AttributeError: Tensor.op 在 Tensorflow federated 中使用多个特征输入时启用急切执行时毫无意义答案

【问题标题】：AttributeError: Tensor.op is meaningless when eager execution is enabled when using multiple feature inputs in Tensorflow federatedAttributeError: Tensor.op 在 Tensorflow federated 中使用多个特征输入时启用急切执行时毫无意义
【发布时间】：2020-11-04 11:07:27
【问题描述】：

输入多个特征输入时出现如下问题：

feature_layer = tf.keras.layers.DenseFeatures(feature_cols)

feature_layer_inputs = {}
feature_layer_inputs['a'] = tf.keras.Input(shape=(1,), name='a', dtype=tf.int32)
feature_layer_inputs['b'] = tf.keras.Input(shape=(1,), name='b', dtype=tf.int32)

model = feature_layer(feature_layer_inputs)

for units in [64, 64]:
    model = tf.keras.layers.Dense(units, activation='relu')(model)
c_pred = tf.keras.layers.Dense(3, activation='softmax')(model) #classification 3 outputs

keras_model = tf.keras.Model(inputs=[v for v in feature_layer_inputs.values()], outputs=c_pred)

input_spec = collections.OrderedDict(
        x=collections.OrderedDict(
            a=tf.TensorSpec(shape=[None,], dtype=tf.int32),
            b=tf.TensorSpec(shape=[None,], dtype=tf.int32),
        y=tf.TensorSpec(shape=[None,], dtype=tf.int32))

def model_fn():
  # We _must_ create a new model here, and _not_ capture it from an external
  # scope. TFF will call this within different graph contexts.

  return tff.learning.from_keras_model(
      keras_model,
      #input_spec=preprocessed_sample_dataset.element_spec,
      input_spec=input_spec,
      loss=losses.SparseCategoricalCrossentropy(),
      metrics=[metrics.SparseCategoricalAccuracy()])


iterative_process = tff.learning.build_federated_averaging_process(
    model_fn,
    client_optimizer_fn=lambda: optimizers.Adam(learning_rate=client_lr),
    server_optimizer_fn=lambda: optimizers.SGD(learning_rate=server_lr))

state = iterative_process.initialize()

for round_num in range(1, NUM_ROUNDS+1):
    #state, tff_metrics = iterative_process.next(state, federated_train_data)
    state, tff_metrics = iterative_process.next(state, train_data)
    eval_model = keras_model
    eval_model.compile(optimizer=optimizers.Adam(learning_rate=client_lr),
                       loss=losses.SparseCategoricalCrossentropy(),
                       metrics=[metrics.SparseCategoricalAccuracy()])

    tff.learning.assign_weights_to_keras_model(eval_model, state.model)

    ev_result = eval_model.evaluate(x_test, y_test, verbose=0)

我从这个solution 得到这个代码。但是，我得到了完整的回溯如下。最终，我收到一个错误“AttributeError：Tensor.op 在急切执行时毫无意义”。构建的模型似乎有问题，尤其是 tf.keras.model 函数中的输入。

