multiple gpu. PL does not support return preds with multiple gpu.
pred_trainer_kwargs = self.trainer_kwargs.copy()
if (pred_trainer_kwargs.get("accelerator", None) == "gpu") and (
torch.cuda.device_count() > 1
):
pred_trainer_kwargs["devices"] = [0]
trainer = pl.Trainer(**pred_trainer_kwargs)
fcsts = trainer.predict(self, datamodule=datamodule)
fcsts = torch.vstack(fcsts).numpy().flatten()
fcsts = fcsts.reshape(-1, len(self.loss.output_names))
return fcsts
def decompose(self, dataset, step_size=1, random_seed=None, **data_module_kwargs):
"""Decompose Predictions.
Decompose the predictions through the network's layers.
Available methods are `ESRNN`, `NHITS`, `NBEATS`, and `NBEATSx`.
**Parameters:**<br>
`dataset`: NeuralForecast's `TimeSeriesDataset`, see [documentation here](https://nixtla.github.io/neuralforecast/tsdataset.html).<br>
`step_size`: int=1, step size between each window of temporal data.<br>
`**data_module_kwargs`: PL's TimeSeriesDataModule args, see [documentation](https://pytorch-lightning.readthedocs.io/en/1.6.1/extensions/datamodules.html#using-a-datamodule).
"""
# Restart random seed
if random_seed is None:
random_seed = self.random_seed
torch.manual_seed(random_seed)
self.predict_step_size = step_size
self.decompose_forecast = True
datamodule = TimeSeriesDataModule(
dataset=dataset,
valid_batch_size=self.valid_batch_size,
**data_module_kwargs,
)
trainer = pl.Trainer(**self.trainer_kwargs)
fcsts = trainer.predict(self, datamodule=datamodule)
self.decompose_forecast = False # Default decomposition back to false
return torch.vstack(fcsts).numpy()
neuralforecast/common/_modules.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.modules.ipynb.
# %% auto 0
__all__ = ['ACTIVATIONS', 'MLP', 'Chomp1d', 'CausalConv1d', 'TemporalConvolutionEncoder', 'TransEncoderLayer', 'TransEncoder',
'TransDecoderLayer', 'TransDecoder', 'AttentionLayer', 'PositionalEmbedding', 'TokenEmbedding',
'TimeFeatureEmbedding', 'DataEmbedding']
# %% ../../nbs/common.modules.ipynb 3
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
# %% ../../nbs/common.modules.ipynb 5
ACTIVATIONS = ["ReLU", "Softplus", "Tanh", "SELU", "LeakyReLU", "PReLU", "Sigmoid"]
# %% ../../nbs/common.modules.ipynb 7
class MLP(nn.Module):
"""Multi-Layer Perceptron Class
**Parameters:**<br>
`in_features`: int, dimension of input.<br>
`out_features`: int, dimension of output.<br>
`activation`: str, activation function to use.<br>
`hidden_size`: int, dimension of hidden layers.<br>
`num_layers`: int, number of hidden layers.<br>
`dropout`: float, dropout rate.<br>
"""
def __init__(
self, in_features, out_features, activation, hidden_size, num_layers, dropout
):
super().__init__()
assert activation in ACTIVATIONS, f"{activation} is not in {ACTIVATIONS}"
self.activation = getattr(nn, activation)()
# MultiLayer Perceptron
# Input layer
layers = [
nn.Linear(in_features=in_features, out_features=hidden_size),
self.activation,
nn.Dropout(dropout),
]
# Hidden layers
for i in range(num_layers - 2):
layers += [
nn.Linear(in_features=hidden_size, out_features=hidden_size),
self.activation,
nn.Dropout(dropout),
]
# Output layer
layers += [nn.Linear(in_features=hidden_size, out_features=out_features)]
# Store in layers as ModuleList
self.layers = nn.Sequential(*layers)
def forward(self, x):
return self.layers(x)
# %% ../../nbs/common.modules.ipynb 9
class Chomp1d(nn.Module):
"""Chomp1d
Receives `x` input of dim [N,C,T], and trims it so that only
'time available' information is used.
