IIS/assets/mlflow/research.ipynb

import os
import mlflow

from sklearn.model_selection import train_test_split
import pandas as pd
import numpy

from sklearn.preprocessing import StandardScaler, OrdinalEncoder, OneHotEncoder

from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestRegressor
from sklearn.linear_model import LinearRegression
from catboost import CatBoostRegressor

from sklearn.metrics import mean_absolute_percentage_error, mean_absolute_error, mean_squared_error

df = pd.read_pickle('data/clean_data.pkl').sample(frac=0.1, random_state = 2) # Уменьшаем размер чтобы модель быстрее обучалась на лекции
df.info()

df = df.rename(columns={'price': 'target'})
df = df.drop(columns=['date', 'time'])

df

X_train, X_test, y_train, y_test = train_test_split(df.drop('target', axis=1), df['target'], test_size=0.25, random_state=2)

cat_features = X_train.select_dtypes(include=['category','object']).columns.to_list()
cat_features

num_features = X_train.select_dtypes(include=['number']).columns.to_list()
num_features

https://scikit-learn.org/stable/api/sklearn.preprocessing.html - разные способы кодирования и скалирования

s_scaler = StandardScaler()
l_encoder = OrdinalEncoder(handle_unknown='use_encoded_value', unknown_value=99999999) # unknown_value нужно выбирать с умом
regressor = CatBoostRegressor()

Column transformer

# Для удобной работы со столбцами
preprocessor = ColumnTransformer(
    transformers=[
        ('num', s_scaler, num_features),  # преобразования для числовых признаков
        ('cat', l_encoder, cat_features), # преобразования для категориальных признаков
    ],
    remainder='drop' ) # Удаляем столбцы, которые не затронуты преобразования

pipeline = Pipeline(steps=[('preprocessor', preprocessor), 
                           ('model', regressor)])

pipeline.fit(X_train, y_train)

predictions = pipeline.predict(X_test) 

metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions)
metrics["mse"] = mean_squared_error(y_test, predictions)

metrics

# Работаем с MLflow локально
TRACKING_SERVER_HOST = "127.0.0.1"
TRACKING_SERVER_PORT = 5000

registry_uri = f"http://{TRACKING_SERVER_HOST}:{TRACKING_SERVER_PORT}"
tracking_uri = f"http://{TRACKING_SERVER_HOST}:{TRACKING_SERVER_PORT}"

mlflow.set_tracking_uri(tracking_uri)   
mlflow.set_registry_uri(registry_uri)   

# название тестового эксперимента, запуска (run) внутри него, имени, под которым модель будет регистрироваться
EXPERIMENT_NAME = "estate_project"
RUN_NAME = "baseline model"
REGISTRY_MODEL_NAME = "estate_model_rf"

Логируем вручную

# Обязательно логируем сигнатуру модели и пример входных данных. Подготовим их
from mlflow.models import infer_signature

signature =  infer_signature(model_input = X_train.head(5))
input_example = X_train.head(5)

# Будем логировать requirements и артефакт - текстовый файл
req_file = 'requirements.txt'
art = 'comment.txt'

# Параметры, котороые будут залогированы, можем задавать вручную или полностью взять из модели
#params_dict = {'n_estimators': 10, 'max_depth': 10}
params_dict = pipeline.get_params()

# Когда создаем новый эксперимент, то: 
experiment_id = mlflow.create_experiment(EXPERIMENT_NAME)

# Впоследствии. чтобы добавлять запуски в этот же эксепримент мы должны получить его id:
#experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(params_dict)

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

Удаление runs, experiments

Использовать осторожно

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id
#mlflow.delete_experiment(experiment_id)

mlflow.search_runs(
    #experiment_ids=[experiment_id],
    experiment_names=[EXPERIMENT_NAME],
    # filter_string='status = "FAILED"'
    #filter_string='metrics.mae > 1'
    
)

#mlflow.delete_run('74d2a7a40c07413c9cf65df841164356')

Автологирование

После включения будет срабатывать на каждом обучении модели (на методе fit()).

Есть плюсы, есть и минусы. Предлагается сделать прогон и сравнить с результатами вручную

mlflow.sklearn.autolog()

with mlflow.start_run(run_name='auto', experiment_id=experiment_id) as run:
    pipeline.fit(X_train, y_train)

# Отключаем автологирование
mlflow.sklearn.autolog(disable=True)

Model #2

Обучим вторую "маленькую" модель

regressor2 = RandomForestRegressor(n_estimators=10, max_depth=6)

pipeline = Pipeline(steps=[('preprocessor', preprocessor), 
                           ('model', regressor2)])

pipeline.fit(X_train, y_train)

predictions = pipeline.predict(X_test) 
metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions)
metrics["mse"] = mean_squared_error(y_test, predictions)

metrics

# !!! Проверить название прогона а также все логируемые параметры и артефакты, что они соответствуют второй "маленькой" модели. 


