如何将CNN输出作为RNN中的输入序列进行馈送（使用Tensorflow）

Question

I have implemented a simple CNN using tensorflow. Up until now I used to put a fully connected layer of size 2 at the last convolutional layer to distinct between my 2 classes. Now I want to feed that last convolutional layer to an RNN and then perform the classification.

The output of the last convolutional layer is called "cnn_output" and is a tensor of that form
<tf.Tensor 'Sigmoid_1:0' shape=(?, 168, 32) dtype=float32>

before feeding it to the rnn I split it using this command:
input_series = tf.split(axis=2, num_or_size_splits= 32, value = cnn_output)

the result of the split is 32 tensor of that type:
<tf.Tensor 'split:0' shape=(?, 168, 1) dtype=float32>

the variables for the rnn are defined as below:
cell = tf.contrib.rnn.BasesLSTMCell(32, state_is_tuple=True)
cell_state = tf.placeholder(tf.float32, [168,32])
hidden_state = tf.placeholder(tf.float32, [168,32])
init_state = tf.contrib.rnn.LSTMStateTuple( cell_state, hidden_state)

to feed the output of the cnn to the rnn I use this command
states_series, current_state = tf.nn.static_rnn(cell, input_series, init_state)

At this line I get the following error

ValueError: linear is expecting 2D arguments: [TensorShape([Dimension(None), Dimension(168), Dimension(1)]), TensorShape([Dimension(168), Dimension(32)])]
To me it seems that there is something wrong with that first unknown dimension of the tensors but I am not sure how to deal with it.
I tried change the shape of the cell_state and hidden_state to [None,168,32] but didn't seem to help.
I also tried to use the dynamic_rnn (not sure about the difference with static_rnn) but didn't work either

EDIT UPDATE: I manage to run the code without error by reshaping the input_series of the rnn with this command:

input_series = [tf.reshape(ipt,[-1,168 ])for ipt in input_series]

which resulted in tensors like this

<tf.Tensor 'Reshape_21:0' shape=(?, 168) dtype=float32>

Although I am not sure yet if the rnn unit works properly. I will check it out and update

Answer 1

I think your error comes from the fact that your placeholders are 2D and not 3D. You forgot to add a batch dimension.

Try to define them like this:

cell_state = tf.placeholder(tf.float32, [None,168,32])
hidden_state = tf.placeholder(tf.float32, [None,168,32])

(You might have to change the 32 to 1, given that your inputs have 1 channel each)