Used by one dimensional causal convolutions `CausalConv1d`.
**Parameters:**<br>
`horizon`: int, length of outsample values to skip.
"""
def __init__(self, horizon):
super(Chomp1d, self).__init__()
self.horizon = horizon
def forward(self, x):
return x[:, :, : -self.horizon].contiguous()
class CausalConv1d(nn.Module):
"""Causal Convolution 1d
Receives `x` input of dim [N,C_in,T], and computes a causal convolution
in the time dimension. Skipping the H steps of the forecast horizon, through
its dilation.
Consider a batch of one element, the dilated convolution operation on the
$t$ time step is defined:
$\mathrm{Conv1D}(\mathbf{x},\mathbf{w})(t) = (\mathbf{x}_{[*d]} \mathbf{w})(t) = \sum^{K}_{k=1} w_{k} \mathbf{x}_{t-dk}$
where $d$ is the dilation factor, $K$ is the kernel size, $t-dk$ is the index of
the considered past observation. The dilation effectively applies a filter with skip
connections. If $d=1$ one recovers a normal convolution.
**Parameters:**<br>
`in_channels`: int, dimension of `x` input's initial channels.<br>
`out_channels`: int, dimension of `x` outputs's channels.<br>
`activation`: str, identifying activations from PyTorch activations.
select from 'ReLU','Softplus','Tanh','SELU', 'LeakyReLU','PReLU','Sigmoid'.<br>
`padding`: int, number of zero padding used to the left.<br>
`kernel_size`: int, convolution's kernel size.<br>
`dilation`: int, dilation skip connections.<br>
**Returns:**<br>
`x`: tensor, torch tensor of dim [N,C_out,T] activation(conv1d(inputs, kernel) + bias). <br>
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
padding,
dilation,
activation,
stride: int = 1,
):
super(CausalConv1d, self).__init__()
assert activation in ACTIVATIONS, f"{activation} is not in {ACTIVATIONS}"
self.conv = nn.Conv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
)
self.chomp = Chomp1d(padding)
self.activation = getattr(nn, activation)()
self.causalconv = nn.Sequential(self.conv, self.chomp, self.activation)
def forward(self, x):
return self.causalconv(x)
# %% ../../nbs/common.modules.ipynb 11
class TemporalConvolutionEncoder(nn.Module):
"""Temporal Convolution Encoder
Receives `x` input of dim [N,T,C_in], permutes it to [N,C_in,T]
applies a deep stack of exponentially dilated causal convolutions.
The exponentially increasing dilations of the convolutions allow for
the creation of weighted averages of exponentially large long-term memory.
**Parameters:**<br>
`in_channels`: int, dimension of `x` input's initial channels.<br>
`out_channels`: int, dimension of `x` outputs's channels.<br>
`kernel_size`: int, size of the convolving kernel.<br>
`dilations`: int list, controls the temporal spacing between the kernel points.<br>
`activation`: str, identifying activations from PyTorch activations.
