使用BERT文本分类遇数据适配器错误，请求技术支持

问题背景 #

使用BERT进行文本分类时，执行训练代码触发数据适配器错误，无法正常启动训练流程。

数据集示例 #

现有代码 #

分词器代码 #

def tokenize_function(text):
    return tokenizer(text.numpy(), padding=True, truncation=True, return_tensors='tf')

def tf_tokenize(text):
    result = tf.py_function(tokenize_function, [text], Tout=tf.int32)
    result.set_shape([None, None])
    return result

网络层定义 #

text_input = tf.keras.layers.Input(shape=(), dtype=tf.float32, name='input_ids')
tokenized_input = tf.keras.layers.Lambda(tf_tokenize)(text_input)
outputs = bert_encoder(tokenized_input)
pooled_output = outputs[0][:, 0]

# Neural network layers
l = tf.keras.layers.Dropout(0.1, name="dropout")(pooled_output)
l = tf.keras.layers.Dense(4, activation='sigmoid', name="output")(l)

# Use inputs and outputs to construct a final model
model = tf.keras.Model(inputs=[text_input], outputs = [l]) 

优化器与编译 #

optimizer = tf.keras.optimizers.Adam()
loss = tf.keras.losses.CategoricalCrossentropy()

model.compile(optimizer=optimizer,
              loss=loss,
              metrics='accuracy')

训练代码 #

model.fit(X_train, Y_train, epochs=10)

错误信息 #

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[23], line 1
----> 1 model.fit(X_train, Y_train, epochs=10)

File ~\AppData\Roaming\Python\Python310\site-packages\keras\utils\traceback_utils.py:70, in filter_traceback..error_handler(*args, **kwargs)
     67     filtered_tb = _process_traceback_frames(e.__traceback__)
     68     # To get the full stack trace, call:
     69     # `tf.debugging.disable_traceback_filtering()`
---> 70     raise e.with_traceback(filtered_tb) from None
     71 finally:
     72     del filtered_tb

File ~\AppData\Roaming\Python\Python310\site-packages\keras\engine\data_adapter.py:1082, in select_data_adapter(x, y)
   1079 adapter_cls = [cls for cls in ALL_ADAPTER_CLS if cls.can_handle(x, y)]
   1080 if not adapter_cls:
   1081     # TODO(scottzhu): This should be a less implementation-specific error.
---> 1082     raise ValueError(
   1083         "Failed to find data adapter that can handle input: {}, {}".format(
   1084             _type_name(x), _type_name(y)
   1085         )
   1086     )
   1087 elif len(adapter_cls) > 1:
   1088     raise RuntimeError(
   1089         "Data adapters should be mutually exclusive for "
   1090         "handling inputs. Found multiple adapters {} to handle "
   1091         "input: {}, {}".format(adapter_cls, _type_name(x), _type_name(y))
   1092     )

ValueError: Failed to find data adapter that can handle input: , ( containing values of types {""})

解决方案 #

1. 修正输入层数据类型错误 #

输入层定义时将dtype=tf.float32改为tf.string，因为输入是文本字符串：

text_input = tf.keras.layers.Input(shape=(), dtype=tf.string, name='input_text')

2. 标准化数据集格式 #

• X_train：必须转换为TensorFlow可识别的张量或tf.data.Dataset对象，不能直接传入原始Pandas DataFrame。转换方式：

  X_train_tensor = tf.convert_to_tensor(X_train.values, dtype=tf.string)
  

• Y_train：使用CategoricalCrossentropy损失要求标签为独热编码格式，需先完成标签编码与独热转换：

  from sklearn.preprocessing import LabelEncoder, OneHotEncoder
  import numpy as np
  
  # 标签编码：将字符串标签转为整数
  le = LabelEncoder()
  y_train_encoded = le.fit_transform(Y_train)
  # 独热编码：转为模型可接受的格式
  ohe = OneHotEncoder(sparse_output=False)
  Y_train = ohe.fit_transform(y_train_encoded.reshape(-1, 1))
  

3. 匹配激活函数与损失函数 #

CategoricalCrossentropy损失对应softmax激活函数，而非sigmoid，修改输出层：

l = tf.keras.layers.Dense(4, activation='softmax', name="output")(l)

4. 优化分词流程（推荐） #

将分词逻辑提前到数据集处理阶段，避免在模型内部使用tf.py_function（性能差且易出问题）：

def tokenize_text(text, label):
    # 解码字符串并分词
    tokens = tokenizer(text.numpy().decode('utf-8'), padding='max_length', truncation=True, max_length=128)
    return (tokens['input_ids'], tokens['attention_mask']), label

def tf_tokenize_text(text, label):
    return tf.py_function(tokenize_text, [text, label], Tout=((tf.int32, tf.int32), tf.float32))

# 构建训练数据集
train_dataset = tf.data.Dataset.from_tensor_slices((X_train_tensor, Y_train))
train_dataset = train_dataset.map(tf_tokenize_text).batch(32)

# 调整模型输入为BERT所需的input_ids和attention_mask
input_ids = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name='input_ids')
attention_mask = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name='attention_mask')
outputs = bert_encoder({'input_ids': input_ids, 'attention_mask': attention_mask})
pooled_output = outputs[1]  # BERT的pooled_output通常对应outputs[1]

l = tf.keras.layers.Dropout(0.1)(pooled_output)
l = tf.keras.layers.Dense(4, activation='softmax')(l)

model = tf.keras.Model(inputs=[input_ids, attention_mask], outputs=l)

5. 执行训练 #

如果使用优化后的数据集，训练代码改为：

model.fit(train_dataset, epochs=10)

核心问题总结 #

• 输入层数据类型错误，将文本输入设为float32导致TensorFlow无法识别数据格式。
• 标签格式不匹配，未将字符串标签转为独热编码，无法适配CategoricalCrossentropy损失。
• 激活函数与损失函数不匹配，sigmoid不适用于多分类场景的CategoricalCrossentropy。

内容的提问来源于stack exchange，提问作者Ashok Kumar Reddy Kummetha