Transformed variables causing errors with NUTS but not with SVGD

Questions
6

While trying to build a simpler example to reproduce this error, I also managed to find what’s causing this error. Below code will reproduce the error:

import numpy as np
import pymc3 as pm
import theano.tensor as tt

class Linear(pm.distributions.transforms.ElemwiseTransform):
    name = "linear"

    def __init__(self, alpha, beta):
        self.alpha = tt.as_tensor_variable(alpha)
        self.beta = tt.as_tensor_variable(beta)

    def forward(self, x):
        return self.alpha * x + self.beta

    def forward_val(self, x, point=None):
        return self.alpha * x + self.beta

    def backward(self, x):
        return (1 / self.alpha) * (x - self.beta)

    def jacobian_det(self, x):
        return -tt.log(self.alpha)

obs = dict()
sig = 1.0
obs['x'] = np.random.normal(loc=10.0, scale=sig, size=1000)
alpha = 10
beta = 0
params_init = {'mu':0}

with pm.Model() as mdl:
    mu = pm.Normal('mu', mu=0.0, sd=1.0, transform=Linear(alpha, beta))
    x = pm.Normal('x', mu=mu, sd=1.0, observed=obs['x'])
    trace = pm.sample(start=params_init, chains=1, cores=1)

If the start values are given for variable mu instead of mu_linear__ then this error occurs. Is this the expected behaviour? I understand that sampling is done on transformed variables but it is a bit confusing that initial values are to be defined for variable names which are never explicitly defined in the model.