models.nbeats.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "524620c1",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| default_exp models.nbeats"
   ]
  },
  {
   "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": [
    "# NBEATS"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "376a8a3a",
   "metadata": {},
   "source": [
    "The Neural Basis Expansion Analysis (`NBEATS`) is an `MLP`-based deep neural architecture with backward and forward residual links. The network has two variants: (1) in its interpretable configuration, `NBEATS` sequentially projects the signal into polynomials and harmonic basis to learn trend and seasonality components; (2) in its generic configuration, it substitutes the polynomial and harmonic basis for identity basis and larger network's depth. The Neural Basis Expansion Analysis with Exogenous (`NBEATSx`), incorporates projections to exogenous temporal variables available at the time of the prediction.\n",
    "\n",
    "This method proved state-of-the-art performance on the M3, M4, and Tourism Competition datasets, improving accuracy by 3% over the `ESRNN` M4 competition winner.\n",
    "\n",
    "**References**<br>\n",
    "-[Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, Yoshua Bengio (2019). \"N-BEATS: Neural basis expansion analysis for interpretable time series forecasting\".](https://arxiv.org/abs/1905.10437)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bddd17a6",
   "metadata": {},
   "source": [
    "![Figure 1. Neural Basis Expansion Analysis.](imgs_models/nbeats.png)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "44065066-e72a-431f-938f-1528adef9fe8",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| export\n",
    "from typing import Tuple, Optional\n",
    "\n",
    "import numpy as np\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "\n",
    "from neuralforecast.losses.pytorch import MAE\n",
    "from neuralforecast.common._base_windows import BaseWindows"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4a77bb35",
   "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\n",
    "\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9b7a9fae-2c29-47e2-874e-ca1f20bf7040",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| exporti\n",
    "class IdentityBasis(nn.Module):\n",
    "    def __init__(self, backcast_size: int, forecast_size: int,\n",
    "                 out_features: int=1):\n",
    "        super().__init__()\n",
    "        self.out_features = out_features\n",
    "        self.forecast_size = forecast_size\n",
    "        self.backcast_size = backcast_size\n",
    " \n",
    "    def forward(self, theta: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:\n",
    "        backcast = theta[:, :self.backcast_size]\n",
    "        forecast = theta[:, self.backcast_size:]\n",
    "        forecast = forecast.reshape(len(forecast), -1, self.out_features)\n",
    "        return backcast, forecast\n",
    "\n",
    "class TrendBasis(nn.Module):\n",
    "    def __init__(self, degree_of_polynomial: int,\n",
    "                 backcast_size: int, forecast_size: int,\n",
    "                 out_features: int=1):\n",
    "        super().__init__()\n",
    "        self.out_features = out_features\n",
    "        polynomial_size = degree_of_polynomial + 1\n",
    "        self.backcast_basis = nn.Parameter(\n",
    "            torch.tensor(np.concatenate([np.power(np.arange(backcast_size, dtype=float) / backcast_size, i)[None, :]\n",
    "                                    for i in range(polynomial_size)]), dtype=torch.float32), requires_grad=False)\n",
    "        self.forecast_basis = nn.Parameter(\n",
    "            torch.tensor(np.concatenate([np.power(np.arange(forecast_size, dtype=float) / forecast_size, i)[None, :]\n",
    "                                    for i in range(polynomial_size)]), dtype=torch.float32), requires_grad=False)\n",
    "    \n",
    "    def forward(self, theta: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:\n",
    "        polynomial_size = self.forecast_basis.shape[0] # [polynomial_size, L+H]\n",
    "        backcast_theta = theta[:, :polynomial_size]\n",
    "        forecast_theta = theta[:, polynomial_size:]\n",
    "        forecast_theta = forecast_theta.reshape(len(forecast_theta),polynomial_size,-1)\n",
    "        backcast = torch.einsum('bp,pt->bt', backcast_theta, self.backcast_basis)\n",