select from 'ReLU','Softplus','Tanh','SELU', 'LeakyReLU','PReLU','Sigmoid'.<br>
**Returns:**<br>
`x`: tensor, torch tensor of dim [N,T,C_out].<br>
"""
# TODO: Add dilations parameter and change layers declaration to for loop
def __init__(
self,
in_channels,
out_channels,
kernel_size,
dilations,
activation: str = "ReLU",
):
super(TemporalConvolutionEncoder, self).__init__()
layers = []
for dilation in dilations:
layers.append(
CausalConv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
padding=(kernel_size - 1) * dilation,
activation=activation,
dilation=dilation,
)
)
in_channels = out_channels
self.tcn = nn.Sequential(*layers)
def forward(self, x):
# [N,T,C_in] -> [N,C_in,T] -> [N,T,C_out]
x = x.permute(0, 2, 1).contiguous()
x = self.tcn(x)
x = x.permute(0, 2, 1).contiguous()
return x
# %% ../../nbs/common.modules.ipynb 15
class TransEncoderLayer(nn.Module):
def __init__(
self,
attention,
hidden_size,
conv_hidden_size=None,
dropout=0.1,
activation="relu",
):
super(TransEncoderLayer, self).__init__()
conv_hidden_size = conv_hidden_size or 4 * hidden_size
self.attention = attention
self.conv1 = nn.Conv1d(
in_channels=hidden_size, out_channels=conv_hidden_size, kernel_size=1
)
self.conv2 = nn.Conv1d(
in_channels=conv_hidden_size, out_channels=hidden_size, kernel_size=1
)
self.norm1 = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.dropout = nn.Dropout(dropout)
self.activation = F.relu if activation == "relu" else F.gelu
def forward(self, x, attn_mask=None):
new_x, attn = self.attention(x, x, x, attn_mask=attn_mask)
x = x + self.dropout(new_x)
y = x = self.norm1(x)
y = self.dropout(self.activation(self.conv1(y.transpose(-1, 1))))
y = self.dropout(self.conv2(y).transpose(-1, 1))
return self.norm2(x + y), attn
class TransEncoder(nn.Module):
def __init__(self, attn_layers, conv_layers=None, norm_layer=None):
super(TransEncoder, self).__init__()
self.attn_layers = nn.ModuleList(attn_layers)
self.conv_layers = (
nn.ModuleList(conv_layers) if conv_layers is not None else None
)
self.norm = norm_layer
def forward(self, x, attn_mask=None):
# x [B, L, D]
attns = []
if self.conv_layers is not None:
for attn_layer, conv_layer in zip(self.attn_layers, self.conv_layers):
x, attn = attn_layer(x, attn_mask=attn_mask)
x = conv_layer(x)
attns.append(attn)
x, attn = self.attn_layers[-1](x)
attns.append(attn)
else:
for attn_layer in self.attn_layers:
x, attn = attn_layer(x, attn_mask=attn_mask)
attns.append(attn)
if self.norm is not None:
x = self.norm(x)
return x, attns
# %% ../../nbs/common.modules.ipynb 16
class TransDecoderLayer(nn.Module):
def __init__(
self,
self_attention,
cross_attention,
hidden_size,
conv_hidden_size=None,
dropout=0.1,
activation="relu",
):
super(TransDecoderLayer, self).__init__()
conv_hidden_size = conv_hidden_size or 4 * hidden_size
self.self_attention = self_attention
self.cross_attention = cross_attention
self.conv1 = nn.Conv1d(
in_channels=hidden_size, out_channels=conv_hidden_size, kernel_size=1
)
self.conv2 = nn.Conv1d(
in_channels=conv_hidden_size, out_channels=hidden_size, kernel_size=1
)
self.norm1 = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.norm3 = nn.LayerNorm(hidden_size)
self.dropout = nn.Dropout(dropout)
self.activation = F.relu if activation == "relu" else F.gelu
def forward(self, x, cross, x_mask=None, cross_mask=None):
x = x + self.dropout(self.self_attention(x, x, x, attn_mask=x_mask)[0])
x = self.norm1(x)
x = x + self.dropout(
self.cross_attention(x, cross, cross, attn_mask=cross_mask)[0]
)
y = x = self.norm2(x)
y = self.dropout(self.activation(self.conv1(y.transpose(-1, 1))))
y = self.dropout(self.conv2(y).transpose(-1, 1))
return self.norm3(x + y)
class TransDecoder(nn.Module):
def __init__(self, layers, norm_layer=None, projection=None):
super(TransDecoder, self).__init__()
self.layers = nn.ModuleList(layers)
self.norm = norm_layer
self.projection = projection
def forward(self, x, cross, x_mask=None, cross_mask=None):
for layer in self.layers:
x = layer(x, cross, x_mask=x_mask, cross_mask=cross_mask)
if self.norm is not None:
x = self.norm(x)
if self.projection is not None:
x = self.projection(x)
return x
# %% ../../nbs/common.modules.ipynb 17
class AttentionLayer(nn.Module):
def __init__(self, attention, hidden_size, n_head, d_keys=None, d_values=None):
super(AttentionLayer, self).__init__()
d_keys = d_keys or (hidden_size // n_head)
d_values = d_values or (hidden_size // n_head)
self.inner_attention = attention
self.query_projection = nn.Linear(hidden_size, d_keys * n_head)
self.key_projection = nn.Linear(hidden_size, d_keys * n_head)
self.value_projection = nn.Linear(hidden_size, d_values * n_head)
self.out_projection = nn.Linear(d_values * n_head, hidden_size)
self.n_head = n_head
def forward(self, queries, keys, values, attn_mask):
B, L, _ = queries.shape
_, S, _ = keys.shape
H = self.n_head
queries = self.query_projection(queries).view(B, L, H, -1)
keys = self.key_projection(keys).view(B, S, H, -1)
values = self.value_projection(values).view(B, S, H, -1)
out, attn = self.inner_attention(queries, keys, values, attn_mask)
out = out.view(B, L, -1)
return self.out_projection(out), attn
# %% ../../nbs/common.modules.ipynb 18
class PositionalEmbedding(nn.Module):
def __init__(self, hidden_size, max_len=5000):
super(PositionalEmbedding, self).__init__()