RUN_NAME = 'smaller_model'

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(pipeline.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

# No model
# Логировать можно только артефакты, без модели. Например, залогироавть графики после этапа EDA

RUN_NAME = 'no_model'
experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    run_id = run.info.run_id 
    mlflow.log_artifact(art)


run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

run_id = '06fa7ec1f1b74aedb3509c88dc4ee1c0' # Указываем run id
mlflow.register_model(f"runs:/{run_id}/models", REGISTRY_MODEL_NAME)

# Можно регистрировать сразу при создании прогона

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id

with mlflow.start_run(run_name='register_at_run', experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file,
                             registered_model_name = REGISTRY_MODEL_NAME # Указываем для какой модели регистрируем
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(pipeline.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

# Можно найти зарегистрированные модели
model_reg = mlflow.search_registered_models()
model_reg[0]

model_name = REGISTRY_MODEL_NAME
model_version = 1

model_loaded = mlflow.sklearn.load_model(model_uri=f"models:/{model_name}/{model_version}")

model_loaded.predict(X_test.iloc[0:1])

y_test.iloc[0]

Feature engineering

Sklearn

from sklearn.preprocessing import QuantileTransformer, SplineTransformer, PolynomialFeatures, MinMaxScaler

X_train_sklearn = X_train.copy()

PolynomialFeatures

Создает полином степени degree из указанных признаков

pf = PolynomialFeatures(degree=2)

X_train_sklearn

pf.fit_transform(X_train_sklearn[['area','kitchen_area']])

SplineTransformer

Cоздаёт новую матрицу признаков, состоящую из сплайнов порядка degree. Количество сгенерированных сплайнов равно n_splines=n_knots + degree - 1 для каждого признака, где

n_knots определяет количество узлов (точек, в которых сопрягаются сплайны) для каждого признака.

degree определяет порядок полинома, используемого для построения сплайнов.

sp = SplineTransformer(n_knots=3, degree=3)

sp.fit_transform(X_train_sklearn[['area']])

QuantileTransformer

Этот метод преобразует признаки, чтобы они распределялись равномерно или нормально — так данные меньше подвергаются влиянию выбросов. Преобразование применяется к каждому признаку независимо. Идея метода такова: оценить функцию распределения признака, чтобы преобразовать исходные значения в равномерное или нормальное распределение.

output_distribution='uniform' или output_distribution='normal' соответственно

Пример использования: если у вас есть данные о доходах с широким диапазоном значений, квантильное преобразование сделает их более сопоставимыми и устойчивыми к выбросам.

qt = QuantileTransformer()

qt.fit_transform(X_train_sklearn[['area']])

Объединяем в ColumnTransformer и создаем Pipeline

pf = PolynomialFeatures(degree=2)
qt = QuantileTransformer()
sp = SplineTransformer(n_knots=3, degree=3)

# Значения преобразованных признаков нужно отскейлить, поэтому создаем pipeline из двух шагов - преобразование и скейлинг
pf_pipeline = Pipeline(steps=[
    ('poly', pf),
    ('scale', StandardScaler())
])

preprocessor_sklearn = ColumnTransformer(
    transformers=[
        ('num', s_scaler, num_features),  # преобразования для числовых признаков
        ('cat', l_encoder, cat_features), # преобразования для категориальных признаков
        ('quantile', qt,num_features),
        ('poly', pf_pipeline, ['area', 'kitchen_area']), # В преобразования добавляем созданный ранее pipeline
        ('spline', sp, ['area'])
    ],
    remainder='drop',
    ) # Удаляем столбцы, которые не затронуты преобразования

Посмотрим что из себя теперь представляет датафрейм

## не влезаем в float64 в полиномальном преобразовании. Использовать его нужно с умом!
X_train_sklearn[['area', 'kitchen_area']] = X_train_sklearn[['area', 'kitchen_area']].astype('float128')
X_train_sklearn[['area', 'kitchen_area']] = X_train_sklearn[['area', 'kitchen_area']].astype('float128')

X_train_sklearn_raw = preprocessor_sklearn.fit_transform(X_train_sklearn)
X_train_sklearn = pd.DataFrame(X_train_sklearn_raw, columns=preprocessor_sklearn.get_feature_names_out())

