使用RAPIDS API调用grid_search.fit()时触发隐式转换异常求助

环境信息 #

• Win10 + WSL2
• RAPIDS 24.02

报错信息 #

不允许通过array隐式转换为主机端NumPy数组，若要显式构建GPU矩阵，请考虑使用.to_cupy()；若要显式构建主机端矩阵，请考虑使用.to_numpy()。

问题代码 #

import cudf
import cupy as cp
import numpy as np
from cuml.datasets import make_classification
from cuml.svm import SVC
from cuml.preprocessing import StandardScaler
from cuml.model_selection import GridSearchCV

# Set seeds
cp.random.seed(42)
np.random.seed(42)

# Generate synthetic classification dataset (CuPy format)
X_cupy, y_cupy = make_classification(n_samples=500, n_features=10,
                                     n_informative=5, n_classes=2,
                                     random_state=42)

# Convert to cuDF directly without using .asnumpy()
X_cudf = cudf.DataFrame({f'feat_{i}': X_cupy[:, i] for i in range(X_cupy.shape[1])})
y_cudf = cudf.Series(y_cupy)

# Use cuML StandardScaler (keeps everything on GPU)
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X_cudf)  # stays cuDF

# Define cuML SVC
svc = SVC(class_weight='balanced')

# Define parameter grid
param_grid = {
    'C': [0.1, 1.0],
    'gamma': ['scale', 0.1],
    'kernel': ['rbf']
}

# cuML GridSearchCV — must use cuDF input only
grid_search = GridSearchCV(
    estimator=svc,
    param_grid=param_grid,
    cv=3,
    scoring='accuracy',
    verbose=2
)

# Fit the grid search using cuDF inputs only
grid_search.fit(X_scaled, y_cudf)

# Print best result
print("✅ Best Parameters:", grid_search.best_params_)
print("✅ Best Score:", grid_search.best_score_)

问题原因 #

报错核心是GridSearchCV使用字符串类型的scoring='accuracy'时，会默认调用Scikit-learn的评估函数，而Scikit-learn无法直接处理GPU端的cuDF/CuPy数据，会触发隐式转换为NumPy数组的操作，这被RAPIDS的安全机制禁止。此外，参数网格中的原生Python数值（如0.1）在部分场景下也可能引发类型转换风险。

修复方案 #

方案1：使用cuML原生评估函数替换字符串评分指标 #

显式导入cuML的accuracy_score作为评分函数，避免调用Scikit-learn的主机端评估逻辑：

import cudf
import cupy as cp
import numpy as np
from cuml.datasets import make_classification
from cuml.svm import SVC
from cuml.preprocessing import StandardScaler
from cuml.model_selection import GridSearchCV
# 导入cuML原生评估函数
from cuml.metrics import accuracy_score

# Set seeds
cp.random.seed(42)
np.random.seed(42)

# Generate synthetic classification dataset (CuPy format)
X_cupy, y_cupy = make_classification(n_samples=500, n_features=10,
                                     n_informative=5, n_classes=2,
                                     random_state=42)

# 更高效的cuDF转换方式
X_cudf = cudf.DataFrame(X_cupy, columns=[f'feat_{i}' for i in range(X_cupy.shape[1])])
y_cudf = cudf.Series(y_cupy)

# Use cuML StandardScaler (keeps everything on GPU)
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X_cudf)

# Define cuML SVC
svc = SVC(class_weight='balanced')

# Define parameter grid
param_grid = {
    'C': [0.1, 1.0],
    'gamma': ['scale', 0.1],
    'kernel': ['rbf']
}

# 使用cuML原生accuracy_score作为评分函数
grid_search = GridSearchCV(
    estimator=svc,
    param_grid=param_grid,
    cv=3,
    scoring=accuracy_score,  # 替换为cuML的评估函数
    verbose=2
)

# Fit the grid search
grid_search.fit(X_scaled, y_cudf)

# Print best result
print("✅ Best Parameters:", grid_search.best_params_)
print("✅ Best Score:", grid_search.best_score_)

方案2：参数网格使用CuPy标量（可选增强） #

将参数中的原生Python数值替换为CuPy标量，彻底避免类型转换风险：

param_grid = {
    'C': [cp.float32(0.1), cp.float32(1.0)],
    'gamma': ['scale', cp.float32(0.1)],
    'kernel': ['rbf']
}

方案3：确保标准化输出为cuDF（可选验证） #

如果StandardScaler.fit_transform返回的是CuPy数组，显式转换为cuDF：

X_scaled = cudf.DataFrame(scaler.fit_transform(X_cudf), columns=X_cudf.columns)

验证说明 #

修改后代码会全程在GPU端执行，避免任何隐式的主机端转换，符合RAPIDS的设计规范，能够正常运行并输出最优参数与得分。

内容的提问来源于stack exchange，提问作者user2718067