Attempt fixing bnn tests in Ezyme PR

[torfjelde]

Turing.jl/test/mcmc/hmc.jl

Lines 106 to 121 in cbd5d79

	@model function bnn(ts)
	b1 ~ MvNormal([0. ;0.; 0.],
	[var_prior 0. 0.; 0. var_prior 0.; 0. 0. var_prior])
	w11 ~ MvNormal([0.; 0.], [var_prior 0.; 0. var_prior])
	w12 ~ MvNormal([0.; 0.], [var_prior 0.; 0. var_prior])
	w13 ~ MvNormal([0.; 0.], [var_prior 0.; 0. var_prior])
	bo ~ Normal(0, var_prior)
	wo ~ MvNormal([0.; 0; 0],
	[var_prior 0. 0.; 0. var_prior 0.; 0. 0. var_prior])
	for i = rand(1:N, 10)
	y = nn(xs[i], b1, w11, w12, w13, bo, wo)
	ts[i] ~ Bernoulli(y)
	end
	b1, w11, w12, w13, bo, wo
	end

[torfjelde]

@wsmoses explained that the hvcat call hit due to the syntactic sugar we use in the bnn model won't be an issue if the entire thing is type-stable, so it might be fixed by simply moving the var_prior into the arguments of the model or something. Giviing that a go now.

[torfjelde]

@wsmoses @mhauru

[yebai]

Fixed by 5bfd06d