syropiatovvv/iis-project
1
0
Форкнуть 0
Код Задачи Запросы на слияние Пакеты Проекты Релизы Активность

добавить в блокнот research часть про оптимизацию гиперпараметров, обобщить некоторый код (в особенности фильтры параметров моделей)

Просмотр исходного кода
Этот коммит содержится в:
syropiatovvv
2025-11-01 19:34:52 +03:00
родитель 543a4c6571
Коммит 22ef4d303c
2 изменённых файлов: 185 добавлений и 60 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

244
research/research.py
Просмотреть файл

@@ -50,6 +50,8 @@ mlflow_feateng_run_name: str = 'Model with engineered features'
# Имя ноговго прогона MLFlow для модели, использующей дополнительные признаки
mlflow_feateng_filtered_run_name: str = 'Model with filtered engineered features'
# Имя ноговго прогона MLFlow для модели, использующей дополнительные признаки и фильтрацию признаков
mlflow_optimized_feateng_filtered_run_name: str = 'Optimized model with filtered engineered features'
# Имя ноговго прогона MLFlow для модели с оптимизированными гиперпараметрами, использующей дополнительные признаки и фильтрацию признаков
# %%
from collections.abc import Sequence
@@ -65,6 +67,7 @@ import mlflow.models
import mlflow.sklearn
import mlxtend.feature_selection
import mlxtend.plotting
import optuna
import sklearn.compose
import sklearn.ensemble
import sklearn.metrics
@@ -257,13 +260,16 @@ def build_categorical_features_encoder_target(*, random_state=None):
# Регрессор — небольшой случайный лес, цель — минимизация квадрата ошибки предсказания:
# %%
def build_regressor(*, random_state=None):
def build_regressor(n_estimators, *, max_depth=None, max_features='sqrt', random_state=None):
return sklearn.ensemble.RandomForestRegressor(
10, criterion='squared_error',
max_depth=8, max_features='sqrt',
n_estimators, criterion='squared_error',
max_depth=max_depth, max_features=max_features,
random_state=random_state,
)
def build_regressor_baseline(*, random_state=None):
return build_regressor(16, max_depth=8, max_features='sqrt')
# %%
def score_predictions(target_test, target_test_predicted):
@@ -327,7 +333,7 @@ preprocess_transformer = sklearn.compose.ColumnTransformer(
)
# %%
regressor = build_regressor(random_state=0x016B)
regressor = build_regressor_baseline(random_state=0x016B)
regressor
# %% [markdown]
@@ -408,7 +414,7 @@ features_to_extend_as_spline = ('age',)
# %%
def build_preprocess_transformer():
def build_preprocess_augmenting_transformer():
assert set(features_to_extend_as_polynomial) <= {*features_to_scale_to_standard_columns}
assert set(features_to_extend_as_spline) <= {*features_to_scale_to_standard_columns}
return sklearn.compose.ColumnTransformer(
@@ -445,14 +451,34 @@ def build_preprocess_transformer():
remainder='drop',
)
preprocess_transformer = build_preprocess_transformer()
# %%
PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_INCLUDE = {
**{k: True for k in COLUMN_TRANSFORMER_PARAMS_COMMON_INCLUDE},
'extend_features_as_polynomial': {
'extend_features': True,
'scale_to_standard': True,
},
'extend_features_as_spline': True,
'scale_to_standard': True,
'encode_categorical_wrt_target': True,
}
PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_EXCLUDE = {
'extend_features_as_polynomial': {
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
},
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
}
# %%
preprocess_transformer = build_preprocess_augmenting_transformer()
preprocess_transformer
# %% [markdown]
# Демонстрация предобработки данных:
# %%
preprocess_transformer_tmp = build_preprocess_transformer()
preprocess_transformer_tmp = build_preprocess_augmenting_transformer()
df_augd_features_matrix_train = preprocess_transformer_tmp.fit_transform(df_orig_features_train, df_target_train.iloc[:, 0])
df_augd_features_train = pandas_dataframe_from_transformed_artifacts(df_augd_features_matrix_train, preprocess_transformer_tmp)
del preprocess_transformer_tmp
@@ -467,7 +493,7 @@ df_augd_features_train.info()
df_augd_features_train.head(0x8)
# %%
regressor = build_regressor(random_state=0x3AEF)
regressor = build_regressor_baseline(random_state=0x3AEF)
regressor
# %% [markdown]
@@ -484,28 +510,11 @@ pipeline
