TensorFlow ValueError：NumPy数组转Tensor失败（不支持Timestamp）求助

问题场景 #

使用Megasena结果CSV文件训练模型时，执行model.fit()出现以下错误：

Tensorflow ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type Timestamp)

原代码片段：

data['Data'] = pd.to_datetime(data['Data'], format='%d/%m/%Y')
features = data[['Data', 'Conc']]
labels = data[['NR1', 'NR2', 'NR3', 'NR4', 'NR5', 'NR6']]

# convert date column to numerical features
features['Day'] = features['Data'].dt.day 
features['Month'] = features['Data'].dt.month 
features['Year'] = features['Data'].dt.year  

# normalize the features (year should have more weightage than day and month)
features['Day'] = features['Day'] / 31.0
features['Month'] = features['Month'] / 12.0
features['Year'] = (features['Year'] - 2000) / 20.0

# split the data into training and testing sets
train_features = features[features['Conc'] <= 2601].values
train_labels = labels[labels.index <= 2601].values
test_features = features[features['Conc'] > 2601].values

# define the model architecture
model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(64, input_shape=[5], activation='relu'),
    tf.keras.layers.Dense(32, activation='relu'),
    tf.keras.layers.Dense(6, activation='softmax')
])

# train the model
model.fit(train_features, train_labels, epochs=100)

错误原因 #

核心问题在于features数据框中保留了原始的Data列（Timestamp类型），当调用.values转换为NumPy数组时，数组中混合了Timestamp对象和数值类型，TensorFlow无法将这种混合类型数组转换为张量。你之前尝试转换Day/Month/Year的类型，但并没有处理Data列，所以错误依然存在。

解决方案 #

1. 移除非数值特征列 #

从features中删除Data列，只保留处理后的数值特征（Conc、Day、Month、Year）。

2. 修正模型输入维度 #

移除Data列后，特征数量从5个变为4个，因此模型的input_shape需要从[5]改为[4]。

3. 补充模型编译步骤 #

训练前必须调用model.compile()配置优化器、损失函数等参数，否则无法启动训练。

修正后的完整代码 #

import pandas as pd
import tensorflow as tf

# 读取数据（请替换为你的CSV路径）
# data = pd.read_csv('megasena_results.csv')

# 日期处理
data['Data'] = pd.to_datetime(data['Data'], format='%d/%m/%Y')

# 初始化特征集，只保留数值型的Conc列
features = data[['Conc']]
labels = data[['NR1', 'NR2', 'NR3', 'NR4', 'NR5', 'NR6']]

# 生成日期衍生特征并归一化，直接基于data['Data']计算
features['Day'] = data['Data'].dt.day / 31.0
features['Month'] = data['Data'].dt.month / 12.0
features['Year'] = (data['Data'].dt.year - 2000) / 20.0

# 用统一掩码拆分训练测试集，确保索引对齐
train_mask = data['Conc'] <= 2601
train_features = features[train_mask].values
train_labels = labels[train_mask].values
test_features = features[~train_mask].values

# 定义模型，输入维度改为4（Conc、Day、Month、Year）
model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(64, input_shape=[4], activation='relu'),
    tf.keras.layers.Dense(32, activation='relu'),
    tf.keras.layers.Dense(6, activation='softmax')
])

# 编译模型：根据标签类型选择损失函数
# 若标签是整数形式（如1-60的号码），使用sparse_categorical_crossentropy
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# 启动训练
model.fit(train_features, train_labels, epochs=100)

额外说明 #

• 损失函数选择：如果彩票号码标签已转为one-hot编码，将loss改为categorical_crossentropy。
• 索引对齐：使用统一的掩码（train_mask）拆分特征和标签，避免因索引不一致导致数据错位。

内容的提问来源于stack exchange，提问作者Rodrigo de Souza Ferreira