# Compute the positional encodings once in log space.
pe = torch.zeros(max_len, hidden_size).float()
pe.require_grad = False
position = torch.arange(0, max_len).float().unsqueeze(1)
div_term = (
torch.arange(0, hidden_size, 2).float() * -(math.log(10000.0) / hidden_size)
).exp()
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0)
self.register_buffer("pe", pe)
def forward(self, x):
return self.pe[:, : x.size(1)]
class TokenEmbedding(nn.Module):
def __init__(self, c_in, hidden_size):
super(TokenEmbedding, self).__init__()
padding = 1 if torch.__version__ >= "1.5.0" else 2
self.tokenConv = nn.Conv1d(
in_channels=c_in,
out_channels=hidden_size,
kernel_size=3,
padding=padding,
padding_mode="circular",
bias=False,
)
for m in self.modules():
if isinstance(m, nn.Conv1d):
nn.init.kaiming_normal_(
m.weight, mode="fan_in", nonlinearity="leaky_relu"
)
def forward(self, x):
x = self.tokenConv(x.permute(0, 2, 1)).transpose(1, 2)
return x
class TimeFeatureEmbedding(nn.Module):
def __init__(self, input_size, hidden_size):
super(TimeFeatureEmbedding, self).__init__()
self.embed = nn.Linear(input_size, hidden_size, bias=False)
def forward(self, x):
return self.embed(x)
class DataEmbedding(nn.Module):
def __init__(
self, c_in, exog_input_size, hidden_size, pos_embedding=True, dropout=0.1
):
super(DataEmbedding, self).__init__()
self.value_embedding = TokenEmbedding(c_in=c_in, hidden_size=hidden_size)
if pos_embedding:
self.position_embedding = PositionalEmbedding(hidden_size=hidden_size)
else:
self.position_embedding = None
if exog_input_size > 0:
self.temporal_embedding = TimeFeatureEmbedding(
input_size=exog_input_size, hidden_size=hidden_size
)
else:
self.temporal_embedding = None
self.dropout = nn.Dropout(p=dropout)
def forward(self, x, x_mark=None):
# Convolution
x = self.value_embedding(x)
# Add positional (relative withing window) embedding with sines and cosines
if self.position_embedding is not None:
x = x + self.position_embedding(x)
# Add temporal (absolute in time series) embedding with linear layer
if self.temporal_embedding is not None:
x = x + self.temporal_embedding(x_mark)
return self.dropout(x)
neuralforecast/common/_scalers.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/common.scalers.ipynb.
# %% auto 0
__all__ = ['masked_median', 'masked_mean', 'minmax_statistics', 'minmax1_statistics', 'std_statistics', 'robust_statistics',
'invariant_statistics', 'identity_statistics', 'TemporalNorm']
# %% ../../nbs/common.scalers.ipynb 6
import torch
import torch.nn as nn
# %% ../../nbs/common.scalers.ipynb 10
def masked_median(x, mask, dim=-1, keepdim=True):
"""Masked Median
Compute the median of tensor `x` along dim, ignoring values where
`mask` is False. `x` and `mask` need to be broadcastable.