model_params = filter_params(
pipeline.get_params(),
include={
'preprocess': (
False,
{
**{k: True for k in COLUMN_TRANSFORMER_PARAMS_COMMON_INCLUDE},
'extend_features_as_polynomial': {
'extend_features': True,
'scale_to_standard': True,
},
'extend_features_as_spline': True,
'scale_to_standard': True,
'encode_categorical_wrt_target': True,
},
),
'preprocess': (False, PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_INCLUDE.copy()),
'regress': (False, True),
},
exclude={
'preprocess': {
'extend_features_as_polynomial': {
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
},
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
},
'preprocess': PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_EXCLUDE.copy(),
'regress': RANDOM_FOREST_REGRESSOR_PARAMS_COMMON_EXCLUDE,
},
)
@@ -550,7 +559,7 @@ mlflow_log_model(
# ### Модель с дополнительными и отфильтрованными признаками
# %%
regressor = build_regressor(random_state=0x8EDD)
regressor = build_regressor_baseline(random_state=0x8EDD)
regressor
# %% [markdown]
@@ -562,11 +571,24 @@ len(df_augd_features_train.columns)
# %%
FILTERED_FEATURES_NUM = (4, 8)
# %%
feature_selector = build_sequential_feature_selector(
regressor, k_features=FILTERED_FEATURES_NUM, forward=True, floating=True, cv=4, scoring='neg_mean_absolute_percentage_error',
verbose=1,
)
def build_feature_selector(*, verbose=0):
return build_sequential_feature_selector(
regressor, k_features=FILTERED_FEATURES_NUM, forward=True, floating=True, cv=4, scoring='neg_mean_absolute_percentage_error',
verbose=verbose,
)
# %%
FEATURE_SELECTOR_PARAMS_COMMON_INCLUDE = {
**{k: True for k in SEQUENTIAL_FEATURE_SELECTOR_PARAMS_COMMON_INCLUDE},
'estimator': False,
}
FEATURE_SELECTOR_PARAMS_COMMON_EXCLUDE = () # TODO: ай-яй-яй
# %%
feature_selector = build_feature_selector(verbose=1)
feature_selector
# %%
@@ -595,7 +617,7 @@ _ = ax.set_ylim((None, 0.))
# %%
pipeline = sklearn.pipeline.Pipeline([
('preprocess', preprocess_transformer),
('preprocess', build_preprocess_augmenting_transformer()),
('select_features', feature_selector),
('regress', regressor),
])
@@ -605,36 +627,13 @@ pipeline
model_params = filter_params(
pipeline.get_params(),
include={
'preprocess': (
False,
{
**{k: True for k in COLUMN_TRANSFORMER_PARAMS_COMMON_INCLUDE},
'extend_features_as_polynomial': {
'extend_features': True,
'scale_to_standard': True,
},
'extend_features_as_spline': True,
'scale_to_standard': True,
'encode_categorical_wrt_target': True,
},
),
'select_features': (
False,
{
**{k: True for k in SEQUENTIAL_FEATURE_SELECTOR_PARAMS_COMMON_INCLUDE},
'estimator': False,
},
),
'preprocess': (False, PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_INCLUDE.copy()),
'select_features': (False, FEATURE_SELECTOR_PARAMS_COMMON_INCLUDE.copy()),
'regress': (False, True),
},
exclude={
'preprocess': {
'extend_features_as_polynomial': {
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
},
'scale_to_standard': STANDARD_SCALER_PARAMS_COMMON_EXCLUDE,
},
'select_features': (), # TODO: ай-яй-яй
'preprocess': PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_EXCLUDE.copy(),
'select_features': FEATURE_SELECTOR_PARAMS_COMMON_EXCLUDE,
'regress': RANDOM_FOREST_REGRESSOR_PARAMS_COMMON_EXCLUDE,
},
)
@@ -676,4 +675,129 @@ mlflow_log_model(
),
)
# %% [markdown]
# ### Автоматический подбор гиперпараметров модели
# %% [markdown]
# Составной пайплайн:
# %%
def build_pipeline(regressor_n_estimators, regressor_max_depth=None, regressor_max_features='sqrt'):
return sklearn.pipeline.Pipeline([
('preprocess', build_preprocess_augmenting_transformer()),
('select_features', build_feature_selector()),
('regress', build_regressor(regressor_n_estimators, max_depth=regressor_max_depth, max_features=regressor_max_features)),
])
# %% [markdown]
# Целевая функция для оптимизатора гиперпараметров (подбирает параметры `RandomForestRegressor`: `n_estimators`, `max_depth`, `max_features`):
# %%
def regressor_hyperparams_objective(trial):
n_estimators = trial.suggest_int('n_estimators', 1, 256, log=True)
max_depth = trial.suggest_int('max_depth', 1, 16, log=True)
max_features = trial.suggest_float('max_features', 0.1, 1.)
