fix: PhysicalFeatureExtractor 修复 ndarray 输入时波长丢失问题

- SimpleImputer 把 DataFrame → ndarray,PhysicalFeatureExtractor.fit()
  收不到列名导致 ValueError: 必须在 __init__ 中提供 wavelengths
- get_preprocessing_transformer 新增 wavelengths 可选参数
- train_single_model 在 DualStream_MNF 时提取 X_raw 波长列表传入工厂
- PhysicalFeatureExtractor 已有 ndarray + wavelengths 参数的处理分支
This commit is contained in:
duxin
2026-07-29 10:11:01 +08:00
parent ef9187fc7b
commit 09f8d89b56
2 changed files with 9 additions and 3 deletions

View File

@ -611,7 +611,11 @@ class WaterQualityModelingBatch:
base_model.set_params(verbose=-1)
# ============ 关键:把预处理器塞进 Pipeline ============
preproc = get_preprocessing_transformer(preprocess_method)
# DualStream_MNF 需要传入波长列表,供 PhysicalFeatureExtractor 定位波段
_wl_list = None
if preprocess_method == "DualStream_MNF":
_wl_list = self._extract_train_wavelengths(X_raw.columns)
preproc = get_preprocessing_transformer(preprocess_method, wavelengths=_wl_list)
pipeline = Pipeline([
('imputer', SimpleImputer(strategy='median')),
('preproc', preproc),

View File

@ -501,7 +501,7 @@ _PREPROCESSING_TRANSFORMERS = {
}
def get_preprocessing_transformer(method: str):
def get_preprocessing_transformer(method: str, wavelengths=None):
"""根据预处理方法名返回 sklearn 兼容的 Transformer 实例。
- method 为 "None" 或 None返回 IdentityTransformer
@ -512,6 +512,8 @@ def get_preprocessing_transformer(method: str):
Args:
method: 预处理方法名
wavelengths: DualStream_MNF 时需要传入波长列表(float)
供 PhysicalFeatureExtractor 定位波段列
Returns:
sklearn 兼容的 Transformer 实例(可直接放入 Pipeline
@ -521,7 +523,7 @@ def get_preprocessing_transformer(method: str):
if method == "DualStream_MNF":
from sklearn.pipeline import FeatureUnion
return FeatureUnion([
('physical', PhysicalFeatureExtractor()),
('physical', PhysicalFeatureExtractor(wavelengths=wavelengths)),
('mnf', MNFTransformer(n_components=10)),
])
if method not in _PREPROCESSING_TRANSFORMERS: