Flask部署MultinomialNB模型时特征数不匹配错误的解决方法

问题概述 #

本地训练MultinomialNB模型一切正常，但通过Flask API调用预测时，出现错误：

ValueError: X has 25 features, but MultinomialNB is expecting 26 features as input

训练代码（model.py）和部署代码（app.py）如下：

model.py 代码 #

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.preprocessing import MinMaxScaler, LabelEncoder
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import OrdinalEncoder
from sklearn.metrics import confusion_matrix , classification_report
import pickle

train = pd.read_csv("train_data_evaluation_part_2.csv", index_col=False)
test = pd.read_csv("test_data_evaluation_part2.csv", index_col=False)
df = pd.concat([train,test],axis=0)

def unique_obj_col_value(df):
  for column in df:
    if df[column].dtype == 'object':
      print(f'{column}: {df[column].unique()}')

unique_obj_col_value(df)

df = df.drop('Nationality', axis=1)

df['BookingsCheckedIn'] = df['BookingsCheckedIn'].replace([ 3,  1, 9,  2, 11, 12,  7,  8,  5,  6,  4, 66, 15, 29, 25, 10,
   17, 13, 26, 23, 57, 40, 18, 14, 24, 19, 20, 34], 1)

cols = ['DistributionChannel', 'MarketSegment']
le = LabelEncoder()
for col in cols:
    df[col] = le.fit_transform(df[col])

pickle.dump(le, open('transform.pkl', 'wb'))

df.drop(['Unnamed: 0', 'ID'], axis=1, inplace=True)
df['Age'].fillna(np.mean(df['Age']),inplace=True)

X = df.drop('BookingsCheckedIn', axis=1)
y = df['BookingsCheckedIn']

cols_to_scale = ['Age','DaysSinceCreation', 'AverageLeadTime', 'LodgingRevenue', 'OtherRevenue','DaysSinceLastStay', 'DaysSinceFirstStay']
scaler = MinMaxScaler()
X[cols_to_scale] = scaler.fit_transform(X[cols_to_scale])
print(X.shape)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
print(X_train.shape)
from sklearn.naive_bayes import MultinomialNB

clf = MultinomialNB()
clf.fit(X_train,y_train)
clf.score(X_test,y_test)

yp = clf.predict(X_test)
yp[:5]

y_pred = []
for element in yp:
   if element > 0.5:
      y_pred.append(1)
   else:
      y_pred.append(0)


print(classification_report(y_test,y_pred))

pickle.dump(clf, open('model1.pkl', 'wb'))
model = pickle.load(open('model1.pkl', 'rb'))

app.py 代码 #

import numpy as np
from flask import Flask, request, jsonify, render_template
import pickle

app = Flask(__name__)
model = pickle.load(open('model1.pkl','rb'))
trans = pickle.load(open('transform.pkl', 'rb'))

@app.route('/')
def home():
    return render_template('index.html')

@app.route('/predict',methods=['POST'])
def predict():
    '''
    For rendering results on HTML GUI
    '''
    features = [x for x in request.form.values()]
    final_features = [np.array(features, dtype=float)]
    prediction = model.predict(final_features)

    output = round(prediction[0], 2)

    return render_template('index.html', prediction_text='Booked: ${}'.format(output))


if __name__ == "__main__":
    app.run(debug=True)

错误原因 #

1. 预处理不完整且未复用：训练时对分类列做了LabelEncoder转换、数值列做了MinMaxScaler缩放，但部署时仅加载了LabelEncoder且未正确应用，也未加载Scaler做缩放，导致输入特征不符合模型要求。
2. LabelEncoder使用错误：用同一个LabelEncoder连续处理两个分类列，最终保存的编码器仅适配最后一个列（MarketSegment），无法正确转换第一个列（DistributionChannel）。
3. 输入特征数量/顺序不匹配：网页表单输入的特征数量可能少于训练时X的26列，或特征顺序与训练时不一致，导致模型接收的特征数不足。

修复步骤 #

步骤1：修正训练代码，保存所有必要的预处理组件 #

修改model.py中的预处理部分，单独保存每个分类列的编码器和数值缩放器，并记录特征列顺序：

# 替换原LabelEncoder循环，为每个分类列单独创建并保存编码器
cols = ['DistributionChannel', 'MarketSegment']
# 处理DistributionChannel
le_dist = LabelEncoder()
df['DistributionChannel'] = le_dist.fit_transform(df['DistributionChannel'])
pickle.dump(le_dist, open('le_dist.pkl', 'wb'))
# 处理MarketSegment
le_market = LabelEncoder()
df['MarketSegment'] = le_market.fit_transform(df['MarketSegment'])
pickle.dump(le_market, open('le_market.pkl', 'wb'))

# 保存数值缩放器
scaler = MinMaxScaler()
X[cols_to_scale] = scaler.fit_transform(X[cols_to_scale])
pickle.dump(scaler, open('scaler.pkl', 'wb'))

# 保存训练时的特征列顺序，确保部署时输入顺序一致
pickle.dump(X.columns.tolist(), open('feature_names.pkl', 'wb'))

步骤2：修改部署代码，正确应用预处理 #

更新app.py，加载所有预处理组件，并对输入特征做与训练时完全一致的转换：

import numpy as np
from flask import Flask, request, jsonify, render_template
import pickle

app = Flask(__name__)
# 加载模型和所有预处理组件
model = pickle.load(open('model1.pkl','rb'))
le_dist = pickle.load(open('le_dist.pkl', 'rb'))
le_market = pickle.load(open('le_market.pkl', 'rb'))
scaler = pickle.load(open('scaler.pkl', 'rb'))
feature_names = pickle.load(open('feature_names.pkl', 'rb'))
cols_to_scale = ['Age','DaysSinceCreation', 'AverageLeadTime', 'LodgingRevenue', 'OtherRevenue','DaysSinceLastStay', 'DaysSinceFirstStay']
# 获取需要缩放的特征在列表中的索引
scale_indices = [feature_names.index(col) for col in cols_to_scale]

@app.route('/')
def home():
    return render_template('index.html')

@app.route('/predict',methods=['POST'])
def predict():
    # 获取表单输入，顺序必须与feature_names完全一致
    features = [x for x in request.form.values()]
    
    # 转换分类特征
    features[feature_names.index('DistributionChannel')] = le_dist.transform([features[feature_names.index('DistributionChannel')]])[0]
    features[feature_names.index('MarketSegment')] = le_market.transform([features[feature_names.index('MarketSegment')]])[0]
    
    # 转换为数组
    processed_features = np.array(features, dtype=float)
    
    # 缩放数值特征
    processed_features[scale_indices] = scaler.transform([processed_features[scale_indices]])[0]
    
    # 转为模型要求的2D数组格式
    final_features = [processed_features]
    
    prediction = model.predict(final_features)
    # 分类问题直接输出类别，无需round
    output = "Yes" if prediction[0] == 1 else "No"

    return render_template('index.html', prediction_text=f'Booked: {output}')

if __name__ == "__main__":
    app.run(debug=True)

步骤3：确认网页表单的输入字段 #

检查index.html中的表单，确保输入字段的数量为26个，且顺序与feature_names完全一致，每个字段对应训练时X的一列。

验证 #

重新运行model.py生成新的模型和预处理文件，启动Flask服务后测试预测，即可解决特征数不匹配的问题。

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