**Parameters:**<br>
`x`: torch.Tensor to compute median of along `dim` dimension.<br>
`mask`: torch Tensor bool with same shape as `x`, where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`dim` (int, optional): Dimension to take median of. Defaults to -1.<br>
`keepdim` (bool, optional): Keep dimension of `x` or not. Defaults to True.<br>
**Returns:**<br>
`x_median`: torch.Tensor with normalized values.
"""
x_nan = x.float().masked_fill(mask < 1, float("nan"))
x_median, _ = x_nan.nanmedian(dim=dim, keepdim=keepdim)
x_median = torch.nan_to_num(x_median, nan=0.0)
return x_median
def masked_mean(x, mask, dim=-1, keepdim=True):
"""Masked Mean
Compute the mean of tensor `x` along dimension, ignoring values where
`mask` is False. `x` and `mask` need to be broadcastable.
**Parameters:**<br>
`x`: torch.Tensor to compute mean of along `dim` dimension.<br>
`mask`: torch Tensor bool with same shape as `x`, where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`dim` (int, optional): Dimension to take mean of. Defaults to -1.<br>
`keepdim` (bool, optional): Keep dimension of `x` or not. Defaults to True.<br>
**Returns:**<br>
`x_mean`: torch.Tensor with normalized values.
"""
x_nan = x.float().masked_fill(mask < 1, float("nan"))
x_mean = x_nan.nanmean(dim=dim, keepdim=keepdim)
x_mean = torch.nan_to_num(x_mean, nan=0.0)
return x_mean
# %% ../../nbs/common.scalers.ipynb 14
def minmax_statistics(x, mask, eps=1e-6, dim=-1):
"""MinMax Scaler
Standardizes temporal features by ensuring its range dweels between
[0,1] range. This transformation is often used as an alternative
to the standard scaler. The scaled features are obtained as:
$$
\mathbf{z} = (\mathbf{x}_{[B,T,C]}-\mathrm{min}({\mathbf{x}})_{[B,1,C]})/
(\mathrm{max}({\mathbf{x}})_{[B,1,C]}- \mathrm{min}({\mathbf{x}})_{[B,1,C]})
$$
**Parameters:**<br>
`x`: torch.Tensor input tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute min and max. Defaults to -1.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
mask = mask.clone()
mask[mask == 0] = torch.inf
mask[mask == 1] = 0
x_max = torch.max(
torch.nan_to_num(x - mask, nan=-torch.inf), dim=dim, keepdim=True
)[0]
x_min = torch.min(torch.nan_to_num(x + mask, nan=torch.inf), dim=dim, keepdim=True)[
0
]
x_max = x_max.type(x.dtype)
x_min = x_min.type(x.dtype)
# x_range and prevent division by zero
x_range = x_max - x_min
x_range[x_range == 0] = 1.0
x_range = x_range + eps
return x_min, x_range
# %% ../../nbs/common.scalers.ipynb 15
def minmax_scaler(x, x_min, x_range):
return (x - x_min) / x_range
def inv_minmax_scaler(z, x_min, x_range):
return z * x_range + x_min
# %% ../../nbs/common.scalers.ipynb 17
def minmax1_statistics(x, mask, eps=1e-6, dim=-1):
"""MinMax1 Scaler
Standardizes temporal features by ensuring its range dweels between
[-1,1] range. This transformation is often used as an alternative
to the standard scaler or classic Min Max Scaler.