# Удобно использовать для отображения всех строк\столбцов в DataFrame
with pd.option_context('display.max_rows', 5, 'display.max_columns', None):
    display (X_train_sklearn)

	num__geo_lat	num__geo_lon	num__level	num__levels	num__rooms	num__area	num__kitchen_area	cat__region	cat__building_type	cat__object_type	quantile__geo_lat	quantile__geo_lon	quantile__level	quantile__levels	quantile__rooms	quantile__area	quantile__kitchen_area	poly__1	poly__area	poly__kitchen_area	poly__area^2	poly__area kitchen_area	poly__kitchen_area^2	spline__area_sp_0	spline__area_sp_1	spline__area_sp_2	spline__area_sp_3	spline__area_sp_4
0	0.495902	-0.449742	0.359235	-0.214789	0.253413	0.063735	-0.186285	20.0	1.0	0.0	0.766257	0.511028	0.717217	0.536537	0.600601	0.623624	0.374875	0.0	0.063735	-0.186285	-0.010002	-0.132188	-0.002792	0.155806	0.666179	0.178013	0.000002	0.0
1	0.177806	1.433673	-0.246529	-0.367718	0.253413	-0.114293	-0.186285	70.0	1.0	0.0	0.297142	0.867999	0.522022	0.386887	0.600601	0.541542	0.374875	0.0	-0.114293	-0.186285	-0.017375	-0.169370	-0.002792	0.156921	0.666275	0.176803	0.000001	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
410773	-0.748366	-0.804077	-0.650371	0.702788	0.253413	1.365441	1.501833	52.0	3.0	0.0	0.193143	0.114753	0.309810	0.741742	0.600601	0.961367	0.984535	0.0	1.365441	1.501833	0.068438	1.570163	0.008616	0.147820	0.665159	0.187011	0.000010	0.0
410774	1.257769	-1.101815	-0.044608	0.091070	1.175911	0.553789	-0.142544	14.0	1.0	0.0	0.908036	0.075725	0.604605	0.645646	0.867367	0.841842	0.436436	0.0	0.553789	-0.142544	0.014463	-0.002742	-0.002649	0.152767	0.665860	0.181370	0.000004	0.0

410775 rows × 28 columns

Создаем пайплайн с препроцессингом и моделью

pipeline_sklearn = Pipeline(steps=[
    ('transform', preprocessor_sklearn),
    ('model', regressor)
])

model_sklearn = pipeline_sklearn.fit(X_train, y_train)

model_sklearn

predictions = model_sklearn.predict(X_test) 
metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions)
metrics["mse"] = mean_squared_error(y_test, predictions)

metrics

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id
RUN_NAME = 'fe_sklearn'

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(model_sklearn, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(model_sklearn.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

Autofeat

from autofeat import AutoFeatRegressor
transformations = ["1/", "exp", "log", "abs", "sqrt", "^2", "^3", "1+", "1-", "sin", "cos", "exp-", "2^"] 

afreg = AutoFeatRegressor(verbose=1, feateng_steps=2, max_gb=8, transformations=["log", "sqrt"],feateng_cols=num_features)
X_train_arf = afreg.fit_transform(X_train,y_train)
X_train_arf

# Создаем обертку, в которой добавляем метод get_feature_names_out() для получения названий признаков
import numpy as np

class AutoFeatWrapper():
    def __init__(self, feateng_cols, feateng_steps=1, max_gb=16, transformations=["1/", "exp", "log"], n_jobs=-1, verbose=1):
        self.feateng_cols = feateng_cols
        self.feateng_steps = feateng_steps
        self.max_gb = max_gb
        self.transformations = transformations
        self.n_jobs = n_jobs
        self.afreg = AutoFeatRegressor(feateng_cols=self.feateng_cols,
                                     feateng_steps=self.feateng_steps,
                                     max_gb=self.max_gb,
                                     transformations=self.transformations,
                                     n_jobs=self.n_jobs)
        
    def fit(self, X, y=None):
        self.afreg.fit(X, y)
        return self
    
    def transform(self, X):
        return self.afreg.transform(X)
    
    def get_feature_names_out(self, input_features=None):
        # Преобразуем данные и возвращаем имена фичей из DataFrame
        transformed_X = self.afreg.transform(pd.DataFrame(np.zeros((1, len(self.feateng_cols))), columns=self.feateng_cols))
        return transformed_X.columns.tolist()

afreg_pipeline = Pipeline(steps=[
    ('autofeat', AutoFeatWrapper( feateng_steps=2, max_gb=16, transformations=["log", "sqrt"],feateng_cols=num_features)),
    ('scaler', StandardScaler()),
])

preprocessor_afr = ColumnTransformer(
    transformers=[
        ('num', s_scaler, num_features),  # преобразования для числовых признаков
        ('cat', l_encoder, cat_features), # преобразования для категориальных признаков
        ('afr', afreg_pipeline, num_features), # преобразования autofeat
    ],
    remainder='drop', # Удаляем столбцы, которые не затронуты преобразованиями
    ) 

