Fix off-by-one NUTS adaptations (#2751)

Closes #2750.

Prior to #2674 the counter
`i` would be initialised to `1` after the first adaptation had happened.
That PR, first released in v0.41, changed it such that the counter `i`
was initialised to `1` before the first adaptation happened. That
should, of course, be zero, which this PR fixes.

I've checked locally and this change makes it such that results from
pre-v0.41 are exactly reproducible now, as long as all else is equal
(e.g. initialisation, discard_initial, etc).

Author: penelopeysm

HISTORY.md

@@ -1,4 +1,8 @@
-# 0.43.3
+# 0.42.4
+
+Fixes a typo that caused NUTS to perform one less adaptation step than in versions prior to 0.41.
+
+# 0.42.3
 
 Removes some dead code.
 

Project.toml

@@ -1,6 +1,6 @@
 name = "Turing"
 uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
-version = "0.42.3"
+version = "0.42.4"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"

src/mcmc/hmc.jl

@@ -228,7 +228,7 @@ function Turing.Inference.initialstep(
     adaptor = AHMCAdaptor(spl, hamiltonian.metric, nadapts; ϵ=ϵ)
 
     transition = DynamicPPL.ParamsWithStats(theta, ldf, NamedTuple())
-    state = HMCState(vi, 1, kernel, hamiltonian, z, adaptor, ldf)
+    state = HMCState(vi, 0, kernel, hamiltonian, z, adaptor, ldf)
 
     return transition, state
 end