Cannot have an odd latent_size (working with 2, 4, etc. , but not 3 or 5), shape problem about normalizing-flows HOT 2 CLOSED

Hi @noetits,

It is possible to use real NVP with an odd number of dimensions. In this case, the variable is split into two parts of size (latent_size // 2 +1 , latent_size // 2). Since the first part goes into the parameter map, which computes the parameters for the second part, we need to change the following:

param_map = nf.nets.MLP([(latent_size + 1) // 2, 64, 64, latent_size // 2 * 2], init_zeros=True)

This should work for both even and odd dimensions.

Note that this depends on the kind of Split you use. Also, if you do not use a scale, as in NICE, you need to drop the factor of 2 in the output units, since there will be only one parameter per dimension, i.e. the shift.

Best regards,
Vincent

from normalizing-flows.

I just realized this might be a theoretical constraint

https://en.wikipedia.org/wiki/Flow-based_generative_model

Maybe it would be be good to have an error message disclosing this?

Now I am wondering if there would be a workaround on this, like adding a dummy variable that it e.g. a uniform distribution (or other if better), or if this would not be a good idea for some reason

from normalizing-flows.