    "        forecast = torch.einsum('bpq,pt->btq', forecast_theta, self.forecast_basis)\n",
    "        return backcast, forecast\n",
    "\n",
    "class SeasonalityBasis(nn.Module):\n",
    "    def __init__(self, harmonics: int, \n",
    "                 backcast_size: int, forecast_size: int,\n",
    "                 out_features: int=1):\n",
    "        super().__init__()\n",
    "        self.out_features = out_features\n",
    "        frequency = np.append(np.zeros(1, dtype=float),\n",
    "                                        np.arange(harmonics, harmonics / 2 * forecast_size,\n",
    "                                                    dtype=float) / harmonics)[None, :]\n",
    "        backcast_grid = -2 * np.pi * (\n",
    "                np.arange(backcast_size, dtype=float)[:, None] / forecast_size) * frequency\n",
    "        forecast_grid = 2 * np.pi * (\n",
    "                np.arange(forecast_size, dtype=float)[:, None] / forecast_size) * frequency\n",
    "\n",
    "        backcast_cos_template = torch.tensor(np.transpose(np.cos(backcast_grid)), dtype=torch.float32)\n",
    "        backcast_sin_template = torch.tensor(np.transpose(np.sin(backcast_grid)), dtype=torch.float32)\n",
    "        backcast_template = torch.cat([backcast_cos_template, backcast_sin_template], dim=0)\n",
    "\n",
    "        forecast_cos_template = torch.tensor(np.transpose(np.cos(forecast_grid)), dtype=torch.float32)\n",
    "        forecast_sin_template = torch.tensor(np.transpose(np.sin(forecast_grid)), dtype=torch.float32)\n",
    "        forecast_template = torch.cat([forecast_cos_template, forecast_sin_template], dim=0)\n",
    "\n",
    "        self.backcast_basis = nn.Parameter(backcast_template, requires_grad=False)\n",
    "        self.forecast_basis = nn.Parameter(forecast_template, requires_grad=False)\n",
    "\n",
    "    def forward(self, theta: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:\n",
    "        harmonic_size = self.forecast_basis.shape[0] # [harmonic_size, L+H]\n",
    "        backcast_theta = theta[:, :harmonic_size]\n",
    "        forecast_theta = theta[:, harmonic_size:]\n",
    "        forecast_theta = forecast_theta.reshape(len(forecast_theta),harmonic_size,-1)\n",
    "        backcast = torch.einsum('bp,pt->bt', backcast_theta, self.backcast_basis)\n",
    "        forecast = torch.einsum('bpq,pt->btq', forecast_theta, self.forecast_basis)\n",
    "        return backcast, forecast"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "17382790-7d84-4a89-959b-5676afa46392",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| exporti\n",
    "ACTIVATIONS = ['ReLU',\n",
    "               'Softplus',\n",
    "               'Tanh',\n",
    "               'SELU',\n",
    "               'LeakyReLU',\n",
    "               'PReLU',\n",
    "               'Sigmoid']\n",
    "\n",
    "class NBEATSBlock(nn.Module):\n",
    "    \"\"\"\n",
    "    N-BEATS block which takes a basis function as an argument.\n",
    "    \"\"\"\n",
    "    def __init__(self, \n",
    "                 input_size: int,\n",
    "                 n_theta: int, \n",
    "                 mlp_units: list,\n",
    "                 basis: nn.Module, \n",
    "                 dropout_prob: float, \n",
    "                 activation: str):\n",
    "        \"\"\"\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "\n",
    "        self.dropout_prob = dropout_prob\n",
    "        \n",
    "        assert activation in ACTIVATIONS, f'{activation} is not in {ACTIVATIONS}'\n",
    "        activ = getattr(nn, activation)()\n",
    "        \n",
    "        hidden_layers = [nn.Linear(in_features=input_size, \n",
    "                                   out_features=mlp_units[0][0])]\n",
    "        for layer in mlp_units:\n",
    "            hidden_layers.append(nn.Linear(in_features=layer[0], \n",
    "                                           out_features=layer[1]))\n",
    "            hidden_layers.append(activ)\n",
    "\n",
    "            if self.dropout_prob>0:\n",
    "                raise NotImplementedError('dropout')\n",
    "                #hidden_layers.append(nn.Dropout(p=self.dropout_prob))\n",
    "\n",
    "        output_layer = [nn.Linear(in_features=mlp_units[-1][1], out_features=n_theta)]\n",
    "        layers = hidden_layers + output_layer\n",
    "        self.layers = nn.Sequential(*layers)\n",
