Reproducible example for custom subsampling with LightGBM

custom_subsampling.py

import lightgbm as lgb
import numpy as np
import pandas as pd
from sklearn.metrics import mean_squared_error

def custom_subsample(n, frac):
    """Subsample frac*n indices."""
    return np.random.choice(n, int(n*frac), replace=False)

# load data
print('Load data...')
data_dir = '..'
df_train = pd.read_csv('{}/regression/regression.train'.format(data_dir), header=None, sep='\t')
df_test = pd.read_csv('{}/regression/regression.test'.format(data_dir), header=None, sep='\t')

y_train = df_train[0]
y_test = df_test[0]
X_train = df_train.drop(0, axis=1)
X_test = df_test.drop(0, axis=1)

"""
Built-in subsampling
"""
params = dict(
    boosting_type="gbdt",
    num_leaves=31,
    max_depth=-1,
    learning_rate=0.01,
    max_bin=255,
    objective="regression",
    subsample=0.25,
    subsample_freq=1,
)
# Create dataset
train_set = lgb.Dataset(X_train, label=y_train, params=params)
# Create Booster object
booster = lgb.Booster(params=params, train_set=train_set)
# Train
for _ in range(100):
    booster.update()
y_pred_builtin = booster.predict(X_test)
print('The rmse of built-in prediction is:', mean_squared_error(y_test, y_pred_builtin) ** 0.5)


"""
Custom subsampling
"""
params = dict(
    boosting_type="gbdt",
    num_leaves=31,
    max_depth=-1,
    learning_rate=0.01,
    max_bin=255,
    objective="regression",
    subsample=1.,
    subsample_freq=0,
)
# Create dataset
train_set = lgb.Dataset(X_train, label=y_train, params=params)
# Create Booster object
booster = lgb.Booster(params=params, train_set=train_set)
# Train
ts = []
for _ in range(100):
    subsample = custom_subsample(X_train.shape[0], frac=0.25)
    ts.append(train_set.subset(subsample))
    booster.update(ts[-1])
y_pred_custom = booster.predict(X_test)
print('The rmse of custom prediction is:', mean_squared_error(y_test, y_pred_custom) ** 0.5)

print('The rmse between the two predictions is:', mean_squared_error(y_pred_builtin, y_pred_custom) ** 0.5)