Tensorflow, Reshape like a convolution - tensorflow

Question

I have a matrix [3,3,256], my final output must be [4,2,2,256], I have to use a reshape like a 'convolution' without changing the values. (In this case using a filter 2x2). Is there a method to do this using tensorflow?

Answer 1

If I understand your question correctly, you want to store the original values redundantly in the new structure, like this (without the last dim of 256):
[ [ 1 2 3 ] [ [ 1 2 ] [ [ 2 3 ] [ [ 4 5 ] [ [ 5 6 ]
[ 4 5 6 ] => [ 4 5 ] ], [ 5 6 ] ], [ 7 8 ] ], [ 8 9 ] ]
[ 7 8 9 ] ]
If yes, you can use indexing, like this, with x being the original tensor, and then stack them:
x2 = []
for i in xrange( 2 ):
for j in xrange( 2 ):
x2.append( x[ i : i + 2, j : j + 2, : ] )
y = tf.stack( x2, axis = 0 )
Based on your comment, if you really want to avoid using any loops, you might utilize the tf.extract_image_patches, like below (tested code) but you should run some tests because this might actually be much worse than the above in terms of efficiency and perfomance:
import tensorflow as tf
sess = tf.Session()
x = tf.constant( [ [ [ 1, -1 ], [ 2, -2 ], [ 3, -3 ] ],
[ [ 4, -4 ], [ 5, -5 ], [ 6, -6 ] ],
[ [ 7, -7 ], [ 8, -8 ], [ 9, -9 ] ] ] )
xT = tf.transpose( x, perm = [ 2, 0, 1 ] ) # have to put channel dim as batch for tf.extract_image_patches
xTE = tf.expand_dims( xT, axis = -1 ) # extend dims to have fake channel dim
xP = tf.extract_image_patches( xTE, ksizes = [ 1, 2, 2, 1 ],
strides = [ 1, 1, 1, 1 ], rates = [ 1, 1, 1, 1 ], padding = "VALID" )
y = tf.transpose( xP, perm = [ 3, 1, 2, 0 ] ) # move dims back to original and new dim up front
print( sess.run(y) )
Output (horizontal separator lines added manually for readability):
[[[[ 1 -1]
[ 2 -2]]
[[ 4 -4]
[ 5 -5]]]
[[[ 2 -2]
[ 3 -3]]
[[ 5 -5]
[ 6 -6]]]
[[[ 4 -4]
[ 5 -5]]
[[ 7 -7]
[ 8 -8]]]
[[[ 5 -5]
[ 6 -6]]
[[ 8 -8]
[ 9 -9]]]]

Answer 2

I have a similar problem with you and I found that in tf.contrib.kfac.utils there is a function called extract_convolution_patches. Suppose you have a tensor X with shape (1, 3, 3, 256) where the initial 1 marks batch size, you can call
Y = tf.contrib.kfac.utils.extract_convolution_patches(X, (2, 2, 256, 1), padding='VALID')
Y.shape # (1, 2, 2, 2, 2, 256)
The first two 2's will be the number of your output filters (makes up the 4 in your description). The latter two 2's will be the shape of the filters. You can then call
Y = tf.reshape(Y, [4,2,2,256])
to get your final result.