Variable size Convolutional Neural Network Input and Fixed output

Question

I am trying to make a CNN model that takes variable size input (sentence matrix) and produce a fixed size output for a subsequent fully connected layer (similar to this paper).

I am trying to implement a dynamic kernel size for a max pooling layer so I need the shape of the input at runtime to achieve this.

input = tf.placeholder(tf.float32)
# convolution layer here .... 

tf.nn.max_pool(convolution_output, ksize=[1, s, 1, 1],
                      strides=[1, 1, 1, 1], padding='VALID')

s in ksize=[1, s, 1, 1] should be inferred from the input shape.

However, I can't find a way to do it with Tensorflow.

Anyone knows a way to do it?

This sounds similar to what they do here: (arxiv.org/abs/1412.6071). But unfortunately this is not supported directly by tensorflow yet (github.com/tensorflow/tensorflow/issues/2953). Once added, what you want to do could be just an extension. — Lemm Ras
– Lemm Ras, Commented Jul 28, 2016 at 13:27
Maybe tf.reduce_max could work in this case? Not sure if this will for the paper you are implementing, but at least it works for simple CNN with variable input size. — Konstantin Lopuhin
– Konstantin Lopuhin, Commented Oct 12, 2016 at 13:23
tf.reduce_max made the job for what I wanted to do. I wanted to pool depending of the size of the batch size. (Not for every sentences though) Thanks! — rAyyy
– rAyyy, Commented Jun 8, 2017 at 2:35

AhlyM · Accepted Answer · 2017-10-18 20:59:54Z

4

I know it's an old thread, but for people who are looking for a solution. It has been implemented in tensorflow 1.4.0

tf.nn.max_pool() now takes 1d tensor as an input as opposed to a list of ints in the older versions. So you can use a placeholder as the argument of ksize.

answered Oct 18, 2017 at 20:59

AhlyM

5201 gold badge6 silver badges16 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

yazfield · Accepted Answer · 2016-08-19 04:46:09Z

2

I didn't find a way to do it in Tensorflow and had to reimplement my model in theano and use a different solution: Dynamic KMax Pooling.

answered Aug 19, 2016 at 4:46

yazfield

1,28312 silver badges19 bronze badges

Collectives™ on Stack Overflow

Variable size Convolutional Neural Network Input and Fixed output

2 Answers 2

Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related