    "        self.basis = basis\n",
    "\n",
    "    def forward(self, insample_y: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:\n",
    "        # Compute local projection weights and projection\n",
    "        theta = self.layers(insample_y)\n",
    "        backcast, forecast = self.basis(theta)\n",
    "        return backcast, forecast"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "be997aeb-778f-442d-a97a-ff47de2deab6",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| export\n",
    "class NBEATS(BaseWindows):\n",
    "    \"\"\" NBEATS\n",
    "\n",
    "    The Neural Basis Expansion Analysis for Time Series (NBEATS), is a simple and yet\n",
    "    effective architecture, it is built with a deep stack of MLPs with the doubly \n",
    "    residual connections. It has a generic and interpretable architecture depending\n",
    "    on the blocks it uses. Its interpretable architecture is recommended for scarce\n",
    "    data settings, as it regularizes its predictions through projections unto harmonic\n",
    "    and trend basis well-suited for most forecasting tasks.\n",
    "\n",
    "    **Parameters:**<br>\n",
    "    `h`: int, forecast horizon.<br>\n",
    "    `input_size`: int, considered autorregresive inputs (lags), y=[1,2,3,4] input_size=2 -> lags=[1,2].<br>\n",
    "    `n_harmonics`: int, Number of harmonic terms for seasonality stack type. Note that len(n_harmonics) = len(stack_types). Note that it will only be used if a seasonality stack is used.<br>\n",
    "    `n_polynomials`: int, polynomial degree for trend stack. Note that len(n_polynomials) = len(stack_types). Note that it will only be used if a trend stack is used.<br>\n",
    "    `stack_types`: List[str], List of stack types. Subset from ['seasonality', 'trend', 'identity'].<br>\n",
    "    `n_blocks`: List[int], Number of blocks for each stack. Note that len(n_blocks) = len(stack_types).<br>\n",
    "    `mlp_units`: List[List[int]], Structure of hidden layers for each stack type. Each internal list should contain the number of units of each hidden layer. Note that len(n_hidden) = len(stack_types).<br>\n",
    "    `dropout_prob_theta`: float, Float between (0, 1). Dropout for N-BEATS basis.<br>\n",
    "    `shared_weights`: bool, If True, all blocks within each stack will share parameters. <br>\n",
    "    `activation`: str, activation from ['ReLU', 'Softplus', 'Tanh', 'SELU', 'LeakyReLU', 'PReLU', 'Sigmoid'].<br>\n",
    "    `loss`: PyTorch module, instantiated train loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).<br>\n",
    "    `valid_loss`: PyTorch module=`loss`, instantiated valid loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).<br>\n",
    "    `max_steps`: int=1000, maximum number of training steps.<br>\n",
    "    `learning_rate`: float=1e-3, Learning rate between (0, 1).<br>\n",
    "    `num_lr_decays`: int=3, Number of learning rate decays, evenly distributed across max_steps.<br>\n",
    "    `early_stop_patience_steps`: int=-1, Number of validation iterations before early stopping.<br>\n",
    "    `val_check_steps`: int=100, Number of training steps between every validation loss check.<br>\n",
    "    `batch_size`: int=32, number of different series in each batch.<br>\n",
    "    `valid_batch_size`: int=None, number of different series in each validation and test batch, if None uses batch_size.<br>\n",
    "    `windows_batch_size`: int=1024, number of windows to sample in each training batch, default uses all.<br>\n",
    "    `inference_windows_batch_size`: int=-1, number of windows to sample in each inference batch, -1 uses all.<br>\n",
    "    `start_padding_enabled`: bool=False, if True, the model will pad the time series with zeros at the beginning, by input size.<br>\n",
    "    `step_size`: int=1, step size between each window of temporal data.<br>\n",
    "    `scaler_type`: str='identity', type of scaler for temporal inputs normalization see [temporal scalers](https://nixtla.github.io/neuralforecast/common.scalers.html).<br>\n",
    "    `random_seed`: int, random_seed for pytorch initializer and numpy generators.<br>\n",
    "    `num_workers_loader`: int=os.cpu_count(), workers to be used by `TimeSeriesDataLoader`.<br>\n",
    "    `drop_last_loader`: bool=False, if True `TimeSeriesDataLoader` drops last non-full batch.<br>\n",
    "    `alias`: str, optional,  Custom name of the model.<br>\n",