X_train_afr_raw =  preprocessor_afr.fit_transform(X_train,y_train)
X_train_afr = pd.DataFrame(X_train_afr_raw, columns=preprocessor_afr.get_feature_names_out())

with pd.option_context('display.max_rows', 5, 'display.max_columns', None):
    display (X_train_afr)

	num__geo_lat	num__geo_lon	num__level	num__levels	num__rooms	num__area	num__kitchen_area	cat__region	cat__building_type	cat__object_type	afr__geo_lat	afr__geo_lon	afr__level	afr__levels	afr__rooms	afr__area	afr__kitchen_area	afr__area*rooms	afr__area*geo_lon	afr__levels*rooms	afr__area*kitchen_area	afr__sqrt(area)*geo_lat	afr__sqrt(area)*log(level)	afr__kitchen_area*log(level)	afr__sqrt(area)*kitchen_area	afr__geo_lon*log(kitchen_area)	afr__sqrt(area)*sqrt(kitchen_area)	afr__sqrt(geo_lon)*sqrt(kitchen_area)	afr__log(area)	afr__rooms*log(level)	afr__kitchen_area*rooms	afr__kitchen_area*levels	afr__sqrt(geo_lon)*sqrt(level)	afr__area**(3/2)	afr__geo_lat*log(kitchen_area)	afr__geo_lat*log(geo_lon)
0	0.495902	-0.449742	0.359235	-0.214789	0.253413	0.063735	-0.186285	20.0	1.0	0.0	0.495902	-0.449742	0.359235	-0.214789	0.253413	0.063735	-0.186285	0.006208	-0.195129	0.060916	-0.132188	0.373151	0.688076	0.044178	-0.211335	-0.481294	-0.153548	-0.490805	0.307835	0.690329	-0.132529	-0.352834	0.323880	-0.008748	-0.031529	0.068167
1	0.177806	1.433673	-0.246529	-0.367718	0.253413	-0.114293	-0.186285	70.0	1.0	0.0	0.177806	1.433673	-0.246529	-0.367718	0.253413	-0.114293	-0.186285	-0.083402	0.655053	-0.054279	-0.169370	0.005114	0.071369	-0.173647	-0.252775	1.191304	-0.267268	0.615798	0.031907	0.282625	-0.132529	-0.418643	0.552794	-0.056540	-0.143829	1.129118
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
410773	-0.748366	-0.804077	-0.650371	0.702788	0.253413	1.365441	1.501833	52.0	3.0	0.0	-0.748366	-0.804077	-0.650371	0.702788	0.253413	1.365441	1.501833	0.661427	0.375199	0.752088	1.570163	1.274445	-0.002521	0.745507	2.382258	0.071599	2.828890	1.431272	1.729715	-0.160491	1.581436	2.432437	-0.843150	0.411475	1.671069	-1.052343
410774	1.257769	-1.101815	-0.044608	0.091070	1.175911	0.553789	-0.142544	14.0	1.0	0.0	1.257769	-1.101815	-0.044608	0.091070	1.175911	0.553789	-0.142544	0.807887	-0.330070	0.982478	-0.002742	1.338996	0.635065	-0.040302	-0.055435	-1.025588	0.202136	-0.916054	0.940624	1.217910	0.311575	-0.174762	-0.415359	0.135617	0.359680	-0.246790

410775 rows × 36 columns

pipeline_afr = Pipeline(steps=[('preprocessor', preprocessor_afr), 
                               ('model', regressor)])

pipeline_afr.fit(X_train, y_train)

predictions = pipeline_afr.predict(X_test) 

metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions)
metrics["mse"] = mean_squared_error(y_test, predictions)

metrics

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id

with mlflow.start_run(run_name='autofeat', experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(pipeline_afr, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(pipeline_afr.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

