[theano-users] Segnet and pooling indices.

[Kristoffer Wickstrøm]

Hi everyone,

I have been trying out different Fully Convolutional Networks recently, and
is currently working on Segnet. One of the key point is the way upsampling
is performed, namely that you keep the pooling indices from the encoder
part of the network. Fine in theory, but i am struggling with implementing
this in a good way in theano.

I am aware that pooling indices have been addressed previously here ( Max
Pooling Indices post), and i thought that worked. But the problem occurs
when you have an area where all the values are equal. Take a look at this
example:

import theano
import numpy as np
import theano.tensor as T

inp = T.tensor4()
from theano.tensor.signal.pool import pool_2d
f = theano.function([inp], [T.grad(T.sum(max_pool_2d(inp, ds=(2,2))),
wrt=inp)])

a = np.arange(16).reshape(1, 1, 4, 4)
b = np.ones(16).reshape(1,1,4,4)

a = array([[[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]]]])

b= array([[[[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.]]]])

f(a) = [array([[[[ 0., 0., 0., 0.],
[ 0., 1., 0., 1.],
[ 0., 0., 0., 0.],
[ 0., 1., 0., 1.]]]])]

f(b) = [array([[[[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.]]]])]


As you see it works as intended for a, but for b it just outputs ones
everywhere. I understand that when there is a tie it's not clear which
value to pick, but it would satosfactory if it just outputted one index
randomly form the area, like such:

b= array([[[[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.]]]])

g = some function that works

g(b) =
[array([[[[ 0., 0, 1., 0.],
[ 1., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 1., 0., 1.]]]])]


That way, i would at least get the correct number of indices outputted. Has
anyone done this or got any idea how to solve the problem ?

Thanks in advance, Kristoffer.