TensorFlow 2.16+中keras.layers.DenseFeatures的替代方案是什么？

在TensorFlow 2.16.1及更高版本中，keras.layers.DenseFeatures类已被移除。由于使用Python 3.11.7无法安装TensorFlow 2.15或更早版本，需要找到该类的替代方案来为DNN准备输入层。

原代码示例：

INPUT_COLS = [
    "pickup_longitude",
    "pickup_latitude",
    "dropoff_longitude",
    "dropoff_latitude",
    "passenger_count",
]

feature_columns = {colname: tf.feature_column.numeric_column(colname) for colname in INPUT_COLS}

# Build a keras DNN model using Sequential API
model = Sequential(
    [
        keras.layers.DenseFeatures(feature_columns=feature_columns.values()),
        keras.layers.Dense(units=32, activation="relu", name="h1"),
        keras.layers.Dense(units=8, activation="relu", name="h2"),
        keras.layers.Dense(units=1, activation="linear", name="output"),
    ]
)

针对你的全数值型特征场景，以下是几种实用的替代方案：

方案1：直接使用Keras Input层（最简便） #

既然所有输入都是数值型特征，无需通过特征列做额外处理，可以直接定义Input层接收每个特征，拼接后送入全连接层：

import tensorflow as tf
from tensorflow.keras import Sequential, layers

INPUT_COLS = [
    "pickup_longitude",
    "pickup_latitude",
    "dropoff_longitude",
    "dropoff_latitude",
    "passenger_count",
]

# 定义每个特征的Input层
inputs = {col: layers.Input(shape=(1,), name=col) for col in INPUT_COLS}
# 拼接所有输入特征
x = layers.concatenate(list(inputs.values()))
# 构建后续网络
x = layers.Dense(units=32, activation="relu", name="h1")(x)
x = layers.Dense(units=8, activation="relu", name="h2")(x)
output = layers.Dense(units=1, activation="linear", name="output")(x)

# 构建模型
model = tf.keras.Model(inputs=inputs, outputs=output)

如果你的数据是以张量形式直接输入（而非字典形式），可简化为单个Input层：

model = Sequential(
    [
        layers.Input(shape=(len(INPUT_COLS),)),  # 输入形状为特征数量
        layers.Dense(units=32, activation="relu", name="h1"),
        layers.Dense(units=8, activation="relu", name="h2"),
        layers.Dense(units=1, activation="linear", name="output"),
    ]
)

方案2：使用FeatureSpace API（官方推荐的特征处理新方式） #

TensorFlow推出的FeatureSpace是特征列的替代方案，更简洁且整合了预处理逻辑，适合复杂特征场景：

import tensorflow as tf
from tensorflow.keras import Sequential, layers

INPUT_COLS = [
    "pickup_longitude",
    "pickup_latitude",
    "dropoff_longitude",
    "dropoff_latitude",
    "passenger_count",
]

# 定义特征空间，所有特征均为数值型
feature_space = layers.experimental.preprocessing.FeatureSpace(
    features={col: "numeric" for col in INPUT_COLS},
    output_mode="concat"  # 拼接所有特征输出
)

# 适配输入数据（只需执行一次，用于确定特征形状）
sample_input = {col: tf.random.uniform((1,)) for col in INPUT_COLS}
feature_space.adapt(sample_input)

# 构建模型
model = Sequential(
    [
        feature_space,
        layers.Dense(units=32, activation="relu", name="h1"),
        layers.Dense(units=8, activation="relu", name="h2"),
        layers.Dense(units=1, activation="linear", name="output"),
    ]
)

方案3：保留特征列定义，用Lambda层封装input_layer #

如果你想保留原有的feature_columns定义，可以用tf.feature_column.input_layer结合Lambda层替代DenseFeatures：

import tensorflow as tf
from tensorflow.keras import Sequential, layers

INPUT_COLS = [
    "pickup_longitude",
    "pickup_latitude",
    "dropoff_longitude",
    "dropoff_latitude",
    "passenger_count",
]

feature_columns = {colname: tf.feature_column.numeric_column(colname) for colname in INPUT_COLS}
feature_columns_list = list(feature_columns.values())

# 用Lambda层封装input_layer逻辑
def dense_features_wrapper(inputs):
    return tf.feature_column.input_layer(inputs, feature_columns_list)

model = Sequential(
    [
        layers.Lambda(dense_features_wrapper, name="dense_features"),
        layers.Dense(units=32, activation="relu", name="h1"),
        layers.Dense(units=8, activation="relu", name="h2"),
        layers.Dense(units=1, activation="linear", name="output"),
    ]
)

# 注意：模型输入需为字典格式，需先适配输入形状
sample_input = {col: tf.random.uniform((1,)) for col in INPUT_COLS}
model(sample_input)

内容的提问来源于stack exchange，提问作者shahramy