Turing v0.48.0 Diff since v0.47.4 Breaking changes The SGLD and SGHMC samplers, which did not support minibatching, have been removed along with PolynomialStepsize (#2890). MH(var => proposal, ...) and LinkedRW have been removed. Use MH() for prior proposals, or MH(cov_matrix) for a Gaussian random walk over the complete linked parameter vector. In Gibbs, assign MH(cov_matrix) to the target variable block (#2883). Gibbs now conditions its components with DynamicPPL.condition. GibbsContext and Turing.Inference.make_conditional are removed. Rename isgibbscomponent to supports_gibbs and gibbs_get_raw_values to gibbs_get_parameter_values; the old names are deprecated. Gibbs no longer supports model arguments containing missing: declare the latent variable inside the model and condition on observations instead (#2863). Every variable reached by a Gibbs model must belong to a component. A component may not change the dimension or existence of a variable owned only by another component; put the variable and whatever decides its shape in one block. Changes to another block's support or distributional form remain unchecked and can make the chain reducible; see the Gibbs docstring (#2863). Other changes ESS Gibbs components now use the current conditional prior after another component changes its parameters (#2885). Gibbs chains now include component-sampler statistics (#2863). NUTS and HMCDA now accept a NamedTuple or Dict{VarName} initial_params, which HMC already did (#2878). Emcee now honours chain_type=MCMCChains.Chains, and refuses walkers that start in different parameter layouts instead of failing inside the decode or the stretch proposal (#2879). estimate_mode now represents indexed bounds using the model's parameter shapes, avoiding growable-array warnings during optimisation (#2888). estimate_mode now throws on a bound that does not cover a whole variable, where it previously either dropped it and returned the unconstrained mode or failed with a bare DimensionMismatch, and warns on a bound no variable can use rather than ignoring it in silence (#2880). Prior() now warns that initial_params has no effect instead of discarding it silently (#2881). Gibbs now rejects an external sampler whose state inherits AbstractMCMC's model-dropping three-argument setparams!! fallback, rather than running with stale model-dependent caches (#2891). GibbsConditional now distinguishes a ranged tilde statement from multiple element-wise statements when given one conditional distribution (#2891). Merged pull requests: Gibbs: condition components instead of GibbsContext (#2863) (@yebai) hmc: Convert initial_params for NUTS/HMCDA and delete a dead adaptor branch (#2878) (@yebai) emcee: Honour chain_type and refuse walkers in different parameter layouts (#2879) (@yebai) optimisation: Stop silently discarding lb/ub bounds (#2880) (@yebai) prior: Warn that initial_params has no effect (#2881) (@yebai) mh: remove per-variable proposals (#2883) (@yebai) ESS retains stale prior means after Gibbs conditioning (#2885) (@yebai) Reject model-dropping setparams!! in external Gibbs components (#2886) (@yebai) optimisation: eliminate growable constraint warnings (#2888) (@yebai) docs: clarify README (#2889) (@yebai) Remove SGLD and SGHMC (#2890) (@yebai) Fix Gibbs handling of wrapped external samplers and conditional sites (#2891) (@yebai) Closed issues: remove SGHMC / SGLD (#2270) RepeatSampler should be in AbstractMCMC (#2671) Gibbs does not use initialisation strategies for component samplers (#2693) Stats from component samplers are lost with Gibbs (#2766) Gibbs third branch (#2810) Use Optim.jl for optimisation (#2814) Rework Gibbs: replace GibbsContext with condition, complete the component interface (#2860) Require samplers to opt in to varying-dimensional targets (#2866) ESS in Gibbs retains stale conditional-prior means (#2873) Reject model-dropping setparams!! before running external Gibbs samplers (#2875) Misspecified MH proposals can bias sampling (#2876)