Neural Spline Flow (NSF) based on coupling layers

[austin-hoover]

Hi all,

Have any of you implemented the Neural Spline Flow (NSF) model in Zuko using coupling transformations rather than autoregressive transformations?

Thanks,
Austin

[francois-rozet]

Hi @austin-hoover,

There are several ways to get a NSF with coupling transformations.

1. Use the zuko.flows.NSF class with passes=2 (which makes the autoregressive transform equivalent to a coupling one).

zuko.flows.NSF(5, 3, transforms=4, passes=2, bins=8)

2. Use the zuko.flows.NICE class and set the univariate function to MonotonicRQSTransform.

bins = 8
zuko.flows.NICE(
    5,
    3,
    transforms=4,
    univariate=zuko.transforms.MonotonicRQSTransform,
    shapes=[[bins], [bins], [bins - 1]],
)

3. Build your own flow with zuko.lazy.Flow and zuko.flows.GeneralCouplingTransform.

[austin-hoover]

Thanks for the help!

…

On Tue, Sep 9, 2025 at 8:41 AM François Rozet ***@***.***> wrote: Hi @austin-hoover <https://github.com/austin-hoover>, There are several ways to get a NSF with coupling transformations. 1. Use the zuko.flows.NSF class with passes=2 (which makes the autoregressive transform equivalent to a coupling one). zuko.flows.NSF(5, 3, transforms=4, passes=2, bins=8) 2. Use the zuko.flows.NICE class but change the univariate function to MonotonicRQSTransform. bins = 8zuko.flows.NICE( 5, 3, transforms=4, univariate=zuko.transforms.MonotonicRQSTransform, shapes=[[bins], [bins], [bins - 1]], ) 3. Build your own flow with zuko.lazy.Flow and zuko.flows.GeneralCouplingTransform. — Reply to this email directly, view it on GitHub <#68 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN2TFLS4ODRB75CI6HWMVMD3R3DFXAVCNFSM6AAAAACF6ISP76VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZVGEZTMOA> . You are receiving this because you were mentioned.Message ID: ***@***.***>