The scaled features are obtained as:
$$\mathbf{z} = 2 (\mathbf{x}_{[B,T,C]}-\mathrm{min}({\mathbf{x}})_{[B,1,C]})/ (\mathrm{max}({\mathbf{x}})_{[B,1,C]}- \mathrm{min}({\mathbf{x}})_{[B,1,C]})-1$$
**Parameters:**<br>
`x`: torch.Tensor input tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute min and max. Defaults to -1.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
# Mask values (set masked to -inf or +inf)
mask = mask.clone()
mask[mask == 0] = torch.inf
mask[mask == 1] = 0
x_max = torch.max(
torch.nan_to_num(x - mask, nan=-torch.inf), dim=dim, keepdim=True
)[0]
x_min = torch.min(torch.nan_to_num(x + mask, nan=torch.inf), dim=dim, keepdim=True)[
0
]
x_max = x_max.type(x.dtype)
x_min = x_min.type(x.dtype)
# x_range and prevent division by zero
x_range = x_max - x_min
x_range[x_range == 0] = 1.0
x_range = x_range + eps
return x_min, x_range
# %% ../../nbs/common.scalers.ipynb 18
def minmax1_scaler(x, x_min, x_range):
x = (x - x_min) / x_range
z = x * (2) - 1
return z
def inv_minmax1_scaler(z, x_min, x_range):
z = (z + 1) / 2
return z * x_range + x_min
# %% ../../nbs/common.scalers.ipynb 20
def std_statistics(x, mask, dim=-1, eps=1e-6):
"""Standard Scaler
Standardizes features by removing the mean and scaling
to unit variance along the `dim` dimension.
For example, for `base_windows` models, the scaled features are obtained as (with dim=1):
$$\mathbf{z} = (\mathbf{x}_{[B,T,C]}-\\bar{\mathbf{x}}_{[B,1,C]})/\hat{\sigma}_{[B,1,C]}$$
**Parameters:**<br>
`x`: torch.Tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute mean and std. Defaults to -1.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
x_means = masked_mean(x=x, mask=mask, dim=dim)
x_stds = torch.sqrt(masked_mean(x=(x - x_means) ** 2, mask=mask, dim=dim))
# Protect against division by zero
x_stds[x_stds == 0] = 1.0
x_stds = x_stds + eps
return x_means, x_stds
# %% ../../nbs/common.scalers.ipynb 21
def std_scaler(x, x_means, x_stds):
return (x - x_means) / x_stds
def inv_std_scaler(z, x_mean, x_std):
return (z * x_std) + x_mean
# %% ../../nbs/common.scalers.ipynb 23
def robust_statistics(x, mask, dim=-1, eps=1e-6):
"""Robust Median Scaler
Standardizes features by removing the median and scaling
with the mean absolute deviation (mad) a robust estimator of variance.
This scaler is particularly useful with noisy data where outliers can
heavily influence the sample mean / variance in a negative way.
In these scenarios the median and amd give better results.
For example, for `base_windows` models, the scaled features are obtained as (with dim=1):
$$\mathbf{z} = (\mathbf{x}_{[B,T,C]}-\\textrm{median}(\mathbf{x})_{[B,1,C]})/\\textrm{mad}(\mathbf{x})_{[B,1,C]}$$
$$\\textrm{mad}(\mathbf{x}) = \\frac{1}{N} \sum_{}|\mathbf{x} - \mathrm{median}(x)|$$
**Parameters:**<br>
`x`: torch.Tensor input tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute median and mad. Defaults to -1.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
x_median = masked_median(x=x, mask=mask, dim=dim)
x_mad = masked_median(x=torch.abs(x - x_median), mask=mask, dim=dim)
# Protect x_mad=0 values
# Assuming normality and relationship between mad and std
x_means = masked_mean(x=x, mask=mask, dim=dim)
x_stds = torch.sqrt(masked_mean(x=(x - x_means) ** 2, mask=mask, dim=dim))
x_mad_aux = x_stds * 0.6744897501960817
x_mad = x_mad * (x_mad > 0) + x_mad_aux * (x_mad == 0)
# Protect against division by zero
x_mad[x_mad == 0] = 1.0
x_mad = x_mad + eps
return x_median, x_mad
# %% ../../nbs/common.scalers.ipynb 24
def robust_scaler(x, x_median, x_mad):
return (x - x_median) / x_mad
def inv_robust_scaler(z, x_median, x_mad):
return z * x_mad + x_median
# %% ../../nbs/common.scalers.ipynb 26
def invariant_statistics(x, mask, dim=-1, eps=1e-6):
"""Invariant Median Scaler
Standardizes features by removing the median and scaling
with the mean absolute deviation (mad) a robust estimator of variance.
Aditionally it complements the transformation with the arcsinh transformation.
