Keras混合数据多输入模型训练报错：输入张量数量不匹配求助

我尝试用Embedding层处理包含分类特征与数值特征的混合数据，采用Kaggle房价数据集保证可复现性，编写的初步代码如下：

from tensorflow.keras.layers import Normalization, Dense, Flatten, Embedding, Input, Reshape, Concatenate, Dropout, Activation
import tensorflow as tf
import os
import pandas as pd
import numpy as np
from keras.utils import plot_model
from IPython.display import Image
from keras.models import Sequential, Model

# 参考并改编自相关教程
class EmbeddingMapping():
    """
    处理分类变量的辅助类
    每个分类变量对应一个该类的实例
    """

    def __init__(self, series, name):
        # 获取唯一值列表
        values = series.unique().tolist()

        # 建立值到整数的映射字典
        self.embedding_dict = {value: int_value + 1 for int_value, value in enumerate(values)}

        # num_values用于定义Embedding层的input_dim，同时作为未见过值的映射结果
        self.num_values = len(values) + 1
        self.name = name

    def get_mapping(self, value):
        # 训练集中见过的值返回对应的整数映射
        if value in self.embedding_dict:
            return self.embedding_dict[value]

        # 未见过的值返回统一的整数
        else:
            return self.num_values
    
    def get_name(self):
        return self.name 
    
    def get_embedding_name(self):
        return 'embedded_' + self.name 
    
    def get_embedding_size(self):    
        return int(min(np.ceil((len(self.embedding_dict))/2), 50))

os.chdir("../data")

train_data = pd.read_csv("train.csv")

target_name = "SalePrice"

categorical_features = [
      "GarageFinish" # 分类特征
    , "Neighborhood" # 分类特征
]

numerical_features = [
      "GrLivArea" 
    , "TotalBsmtSF"
    , "BsmtFinSF1"
    , "YearRemodAdd"
    , "YearBuilt"
    , "GarageCars"
    , "LotArea"   
    , "1stFlrSF"
    , "GarageArea"
    , "Fireplaces"
    , "2ndFlrSF" 
    
    , "OverallQual" # 有序分类特征
    , "OverallCond"  # 有序分类特征
    
    # TODO 将这些有序特征映射为数值
    # , "BsmtQual" # 有序分类特征
    # , "KitchenQual" # 有序分类特征
]

relevant_columns = []
relevant_columns.extend(categorical_features)
relevant_columns.extend(numerical_features)
relevant_columns.append(target_name)

sub_data = train_data[relevant_columns]

embedding_mappings=[]
embedding_inputs=[]
embeddings=[]
for categorical_column in categorical_features:
    
    embedding_mapping = EmbeddingMapping(sub_data[categorical_column], categorical_column)
    input = Input(shape=(1,), dtype='int32')
    embedding_size=embedding_mapping.get_embedding_size()
    embedding = Embedding(output_dim=embedding_size, input_dim=embedding_mapping.num_values , input_length=1, name=categorical_column)(input)
    embedding = Reshape(target_shape=(embedding_size,))(embedding) 
    
    embedding_mappings.append(embedding_mapping)
    embedding_inputs.append(input)
    embeddings.append(embedding)
    
    sub_data = sub_data.assign(temp_name=sub_data[categorical_column].apply(embedding_mapping.get_mapping))
    sub_data.rename(columns={'temp_name': embedding_mapping.get_embedding_name()}, inplace=True)
    
    sub_data.drop(categorical_column, axis = 1, inplace=True)

normalisation_inputs=[]
normalisations=[]
for numerical_feature in numerical_features:
    
    input = Input(shape=(1,))
    norm = Normalization(name=numerical_feature)(input)
    
    normalisation_inputs.append(input)
    normalisations.append(norm)
 
embeddings.extend(normalisations)
embedding_inputs.extend(normalisation_inputs)

output = Concatenate()(embeddings)
output = Dense(512, kernel_initializer="uniform")(output)
output = Activation('relu')(output)
output= Dropout(0.4)(output)
output = Dense(256, kernel_initializer="uniform")(output)
output = Activation('relu')(output)
output= Dropout(0.3)(output)
output = Dense(1, activation='linear')(output)

model = Model(inputs=embedding_inputs, outputs=output)
model.compile(loss=tf.keras.losses.MeanAbsolutePercentageError(), optimizer=tf.keras.optimizers.Adam(learning_rate=0.01))

y_train = np.log(sub_data[target_name])
X_train = sub_data.drop(target_name, axis=1)
model.fit(X_train, y_train, epochs=10, batch_size=30)

报错信息 #

训练时触发如下错误：

ValueError: Layer "model_6" expects 15 input(s), but it received 1 input tensors. Inputs received: [<tf.Tensor 'IteratorGetNext:0' shape=(None, 15) dtype=int64>]

模型结构 #

猜想解决方案 #

我提出了如下解决方案，希望得到验证：

input_train_list = []
for column in X_train.columns:
  input_train_list.append(X_train[column ].values)

model.fit(input_train_list, y_train, epochs=10, batch_size=30)

内容的提问来源于stack exchange，提问作者cs0815