Graph执行报错：无法将符号tf.Tensor作为Python bool使用，求解决

问题重现 #

模型代码 #

learning_rate = 0.1
momentum = 0.9

lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate=learning_rate,
    decay_steps=EPOCH_SIZE,
    decay_rate=0.1
)
optimizer = tf.keras.optimizers.SGD(learning_rate=lr_schedule, momentum=momentum, nesterov=True)

input_correct_responses = tf.keras.Input(shape=(26,), name='input_correct_responses')

pe = tf.keras.losses.categorical_crossentropy(input_correct_responses, z)
ce = tf.keras.metrics.CategoricalAccuracy()(input_correct_responses, z)

model = Model(inputs=[input_obscured_word_seen, input_letters_guessed_previously, input_correct_responses], outputs=z)
model.compile(optimizer=optimizer, loss=pe, metrics=[ce])

progress_printer = tf.keras.callbacks.LambdaCallback(
    on_epoch_end=lambda epoch, logs: print(f"Training - epoch: {epoch + 1}, loss: {logs['loss']}, accuracy: {logs['categorical_accuracy']}"),
    on_train_end=lambda logs: print("Training completed.")
)

错误信息 #

---------------------------------------------------------------------------
OperatorNotAllowedInGraphError            Traceback (most recent call last)
Cell In[83], line 17
     14 ce = tf.keras.metrics.CategoricalAccuracy()(input_correct_responses, z)
     16 model = Model(inputs=[input_obscured_word_seen, input_letters_guessed_previously, input_correct_responses], outputs=z)
---> 17 model.compile(optimizer=optimizer, loss=pe, metrics=[ce])
     19 progress_printer = tf.keras.callbacks.LambdaCallback(
     20     on_epoch_end=lambda epoch, logs: print(f"Training - epoch: {epoch + 1}, loss: {logs['loss']}, accuracy: {logs['categorical_accuracy']}"),
     21     on_train_end=lambda logs: print("Training completed.")
     22 )

File ~/anaconda3/lib/python3.10/site-packages/tensorflow/python/trackable/base.py:205, in no_automatic_dependency_tracking.._method_wrapper(self, *args, **kwargs)
    203 self._self_setattr_tracking = False  # pylint: disable=protected-access
    204 try:
--> 205   result = method(self, *args, **kwargs)
    206 finally:
    207   self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

File ~/anaconda3/lib/python3.10/site-packages/keras/engine/training_v1.py:406, in Model.compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, distribute, **kwargs)
    402 if isinstance(self.optimizer, tf.__internal__.tracking.Trackable):
    403     self._track_trackable(
    404         self.optimizer, name="optimizer", overwrite=True
    405     )
--> 406 self.loss = loss or {}
    407 self.loss_weights = loss_weights
...
--> 544   raise errors.OperatorNotAllowedInGraphError(
    545       f"{task} is not allowed in Graph execution. Use Eager execution or"
    546       " decorate this function with @tf.function.")

OperatorNotAllowedInGraphError: Using a symbolic `tf.Tensor` as a Python `bool` is not allowed in Graph execution. Use Eager execution or decorate this function with @tf.function.

问题根源 #

你提前调用了tf.keras.losses.categorical_crossentropy和tf.keras.metrics.CategoricalAccuracy()()，得到的是符号张量，但model.compile的loss和metrics参数需要的是损失/指标的可调用对象或名称，而非已经计算好的张量。当代码执行loss=pe时，TensorFlow在图模式下尝试将该张量当作Python布尔值处理（如loss or {}逻辑），直接触发了错误。

修复方案 #

修改model.compile的参数传递方式，直接传入损失函数和指标的实例或名称字符串，而非提前计算的张量：

修改后的完整代码 #

learning_rate = 0.1
momentum = 0.9

lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate=learning_rate,
    decay_steps=EPOCH_SIZE,
    decay_rate=0.1
)
optimizer = tf.keras.optimizers.SGD(learning_rate=lr_schedule, momentum=momentum, nesterov=True)

input_correct_responses = tf.keras.Input(shape=(26,), name='input_correct_responses')

# 移除提前计算的损失和指标张量
# pe = tf.keras.losses.categorical_crossentropy(input_correct_responses, z)
# ce = tf.keras.metrics.CategoricalAccuracy()(input_correct_responses, z)

model = tf.keras.Model(inputs=[input_obscured_word_seen, input_letters_guessed_previously, input_correct_responses], outputs=z)
# 直接传递损失函数实例和指标实例
model.compile(
    optimizer=optimizer,
    loss=tf.keras.losses.CategoricalCrossentropy(),  # 也可以用字符串'categorical_crossentropy'
    metrics=[tf.keras.metrics.CategoricalAccuracy()]  # 也可以用字符串'categorical_accuracy'
)

progress_printer = tf.keras.callbacks.LambdaCallback(
    on_epoch_end=lambda epoch, logs: print(f"Training - epoch: {epoch + 1}, loss: {logs['loss']}, accuracy: {logs['categorical_accuracy']}"),
    on_train_end=lambda logs: print("Training completed.")
)

关键说明 #

• loss参数：接受内置损失名称字符串、损失函数类实例，或自定义可调用函数，TensorFlow会在训练过程中自动计算每一步的损失值。
• metrics参数：同理，需要传入指标类实例或内置指标名称字符串，框架会自动处理指标的更新和结果统计。

内容的提问来源于stack exchange，提问作者Anwesh saha