FEATURE SELECTION

RFE

Используем autofeat признаки

Поскольку autofeat дает разные совокупности сгенерированных признаков, мы можем добавить выбор информативных только как шаг пайплайна

from sklearn.feature_selection import RFE
X_train_afr

	num__geo_lat	num__geo_lon	num__level	num__levels	num__rooms	num__area	num__kitchen_area	cat__region	cat__building_type	cat__object_type	...	afr__sqrt(area)*sqrt(kitchen_area)	afr__sqrt(geo_lon)*sqrt(kitchen_area)	afr__log(area)	afr__rooms*log(level)	afr__kitchen_area*rooms	afr__kitchen_area*levels	afr__sqrt(geo_lon)*sqrt(level)	afr__area**(3/2)	afr__geo_lat*log(kitchen_area)	afr__geo_lat*log(geo_lon)
0	0.495902	-0.449742	0.359235	-0.214789	0.253413	0.063735	-0.186285	20.0	1.0	0.0	...	-0.153548	-0.490805	0.307835	0.690329	-0.132529	-0.352834	0.323880	-0.008748	-0.031529	0.068167
1	0.177806	1.433673	-0.246529	-0.367718	0.253413	-0.114293	-0.186285	70.0	1.0	0.0	...	-0.267268	0.615798	0.031907	0.282625	-0.132529	-0.418643	0.552794	-0.056540	-0.143829	1.129118
2	0.440548	0.047222	-0.448450	-0.367718	-0.669085	-0.456947	-0.142544	15.0	3.0	1.0	...	-0.454880	-0.067183	-0.603122	-0.512211	-0.487813	-0.383803	-0.243092	-0.140800	0.063464	0.460495
3	-1.588818	-0.722477	-0.246529	-0.979436	0.253413	-0.181292	-0.142544	18.0	1.0	0.0	...	-0.254514	-0.607607	-0.080304	0.282625	-0.088119	-0.662523	-0.369355	-0.073838	-0.672113	-1.481033
4	1.493662	1.125819	0.157313	0.549858	0.253413	0.615045	-0.011322	10.0	2.0	0.0	...	0.438600	0.891383	1.009612	0.574497	0.045112	0.208478	0.945981	0.154902	0.780855	1.923382
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
410770	0.592011	0.355014	0.561156	1.008646	0.253413	-0.079836	-0.092653	54.0	2.0	0.0	...	-0.120035	0.237580	0.087725	0.792500	-0.037463	0.322797	0.974381	-0.047496	0.243018	0.789871
410771	0.240478	0.392697	-0.650371	-0.979436	0.253413	-0.334434	-0.404989	45.0	3.0	0.0	...	-0.716150	-0.510766	-0.357277	-0.160491	-0.354582	-0.778657	-0.406361	-0.111897	-0.808157	0.574534
410772	-1.936771	-0.688830	0.359235	0.855717	-0.669085	-0.456947	-0.142544	18.0	0.0	1.0	...	-0.454880	-0.581851	-0.603122	-0.211576	-0.487813	0.173638	0.170166	-0.140800	-0.798234	-1.663294
410773	-0.748366	-0.804077	-0.650371	0.702788	0.253413	1.365441	1.501833	52.0	3.0	0.0	...	2.828890	1.431272	1.729715	-0.160491	1.581436	2.432437	-0.843150	0.411475	1.671069	-1.052343
410774	1.257769	-1.101815	-0.044608	0.091070	1.175911	0.553789	-0.142544	14.0	1.0	0.0	...	0.202136	-0.916054	0.940624	1.217910	0.311575	-0.174762	-0.415359	0.135617	0.359680	-0.246790

410775 rows × 36 columns

rfe_selector = RFE(estimator=regressor, n_features_to_select=12, step = 0.2) #drop 20% of features each iteration
X_train_rfe = rfe_selector.fit_transform(X_train_afr,y_train)

X_train_afr_rfe = pd.DataFrame(X_train_rfe, columns=rfe_selector.get_feature_names_out())
X_train_afr_rfe

	num__geo_lat	num__geo_lon	afr__geo_lon	afr__area*kitchen_area	afr__sqrt(area)*geo_lat	afr__sqrt(area)*log(level)	afr__kitchen_area*log(level)	afr__sqrt(area)*sqrt(kitchen_area)	afr__rooms*log(level)	afr__kitchen_area*rooms	afr__sqrt(geo_lon)*sqrt(level)	afr__geo_lat*log(geo_lon)
0	0.495902	-0.449742	-0.449742	-0.132188	0.373151	0.688076	0.044178	-0.153548	0.690329	-0.132529	0.323880	0.068167
1	0.177806	1.433673	1.433673	-0.169370	0.005114	0.071369	-0.173647	-0.267268	0.282625	-0.132529	0.552794	1.129118
2	0.440548	0.047222	0.047222	-0.226261	-0.425530	-0.335537	-0.239271	-0.454880	-0.512211	-0.487813	-0.243092	0.460495
3	-1.588818	-0.722477	-0.722477	-0.165302	-0.723225	0.034116	-0.129771	-0.254514	0.282625	-0.088119	-0.369355	-1.481033
4	1.493662	1.125819	1.125819	0.094342	1.522265	0.862773	0.194490	0.438600	0.574497	0.045112	0.945981	1.923382
...	...	...	...	...	...	...	...	...	...	...	...	...
410770	0.592011	0.355014	0.355014	-0.120841	0.206926	0.714499	0.226990	-0.120035	0.792500	-0.037463	0.974381	0.789871
410771	0.240478	0.392697	0.392697	-0.296252	-0.297209	-0.551021	-0.560144	-0.716150	-0.160491	-0.354582	-0.406361	0.574534
410772	-1.936771	-0.688830	-0.688830	-0.226261	-1.192706	0.306280	0.100868	-0.454880	-0.211576	-0.487813	0.170166	-1.663294
410773	-0.748366	-0.804077	-0.804077	1.570163	1.274445	-0.002521	0.745507	2.828890	-0.160491	1.581436	-0.843150	-1.052343
410774	1.257769	-1.101815	-1.101815	-0.002742	1.338996	0.635065	-0.040302	0.202136	1.217910	0.311575	-0.415359	-0.246790

410775 rows × 12 columns

rfe_pipeline = Pipeline(steps=[
    ('preprocessor', preprocessor_afr), 
    ('rfe_extractor', RFE(estimator=regressor, n_features_to_select=12, step = 0.2)),
    ('model', regressor)
])

rfe_pipeline.fit(X_train, y_train)

predictions_rfe = rfe_pipeline.predict(X_test)

metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions_rfe)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions_rfe)
metrics["mse"] = mean_squared_error(y_test, predictions_rfe)

metrics

{'mae': 1431925.3203264712,
 'mape': 1.239752923791043e+18,
 'mse': 261947924998018.2}

experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id
RUN_NAME = 'rfe_feature_selection'

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(rfe_pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact(art)
    mlflow.log_params(model_sklearn.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

Downloading artifacts: 100%|██████████| 7/7 [00:00<00:00, 40.15it/s]
2024/10/17 14:26:50 INFO mlflow.tracking._tracking_service.client: 🏃 View run rfe_feature_selection at: http://127.0.0.1:5000/#/experiments/1/runs/96f0bbcd6d88466abcf38f3b53f06ff1.
2024/10/17 14:26:50 INFO mlflow.tracking._tracking_service.client: 🧪 View experiment at: http://127.0.0.1:5000/#/experiments/1.

Используем sklearn признаки

Тут мы можем отобрать признаки один раз на обучении, а далее в качестве шага пайплайна использовать написанный класс ColumnExtractor для выбора нуных столбцов

rfe_skl_selector = RFE(estimator=regressor, n_features_to_select=12, step = 0.2) #drop 20% of features each iteration
X_train_skl_rfe = rfe_skl_selector.fit_transform(X_train_sklearn,y_train)

X_train_skl_rfe = pd.DataFrame(X_train_skl_rfe, columns=rfe_skl_selector.get_feature_names_out())
X_train_skl_rfe

	num__geo_lat	num__geo_lon	num__level	num__rooms	num__kitchen_area	cat__region	quantile__geo_lat	quantile__geo_lon	quantile__level	poly__area kitchen_area	spline__area_sp_0	spline__area_sp_2
0	0.495902	-0.449742	0.359235	0.253413	-0.186285	20.0	0.766257	0.511028	0.717217	-0.132188	0.155806	0.178013
1	0.177806	1.433673	-0.246529	0.253413	-0.186285	70.0	0.297142	0.867999	0.522022	-0.169370	0.156921	0.176803
2	0.440548	0.047222	-0.448450	-0.669085	-0.142544	15.0	0.732330	0.629984	0.417417	-0.226261	0.159080	0.174488
3	-1.588818	-0.722477	-0.246529	0.253413	-0.142544	18.0	0.148789	0.295262	0.522022	-0.165302	0.157341	0.176349
4	1.493662	1.125819	0.157313	0.253413	-0.011322	10.0	0.985937	0.758363	0.662663	0.094342	0.152390	0.181792
...	...	...	...	...	...	...	...	...	...	...	...	...
410770	0.592011	0.355014	0.561156	0.253413	-0.092653	54.0	0.788393	0.686728	0.771271	-0.120841	0.156705	0.177037
410771	0.240478	0.392697	-0.650371	0.253413	-0.404989	45.0	0.494062	0.717240	0.309810	-0.296252	0.158306	0.175314
410772	-1.936771	-0.688830	0.359235	-0.669085	-0.142544	18.0	0.131352	0.327613	0.717217	-0.226261	0.159080	0.174488
410773	-0.748366	-0.804077	-0.650371	0.253413	1.501833	52.0	0.193143	0.114753	0.309810	1.570163	0.147820	0.187011
410774	1.257769	-1.101815	-0.044608	1.175911	-0.142544	14.0	0.908036	0.075725	0.604605	-0.002742	0.152767	0.181370