    "    `optimizer`: Subclass of 'torch.optim.Optimizer', optional, user specified optimizer instead of the default choice (Adam).<br>\n",
    "    `optimizer_kwargs`: dict, optional, list of parameters used by the user specified `optimizer`.<br>\n",
    "    `**trainer_kwargs`: int,  keyword trainer arguments inherited from [PyTorch Lighning's trainer](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.trainer.trainer.Trainer.html?highlight=trainer).<br>\n",
    "\n",
    "    **References:**<br>\n",
    "    -[Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, Yoshua Bengio (2019). \n",
    "    \"N-BEATS: Neural basis expansion analysis for interpretable time series forecasting\".](https://arxiv.org/abs/1905.10437)\n",
    "    \"\"\"\n",
    "    # Class attributes\n",
    "    SAMPLING_TYPE = 'windows'\n",
    "    \n",
    "    def __init__(self,\n",
    "                 h,\n",
    "                 input_size,\n",
    "                 n_harmonics: int = 2,\n",
    "                 n_polynomials: int = 2,\n",
    "                 stack_types: list = ['identity', 'trend', 'seasonality'],\n",
    "                 n_blocks: list = [1, 1, 1],\n",
    "                 mlp_units: list = 3 * [[512, 512]],\n",
    "                 dropout_prob_theta: float = 0.,\n",
    "                 activation: str = 'ReLU',\n",
    "                 shared_weights: bool = False,                 \n",
    "                 loss = MAE(),\n",
    "                 valid_loss = None,\n",
    "                 max_steps: int = 1000,\n",
    "                 learning_rate: float = 1e-3,\n",
    "                 num_lr_decays: int = 3,\n",
    "                 early_stop_patience_steps: int =-1,\n",
    "                 val_check_steps: int = 100,\n",
    "                 batch_size: int = 32,\n",
    "                 valid_batch_size: Optional[int] = None,\n",
    "                 windows_batch_size: int = 1024,\n",
    "                 inference_windows_batch_size: int = -1,\n",
    "                 start_padding_enabled = False,\n",
    "                 step_size: int = 1,\n",
    "                 scaler_type: str ='identity',\n",
    "                 random_seed: int = 1,\n",
    "                 num_workers_loader: int = 0,\n",
    "                 drop_last_loader: bool = False,\n",
    "                 optimizer = None,\n",
    "                 optimizer_kwargs = None,\n",
    "                 **trainer_kwargs):\n",
    "        \n",
    "        # Protect horizon collapsed seasonality and trend NBEATSx-i basis\n",
    "        if h == 1 and ((\"seasonality\" in stack_types) or (\"trend\" in stack_types)):\n",
    "            raise Exception(\n",
    "                \"Horizon `h=1` incompatible with `seasonality` or `trend` in stacks\"\n",
    "            )\n",
    "\n",
    "        # Inherit BaseWindows class\n",
    "        super(NBEATS, self).__init__(h=h,\n",
    "                                     input_size=input_size,\n",
    "                                     loss=loss,\n",
    "                                     valid_loss=valid_loss,\n",
    "                                     max_steps=max_steps,\n",
    "                                     learning_rate=learning_rate,\n",
    "                                     num_lr_decays=num_lr_decays,\n",
    "                                     early_stop_patience_steps=early_stop_patience_steps,\n",
    "                                     val_check_steps=val_check_steps,\n",
    "                                     batch_size=batch_size,\n",
    "                                     windows_batch_size=windows_batch_size,\n",
    "                                     valid_batch_size=valid_batch_size,\n",
    "                                     inference_windows_batch_size=inference_windows_batch_size,\n",
    "                                     start_padding_enabled=start_padding_enabled,\n",
    "                                     step_size=step_size,\n",
    "                                     scaler_type=scaler_type,\n",
    "                                     num_workers_loader=num_workers_loader,\n",
    "                                     drop_last_loader=drop_last_loader,\n",
    "                                     random_seed=random_seed,\n",
    "                                     optimizer=optimizer,\n",
    "                                     optimizer_kwargs=optimizer_kwargs,\n",
