Composing TanhTransform with flows

Misc.
1

I am trying to compose a Normalizing Flow (Sylvester) with a bijective (Tanh) transform to run the following code:

dim = 2
mu = torch.zeros(dim)
log_std = torch.zeros(dim)
transforms = [distributions.transforms.SylvesterFlow(input_dim=dim),
                           distributions.transforms.TanhTransforms()]
distributions.TransformedDistribution(distributions.Normal(mu,log_std.exp()), transforms)

but get the following:

~/sandbox/pyro_fork/pyro/distributions/transforms/sylvester.py in _inverse(self, y)
    127         """
    128 
--> 129         raise KeyError("SylvesterFlow expected to find key in intermediates cache but didn't")
    130 
    131     def log_abs_det_jacobian(self, x, y):

KeyError: "SylvesterFlow expected to find key in intermediates cache but didn't"

From what I understand, only Householder and IAF have their analytical inverse coded, which is causing the error.
AFAIK, the inversion happens by looking up the value y in cache, and if the corresponding x is found, it is returned. However, since Tanh is bijective, shouldn’t this cause no trouble, as the value y=Tanh(NF(x)) will be cached and returned?

2

hello, could you please provide a (fully) reproducible code snippet?