初始化RNNCell

Question

我在使用任何张量流RNNCell的子类时遇到问题。根据tensorflow的来源，任何RNNCell的状态都应该是一个元组，但是当我给它一个元组时，它抛出一个错误，表明它正在尝试询问状态元组的ndims属性，该属性不存在。 / p>

我正在尝试创建一个LSTM，以便可以控制每个单独的输入。

这是我可以做的最简单的代码，即使有同样的问题，所以我希望我做的事情很容易解决。这是简单的代码：

    lstm_layer = tf.contrib.rnn.LSTMCell(num_units = 64)
    initial_state = lstm_layer.zero_state(batch_size=1,dtype=tf.float32)
    initial_input = np.expand_dims(np.array([1,2,3,4,5,6,7,8]),0)
    output_single, state_single = lstm_layer(inputs=initial_input,state=initial_state)

这是我得到的错误：

AttributeError                            Traceback (most recent call last)
<ipython-input-22-1dcce10906e5> in <module>
      2 initial_state = lstm_layer.zero_state(batch_size=1,dtype=tf.float32)
      3 initial_input = np.expand_dims(np.array([1,2,3,4,5,6,7,8]),0)
----> 4 output_single, state_single = lstm_layer(inputs=initial_input,state=initial_state)

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/ops/rnn_cell_impl.py in __call__(self, inputs, state, scope, *args, **kwargs)
    369     # method.  See the class docstring for more details.
    370     return base_layer.Layer.__call__(self, inputs, state, scope=scope,
--> 371                                      *args, **kwargs)
    372 
    373 

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/layers/base.py in __call__(self, inputs, *args, **kwargs)
    528 
    529       # Actually call layer
--> 530       outputs = super(Layer, self).__call__(inputs, *args, **kwargs)
    531 
    532     if not context.executing_eagerly():

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, inputs, *args, **kwargs)
    536       if not self.built:
    537         # Build layer if applicable (if the `build` method has been overridden).
--> 538         self._maybe_build(inputs)
    539         # We must set self.built since user defined build functions are not
    540         # constrained to set self.built.

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py in _maybe_build(self, inputs)
   1589     # Check input assumptions set before layer building, e.g. input rank.
   1590     input_spec.assert_input_compatibility(
-> 1591         self.input_spec, inputs, self.name)
   1592     input_list = nest.flatten(inputs)
   1593     if input_list and self._dtype is None:

~/anaconda3/lib/python3.7/site-packages/tensorflow/python/keras/engine/input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    107         spec.min_ndim is not None or
    108         spec.max_ndim is not None):
--> 109       if x.shape.ndims is None:
    110         raise ValueError('Input ' + str(input_index) + ' of layer ' +
    111                          layer_name + ' is incompatible with the layer: '

AttributeError: 'tuple' object has no attribute 'ndims'

看来keras API中的所有层都有一些基本方法被调用，但是不适用于元组。但是，我感到奇怪的是，这将是一个前所未有的问题。所以我希望只是我犯错了

Answer 1

我发现了问题。在这种情况下，Tensorflow与numpy的配合不好。代替

    initial_input = np.expand_dims(np.array([1,2,3,4,5,6,7,8]),0)

我需要给它

    initial_input = tf.expand_dims(np.array([1,2,3,4,5,6,7,8],dtype=np.float32),0)