Reshape tensor of 3 rank to 2 rank in tensorflow to use it in fully connected layer

Question

I did a lot of search and try but I have not found any solution for the next problem:

I am working on neural network models to test classification of sentences. These sentences are represented in [rows, words_encoded_by_word2vec] format. The first network - fully connected is done. In the second model I am trying to add a conv1d and a max_pool1d layers just before the dense layer. These layers expect a tensor in [batch_size, length, channel] format. Ok. This is not a problem and it is done. /I have only 1 channel/

However the connection between fully connected and max pool layers is extremely difficult to set up because of the unknown batch_size .

self.X = tf.placeholder(tf.float32, shape=(None, self.n_inputs, 1), name="input")

self.convolution = tf.nn.conv1d(self.X, self.filter, stride=50, adding="SAME")
self.max_pool = tf.layers.max_pooling1d(self.convolution, pool_size=2, strides=1, padding="SAME")
self.tensor_vector = tf.reshape(tensor=self.max_pool, shape=(-1, tf.shape(self.max_pool)[1]*tf.shape(self.max_pool)[2]))

This works but the dense layer does not accept it, and gives an error:

ValueError: The last dimension of the inputs to `Dense` should be defined. Found `None`.

The specific dense layer:

hiddens.append(dropout(fully_connected(self.tensor_vector, layers[0], scope="hidden0",weights_initializer=tf.contrib.layers.xavier_initializer(),biases_initializer=tf.random_uniform_initializer(-0.1,0.1)), self.keep_prob, is_training=self.is_training))  

Are there any possibilities to reconcile the two format? Any helps will be appreciated :)

Tanks.

Answer 1

I found an answer that can be applied for my specific problem. tf.squeeze() can remove the 1 sized dimension (deprecated dimension). This is exactly the channel that I want to remove. This transformation did the trick for me.

However I think these shape problems are prevalent in tensorflow development so it would be nice if there were adapters for these problems.