XGBoost、LightGBM、CatBoostにおけるGPU加速の効果を検証する

import time
import numpy as np
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

# データ生成（サンプル数: 80,000, 特徴量: 50）
X, y = make_classification(n_samples=80000, n_features=50, n_informative=30, n_redundant=10, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
print("データ生成完了: X_train={}, X_test={}".format(X_train.shape, X_test.shape))

import xgboost as xgb

# DMatrixに変換
dtrain = xgb.DMatrix(X_train, label=y_train)
dtest  = xgb.DMatrix(X_test, label=y_test)

# XGBoostパラメータ設定
params_cpu = {
    'objective': 'binary:logistic',
    'eval_metric': 'logloss'
}
params_gpu = {
    'objective': 'binary:logistic',
    'eval_metric': 'logloss',
    'device': 'cuda'
}
num_round = 100

# CPUでの学習時間計測
start = time.time()
bst_cpu = xgb.train(params_cpu, dtrain, num_boost_round=num_round, evals=[(dtest, 'test')], verbose_eval=False)
cpu_time = time.time() - start
print("XGBoost CPU training time: {:.4f} seconds".format(cpu_time))

# GPUでの学習時間計測
start = time.time()
bst_gpu = xgb.train(params_gpu, dtrain, num_boost_round=num_round, evals=[(dtest, 'test')], verbose_eval=False)
gpu_time = time.time() - start
print("XGBoost GPU training time: {:.4f} seconds".format(gpu_time))

import lightgbm as lgb

# LightGBM用データセットに変換
lgb_train = lgb.Dataset(X_train, label=y_train)
lgb_test  = lgb.Dataset(X_test, label=y_test, reference=lgb_train)

params_cpu = {
    'objective': 'binary',
    'metric': 'binary_logloss',
    'boosting_type': 'gbdt'
}
params_gpu = {
    'objective': 'binary',
    'metric': 'binary_logloss',
    'boosting_type': 'gbdt',
    'device': 'gpu'
}
num_round = 100

# CPUでの学習時間計測
start = time.time()
gbm_cpu = lgb.train(params_cpu, lgb_train, num_boost_round=num_round, valid_sets=[lgb_test])
cpu_time = time.time() - start
print("LightGBM CPU training time: {:.4f} seconds".format(cpu_time))

# GPUでの学習時間計測
start = time.time()
gbm_gpu = lgb.train(params_gpu, lgb_train, num_boost_round=num_round, valid_sets=[lgb_test])
gpu_time = time.time() - start
print("LightGBM GPU training time: {:.4f} seconds".format(gpu_time))

# サンプル数を800,000に増やしたデータセットの生成
X_large, y_large = make_classification(n_samples=800000, n_features=50, n_informative=30, n_redundant=10, random_state=42)
X_train_large, X_test_large, y_train_large, y_test_large = train_test_split(X_large, y_large, test_size=0.2, random_state=42)
print("大規模データ生成完了: X_train={}, X_test={}".format(X_train_large.shape, X_test_large.shape))

# LightGBM用の大規模データセットに変換
lgb_train_large = lgb.Dataset(X_train_large, label=y_train_large)
lgb_test_large  = lgb.Dataset(X_test_large, label=y_test_large, reference=lgb_train_large)

# CPUでの学習時間計測
start = time.time()
gbm_cpu_large = lgb.train(params_cpu, lgb_train_large, num_boost_round=num_round, valid_sets=[lgb_test_large])
cpu_time_large = time.time() - start
print("LightGBM CPU training time (large data): {:.4f} seconds".format(cpu_time_large))

# GPUでの学習時間計測
start = time.time()
gbm_gpu_large = lgb.train(params_gpu, lgb_train_large, num_boost_round=num_round, valid_sets=[lgb_test_large])
gpu_time_large = time.time() - start
print("LightGBM GPU training time (large data): {:.4f} seconds".format(gpu_time_large))

from catboost import CatBoostClassifier

params_cpu = {
    'iterations': 100,
    'loss_function': 'Logloss',
    'verbose': False,
    'task_type': 'CPU'
}
params_gpu = {
    'iterations': 100,
    'loss_function': 'Logloss',
    'verbose': False,
    'task_type': 'GPU'
}

# CPUでの学習時間計測
start = time.time()
model_cpu = CatBoostClassifier(**params_cpu)
model_cpu.fit(X_train, y_train, eval_set=(X_test, y_test))
cpu_time = time.time() - start
print("CatBoost CPU training time: {:.4f} seconds".format(cpu_time))

# GPUでの学習時間計測
start = time.time()
model_gpu = CatBoostClassifier(**params_gpu)
model_gpu.fit(X_train, y_train, eval_set=(X_test, y_test))
gpu_time = time.time() - start
print("CatBoost GPU training time: {:.4f} seconds".format(gpu_time))