    "                                     **trainer_kwargs)\n",
    "\n",
    "        # Architecture\n",
    "        blocks = self.create_stack(h=h,\n",
    "                                   input_size=input_size,\n",
    "                                   stack_types=stack_types, \n",
    "                                   n_blocks=n_blocks,\n",
    "                                   mlp_units=mlp_units,\n",
    "                                   dropout_prob_theta=dropout_prob_theta,\n",
    "                                   activation=activation,\n",
    "                                   shared_weights=shared_weights,\n",
    "                                   n_polynomials=n_polynomials, \n",
    "                                   n_harmonics=n_harmonics)\n",
    "        self.blocks = torch.nn.ModuleList(blocks)\n",
    "\n",
    "    def create_stack(self, stack_types, \n",
    "                     n_blocks, \n",
    "                     input_size, \n",
    "                     h, \n",
    "                     mlp_units, \n",
    "                     dropout_prob_theta, \n",
    "                     activation, shared_weights,\n",
    "                     n_polynomials, n_harmonics):                     \n",
    "\n",
    "        block_list = []\n",
    "        for i in range(len(stack_types)):\n",
    "            for block_id in range(n_blocks[i]):\n",
    "\n",
    "                # Shared weights\n",
    "                if shared_weights and block_id>0:\n",
    "                    nbeats_block = block_list[-1]\n",
    "                else:\n",
    "                    if stack_types[i] == 'seasonality':\n",
    "                        n_theta = 2 * (self.loss.outputsize_multiplier + 1) * \\\n",
    "                                  int(np.ceil(n_harmonics / 2 * h) - (n_harmonics - 1))\n",
    "                        basis = SeasonalityBasis(harmonics=n_harmonics,\n",
    "                                                 backcast_size=input_size,forecast_size=h,\n",
    "                                                 out_features=self.loss.outputsize_multiplier)\n",
    "\n",
    "                    elif stack_types[i] == 'trend':\n",
    "                        n_theta = (self.loss.outputsize_multiplier + 1) * (n_polynomials + 1)\n",
    "                        basis = TrendBasis(degree_of_polynomial=n_polynomials,\n",
    "                                           backcast_size=input_size,forecast_size=h,\n",
    "                                           out_features=self.loss.outputsize_multiplier)\n",
    "\n",
    "                    elif stack_types[i] == 'identity':\n",
    "                        n_theta = input_size + self.loss.outputsize_multiplier * h\n",
    "                        basis = IdentityBasis(backcast_size=input_size, forecast_size=h,\n",
    "                                              out_features=self.loss.outputsize_multiplier)\n",
    "                    else:\n",
    "                        raise ValueError(f'Block type {stack_types[i]} not found!')\n",
    "\n",
    "                    nbeats_block = NBEATSBlock(input_size=input_size,\n",
    "                                               n_theta=n_theta,\n",
    "                                               mlp_units=mlp_units,\n",
    "                                               basis=basis,\n",
    "                                               dropout_prob=dropout_prob_theta,\n",
    "                                               activation=activation)\n",
    "\n",
    "                # Select type of evaluation and apply it to all layers of block\n",
    "                block_list.append(nbeats_block)\n",
    "                \n",
    "        return block_list\n",
    "\n",
    "    def forward(self, windows_batch):\n",
    "        \n",
    "        # Parse windows_batch\n",
    "        insample_y    = windows_batch['insample_y']\n",
    "        insample_mask = windows_batch['insample_mask']\n",
    "\n",
    "        # NBEATS' forward\n",
    "        residuals = insample_y.flip(dims=(-1,)) # backcast init\n",
    "        insample_mask = insample_mask.flip(dims=(-1,))\n",
    "        \n",
    "        forecast = insample_y[:, -1:, None] # Level with Naive1\n",
    "        block_forecasts = [ forecast.repeat(1, self.h, 1) ]\n",
    "        for i, block in enumerate(self.blocks):\n",