For example, for `base_windows` models, the scaled features are obtained as (with dim=1):
$$\mathbf{z} = (\mathbf{x}_{[B,T,C]}-\\textrm{median}(\mathbf{x})_{[B,1,C]})/\\textrm{mad}(\mathbf{x})_{[B,1,C]}$$
$$\mathbf{z} = \\textrm{arcsinh}(\mathbf{z})$$
**Parameters:**<br>
`x`: torch.Tensor input tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute median and mad. Defaults to -1.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
x_median = masked_median(x=x, mask=mask, dim=dim)
x_mad = masked_median(x=torch.abs(x - x_median), mask=mask, dim=dim)
# Protect x_mad=0 values
# Assuming normality and relationship between mad and std
x_means = masked_mean(x=x, mask=mask, dim=dim)
x_stds = torch.sqrt(masked_mean(x=(x - x_means) ** 2, mask=mask, dim=dim))
x_mad_aux = x_stds * 0.6744897501960817
x_mad = x_mad * (x_mad > 0) + x_mad_aux * (x_mad == 0)
# Protect against division by zero
x_mad[x_mad == 0] = 1.0
x_mad = x_mad + eps
return x_median, x_mad
# %% ../../nbs/common.scalers.ipynb 27
def invariant_scaler(x, x_median, x_mad):
return torch.arcsinh((x - x_median) / x_mad)
def inv_invariant_scaler(z, x_median, x_mad):
return torch.sinh(z) * x_mad + x_median
# %% ../../nbs/common.scalers.ipynb 29
def identity_statistics(x, mask, dim=-1, eps=1e-6):
"""Identity Scaler
A placeholder identity scaler, that is argument insensitive.
**Parameters:**<br>
`x`: torch.Tensor input tensor.<br>
`mask`: torch Tensor bool, same dimension as `x`, indicates where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`dim` (int, optional): Dimension over to compute median and mad. Defaults to -1.<br>
**Returns:**<br>
`x`: original torch.Tensor `x`.
"""
# Collapse dim dimension
shape = list(x.shape)
shape[dim] = 1
x_shift = torch.zeros(shape)
x_scale = torch.ones(shape)
return x_shift, x_scale
# %% ../../nbs/common.scalers.ipynb 30
def identity_scaler(x, x_shift, x_scale):
return x
def inv_identity_scaler(z, x_shift, x_scale):
return z
# %% ../../nbs/common.scalers.ipynb 33
class TemporalNorm(nn.Module):
"""Temporal Normalization
Standardization of the features is a common requirement for many
machine learning estimators, and it is commonly achieved by removing
the level and scaling its variance. The `TemporalNorm` module applies
temporal normalization over the batch of inputs as defined by the type of scaler.
$$\mathbf{z}_{[B,T,C]} = \\textrm{Scaler}(\mathbf{x}_{[B,T,C]})$$
If `scaler_type` is `revin` learnable normalization parameters are added on top of
the usual normalization technique, the parameters are learned through scale decouple
global skip connections. The technique is available for point and probabilistic outputs.
$$\mathbf{\hat{z}}_{[B,T,C]} = \\boldsymbol{\hat{\\gamma}}_{[1,1,C]} \mathbf{z}_{[B,T,C]} +\\boldsymbol{\hat{\\beta}}_{[1,1,C]}$$
**Parameters:**<br>
`scaler_type`: str, defines the type of scaler used by TemporalNorm. Available [`identity`, `standard`, `robust`, `minmax`, `minmax1`, `invariant`, `revin`].<br>
`dim` (int, optional): Dimension over to compute scale and shift. Defaults to -1.<br>
`eps` (float, optional): Small value to avoid division by zero. Defaults to 1e-6.<br>
`num_features`: int=None, for RevIN-like learnable affine parameters initialization.<br>
**References**<br>
- [Kin G. Olivares, David Luo, Cristian Challu, Stefania La Vattiata, Max Mergenthaler, Artur Dubrawski (2023). "HINT: Hierarchical Mixture Networks For Coherent Probabilistic Forecasting". Neural Information Processing Systems, submitted. Working Paper version available at arxiv.](https://arxiv.org/abs/2305.07089)<br>
"""
def __init__(self, scaler_type="robust", dim=-1, eps=1e-6, num_features=None):
super().__init__()
compute_statistics = {
None: identity_statistics,
"identity": identity_statistics,
"standard": std_statistics,
"revin": std_statistics,
"robust": robust_statistics,
"minmax": minmax_statistics,
"minmax1": minmax1_statistics,
"invariant": invariant_statistics,
}
scalers = {
None: identity_scaler,
"identity": identity_scaler,
"standard": std_scaler,
"revin": std_scaler,
"robust": robust_scaler,
"minmax": minmax_scaler,
"minmax1": minmax1_scaler,
"invariant": invariant_scaler,
}
inverse_scalers = {
None: inv_identity_scaler,
"identity": inv_identity_scaler,
"standard": inv_std_scaler,
"revin": inv_std_scaler,
"robust": inv_robust_scaler,
"minmax": inv_minmax_scaler,
"minmax1": inv_minmax1_scaler,
"invariant": inv_invariant_scaler,
}
assert scaler_type in scalers.keys(), f"{scaler_type} not defined"
if (scaler_type == "revin") and (num_features is None):
raise Exception("You must pass num_features for ReVIN scaler.")
