Efficient conversion

Question

I have the following code

output = T.switch(cond, a, b) 

where cond is t (N,1) bool Tensor, and a and b are (N, M) numerical tensors with M quite large. The condition works differently.

I can easily get the switch to work by running T.repeat() on cond , but it's pretty slow. Is there a way that I can effectively make bools in cond decide whether to return a or b ?

Answer 1

Is there a way that I can effectively make bools in cond, decide whether to return a or b?

Yes you could do

 cond * a + (1-cond) * b 

cond will be passed in the form (N, M) .

This should be close to the theoretical limit, which is the memory bandwidth: this operation should read N*M elements and write N*M

Instead, we read 2*N*M , but remove the conditional logic.

(I don’t have Anano in front of me, so I’m not sure that it is faster than T.switch , but it should be about as good as him. Also, I would try casting cond the same dtype as a and b )

---

If you want to update a in place, you can do this using T.set_subtensor :

 a = np.random.uniform(size=(N, M)).astype(np.float32) b = np.random.uniform(size=(N, M)).astype(np.float32) a = theano.shared(a) b = theano.shared(b) c = T.vector() # mostly 0, presumably (1-cond) nz = T.nonzero(c) s = T.set_subtensor(a[nz], b[nz]) fn = theano.function([c], [], updates=[(a, s)]) ... fn(1-cond) 

It may or may not be faster than the first approach, depending on N , M and other factors.