Using a variable for num_splits for tf.split()

Question

Is it possible to use a placeholder input for the num_split parameter to tf.split()?

I would ideally like to do something like:

num_splits = tf.place         


        

Answer 1

A solution to a similar problem in an unrelated circumstance is using a placeholder for a fixed length of maximum size. Suppose I have a sequence of length 40. t = tf.range(40). Now at run time, I get a split, (say x = [6,9,10,5,1]). Now follow the steps

Step 1: Determine the maximum number of splits there can be, say: 19

Step 2:

    num_splits = tf.placeholder(tf.int32, [19]) 
    y= tf.split(t, num_or_size_splits=num_splits, axis=0)

This will break the sequence into run-time determined split sizes

Step 4: At run time :

    x = [6,9,10,5,1] 
    x += [40-sum(x)] + [0 for i in range(19-1-len(x))]

First line means the actual split sizes we need
Split requires the split sizes should sum upto total split size, i.e. 40 in this case, and 0 are the split sizes for the left over splits.

session.run(y, feed_dict={num_splits:x}) will show the results like :

[0, 1, 2, 3, 4, 5]  
[ 6,  7,  8,  9, 10, 11, 12, 13, 14]  
[15, 16, 17, 18, 19, 20, 21, 22, 23, 24]  
[25, 26, 27, 28, 29]  
[30]  
[31, 32, 33, 34, 35, 36, 37, 38, 39]  
[]
.
.
[]  

Step 5: (Optional, preferred) pad with zeros till the max length of the sequence

    def pad_up_to(t, max_in_dims, constant_values):
        s = tf.shape(t)
        paddings = [[0, m-s[i]] for (i,m) in enumerate(max_in_dims)]
        return tf.pad(t, paddings, 'CONSTANT', constant_values=constant_values)


    m = []
    for t1 in y :
      t1=tf.reshape(t1,[1,-1])
      t_padded = pad_up_to(t1, [1,15], 0)
      session.run(t_padded  , feed_dict={num_splits:x})  
      m+=  [t_padded]
    m= tf.concat(m,0)

This will pad the chunks with zeros to create equal sized chunks.

NOTE: The above methodology helped me in converting sequences into sentences (variable number of sentences) for NLP related tasks

funciton : pad_up_to() is taken from Q:42334646