    "            backcast, block_forecast = block(insample_y=residuals)\n",
    "            residuals = (residuals - backcast) * insample_mask\n",
    "            forecast = forecast + block_forecast\n",
    "\n",
    "            if self.decompose_forecast:\n",
    "                block_forecasts.append(block_forecast)\n",
    "\n",
    "        # Adapting output's domain\n",
    "        forecast = self.loss.domain_map(forecast)                \n",
    "\n",
    "        if self.decompose_forecast:\n",
    "            # (n_batch, n_blocks, h, out_features)\n",
    "            block_forecasts = torch.stack(block_forecasts)\n",
    "            block_forecasts = block_forecasts.permute(1,0,2,3)\n",
    "            block_forecasts = block_forecasts.squeeze(-1) # univariate output\n",
    "            return block_forecasts\n",
    "        else:\n",
    "            return forecast"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c57a831f-94bc-4616-b579-c114c3fc57c7",
   "metadata": {},
   "outputs": [],
   "source": [
    "show_doc(NBEATS)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f9013b63-f65b-4a92-913c-b696e6e69914",
   "metadata": {},
   "outputs": [],
   "source": [
    "show_doc(NBEATS.fit, name='NBEATS.fit')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a66184ee-7a71-4598-976c-c79b83089a6f",
   "metadata": {},
   "outputs": [],
   "source": [
    "show_doc(NBEATS.predict, name='NBEATS.predict')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c2bf3e1d-2935-4503-afb6-fe3d6f52622c",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "import logging\n",
    "import warnings\n",
    "\n",
    "import pandas as pd\n",
    "import pytorch_lightning as pl\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from neuralforecast.tsdataset import TimeSeriesDataset, TimeSeriesLoader\n",
    "from neuralforecast.utils import AirPassengersDF as Y_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6bb4c6c6-ef60-47c9-8c90-4002e68410d3",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "Y_train_df = Y_df[Y_df.ds<Y_df['ds'].values[-12]] # 132 train\n",
    "Y_test_df = Y_df[Y_df.ds>=Y_df['ds'].values[-12]]   # 12 test\n",
    "\n",
    "dataset, *_ = TimeSeriesDataset.from_df(df = Y_train_df)\n",
    "nbeats = NBEATS(h=12, input_size=24, windows_batch_size=None, \n",
    "                stack_types=['identity', 'trend', 'seasonality'], max_steps=1)\n",
    "nbeats.fit(dataset=dataset)\n",
    "y_hat = nbeats.predict(dataset=dataset)\n",
    "Y_test_df['N-BEATS'] = y_hat\n",
    "\n",
    "pd.concat([Y_train_df, Y_test_df]).drop('unique_id', axis=1).set_index('ds').plot()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "db94b63e-d82c-423f-8f75-184ae285904d",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "#test we recover the same forecast\n",
    "y_hat2 = nbeats.predict(dataset=dataset)\n",
    "test_eq(y_hat, y_hat2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "46090447-8e67-4f08-8a3d-9547183983f9",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "#test no leakage with test_size\n",
    "dataset, *_ = TimeSeriesDataset.from_df(Y_df)\n",
    "model = NBEATS(input_size=24, h=12, \n",
    "               windows_batch_size=None, max_steps=1)\n",
    "model.fit(dataset=dataset, test_size=12)\n",
    "y_hat_test = model.predict(dataset=dataset, step_size=1)\n",
    "np.testing.assert_almost_equal(y_hat, y_hat_test, decimal=4)\n",
    "#test we recover the same forecast\n",
    "y_hat_test2 = model.predict(dataset=dataset, step_size=1)\n",
    "test_eq(y_hat_test, y_hat_test2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0298fce5-eb13-40dc-9964-b026fd2a8928",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "# test validation step\n",
    "dataset, *_ = TimeSeriesDataset.from_df(Y_train_df)\n",
    "model = NBEATS(input_size=24, h=12, windows_batch_size=None, max_steps=1)\n",
    "model.fit(dataset=dataset, val_size=12)\n",
    "y_hat_w_val = model.predict(dataset=dataset)\n",
    "Y_test_df['N-BEATS'] = y_hat_w_val\n",
    "\n",
    "pd.concat([Y_train_df, Y_test_df]).drop('unique_id', axis=1).set_index('ds').plot()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6f987ed0-ee6e-4f66-bd8f-96acc6fbd56c",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "# test no leakage with test_size and val_size\n",
    "dataset, *_ = TimeSeriesDataset.from_df(Y_train_df)\n",
    "model = NBEATS(input_size=24, h=12, windows_batch_size=None, max_steps=1)\n",
