End to End Walkthrough | NeuralForecast

%%capture
! pip install neuralforecast

import pandas as pd

Y_df = pd.read_parquet('https://datasets-nixtla.s3.amazonaws.com/m4-hourly.parquet')
Y_df.head()

uids = Y_df['unique_id'].unique()[:10] # Select 10 ids to make the example faster
Y_df = Y_df.query('unique_id in @uids').reset_index(drop=True)

from utilsforecast.plotting import plot_series

plot_series(Y_df)

import logging

import optuna
import ray.tune as tune
import torch

from neuralforecast import NeuralForecast
from neuralforecast.auto import AutoNHITS, AutoLSTM
from neuralforecast.losses.pytorch import MQLoss

optuna.logging.set_verbosity(optuna.logging.WARNING)
logging.getLogger('pytorch_lightning').setLevel(logging.ERROR)
torch.set_float32_matmul_precision('high')

def config_nhits(trial):
    return {
        "input_size": trial.suggest_categorical(          # Length of input window
            "input_size", (48, 48*2, 48*3)                
        ),                                                
        "start_padding_enabled": True,                                          
        "n_blocks": 5 * [1],                              # Length of input window
        "mlp_units": 5 * [[64, 64]],                      # Length of input window
        "n_pool_kernel_size": trial.suggest_categorical(  # MaxPooling Kernel size
            "n_pool_kernel_size",
            (5*[1], 5*[2], 5*[4], [8, 4, 2, 1, 1])
        ),     
        "n_freq_downsample": trial.suggest_categorical(   # Interpolation expressivity ratios
            "n_freq_downsample",
            ([8, 4, 2, 1, 1],  [1, 1, 1, 1, 1])
        ),     
        "learning_rate": trial.suggest_float(             # Initial Learning rate
            "learning_rate",
            low=1e-4,
            high=1e-2,
            log=True,
        ),            
        "scaler_type": None,                              # Scaler type
        "max_steps": 1000,                                # Max number of training iterations
        "batch_size": trial.suggest_categorical(          # Number of series in batch
            "batch_size",
            (1, 4, 10),
        ),                   
        "windows_batch_size": trial.suggest_categorical(  # Number of windows in batch
            "windows_batch_size",
            (128, 256, 512),
        ),      
        "random_seed": trial.suggest_int(                 # Random seed   
            "random_seed",
            low=1,
            high=20,
        ),                      
    }

def config_lstm(trial):
    return {
        "input_size": trial.suggest_categorical(           # Length of input window
            "input_size",
            (48, 48*2, 48*3)
        ),   
        "encoder_hidden_size": trial.suggest_categorical(  # Hidden size of LSTM cells
            "encoder_hidden_size",
            (64, 128),
        ),  
        "encoder_n_layers": trial.suggest_categorical(     # Number of layers in LSTM
            "encoder_n_layers",
            (2,4),
        ),        
        "learning_rate": trial.suggest_float(              # Initial Learning rate
            "learning_rate",
            low=1e-4,
            high=1e-2,
            log=True,
        ),   
        "scaler_type": 'robust',                           # Scaler type
        "max_steps": trial.suggest_categorical(           # Max number of training iterations
            "max_steps",
            (500, 1000)
        ),          
        "batch_size": trial.suggest_categorical(           # Number of series in batch
            "batch_size",
            (1, 4)
        ),              
        "random_seed": trial.suggest_int(                  # Random seed
            "random_seed",
            low=1,
            high=20
        ),             
    }

nf = NeuralForecast(
    models=[
        AutoNHITS(h=48, config=config_nhits, loss=MQLoss(), backend='optuna', num_samples=5),
        AutoLSTM(h=48, config=config_lstm, loss=MQLoss(), backend='optuna', num_samples=2),
    ],
    freq=1,
)

%%capture
nf.fit(df=Y_df)

fcst_df = nf.predict()
fcst_df.columns = fcst_df.columns.str.replace('-median', '')
fcst_df.head()

plot_series(Y_df, fcst_df, plot_random=False, max_insample_length=48 * 3, level=[80, 90])

# Plot to unique_ids and some selected models
plot_series(Y_df, fcst_df, models=["AutoLSTM"], ids=["H107", "H104"], level=[80, 90])

# Explore other models 
plot_series(Y_df, fcst_df, models=["AutoNHITS"], ids=["H10", "H105"], level=[80, 90])

from neuralforecast.auto import AutoNHITS, AutoLSTM

nf = NeuralForecast(
    models=[
        AutoNHITS(h=48, config=config_nhits, loss=MQLoss(), num_samples=5, backend="optuna"),
        AutoLSTM(h=48, config=config_lstm, loss=MQLoss(), num_samples=2, backend="optuna"), 
    ],
    freq=1,
)

%%capture
cv_df = nf.cross_validation(Y_df, n_windows=2)

cv_df.columns = cv_df.columns.str.replace('-median', '')

cv_df.head()

from IPython.display import display

for cutoff in cv_df['cutoff'].unique():
    display(
        plot_series(
            Y_df,
            cv_df.query('cutoff == @cutoff').drop(columns=['y', 'cutoff']),
            max_insample_length=48 * 4, 
            ids=['H102'],
        )
    )

from utilsforecast.evaluation import evaluate
from utilsforecast.losses import mse, mae, rmse

evaluation_df = evaluate(cv_df.drop(columns='cutoff'), metrics=[mse, mae, rmse])
evaluation_df['best_model'] = evaluation_df.drop(columns=['metric', 'unique_id']).idxmin(axis=1)
evaluation_df.head()

summary_df = evaluation_df.groupby(['metric', 'best_model']).size().sort_values().to_frame()
summary_df = summary_df.reset_index()
summary_df.columns = ['metric', 'model', 'nr. of unique_ids']
summary_df

summary_df.query('metric == "mse"')

nhits_ids = evaluation_df.query('best_model == "AutoNHITS" and metric == "mse"')['unique_id'].unique()

plot_series(Y_df, fcst_df, ids=nhits_ids)

def get_best_model_forecast(forecasts_df, evaluation_df, metric):
    metric_eval = evaluation_df.loc[evaluation_df['metric'] == metric, ['unique_id', 'best_model']]
    with_best = forecasts_df.merge(metric_eval)
    res = with_best[['unique_id', 'ds']].copy()
    for suffix in ('', '-lo-90', '-hi-90'):
        res[f'best_model{suffix}'] = with_best.apply(lambda row: row[row['best_model'] + suffix], axis=1)
    return res

prod_forecasts_df = get_best_model_forecast(fcst_df, evaluation_df, metric='mse')
prod_forecasts_df

plot_series(Y_df, prod_forecasts_df, level=[90])