self.compute_statistics = compute_statistics[scaler_type]
self.scaler = scalers[scaler_type]
self.inverse_scaler = inverse_scalers[scaler_type]
self.scaler_type = scaler_type
self.dim = dim
self.eps = eps
if scaler_type == "revin":
self._init_params(num_features=num_features)
def _init_params(self, num_features):
# Initialize RevIN scaler params to broadcast:
if self.dim == 1: # [B,T,C] [1,1,C]
self.revin_bias = nn.Parameter(torch.zeros(1, 1, num_features))
self.revin_weight = nn.Parameter(torch.ones(1, 1, num_features))
elif self.dim == -1: # [B,C,T] [1,C,1]
self.revin_bias = nn.Parameter(torch.zeros(1, num_features, 1))
self.revin_weight = nn.Parameter(torch.ones(1, num_features, 1))
# @torch.no_grad()
def transform(self, x, mask):
"""Center and scale the data.
**Parameters:**<br>
`x`: torch.Tensor shape [batch, time, channels].<br>
`mask`: torch Tensor bool, shape [batch, time] where `x` is valid and False
where `x` should be masked. Mask should not be all False in any column of
dimension dim to avoid NaNs from zero division.<br>
**Returns:**<br>
`z`: torch.Tensor same shape as `x`, except scaled.
"""
x_shift, x_scale = self.compute_statistics(
x=x, mask=mask, dim=self.dim, eps=self.eps
)
self.x_shift = x_shift
self.x_scale = x_scale
# Original Revin performs this operation
# z = self.revin_weight * z
# z = z + self.revin_bias
# However this is only valid for point forecast not for
# distribution's scale decouple technique.
if self.scaler_type == "revin":
self.x_shift = self.x_shift + self.revin_bias
self.x_scale = self.x_scale * (torch.relu(self.revin_weight) + self.eps)
z = self.scaler(x, x_shift, x_scale)
return z
# @torch.no_grad()
def inverse_transform(self, z, x_shift=None, x_scale=None):
"""Scale back the data to the original representation.
**Parameters:**<br>
`z`: torch.Tensor shape [batch, time, channels], scaled.<br>
**Returns:**<br>
`x`: torch.Tensor original data.
"""
if x_shift is None:
x_shift = self.x_shift
if x_scale is None:
x_scale = self.x_scale
# Original Revin performs this operation
# z = z - self.revin_bias
# z = (z / (self.revin_weight + self.eps))
# However this is only valid for point forecast not for
# distribution's scale decouple technique.
x = self.inverse_scaler(z, x_shift, x_scale)
return x
def forward(self, x):
# The gradients are optained from BaseWindows/BaseRecurrent forwards.
pass
neuralforecast/compat.py 0 → 100644
View file @ e3f7f7b3
# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/compat.ipynb.
# %% auto 0
__all__ = []
# %% ../nbs/compat.ipynb 1
try:
from pyspark.sql import DataFrame as SparkDataFrame
except ImportError:
class SparkDataFrame: ...
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment