Question

对于MNIST识别，我没有一个非常复杂的模型，需要研究其体系结构并试图可视化内部发生的事情。添加了用于关闭网络某些部分的变量（towerWeights…）后，尝试获取图层的渐变时我开始出现错误

Traceback (most recent call last):
  File "VisQtMain.py", line 1021, in onGradientsByImagesPressed
    dataList.append(self.netWrapper.getGradients(layerName, imageNum, 1, epochNum, True))
  File "E:\Projects\Python\Visualiz_Zeiler\MnistNetVisWrapper.py", line 241, in getGradients
    outputs=self.gradientTensors)
  File "C:\Program Files\Python35\lib\site-packages\keras\backend\tensorflow_backend.py", line 3009, in function
    **kwargs)
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\keras\backend.py", line 3758, in function
    return EagerExecutionFunction(inputs, outputs, updates=updates, name=name)
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\keras\backend.py", line 3655, in __init__
    base_graph=source_graph)
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\eager\lift_to_graph.py", line 249, in lift_to_graph
    visited_ops = set([x.op for x in sources])
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\eager\lift_to_graph.py", line 249, in <listcomp>
    visited_ops = set([x.op for x in sources])
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\ops\resource_variable_ops.py", line 559, in op
    return self._handle.op
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\framework\ops.py", line 1098, in op
    "Tensor.op is meaningless when eager execution is enabled.")
AttributeError: Tensor.op is meaningless when eager execution is enabled.

网络就像

inputs = Input(shape=(28, 28, 1))

conv_1 = Conv2D(16, 5, strides=(2, 2), activation='relu', name='conv_1_common')(inputs)

towerWeightsKerasVar = tf.compat.v1.Variable(np.ones([4]), dtype=tf.float32, name='tower_weights')
# towerWeightsKerasVar._trainable = False
**towerWeights = Input(shape=(4, ), tensor=towerWeightsKerasVar)**
…
dense_2 = Dropout(0.3)(conv _1)
dense_2 = Dense(10, name='dense_2')(dense_2)
prediction = Activation("softmax", name="softmax")(dense_2)

model = Model(inputs=[inputs, towerWeights], outputs=prediction)

，并且可以在此cut变体上复制。 TensorFlow 2.0 RC，2.0.0和今天晚上的tf –结果是相同的。 这是TF中的错误还是我做错了什么？

TowerWeights用于将网络的某些部分乘以其组件。但是在cut变体中，它只是存在而无所事事。

我不太了解正在发生的事情，渴望模式在起作用，在不起作用的情况下我尝试使用tf.compat.v1.disable_eager_execution（）禁用急切执行，但后来我得到

  File "VisQtMain.py", line 1021, in onGradientsByImagesPressed
    dataList.append(self.netWrapper.getGradients(layerName, imageNum, 1, epochNum, True))
  File "E:\Projects\Python\Visualiz_Zeiler\MnistNetVisWrapper.py", line 246, in getGradients
    gradients = self.gradientKerasFunc(inp)
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\keras\backend.py", line 3565, in __call__
    run_metadata=self.run_metadata)
  File "C:\Program Files\Python35\lib\site-packages\tensorflow_core\python\client\session.py", line 1470, in __call__
    run_metadata_ptr)
tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'softmax_target' with dtype float and shape [?,?]
     [[{{node softmax_target}}]]

所以在这种情况下也有问题...

Python 3.5，Windows 7 x64

Answer 1

我还没有找到解决方案，但是找到了解决方法，就像这样

def get_tensor_array_element(x):
    def f(X):                               # Here we have e.g. np.ones([4])
        return X[x : x + 1]

    def g(input_shape):
        output_shape = list(input_shape)    # Here - e.g. (None, 4)
        output_shape[-1] = 1
        return tuple(output_shape)

    return Lambda(f, output_shape=lambda input_shape: g(input_shape))


towerWeightsKerasVar = tf.compat.v1.Variable(np.ones([towerCount]),
             dtype=tf.float32, name='tower_weights')
towerWeightsKerasVar._trainable = False    # "Doesn't help against Tensor.op is meaningless when eager execution is enabled."
towerWeights = Input(shape=(towerCount, ), tensor=towerWeightsKerasVar)
    # Looks like input, should be declared in model's input, but (if the tensor parameter is present)
    # should not be supplied to the net (in model.fit and so on)
...
conv = Multiply()([conv, get_tensor_array_element(i)(towerWeights)])

TensorFlow AttributeError：启用急切执行时，Tensor.op没有意义

1 个答案: