How to handle empty sequence with theano.scan()?

Question

The code snippet is like:

t = tensor.arange(1, K)
results, updates = theano.scan(fn=updatefunc, sequences=t, ...)

the scan process will iterate along t. However, when K<=1, t will be an empty range, then theano.scan() will crash. Is there any way to fix this?

Answer 1

You can use theano.ifelse.ifelse to compute the scan only when the sequence has some elements in it. For example:

import theano
import theano.tensor as tt
import theano.ifelse


def step(x_t, s_tm1):
    return s_tm1 + x_t


def compile():
    K = tt.lscalar()
    t = tt.arange(1, K)
    zero = tt.constant(0, dtype='int64')
    outputs, _ = theano.scan(step, sequences=[t], outputs_info=[zero])
    output = theano.ifelse.ifelse(tt.gt(K, 1), outputs[-1], zero)
    return theano.function([K], outputs=[output])


def main():
    f = compile()
    print f(3)
    print f(2)
    print f(1)
    print f(0)
    print f(-1)


main()

prints

[array(3L, dtype=int64)]
[array(1L, dtype=int64)]
[array(0L, dtype=int64)]
[array(0L, dtype=int64)]
[array(0L, dtype=int64)]