one-dimensional coupling flows do not work about normalizing-flows HOT 3 CLOSED

I reproduce your result. And your set is not of 1d but a scalar actually.
Increasing the size can resolve this, though I'm not sure if it's proper to use a scalar in this pipeline.

from normalizing-flows.

Hi @jackerschott,

Coupling flows are inherently designed to operate on at least two dimensions, since in a coupling layer the dimensions are split into two disjoint sets, i.e. one that is transformed and another that is kept constant but used for conditioning, see e.g. the Real NVP paper. Hence, you cannot apply a coupling flow to a one dimensional target distribution. Instead, you could use an autoregressive affine or spline flow, which will use constant parameters in the one dimensional case. Moreover, using a residual flow is possible, and this would be perhaps the most expressive flow for a one dimensional problem.

@sipie800 this could indeed be a problem. However, there is still the issue that the inputs will be split into two parts, so one part will be empty and this might not be handled appropriately since coupling flows aren't designed for this case.

Best regards,
Vincent

from normalizing-flows.

Hi all,

I'm closing this issue due to inactivity, so I assume my reply solved it for you.

Best regards,
Vincent

from normalizing-flows.