# составной пайплайн:
pipeline = build_pipeline(n_estimators, regressor_max_depth=max_depth, regressor_max_features=max_features)
# обучение модели:
_ = pipeline.fit(df_orig_features_train, df_target_train.iloc[:, 0])
# оценка качества:
target_test_predicted = pipeline.predict(df_orig_features_test)
# метрика качества (MAPE):
mape = sklearn.metrics.mean_absolute_percentage_error(df_target_test, target_test_predicted)
return mape
# %% [markdown]
# optuna study:
# %%
optuna_study = optuna.create_study(direction='minimize')
optuna_study.optimize(regressor_hyperparams_objective, n_trials=64, timeout=120.)
# %% [markdown]
# Количество выполненных trials:
# %%
len(optuna_study.trials)
# %% [markdown]
# Лучшие найдённые гиперпараметры (недетерминированы, один из результатов записан явно):
# %%
optuna_study.best_params
# %%
regressor_best_params = {
#'n_estimators': 51,
'n_estimators': 50,
'max_depth': 11,
#'max_features': 0.44655290756636146,
'max_features': 0.45,
}
# %% [markdown]
# Составной пайплайн:
# %%
pipeline = build_pipeline(
regressor_best_params['n_estimators'],
regressor_max_depth=regressor_best_params['max_depth'],
regressor_max_features=regressor_best_params['max_features'],
)
pipeline
# %%
model_params = filter_params(
pipeline.get_params(),
include={
'preprocess': (False, PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_INCLUDE.copy()),
'select_features': (False, FEATURE_SELECTOR_PARAMS_COMMON_INCLUDE.copy()),
'regress': (False, True),
},
exclude={
'preprocess': PREPROCESS_AUGMENTING_TRANSFORMER_PARAMS_COMMON_EXCLUDE.copy(),
'select_features': FEATURE_SELECTOR_PARAMS_COMMON_EXCLUDE,
'regress': RANDOM_FOREST_REGRESSOR_PARAMS_COMMON_EXCLUDE,
},
)
model_params
# %% [markdown]
# Обучение модели:
# %%
_ = pipeline.fit(df_orig_features_train, df_target_train.iloc[:, 0])
# %% [markdown]
# Оценка качества:
# %%
target_test_predicted = pipeline.predict(df_orig_features_test)
# %% [markdown]
# Метрики качества (MAPE, а также MSE, MAE):
# %%
metrics = score_predictions(df_target_test, target_test_predicted)
metrics
# %%
mlflow_log_model(
pipeline,
model_params=model_params,
metrics={k: float(v) for k, v in metrics.items()},
run_name=mlflow_optimized_feateng_filtered_run_name,
model_signature=mlflow_model_signature,
input_example=df_orig_features.head(MODEL_INOUT_EXAMPLE_SIZE),
#pip_requirements=str(MODEL_PIP_REQUIREMENTS_PATH),
comment_file_path=(
model_comment_path
if model_comment_path is not None
else (BASE_PATH / 'research' / model_comment_relpath)
),
)
# %%