如何在 Flask 上的 sklearn 管道中使用自定義轉換？

Question

我無法使用 sklearn 管道在 Flask 中使用picke.dump()進行自定義轉換。

假設我想在泡菜中使用以下管道來服務 Flask。 它選擇變量，創建虛擬變量並將 go 的變量分離到 model 中。

# Custom Transformer that extracts columns passed as argument to its constructor
class FeatureSelector(BaseEstimator, TransformerMixin):
    # Class Constructor
    def __init__(self, feature_names):
        self._feature_names = feature_names

    # Return self nothing else to do here
    def fit(self, X, y=None):
        return self

    # Method that describes what we neeed this transformer to do
    def transform(self, X, y=None):
        return X[self._feature_names]

# Defining the steps in the categorical pipeline
categorical_features = ['marital', 'contact', 'job']

# Converts certain features to binary
class CategoricalBinary(TransformerMixin):
    # Class Constructor
    #def __init__(self):

    # Return self nothing else to do here
    def fit(self, X, y=None):
        return self

    # Faz as transformações com a função get_dummies
    def transform(self, X, y=None):
        X = pd.get_dummies(X, columns=X.columns.tolist())
        return X

# Custom Transformer that extracts columns passed as argument to its constructor
class ModelFeatureSelector(BaseEstimator, TransformerMixin):
    # Class Constructor
    def __init__(self, feature_names):
        self._feature_names = feature_names

    # Return self nothing else to do here
    def fit(self, X, y=None):
        return self

    # Method that describes what we neeed this transformer to do
    def transform(self, X, y=None):
        return X[self._feature_names]

model_features = ['marital_divorced', 'marital_married', 'marital_single',
                        'contact_cellular', 'job_admin.', 'job_blue-collar',
                        'job_entrepreneur', 'job_housemaid', 'job_management',
                        'job_retired', 'job_self-employed', 'job_services',
                        'job_student', 'job_technician', 'job_unemployed']

categorical_transform = Pipeline(steps=[('feature_selector', FeatureSelector(categorical_features)),
                                        ('categorical_dummy', CategoricalBinary()),
                                        ('model_features', ModelFeatureSelector(model_features)),
                                        ('logreg', LogisticRegression(class_weight='balanced', solver='liblinear'))])

# Fits
categorical_transform.fit(X_train, y_train)

# Save on pickle
with open('categorical_transform.pkl', 'wb') as f:
    pickle.dump(categorical_transform, f)

我編程了 Flask

# Importa as classes necessárias do pacote `flask`
from flask import Flask, request, jsonify

# Importa o pacote de interação com o sistema `os`, o pacote `pandas` para maniulação da informação
# e o pacote `pickle` para carregar nosso modelo.
import os
import pandas as pd
import pickle
from sklearn.pipeline import FeatureUnion, Pipeline
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.impute import SimpleImputer
import outros

# Cria a aplicação
app = Flask(__name__)

# Model
model = pickle.load(open('categorical_transform.pkl', 'rb'))

# TAREFA: PREENCHA O MÉTODO DE REQUISIÇÃO
@app.route('/predict', methods=['POST'])
def predictor():
    # Recebe o conteúdo da postagem no formato json.
    content = request.json
    features = pd.DataFrame([content])

    # Aplica a predição do modelo
    predito = model.predict(features)

    # Cria e envia uma resposta para o 'chamador' da API
    return jsonify(status='completed', predict=float(predito[1][1]))

# Essa linha garante que a aplicação execute no localhost, ou seja, no IP "0.0.0.0"
# e que esteja na porta padrão do sistema ou, caso ela não exista, na porta 8080.
if __name__ == '__main__':
    app.run(debug=True,host='0.0.0.0',port=os.environ.get('PORT', 8080))

引發以下異常：

AttributeError: Can't get attribute 'FeatureSelector'

Answer 1

您已經在一些 python 腳本中創建了一個自定義轉換器FeatureSelector 。 在 Flask 腳本中，您導入自定義 class 的父類，但沒有導入FeatureSelector本身。 因此，在 Flask 腳本中沒有自定義 class 的定義。 因此，pickle 無法重建 object。

假設FeatureSelector在 custom_classes.py 中定義。 那么你的 Flask 腳本應該包含一行：

from custom_classes import FeatureSelector