Tensorflow@tf.function-无法在Tensorflow图函数内获取会话

我试图使用@tf.function指令和Keras函数API，在简单神经网络的训练步骤中创建一个tf图。我使用的是与Python3.7一起安装的TensorFlowV2.1.0。 然而，我得到了标题中的运行时错误，我希望任何提示都能理解其原因

代码如下

import tensorflow as tf
import numpy as np

# import the CIFAR10 dataset and normalise the feature distributions                                                                                                                             
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.cifar10.load_data()                                                                                         
train_images = train_images / np.max(train_images)
test_images  = test_images / np.max(train_images)

# convert the datasets to tf.data, batching the data                                                                                                                    
train_data = tf.data.Dataset.from_tensor_slices((train_images, train_labels)).batch(128)
test_data  = tf.data.Dataset.from_tensor_slices((test_images,  test_labels)).batch(128)

# make a model with a single dense layer
# note that the flatten layer is needed to convert the                                                                                                         
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(units = 10, activation = "relu"))

# compile the model                                                                                                                                  
model.compile(optimizer = tf.keras.optimizers.Adam(learning_rate = 0.001),
              loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits = True),
              metrics = ["accuracy"])

# training step
@tf.function
def train(model, train_data, test_data):
    model.fit(x = train_data,
              validation_data = test_data,
              epochs = 10)

    return

# train the model                                                                                                                                    
train(model = model, train_data = train_data, test_data = test_data)

我在运行时得到的错误如下

2020-04-01 11:33:27.084545: W tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 1228800000 exceeds 10% of system memory.
Traceback (most recent call last):
  File "report.py", line 41, in <module>
    train(model = model, train_data = train_data, test_data = test_data)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 568, in __call__
    result = self._call(*args, **kwds)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 615, in _call
    self._initialize(args, kwds, add_initializers_to=initializers)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 497, in _initialize
    *args, **kwds))
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 2389, in _get_concrete_function_internal_garbage_collected
    graph_function, _, _ = self._maybe_define_function(args, kwargs)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 2703, in _maybe_define_function
    graph_function = self._create_graph_function(args, kwargs)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/function.py", line 2593, in _create_graph_function
    capture_by_value=self._capture_by_value),
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/framework/func_graph.py", line 978, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/eager/def_function.py", line 439, in wrapped_fn
    return weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/framework/func_graph.py", line 968, in wrapper
    raise e.ag_error_metadata.to_exception(e)
RuntimeError: in converted code:

    report.py:34 train  *
        model.fit(x = train_data,
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:819 fit
        use_multiprocessing=use_multiprocessing)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_arrays.py:648 fit
        shuffle=shuffle)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:2346 _standardize_user_data
        all_inputs, y_input, dict_inputs = self._build_model_with_inputs(x, y)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py:2523 _build_model_with_inputs
        inputs, targets, _ = training_utils.extract_tensors_from_dataset(inputs)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_utils.py:1677 extract_tensors_from_dataset
        iterator = get_iterator(dataset)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_utils.py:1658 get_iterator
        initialize_iterator(iterator)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_utils.py:1665 initialize_iterator
        K.get_session((init_op,)).run(init_op)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/backend.py:493 get_session
        session = _get_session(op_input_list)
    /home/alessio/.local/lib/python3.7/site-packages/tensorflow_core/python/keras/backend.py:453 _get_session
        raise RuntimeError('Cannot get session inside Tensorflow graph function.')

    RuntimeError: Cannot get session inside Tensorflow graph function.

1228800000的分配超过了系统内存的10%。 回溯（最近一次呼叫最后一次）： 文件“report.py”，第41行，在 列车（模型=模型，列车数据=列车数据，测试数据=测试数据） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/eager/def_function.py”，第568行，在调用中__ 结果=自身调用（*args，**kwds） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/eager/def_function.py”，第615行，在调用中 self.\u初始化（参数、KWD、添加初始值设定项到=初始值设定项） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/eager/def_function.py”，第497行，在 *args，**科威特第纳尔） 文件“/home/alesio/.local/lib/python3.7/site-packages/tensorflow\u-core/python/eager/function.py”，第2389行，位于“获取”具体“函数”内部“垃圾”收集中 图函数，自我，可能定义函数（args，kwargs） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/eager/function.py”，第2703行，在函数定义中 graph\u function=self.\u create\u graph\u function（args，kwargs） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/eager/function.py”，第2593行，在创建图形函数中 按值捕获=自身。_按值捕获）， 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/framework/func_graph.py”，第978行，在func_graph_中，从_py_func func_outputs=python_func（*func_args，**func_kwargs） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/eager/def_function.py”，第439行，包装为 返回弱_-wrapped_-fn（） 文件“/home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/framework/func_graph.py”，第968行，在包装器中 将e.ag\u错误\u元数据引发到\u异常（e） 运行时错误：在转换的代码中： 报告：py:34列车* 模型拟合（x=列车数据， /home/alesio/.local/lib/python3.7/site packages/tensorflow_core/python/keras/engine/training.py:819-fit 使用多处理=使用多处理） /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training\u arrays.py:648 fit 洗牌 /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training.py:2346\u标准化\u用户\u数据 所有输入，y输入，dict输入=自我。用输入（x，y）建立模型 /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training.py:2523\u build\u model\u with\u input 输入，目标，从数据集中提取张量（输入） /home/alesio/.local/lib/python3.7/site-packages/tensorflow\u-core/python/keras/engine/training\u-utils.py:1677从数据集中提取张量 迭代器=获取迭代器（数据集） /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training\u utils.py:1658 get\u迭代器 初始化迭代器（迭代器） /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training\u utils.py:1665 initialize\u迭代器 K.get_会话（（init_op，）。运行（init_op） /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/backend.py:493 get\u会话 会话=\获取\会话（操作\输入\列表） /home/alesio/.local/lib/python3.7/site packages/tensorflow\u core/python/keras/backend.py:453\u get\u session raise RUNTIMERROR（'无法在Tensorflow图函数内获取会话'） RuntimeError:无法在Tensorflow图函数内获取会话。 请注意，在没有@tf.function指令的情况下，与前面相同的代码可以正常运行。 另一方面，我在不同的数据集和不同的模型上得到相同的错误

---

提前感谢。

查看文档，我不清楚您定义的函数是否可以编译成图形。我认为它应该用于Lambda层或类似层中使用的函数

您已经在模型上调用了compile，它将把模型转换成一个图，无需更多操作

我猜它是抛出的，因为它不知道如何从

model.fit

调用构建图形，但是错误消息非常混乱

如果你尝试一个简单的算术函数，比如

@tf.function
def add(x, y):
    return x + y

add(1, 2)

现在输出一个张量：

<tf.Tensor: shape=(), dtype=int32, numpy=3>

TensorFlow速度很快。我不会担心性能，除非您真正了解库中发生了什么，并且知道存在问题