TensorFlow项目报错：'str' object has no attribute 'base_dtype'求助

问题详情 #

运行音频序列预测项目时，调用model.fit()触发如下错误：

AttributeError: 'str' object has no attribute 'base_dtype'

原始代码片段 #

print("train_x.shape", train_x.shape)
print("train_y.shape", train_y.shape)
print("Data type of train_x:", train_x.dtype)
print("Data type of train_y:", train_y.dtype)
print("Data type of train_y[0]:", train_y[0].dtype)


model = keras.models.Sequential([
    keras.Input(shape=(train_x.shape[1],), name="Input"),
    keras.layers.Dense(512, activation="linear", name="Entry"),
    keras.layers.Dense(train_y.shape[1], activation="linear", name="Output"),
    #keras.layers.Dense(train_y.shape[1], activation=None, name="Output"),
])

print(model.summary())

model.compile(
    loss='mean_squared_error',
    optimizer=keras.optimizers.Adam(learning_rate=0.0001),
    metrics=["accuracy"],
)
print("Model input shape:", model.input_shape)

model.fit(train_x, train_y, epochs=10, batch_size=32)

控制台输出 #

train_x.shape (22, 5000)
train_y.shape (22, 5000)
Data type of train_x: <dtype: 'float32'>
Data type of train_y: <dtype: 'float32'>
Data type of train_y[0]: <dtype: 'float32'>

Model: "sequential_3"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃ Layer (type)                       ┃ Output Shape                  ┃     Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ Entry (Dense)                      │ (None, 512)                   │   2,560,512 │
├────────────────────────────────────┼───────────────────────────────┼─────────────┤
│ Output (Dense)                     │ (None, 5000)                  │   2,565,000 │
└────────────────────────────────────┴───────────────────────────────┴─────────────┘

 Total params: 5,125,512 (19.55 MB)

 Trainable params: 5,125,512 (19.55 MB)

 Non-trainable params: 0 (0.00 B)

None
Model input shape: (None, 5000)

报错堆栈 #

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[11], line 1
----> 1 model.fit(train_x, train_y, epochs=10, batch_size=32)

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

File ~\anaconda3\envs\tf\lib\site-packages\keras\backend\tensorflow\trainer.py:69, in TensorFlowTrainer.train_step(self, data)
     67 if self.trainable_weights:
     68     trainable_weights = self.trainable_weights
--> 69     gradients = tape.gradient(loss, trainable_weights)
     71     # Update weights
     72     self.optimizer.apply_gradients(zip(gradients, trainable_weights))

AttributeError: 'str' object has no attribute 'base_dtype'

版本信息 #

• TensorFlow版本：2.10.1
• Keras版本：3.0.2

复现简化代码 #

import tensorflow as tf
import keras
print("Tensorflow version:", tf.__version__)
print("Keras version:", keras.__version__)

train_x = tf.constant([[1, 2, 3],[4, 5, 6]])

model = keras.models.Sequential([
    keras.Input(shape=(train_x.shape[1],), name="Input"),
    keras.layers.Dense(512, activation="linear", name="Entry"),
    keras.layers.Dense(train_x.shape[1], activation="linear", name="Output"),
])

print(model.summary())

model.compile(
    loss=keras.losses.mean_squared_error,
    optimizer=keras.optimizers.Adam(learning_rate=0.0001),
    metrics=["accuracy"],
)
print("Model input shape:", model.input_shape)

model.fit(x=train_x, y=train_x, batch_size=2, epochs=10)

问题根源 #

这是TensorFlow 2.10.1与Keras 3.0.2版本不兼容导致的：

• Keras 3.x是独立于TensorFlow的全新版本，API和内部实现与旧版tf.keras差异极大；
• TensorFlow 2.10.x仅适配集成在其内部的tf.keras（对应Keras 2.x版本），混用独立Keras 3.x会引发底层张量类型匹配错误，最终触发该报错。

解决方案 #

方案1：降级Keras到兼容版本 #

卸载当前Keras 3.x，安装适配TensorFlow 2.10的Keras 2.x版本：

pip uninstall keras -y
pip install keras==2.10.0

之后代码改用tf.keras导入所有组件，避免混用独立Keras：

import tensorflow as tf

train_x = tf.constant([[1, 2, 3],[4, 5, 6]], dtype=tf.float32)
train_y = train_x  # 替换为真实数据

model = tf.keras.models.Sequential([
    tf.keras.Input(shape=(train_x.shape[1],), name="Input"),
    tf.keras.layers.Dense(512, activation="linear", name="Entry"),
    tf.keras.layers.Dense(train_y.shape[1], activation="linear", name="Output"),
])

model.compile(
    loss='mean_squared_error',
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),
    metrics=[tf.keras.metrics.MeanAbsoluteError()],  # 回归任务替换accuracy为回归指标
)

model.fit(train_x, train_y, epochs=10, batch_size=32)

方案2：升级TensorFlow到支持Keras 3.x的版本 #

TensorFlow 2.15及以上版本原生支持Keras 3.x，直接升级即可：

pip install tensorflow>=2.15.0

升级后无需修改原有代码，可继续使用独立keras导入。

额外优化建议 #

• 音频序列预测属于回归任务，使用accuracy作为指标完全不合理，建议替换为mean_absolute_error或mean_squared_error；
• 确保输入数据类型为float32（与模型权重默认类型一致），避免不必要的类型转换开销。

内容的提问来源于stack exchange，提问作者Profs