Traversing theano vector slice by slice

Question

Suppose I have a vector x = [0.1, 0.2, 0.1, 0.1, 0.2, 0.4] with length 30. This vector is wrapped by a theano Tensor.

I want to create a new vector, with same size as x, but setting each 3 size slices elements to the sum of them and then do the same for the next 3 elements and so forth.

In the example, the first 3-length slice sum to 0.4, and the second 0.7 so r = [0.4, 0.4, 0.4, 0.7, 0.7, 0.7].

How can I do that? is using scan the only way?

Answer 1

You can do the following in numpy:

import numpy as np
x = np.array([0.1, 0.2, 0.1, 0.1, 0.2, 0.4])
y = (x.reshape(-1, 3).mean(1)[:, np.newaxis] * np.ones(3)).ravel()

print(y)

In theano, you can proceed in a very similar way:

import theano
import theano.tensor as T

xx = T.fvector()
yy = (xx.reshape((-1, 3)).mean(axis=1).reshape((-1, 1)) * T.ones(3)).ravel()

print(yy.eval({xx: x.astype('float32')}))