Traceback (most recent call last):
  File "accident_modify_uk_final3b.py", line 323, in <module>
    server_optimizer_fn=lambda: optimizers.SGD(learning_rate=server_lr))
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/federated_averaging.py", line 212, in build_federated_averaging_process
    stateful_delta_aggregate_fn, stateful_model_broadcast_fn)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/framework/optimizer_utils.py", line 360, in build_model_delta_optimizer_process
    @tff.tf_computation
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/api/computations.py", line 152, in tf_computation
    return computation_wrapper_instances.tensorflow_wrapper(*args)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/impl/wrappers/computation_wrapper.py", line 333, in __call__
    self._wrapper_fn)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/impl/wrappers/computation_wrapper.py", line 91, in _wrap
    concrete_fn = wrapper_fn(fn, parameter_type, unpack=None)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/impl/wrappers/computation_wrapper_instances.py", line 52, in _tf_wrapper_fn
    target_fn, parameter_type, ctx_stack)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/impl/tensorflow_serialization.py", line 275, in serialize_py_fn_as_tf_computation
    result = target(*args)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/core/impl/utils/function_utils.py", line 455, in <lambda>
    return lambda: fn()  # pylint: disable=unnecessary-lambda
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/framework/optimizer_utils.py", line 364, in tf_init_fn
    stateful_model_broadcast_fn.initialize())
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/framework/optimizer_utils.py", line 227, in server_init
    _, optimizer_vars = _build_server_optimizer(model, optimizer)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/framework/optimizer_utils.py", line 123, in _build_server_optimizer
    apply_delta(delta=weights_delta)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py", line 580, in __call__
    result = self._call(*args, **kwds)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py", line 627, in _call
    self._initialize(args, kwds, add_initializers_to=initializers)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py", line 506, in _initialize
    *args, **kwds))
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 2446, in _get_concrete_function_internal_garbage_collected
    graph_function, _, _ = self._maybe_define_function(args, kwargs)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 2777, in _maybe_define_function
    graph_function = self._create_graph_function(args, kwargs)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 2667, in _create_graph_function
    capture_by_value=self._capture_by_value),
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py", line 981, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py", line 441, in wrapped_fn
    return weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py", line 968, in wrapper
    raise e.ag_error_metadata.to_exception(e)
AttributeError: in user code:

    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow_federated/python/learning/framework/optimizer_utils.py:112 apply_delta  *
        optimizer.apply_gradients(grads_and_vars, name='server_update')
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:508 apply_gradients  **
        "name": name,
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/distribute/distribute_lib.py:2420 merge_call
        return self._merge_call(merge_fn, args, kwargs)
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/distribute/distribute_lib.py:2427 _merge_call
        return merge_fn(self._strategy, *args, **kwargs)
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:590 _distributed_apply  **
        "update_" + var.op.name, skip_on_eager=True):
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/ops/resource_variable_ops.py:581 op
        return self._handle.op
    /home/anaconda3/envs/env1_TF2.1/lib/python3.7/site-packages/tensorflow/python/framework/ops.py:1113 op
        "Tensor.op is meaningless when eager execution is enabled.")

    AttributeError: Tensor.op is meaningless when eager execution is enabled.

如何解决这个问题？

【问题讨论】：

请发布完整的回溯。
回溯指向代码中未提及的部分。请贴出相关代码。
@ewong 完成！ ...
如果 Keras 模型直接拟合数据（不使用 TFF），是否会出现错误？例如调用model_fn().compile(...).fit(train_data[0])。另请注意 model_fn() 中的代码注释被忽略，注释指出必须创建模型，但 keras_model 是从全局范围捕获的。
@ZacharyGarrett，感谢您提出代码注释，当使用 def model_fn() 中的另一个变量调用 def keras_model_fn() 时问题消失了

标签： python tensorflow keras tensorflow-federated

【解决方案1】：

我认为问题在于model_fn() 是从全局范围捕获keras_model，而不是在调用它时构造一个新实例。

来自tff.learning.build_federated_averaging_process 的文档：

model_fn：返回 tff.learning.Model 的无参数函数。此方法不得捕获 TensorFlow 张量或变量并使用它们。每次调用都必须完全从头构建模型，每次调用都返回相同的预构建模型会导致错误。

上面的代码中有一段代码注释暗示了这一点（复制到这里）：

def model_fn():
  # We _must_ create a new model here, and _not_ capture it from an external
  # scope. TFF will call this within different graph contexts.

  return tff.learning.from_keras_model(
      keras_model,
      #input_spec=preprocessed_sample_dataset.element_spec,
      input_spec=input_spec,
      loss=losses.SparseCategoricalCrossentropy(),
      metrics=[metrics.SparseCategoricalAccuracy()])

要解决这个问题，请尝试调用tf.keras.Model 构造函数inside model_fn()。很可能模型的层也需要在内部创建，因此拥有一个返回 new tf.keras.Model 的 create_keras_model() 函数并从 @987654330 中调用该助手可能会很有用@。

【讨论】：