如何解决“Failed to convert a NumPy array to a Tensor (Unsupported object type int)”错误

问题根源 #

你的输入特征feat包含未处理的文本列（text，object类型）和数值列（attachment，float类型），但模型第一层是Embedding层——该层仅能接收文本经分词编码后的整数序列，直接传入原始DataFrame会因类型不兼容无法转为Tensor，触发报错。

具体修复方案 #

1. 文本预处理：将text转为整数序列 #

用Tokenizer对文本分词、编码，转成固定长度的整数序列，适配Embedding层输入要求：

from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences

def preprocess_text(texts, max_feature, max_len):
    tokenizer = Tokenizer(num_words=max_feature)
    tokenizer.fit_on_texts(texts)
    sequences = tokenizer.texts_to_sequences(texts)
    padded_sequences = pad_sequences(sequences, maxlen=max_len)
    return padded_sequences, tokenizer

2. 分离并处理数值特征 #

提取attachment列，转成二维numpy数组，方便后续和文本特征拼接：

def process_numeric_features(df):
    numeric_feat = df['attachment'].values.reshape(-1, 1)
    return numeric_feat

3. 修改模型结构：支持多输入（文本+数值） #

改用函数式API构建模型，分别处理文本和数值特征后拼接输出：

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Concatenate

def model(max_feature, embed_size, max_len):
    # 文本分支
    text_input = Input(shape=(max_len,), name='text_input')
    x = Embedding(max_feature, embed_size, input_length=max_len)(text_input)
    x = Conv1D(filters=64, kernel_size=7, padding='same', activation='relu')(x)
    x = MaxPooling1D(pool_size=2, padding='same')(x)
    x = Dropout(0.1)(x)
    x = Bidirectional(LSTM(64, recurrent_dropout=0.02, return_sequences=True))(x)
    x = GlobalMaxPool1D()(x)
    x = Dropout(0.1)(x)
    text_output = Dense(16, activation='relu')(x)

    # 数值分支
    numeric_input = Input(shape=(1,), name='numeric_input')
    numeric_output = Dense(8, activation='relu')(numeric_input)

    # 拼接分支输出
    combined = Concatenate()([text_output, numeric_output])
    final_output = Dense(train_args["total_classes"], activation='softmax')(combined)

    model = Model(inputs=[text_input, numeric_input], outputs=final_output)
    model.compile(optimizer='adam',  
                  loss='categorical_crossentropy',
                  metrics=['accuracy'])
    logger.info(model.summary())
    return model

4. 调整数据拆分与训练流程 #

在数据拆分和训练时，分别传入处理后的文本与数值特征：

def split_data(text_feat, numeric_feat, label):
    x_text_train, x_text_val, x_num_train, x_num_val, y_train, y_val = train_test_split(
        text_feat, numeric_feat, encode_labels(label), test_size=0.20, random_state=42, shuffle=True)
    data = {
        "train": {"text": x_text_train, "numeric": x_num_train, "y": y_train},
        "test": {"text": x_text_val, "numeric": x_num_val, "y": y_val}
    }
    return data

def train_model(data, train_args):    
    logger.info("Starting training Neural Network")
    nn_model = model(train_args["max_feature"], train_args["embed_size"], train_args["max_len"])
    history = nn_model.fit(
        [data["train"]["text"], data["train"]["numeric"]], 
        data["train"]["y"], 
        batch_size=64, 
        epochs=10, 
        validation_data=([data["test"]["text"], data["test"]["numeric"]], data["test"]["y"])
    )
    logger.info("Finished training Neural Network")
    return nn_model

def main():
    data_dir = "data"
    data_file = os.path.join(data_dir, 'data.csv')
    train_df = pd.read_csv(data_file)
    label = train_df["category"]
    
    # 预处理文本和数值特征
    text_feat, tokenizer = preprocess_text(train_df['text'], train_args["max_feature"], train_args["max_len"])
    numeric_feat = process_numeric_features(train_df)
    
    data = split_data(text_feat, numeric_feat, label)
    model = train_model(data, train_args)

    metrics = get_model_metrics(model, data)
    for (k, v) in metrics.items():
        print(f"{k}: {v}")

5. 适配评估函数 #

评估时同样传入多输入：

def get_model_metrics(model, data):
    preds = model.predict([data["test"]["text"], data["test"]["numeric"]])
    accuracy = accuracy_score(np.argmax(preds, axis=1), np.argmax(data["test"]["y"],axis=1))
    metrics = {"accuracy_score": accuracy}
    return metrics

内容的提问来源于stack exchange，提问作者Deepak Soni