    "model.fit(dataset=dataset, val_size=12)\n",
    "y_hat_w_val = model.predict(dataset=dataset)\n",
    "\n",
    "dataset, *_ = TimeSeriesDataset.from_df(Y_df)\n",
    "model = NBEATS(input_size=24, h=12, windows_batch_size=None, max_steps=1)\n",
    "model.fit(dataset=dataset, val_size=12, test_size=12)\n",
    "\n",
    "y_hat_test_w_val = model.predict(dataset=dataset, step_size=1)\n",
    "\n",
    "np.testing.assert_almost_equal(y_hat_test_w_val, y_hat_w_val, decimal=4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ba4e41b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| hide\n",
    "# qualitative decomposition evaluation\n",
    "y_hat = model.decompose(dataset=dataset)\n",
    "\n",
    "fig, ax = plt.subplots(5, 1, figsize=(10, 15))\n",
    "\n",
    "ax[0].plot(Y_test_df['y'].values, label='True', color=\"#9C9DB2\", linewidth=4)\n",
    "ax[0].plot(y_hat.sum(axis=1).flatten(), label='Forecast', color=\"#7B3841\")\n",
    "ax[0].grid()\n",
    "ax[0].legend(prop={'size': 20})\n",
    "for label in (ax[0].get_xticklabels() + ax[0].get_yticklabels()):\n",
    "    label.set_fontsize(18)\n",
    "ax[0].set_ylabel('y', fontsize=20)\n",
    "\n",
    "ax[1].plot(y_hat[0,0], label='level', color=\"#7B3841\")\n",
    "ax[1].grid()\n",
    "ax[1].set_ylabel('Level', fontsize=20)\n",
    "\n",
    "ax[2].plot(y_hat[0,1], label='stack1', color=\"#7B3841\")\n",
    "ax[2].grid()\n",
    "ax[2].set_ylabel('Identity', fontsize=20)\n",
    "\n",
    "ax[3].plot(y_hat[0,2], label='stack2', color=\"#D9AE9E\")\n",
    "ax[3].grid()\n",
    "ax[3].set_ylabel('Trend', fontsize=20)\n",
    "\n",
    "ax[4].plot(y_hat[0,3], label='stack3', color=\"#D9AE9E\")\n",
    "ax[4].grid()\n",
    "ax[4].set_ylabel('Seasonality', fontsize=20)\n",
    "\n",
    "ax[4].set_xlabel('Prediction \\u03C4 \\u2208 {t+1,..., t+H}', fontsize=20)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "cdc17eef",
   "metadata": {},
   "source": [
    "## Usage Example"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3017c43a",
   "metadata": {},
   "outputs": [],
   "source": [
    "#| eval: false\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import pytorch_lightning as pl\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from neuralforecast import NeuralForecast\n",
    "from neuralforecast.models import NBEATS\n",
    "from neuralforecast.losses.pytorch import MQLoss, DistributionLoss\n",
    "from neuralforecast.tsdataset import TimeSeriesDataset\n",
    "from neuralforecast.utils import AirPassengers, AirPassengersPanel, AirPassengersStatic\n",
    "\n",
    "Y_train_df = AirPassengersPanel[AirPassengersPanel.ds<AirPassengersPanel['ds'].values[-12]] # 132 train\n",
    "Y_test_df = AirPassengersPanel[AirPassengersPanel.ds>=AirPassengersPanel['ds'].values[-12]].reset_index(drop=True) # 12 test\n",
    "\n",
    "model = NBEATS(h=12, input_size=24,\n",
    "               loss=DistributionLoss(distribution='Poisson', level=[80, 90]),\n",
    "               stack_types = ['identity', 'trend', 'seasonality'],\n",
    "               max_steps=100,\n",
    "               val_check_steps=10,\n",
    "               early_stop_patience_steps=2)\n",
    "\n",
    "fcst = NeuralForecast(\n",
    "    models=[model],\n",
    "    freq='M'\n",
    ")\n",
    "fcst.fit(df=Y_train_df, static_df=AirPassengersStatic, val_size=12)\n",
    "forecasts = fcst.predict(futr_df=Y_test_df)\n",
    "\n",
    "# Plot quantile predictions\n",
    "Y_hat_df = forecasts.reset_index(drop=False).drop(columns=['unique_id','ds'])\n",
    "plot_df = pd.concat([Y_test_df, Y_hat_df], axis=1)\n",
    "plot_df = pd.concat([Y_train_df, plot_df])\n",
    "\n",
    "plot_df = plot_df[plot_df.unique_id=='Airline1'].drop('unique_id', axis=1)\n",
    "plt.plot(plot_df['ds'], plot_df['y'], c='black', label='True')\n",
    "plt.plot(plot_df['ds'], plot_df['NBEATS-median'], c='blue', label='median')\n",
    "plt.fill_between(x=plot_df['ds'][-12:], \n",
    "                 y1=plot_df['NBEATS-lo-90'][-12:].values, \n",
    "                 y2=plot_df['NBEATS-hi-90'][-12:].values,\n",
    "                 alpha=0.4, label='level 90')\n",
    "plt.grid()\n",
    "plt.legend()\n",
    "plt.plot()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d7cbd9ad",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "python3",
   "language": "python",
   "name": "python3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}