init_strategy does not work with CircularReparam

[19MM97]

Hi!

I am having an issue with initialising a model with the init_by_median() strategy for my MCMC NUTS model.
I am drawing the location parameter of a VonMises distribution from a VonMises prior and I reparam this parameter with CircularReparam().

Working example:

import numpyro
from numpyro import sample, plate
from numpyro.infer import NUTS, MCMC, init_to_median
from numpyro.distributions import VonMises
from numpyro.handlers import reparam
from jax.random import PRNGKey
import jax.numpy as jnp

@numpyro.handlers.reparam(
    config={"loc_1": numpyro.infer.reparam.CircularReparam()}
)
def tst(x):
    loc_1 = sample("loc_1", VonMises(0, 100))
    conc = 2
    
    with plate("plate", len(x)):
        y = sample("y", VonMises(loc_1, conc), obs=x)
    return loc_1

if __name__ == "__main__":
    kernel = NUTS(tst, init_strategy=init_to_median(num_samples=1000))
    mcmc = MCMC(kernel, num_warmup=1, num_samples=10)

    mcmc.warmup(PRNGKey(42), collect_warmup=True, x=jnp.array([1, 2, 3]))

    print(mcmc.get_samples())

Output:

{'loc_1': Array([1.7704213], dtype=float32), 'loc_1_unwrapped': Array([1.7704215], dtype=float32)}

This should be the median of VonMises(0, 100), which is 0. If you remove the reparameterisation, you get exactly this value.

[fehiepsi]

As a context, we have a Pyro thread tracking the issue pyro-ppl/pyro#2878