[theano-users] Separate 2d convolution over different channels

[Peter O'Connor]

Hi. Did anyone figure this out? I really need this op. In my case I want
to compute the per-channel cross correlation of two (512, 20, 20) feature
maps.

Currently, I can't see any way to do this without using scan. My current
solution is:

class ChannelwiseCrossCorr(object):

def __init__(self, border_mode='full', subsample=(1, 1)):
self.border_mode = border_mode
self.subsample = subsample

def __call__(self, (x1, x2)):
"""
:param (x1, x2): are each (n_samples, n_channels, size_y, size_x) images
:return: A (n_samples, n_channels, size_y*2-1, size_x*2-1) image representing the channelwise cross-correlation between
each pair of images.
"""
from theano.tensor.signal.conv import conv2d as sconv2d
x1_flat = x1.reshape((x1.shape[0]*x1.shape[1], x2.shape[2], x2.shape[3]))
x2_flat = x2.reshape((x2.shape[0]*x2.shape[1], x2.shape[2], x2.shape[3]))[:, ::-1, ::-1]
map_flat, _ = theano.scan(partial(sconv2d, border_mode=self.border_mode, subsample=self.subsample), sequences=[x1_flat, x2_flat])
conv_maps = map_flat.reshape((x1.shape[0], x1.shape[1], map_flat.shape[1], map_flat.shape[2]))
return conv_maps


But it runs much slower than I expected, presumably due to the scan op.

[Pascal Lamblin]

I think the grouped convolutions, that have been merged in the
development version this summer, should do what you need.
The speed may not be optimal, but it should be faster than using scan.

Hi.  Did anyone figure this out?  I really need this op.  In my case I
want to compute the per-channel cross correlation of two (512, 20, 20)
feature maps.
Currently, I can't see any way to do this without using scan.  My
self.border_mode = border_mode
self.subsample = subsample
"""
:param(x1, x2): are each (n_samples, n_channels, size_y, size_x) images
:return: A (n_samples, n_channels, size_y*2-1, size_x*2-1) image
representing the channelwise cross-correlation between
each pair of images.
"""
from theano.tensor.signal.convimport conv2das sconv2d
x1_flat = x1.reshape((x1.shape[0]*x1.shape[1], x2.shape[2], x2.shape[3]))
x2_flat = x2.reshape((x2.shape[0]*x2.shape[1], x2.shape[2], x2.shape[3]))[:, ::-1, ::-1]
map_flat, _ = theano.scan(partial(sconv2d, border_mode=self.border_mode, subsample=self.subsample), sequences=[x1_flat, x2_flat])
conv_maps = map_flat.reshape((x1.shape[0], x1.shape[1], map_flat.shape[1], map_flat.shape[2]))
return conv_maps
But it runs much slower than I expected, presumably due to the scan op.
--
---
You received this message because you are subscribed to the Google
Groups "theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send
For more options, visit https://groups.google.com/d/optout.