410775 rows × 12 columns

rfe_cols = X_train_skl_rfe.columns.tolist()
rfe_cols

['num__geo_lat',
 'num__geo_lon',
 'num__level',
 'num__rooms',
 'num__kitchen_area',
 'cat__region',
 'quantile__geo_lat',
 'quantile__geo_lon',
 'quantile__level',
 'poly__area kitchen_area',
 'spline__area_sp_0',
 'spline__area_sp_2']

rfe_idx = rfe_skl_selector.support_
rfe_idx

array([ True,  True,  True, False,  True, False,  True,  True, False,
       False,  True,  True,  True, False, False, False, False, False,
       False, False, False,  True, False,  True, False,  True, False,
       False])

# Отбираемые столбцы нужно залогировать, иначе мы потеряем информацию о том, какие призныки выбраны
with open('rfe_skl_idx.txt', 'w+') as f:
    f.write(str(rfe_idx))
with open('rfe_skl_cols.txt', 'w+') as f:
    f.write(str(rfe_cols))

class ColumnExtractor(object):

    def __init__(self, cols):
        self.cols = cols

    def transform(self, X):
        return X[:,self.cols]
    
    def fit(self, X, y=None):
        return self

rfe_skl_pipeline = Pipeline(steps=[
    ('preprocessor', preprocessor_sklearn), 
    ('rfe_extractor', ColumnExtractor(rfe_idx)),
    ('model', regressor)
])

rfe_skl_pipeline.fit(X_train, y_train)

predictions_rfe_skl = rfe_skl_pipeline.predict(X_test)

metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions_rfe_skl)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions_rfe_skl)
metrics["mse"] = mean_squared_error(y_test, predictions_rfe_skl)

metrics
experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id
RUN_NAME = 'rfe_skl_feature_selection'

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(rfe_pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact('rfe_skl_cols.txt')
    mlflow.log_artifact('rfe_skl_idx.txt')
    mlflow.log_params(model_sklearn.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

Downloading artifacts: 100%|██████████| 7/7 [00:00<00:00, 193.34it/s]
2024/10/17 14:32:07 INFO mlflow.tracking._tracking_service.client: 🏃 View run rfe_skl_feature_selection at: http://127.0.0.1:5000/#/experiments/1/runs/e55206caeb1549e4aa0d98343d5c1d4d.
2024/10/17 14:32:07 INFO mlflow.tracking._tracking_service.client: 🧪 View experiment at: http://127.0.0.1:5000/#/experiments/1.

mlextend

https://github.com/rasbt/mlxtend/blob/master/docs/sources/user_guide/feature_selection/SequentialFeatureSelector.ipynb

from mlxtend.feature_selection import SequentialFeatureSelector 
#from sklearn.feature_selection import SequentialFeatureSelector

sfs = SequentialFeatureSelector(RandomForestRegressor(n_estimators=3), 
                                k_features=3,
                                forward=True,
                                floating=False, # True to drop selected features
                                scoring='neg_mean_absolute_error',
                                cv=2)

sfs.fit(X_train_sklearn,y_train)

selected_features_sfs = X_train_sklearn.loc[:, sfs.k_feature_names_]
selected_features_sfs

	num__geo_lon	quantile__geo_lat	spline__area_sp_3
0	-0.449742	0.766257	1.826008e-06
1	1.433673	0.297142	1.310449e-06
2	0.047222	0.732330	6.098363e-07
3	-0.722477	0.148789	1.144942e-06
4	1.125819	0.985937	4.240047e-06
...	...	...	...
410770	0.355014	0.788393	1.401454e-06
410771	0.392697	0.494062	8.202272e-07
410772	-0.688830	0.131352	6.098363e-07
410773	-0.804077	0.193143	1.004843e-05
410774	-1.101815	0.908036	3.903343e-06

