TensorFlow模型训练报错：output_signature需为tf.TypeSpec子类而非list

问题背景 #

使用微软AutonomousDrivingCookbook中的AirSim训练笔记本及DriveDataGenerator类训练模型时，触发TypeError，提示output_signature必须包含tf.TypeSpec的子类对象，但检测到<class 'list'>。检查生成器代码未发现明显问题，需要修复该错误以完成训练。

报错日志 #

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[26], line 1
----> 1 history = model.fit(train_generator, steps_per_epoch=num_train_examples//batch_size, epochs=500, callbacks=callbacks,
      2                    validation_data=eval_generator, validation_steps=num_eval_examples//batch_size, verbose=1)

File ~\anaconda3\envs\my-conda-env\Lib\site-packages\keras\src\utils\traceback_utils.py:122, in filter_traceback.<locals>.error_handler(*args, **kwargs)
    119     filtered_tb = _process_traceback_frames(e.__traceback__)
    120     # To get the full stack trace, call:
    121     # `keras.config.disable_traceback_filtering()`
--> 122     raise e.with_traceback(filtered_tb) from None
    123 finally:
    124     del filtered_tb

File ~\anaconda3\envs\my-conda-env\Lib\site-packages\tensorflow\python\data\ops\from_generator_op.py:124, in _from_generator(generator, output_types, output_shapes, args, output_signature, name)
    122   for spec in nest.flatten(output_signature):
    123     if not isinstance(spec, type_spec.TypeSpec):
--> 124       raise TypeError(f"`output_signature` must contain objects that are "
    125                       f"subclass of `tf.TypeSpec` but found {type(spec)} "
    126                       f"which is not.")
    127 else:
    128   if output_types is None:

TypeError: `output_signature` must contain objects that are subclass of `tf.TypeSpec` but found <class 'list'> which is not.

报错触发代码 #

history = model.fit(train_generator, steps_per_epoch=num_train_examples//batch_size, epochs=500, callbacks=callbacks,
                   validation_data=eval_generator, validation_steps=num_eval_examples//batch_size, verbose=1)

解决方案 #

该错误源于TensorFlow版本兼容性问题：旧版AutonomousDrivingCookbook代码适配早期TensorFlow版本，新版TF（2.7+）对tf.data.Dataset.from_generator的output_signature格式要求更严格，不允许直接传入列表类型的签名。以下是具体修复步骤：

1. 修改DriveDataGenerator的输出结构 #

找到Generator.py中的__getitem__方法，将返回的列表改为元组或numpy数组：

• 原代码示例：

  return image_array, [steering_angle, throttle]
  

• 修改为：

  # 改为元组
  return image_array, (steering_angle, throttle)
  # 或直接返回numpy数组
  return image_array, np.array([steering_angle, throttle], dtype=np.float32)
  

2. 手动指定Dataset的output_signature #

如果生成器被包装为tf.data.Dataset，需明确指定符合tf.TypeSpec的输出签名：

import tensorflow as tf

# 根据你的数据尺寸调整shape参数
output_signature = (
    tf.TensorSpec(shape=(140, 256, 3), dtype=tf.float32),
    tf.TensorSpec(shape=(2,), dtype=tf.float32)
)

train_dataset = tf.data.Dataset.from_generator(
    lambda: train_generator,
    output_signature=output_signature
)

之后用train_dataset替代train_generator传入model.fit即可。

3. 降级TensorFlow版本（可选） #

如果不想修改代码，可以将TensorFlow降级到2.6或更早版本，适配原Cookbook的代码逻辑：

pip install tensorflow==2.6

内容的提问来源于stack exchange，提问作者D S Raigagla