Skip to content

Latest commit

 

History

History
70 lines (66 loc) · 3.61 KB

quick_start.md

File metadata and controls

70 lines (66 loc) · 3.61 KB

Quick Start

First, install PGBM.

Don't care about GPU training? Use the Scikit-learn backend, as it is faster when using CPU.

Training a model on Torch backend

For example, for the California Housing dataset:

from pgbm.torch import PGBMRegressor
from sklearn.model_selection import train_test_split
from sklearn.datasets import fetch_california_housing
X, y = fetch_california_housing(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=0)
model = PGBMRegressor().fit(X_train, y_train)

Point predictions You can make point predictions by calling the predict attribute of a PGBM or PGBMRegressor model:

yhat_point = model.predict(X_test)

Probabilistic predictions You can make probabilistic predictions by calling the predict_dist attribute of a PGBM or PGBMRegressor model. By default, 100 forecasts will be created for each sample using a normal distribution (see also Parameters).

yhat_dist = model.predict_dist(X_test)

Plotting predictions We can plot the results for both the point and probabilistic estimates for the first 100 samples of the test set:

import matplotlib.pyplot as plt
import numpy as np
n_samples_plot = 100
plt.rcParams.update({'font.size': 20})
plt.plot(y_test[:n_samples_plot], 'o', label='Actual')
plt.plot(yhat_point[:n_samples_plot], 'ko', label='Point prediction PGBM')
plt.fill_between(np.arange(len(y_test[:n_samples_plot])), 
                 yhat_dist.min(0)[:n_samples_plot], 
                 yhat_dist.max(0)[:n_samples_plot], 
                 color="#b9cfe7", label='Uncertainty')
plt.title('California Housing - Probabilistic Prediction with PGBM')
plt.xlabel('Sample')
plt.ylabel('Prediction')
plt.legend(ncol=3)

which should produce something similar to the following: PGBM Quick Start - California Housing Probabilistic Predictions As you can see, the probabilistic forecast covers most of the actual observed values, which is typically desired. Moreover, the indicated forecast range varies per test point, indicating that the model has learned to produce different variances for various samples. Again, this is typically desired behaviour when forecasting. We can improve on this result through hyperparameter tuning.

Training a model on scikit-learn backend

For example, for the California Housing dataset:

from pgbm.sklearn import HistGradientBoostingRegressor, crps_ensemble
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import train_test_split
from sklearn.datasets import fetch_california_housing
import numpy as np
X, y = fetch_california_housing(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=0)
model = HistGradientBoostingRegressor(random_state=0).fit(X_train, y_train)

Point predictions You can make point predictions by calling the predict attribute of a HistGradientBoostingRegressor model, and we also return the standard deviation of our predictions by setting return_std=True:

yhat_point, yhat_point_std = model.predict(X_test, return_std=True)

Probabilistic predictions You can make probabilistic predictions by calling the sample attribute of a HistGradientBoostingRegressor model. By default, 1 forecast will be created for each sample using a normal distribution (see also Parameters).

yhat_dist = model.sample(yhat_point, yhat_point_std, n_estimates=1000, random_state=0)