410775 rows × 3 columns

rfe_sfs_idx = list(sfs.k_feature_idx_)
rfe_sfs_idx
rfe_sfs_col = list(sfs.k_feature_names_)
rfe_sfs_col

['num__geo_lon', 'quantile__geo_lat', 'spline__area_sp_3']

import matplotlib.pyplot as plt
from mlxtend.plotting import plot_sequential_feature_selection as plot_sfs

fig = plot_sfs(sfs.get_metric_dict(), kind='std_dev')

plt.title('Sequential Forward Selection (w. StdDev)')
plt.grid()
plt.show()

rfe_sfs_pipeline = Pipeline(steps=[
    ('preprocessor', preprocessor_sklearn), 
    ('rfe_extractor', ColumnExtractor(rfe_sfs_idx)),
    ('model', regressor)
])

rfe_sfs_pipeline.fit(X_train, y_train)

predictions_sfs = rfe_sfs_pipeline.predict(X_test)

metrics = {}
metrics["mae"] = mean_absolute_error(y_test, predictions_sfs)   
metrics["mape"] = mean_absolute_percentage_error(y_test, predictions_sfs)
metrics["mse"] = mean_squared_error(y_test, predictions_sfs)

metrics
experiment_id = mlflow.get_experiment_by_name(EXPERIMENT_NAME).experiment_id
RUN_NAME = 'rfe_sfs_feature_selection'

with mlflow.start_run(run_name=RUN_NAME, experiment_id=experiment_id) as run:
    # получаем уникальный идентификатор запуска эксперимента
    run_id = run.info.run_id 
    mlflow.sklearn.log_model(rfe_sfs_pipeline, 
                             artifact_path="models",
                             signature=signature,
                             input_example=input_example,
                             pip_requirements=req_file
                             )
    mlflow.log_metrics(metrics)
    mlflow.log_artifact('rfe_skl_cols.txt')
    mlflow.log_artifact('rfe_skl_idx.txt')
    mlflow.log_params(model_sklearn.get_params())

run = mlflow.get_run(run_id) 
assert (run.info.status =='FINISHED')

Можно совмещать признаки, выбранные по sfs и sbs: брать их объединение или пересечение. Можно комбинировать с признаками, выделенными разными подходами - целое поле для исследований

HYPERPARAMS

Gridsearch

from sklearn.model_selection import  GridSearchCV

param_grid = {
    'model__depth': [1,3,5]
}

gs = GridSearchCV(rfe_sfs_pipeline, param_grid, cv=2, scoring='neg_mean_absolute_error')
gs.fit(X_train, y_train)
print("Лучшие гиперпараметры:", gs.best_params_)

gs_pipeline  = Pipeline(steps=[
    ('preprocessor', preprocessor_sklearn), 
    ('rfe_extractor', ColumnExtractor(rfe_sfs_idx)),
    ('model', CatBoostRegressor(depth=5))
])

# Проведем стандартную проверку на тестовом множестве и залогируем run

Вместо GridSearch можно использовать RandomSearch

Optuna

import optuna

def objective(trial):
    # предлагаем гиперпараметры
    depth = trial.suggest_int('depth', 1, 10)
    learning_rate = trial.suggest_float('learning_rate', 0.001, 0.1)

    # создаём и обучаем модель
    opt_pipeline  = Pipeline(steps=[
        ('preprocessor', preprocessor_sklearn), 
        ('rfe_extractor', ColumnExtractor(rfe_sfs_idx)),
        ('model', CatBoostRegressor(depth=depth, learning_rate=learning_rate, verbose=0))
    ])

    opt_pipeline.fit(X_train, y_train)

    # предсказываем и вычисляем RMSE
    preds = opt_pipeline.predict(X_test)
    mae =  mean_absolute_error(y_test, preds)   

    return mae

study = optuna.create_study(direction='minimize')
study.optimize(objective, n_trials=10)

# выводим результаты
print('Number of finished trials:', len(study.trials))
print('Best trial:', study.best_trial.params) 

opt_pipeline  = Pipeline(steps=[
    ('preprocessor', preprocessor_sklearn), 
    ('rfe_extractor', ColumnExtractor(rfe_sfs_idx)),
    ('model', CatBoostRegressor(depth=3, learning_rate=0.02789))
])

# Проведем стандартную проверку на тестовом множестве и залогируем run

Выбираем лучшую модель. Обучаем ее на всей выборке (а не только на train-части). Далее будем деплоить именно её