TF2与Python中BERT预处理器模型报错问题求助

错误信息 #

ValueError                                Traceback (most recent call last)
Cell In[37], line 4
      2 text_input = tf.keras.Input(shape=(), dtype=tf.string, name='text')
      3 bert_preprocess = hub.KerasLayer(preprocess_url, name='preprocessing')
----> 4 preprocessed_text = bert_preprocess(text_input)
      5 bert_encoder = hub.KerasLayer(encoder_url, 
      6                               trainable=True, 
      7                               name='BERT_encoder')
      8 outputs = bert_encoder(preprocessed_text)
ValueError: Exception encountered when calling layer 'preprocessing' (type KerasLayer).
A KerasTensor is symbolic: it's a placeholder for a shape an a dtype. It doesn't have any actual numerical value. You cannot convert it to a NumPy array.

Call arguments received by layer 'preprocessing' (type KerasLayer):
  • inputs=<KerasTensor shape=(None,), dtype=string, sparse=None, name=text>
  • training=None

A KerasTensor is symbolic: it's a placeholder for a shape an a dtype. It doesn't have any actual numerical value. You cannot convert it to a NumPy array.

触发错误的模型代码 #

preprocess_url = 'https://www.kaggle.com/models/tensorflow/bert/frameworks/TensorFlow2/variations/en-uncased-preprocess/versions/3'
encoder_url = 'https://www.kaggle.com/models/tensorflow/bert/frameworks/TensorFlow2/variations/bert-en-uncased-l-12-h-768-a-12/versions/2'

# Bert Layers
text_input = tf.keras.Input(shape=(), dtype=tf.string, name='text')
bert_preprocess = hub.KerasLayer(preprocess_url, name='preprocessing')
preprocessed_text = bert_preprocess(text_input)
bert_encoder = hub.KerasLayer(encoder_url, 
                              trainable=True, 
                              name='BERT_encoder')
outputs = bert_encoder(preprocessed_text)

# Neural network layers
l = tf.keras.layers.Dropout(0.1)(outputs['pooled_output'])
l = tf.keras.layers.Dense(num_classes, activation='softmax', name='output')(l)

# Construct final model
model = tf.keras.Model(inputs=[text_input], outputs=[l])

环境与已尝试方案 #

• 当前环境：tensorflow-gpu-jupyter Docker容器，Tensorflow 2.16.1、tensorflow-text 2.16.1、tensorflow-hub 0.16.1
• 已尝试：更换TensorFlow、tensorflow-text、tensorflow-hub版本；启用tf.config.run_functions_eagerly(True)；参考官方文档及教程复制代码均无效

解决思路 #

• 替换预处理器为TensorFlow Hub官方版本：Kaggle托管的预处理器层可能存在符号张量兼容性问题，改用官方托管的兼容版本：

  preprocess_url = "https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3"
  encoder_url = "https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/4"
  

• 重新安装匹配版本的依赖包：确保GPU环境下依赖包完全兼容，执行以下命令重装：

  pip uninstall -y tensorflow-text tensorflow-hub
  pip install tensorflow-text==2.16.1 tensorflow-hub==0.16.1 --no-cache-dir
  

• 为预处理器层显式指定输入签名：强制层适配符号张量输入：

  bert_preprocess = hub.KerasLayer(
      preprocess_url,
      name='preprocessing',
      input_signature=[tf.TensorSpec(shape=(None,), dtype=tf.string)]
  )
  

• 改用分步式Functional API构建：避免定义时直接调用层，改为链式连接：

  inputs = tf.keras.Input(shape=(), dtype=tf.string)
  x = hub.KerasLayer(preprocess_url)(inputs)
  x = hub.KerasLayer(encoder_url, trainable=True)(x)
  x = tf.keras.layers.Dropout(0.1)(x['pooled_output'])
  outputs = tf.keras.layers.Dense(num_classes, activation='softmax')(x)
  model = tf.keras.Model(inputs=inputs, outputs=outputs)
  

• 降级TensorFlow版本到2.15.x：TensorFlow 2.16.x对KerasLayer的符号张量处理逻辑有变更，降级到稳定版本可规避问题：

  pip install tensorflow==2.15.1 tensorflow-text==2.15.1 tensorflow-hub==0.15.0 --no-cache-dir
  

内容的提问来源于stack exchange，提问作者vmmgame