Profiling skforecast¶

This document shows the profiling of the main classes, methods and functions available in skforecast. Understanding the bottlenecks will help to:

• Use it more efficiently
• Improve the code for future releases

Libraries¶

In [1]:

# Libraries
# ==============================================================================
import time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from skforecast.ForecasterAutoreg import ForecasterAutoreg
from skforecast.ForecasterAutoregCustom import ForecasterAutoregCustom
from skforecast.ForecasterAutoregDirect import ForecasterAutoregDirect
from skforecast.model_selection import grid_search_forecaster
from skforecast.model_selection import backtesting_forecaster

from sklearn.linear_model import LinearRegression
from sklearn.linear_model import Ridge
from sklearn.ensemble import HistGradientBoostingRegressor
from lightgbm import LGBMRegressor

%load_ext pyinstrument

# Libraries # ============================================================================== import time import numpy as np import pandas as pd import matplotlib.pyplot as plt from skforecast.ForecasterAutoreg import ForecasterAutoreg from skforecast.ForecasterAutoregCustom import ForecasterAutoregCustom from skforecast.ForecasterAutoregDirect import ForecasterAutoregDirect from skforecast.model_selection import grid_search_forecaster from skforecast.model_selection import backtesting_forecaster from sklearn.linear_model import LinearRegression from sklearn.linear_model import Ridge from sklearn.ensemble import HistGradientBoostingRegressor from lightgbm import LGBMRegressor %load_ext pyinstrument

Data¶

A time series of length 1000 with random values is created.

In [2]:

# Data
# ==============================================================================
np.random.seed(123)
n = 1_000
data = pd.Series(data = np.random.normal(size=n))

# Data # ============================================================================== np.random.seed(123) n = 1_000 data = pd.Series(data = np.random.normal(size=n))

Dummy regressor¶

To isolate the training process of the regressor from the other parts of the code, a dummy regressor class is created. This dummy regressor has a fit method that does nothing, and a predict method that returns a constant value.

In [3]:

class DummyRegressor(LinearRegression):
    """
    Dummy regressor with dummy fit and predict methods.
    """
    
    def fit(self, X, y):
        pass

    def predict(self, y):
        predictions = np.ones(shape = len(y))
        return predictions

class DummyRegressor(LinearRegression): """ Dummy regressor with dummy fit and predict methods. """ def fit(self, X, y): pass def predict(self, y): predictions = np.ones(shape = len(y)) return predictions

Profiling fit¶

In [4]:

%%pyinstrument

forecaster = ForecasterAutoreg(
                 regressor = DummyRegressor(),
                 lags      = 24
             )

forecaster.fit(y=data)

%%pyinstrument forecaster = ForecasterAutoreg( regressor = DummyRegressor(), lags = 24 ) forecaster.fit(y=data)

c:\Users\jaesc2\Miniconda3\envs\skforecast_py11\Lib\site-packages\sklearn\preprocessing\_discretization.py:278: UserWarning: Feature 0 is constant and will be replaced with 0.
  warnings.warn(

Almost all of the time spent by fit is required by the create_train_X_y method.

In [5]:

%%pyinstrument

forecaster = ForecasterAutoreg(
                 regressor = HistGradientBoostingRegressor(max_iter=10, random_state=123),
                 lags      = 24
             )

forecaster.fit(y=data)

%%pyinstrument forecaster = ForecasterAutoreg( regressor = HistGradientBoostingRegressor(max_iter=10, random_state=123), lags = 24 ) forecaster.fit(y=data)

When training a forecaster with a real machine learning regressor, the time spent by create_train_X_y is negligible compared to the time needed by the fit method of the regressor. Therefore, improving the speed of create_train_X_y will not have much impact.

Profiling create_train_X_y¶

Understand how the create_train_X_y method is influenced by the length of the series and the number of lags.

In [6]:

# Profiling `create_train_X_y` for different length of series and number of lags
# ======================================================================================
series_length = np.linspace(1000, 1000000, num=5, dtype=int)
n_lags = [5, 10, 50, 100, 200]
results = {}

for lags in n_lags:
    execution_time = []
    forecaster = ForecasterAutoreg(
                     regressor = DummyRegressor(),
                     lags      = lags
                 )

    for n in series_length:
        y = pd.Series(data = np.random.normal(size=n))
        tic = time.perf_counter()
        _ = forecaster.create_train_X_y(y=y)
        toc = time.perf_counter()
        execution_time.append(toc-tic)

    results[lags] = execution_time

results = pd.DataFrame(
              data =  results,
              index = series_length
          )

results

# Profiling `create_train_X_y` for different length of series and number of lags # ====================================================================================== series_length = np.linspace(1000, 1000000, num=5, dtype=int) n_lags = [5, 10, 50, 100, 200] results = {} for lags in n_lags: execution_time = [] forecaster = ForecasterAutoreg( regressor = DummyRegressor(), lags = lags ) for n in series_length: y = pd.Series(data = np.random.normal(size=n)) tic = time.perf_counter() _ = forecaster.create_train_X_y(y=y) toc = time.perf_counter() execution_time.append(toc-tic) results[lags] = execution_time results = pd.DataFrame( data = results, index = series_length ) results

Out[6]:

	5	10	50	100	200
1000	0.001045	0.004704	0.007673	0.038916	0.069427
250750	0.007231	0.014726	0.213554	0.364044	0.561126
500500	0.023345	0.052342	0.427991	0.775216	1.522044
750250	0.037870	0.068540	0.666305	1.062016	2.148227
1000000	0.042683	0.092421	0.832870	1.425935	2.702640

In [7]:

fig, ax = plt.subplots(figsize=(7, 4))
results.plot(ax=ax, marker='.')
ax.set_xlabel('length of series')
ax.set_ylabel('time (seconds)')
ax.set_title('Profiling create_train_X_y()')
ax.legend(title='number of lags');

fig, ax = plt.subplots(figsize=(7, 4)) results.plot(ax=ax, marker='.') ax.set_xlabel('length of series') ax.set_ylabel('time (seconds)') ax.set_title('Profiling create_train_X_y()') ax.legend(title='number of lags');

Profiling predict¶

In [8]:

forecaster = ForecasterAutoreg(
                 regressor = DummyRegressor(),
                 lags      = 24
             )

forecaster.fit(y=data)

forecaster = ForecasterAutoreg( regressor = DummyRegressor(), lags = 24 ) forecaster.fit(y=data)

c:\Users\jaesc2\Miniconda3\envs\skforecast_py11\Lib\site-packages\sklearn\preprocessing\_discretization.py:278: UserWarning: Feature 0 is constant and will be replaced with 0.
  warnings.warn(

In [9]:

%%pyinstrument

_ = forecaster.predict(steps=1000)

%%pyinstrument _ = forecaster.predict(steps=1000)

In [10]:

forecaster = ForecasterAutoreg(
                 regressor = HistGradientBoostingRegressor(max_iter=10, random_state=123),
                 lags      = 24
             )

forecaster.fit(y=data)

forecaster = ForecasterAutoreg( regressor = HistGradientBoostingRegressor(max_iter=10, random_state=123), lags = 24 ) forecaster.fit(y=data)

In [11]:

%%pyinstrument

_ = forecaster.predict(steps=1000)

%%pyinstrument _ = forecaster.predict(steps=1000)

Inside the predict method, the append action is the most expensive but, similar to what happen with fit, it is negligible compared to the time need by the predict method of the regressor.