Theano Scan Updates

scan_updates.py

import theano
from theano.tensor.shared_randomstreams import RandomStreams

r = RandomStreams(1)

y, u = theano.scan(lambda : r.binomial(size=(2, 2)), n_steps=3)
f = theano.function([], [y])

# It still gives me three different sequence of random matrices!
print f()
print f()
print f()

Here's an example which shows that passing updates gives you the correct answer, while not passing it gives you the wrong one:

import theano
from theano.tensor.shared_randomstreams import RandomStreams

r = RandomStreams(1)

y, u = theano.scan(lambda : r.normal(), n_steps=3)
f = theano.function([], y)

# This is what you get when don't pass updates
f()
print f()

r = RandomStreams(1)

y, u = theano.scan(lambda : r.normal(), n_steps=3)
f = theano.function([], y, updates=u)

# And this is what you get when you do
f()
print f()

# This is what it should be
r = RandomStreams(1)
f = theano.function([], r.normal(size=(6,))[3:])
print f()

[ 0.46301228 -1.47958577 -1.48030508]
[-1.48030508 -0.47802526  0.38352579]
[-1.48030508 -0.47802526  0.38352579]

Sorry, have not seen this response before. Looks like it actually gives us "better" random :)

We can attach updates to the tag of the Apply node and find it then when exploring the computation graph.

And by the way: in a more complex situation I could not even call a compiled without updates function...