Skip to content

Commit

Permalink
More performance improvement to log density evaluation (#230)
Browse files Browse the repository at this point in the history 
This remove another performance bottleneck: `getindex` from MetaGraph.
This is done by storing the data in arrays.

The size of a `BUGSModel` is increased by 15%. 

And it roughly half the previous time, while not impacting the
ReverseDiff time.

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
  • Loading branch information
@sunxd3 @github-actions
sunxd3 and github-actions[bot] authored Oct 29, 2024
1 parent f9c253b commit b466e8e
Show file tree
Hide file tree
Showing 4 changed files with 93 additions and 37 deletions.
2 changes: 1 addition & 1 deletion ext/JuliaBUGSMCMCChainsExt.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -84,7 +84,7 @@ function JuliaBUGS.gen_chains(
g = model.g

generated_vars = find_generated_vars(g)
generated_vars = [v for v in model.sorted_nodes if v in generated_vars] # keep the order
generated_vars = [v for v in model.eval_cache.sorted_nodes if v in generated_vars] # keep the order

param_vals = []
generated_quantities = []
Expand Down
12 changes: 6 additions & 6 deletions src/gibbs.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ abstract type AbstractGibbsState end
struct GibbsState{T,S,C} <: AbstractGibbsState
values::T
conditioning_schedule::S
sorted_nodes_cache::C
cached_eval_caches::C
end

ensure_vector(x) = x isa Union{Number,VarName} ? [x] : x
Expand All @@ -25,17 +25,17 @@ function AbstractMCMC.step(
model=l_model.logdensity,
kwargs...,
) where {N,S}
sorted_nodes_cache, conditioning_schedule = OrderedDict(), OrderedDict()
cached_eval_caches, conditioning_schedule = OrderedDict(), OrderedDict()
for variable_group in keys(sampler.sampler_map)
variable_to_condition_on = setdiff(model.parameters, ensure_vector(variable_group))
conditioning_schedule[variable_to_condition_on] = sampler.sampler_map[variable_group]
conditioned_model = AbstractPPL.condition(
model, variable_to_condition_on, model.evaluation_env
)
sorted_nodes_cache[variable_to_condition_on] = conditioned_model.sorted_nodes
cached_eval_caches[variable_to_condition_on] = conditioned_model.eval_cache
end
param_values = JuliaBUGS.getparams(model)
return param_values, GibbsState(param_values, conditioning_schedule, sorted_nodes_cache)
return param_values, GibbsState(param_values, conditioning_schedule, cached_eval_caches)
end

function AbstractMCMC.step(
Expand All @@ -50,12 +50,12 @@ function AbstractMCMC.step(
for vs in keys(state.conditioning_schedule)
model = initialize!(model, param_values)
cond_model = AbstractPPL.condition(
model, vs, model.evaluation_env, state.sorted_nodes_cache[vs]
model, vs, model.evaluation_env, state.cached_eval_caches[vs]
)
param_values = gibbs_internal(rng, cond_model, state.conditioning_schedule[vs])
end
return param_values,
GibbsState(param_values, state.conditioning_schedule, state.sorted_nodes_cache)
GibbsState(param_values, state.conditioning_schedule, state.cached_eval_caches)
end

function gibbs_internal end
Expand Down
112 changes: 84 additions & 28 deletions src/model.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,14 +3,53 @@
# instead of https://github.com/TuringLang/AbstractMCMC.jl/blob/d7c549fe41a80c1f164423c7ac458425535f624b/src/logdensityproblems.jl#L90
abstract type AbstractBUGSModel end

"""
EvalCache{TNF,TNA,TV}
Pre-compute the values of the nodes in the model to avoid lookups from MetaGraph.
"""
struct EvalCache{TNF,TNA,TV}
sorted_nodes::Vector{<:VarName}
is_stochastic_vals::Vector{Bool}
is_observed_vals::Vector{Bool}
node_function_vals::TNF
node_args_vals::TNA
loop_vars_vals::TV
end

function EvalCache(sorted_nodes::Vector{<:VarName}, g::BUGSGraph)
is_stochastic_vals = Array{Bool}(undef, length(sorted_nodes))
is_observed_vals = Array{Bool}(undef, length(sorted_nodes))
node_function_vals = []
node_args_vals = []
loop_vars_vals = []
for (i, vn) in enumerate(sorted_nodes)
(; is_stochastic, is_observed, node_function, node_args, loop_vars) = g[vn]
is_stochastic_vals[i] = is_stochastic
is_observed_vals[i] = is_observed
push!(node_function_vals, node_function)
push!(node_args_vals, Val(node_args))
push!(loop_vars_vals, loop_vars)
end
return EvalCache(
sorted_nodes,
is_stochastic_vals,
is_observed_vals,
node_function_vals,
node_args_vals,
loop_vars_vals,
)
end

"""
BUGSModel
The `BUGSModel` object is used for inference and represents the output of compilation. It implements the
[`LogDensityProblems.jl`](https://github.com/tpapp/LogDensityProblems.jl) interface.
"""
struct BUGSModel{base_model_T<:Union{<:AbstractBUGSModel,Nothing},T<:NamedTuple} <:
AbstractBUGSModel
struct BUGSModel{
base_model_T<:Union{<:AbstractBUGSModel,Nothing},T<:NamedTuple,TNF,TNA,TV
} <: AbstractBUGSModel
" Indicates whether the model parameters are in the transformed space. "
transformed::Bool

Expand All @@ -27,8 +66,8 @@ struct BUGSModel{base_model_T<:Union{<:AbstractBUGSModel,Nothing},T<:NamedTuple}
evaluation_env::T
"A vector containing the names of the model parameters (unobserved stochastic variables)."
parameters::Vector{<:VarName}
"A vector containing the names of all the variables in the model, sorted in topological order."
sorted_nodes::Vector{<:VarName}
"An `EvalCache` object containing pre-computed values of the nodes in the model. For each topological order, this needs to be recomputed."
eval_cache::EvalCache{TNF,TNA,TV}

"An instance of `BUGSGraph`, representing the dependency graph of the model."
g::BUGSGraph
Expand Down Expand Up @@ -144,14 +183,15 @@ function BUGSModel(
transformed_var_lengths,
evaluation_env,
parameters,
sorted_nodes,
EvalCache(sorted_nodes, g),
g,
nothing,
)
end

function BUGSModel(
model::BUGSModel,
g::BUGSGraph,
parameters::Vector{<:VarName},
sorted_nodes::Vector{<:VarName},
evaluation_env::NamedTuple=model.evaluation_env,
Expand All @@ -164,8 +204,8 @@ function BUGSModel(
model.transformed_var_lengths,
evaluation_env,
parameters,
sorted_nodes,
model.g,
EvalCache(sorted_nodes, g),
g,
isnothing(model.base_model) ? model : model.base_model,
)
end
Expand All @@ -177,9 +217,13 @@ Initialize the model with a NamedTuple of initial values, the values are expecte
"""
function initialize!(model::BUGSModel, initial_params::NamedTuple)
check_input(initial_params)
for vn in model.sorted_nodes
(; is_stochastic, is_observed, node_function, node_args, loop_vars) = model.g[vn]
args = prepare_arg_values(Val(node_args), model.evaluation_env, loop_vars)
for (i, vn) in enumerate(model.eval_cache.sorted_nodes)
is_stochastic = model.eval_cache.is_stochastic_vals[i]
is_observed = model.eval_cache.is_observed_vals[i]
node_function = model.eval_cache.node_function_vals[i]
node_args = model.eval_cache.node_args_vals[i]
loop_vars = model.eval_cache.loop_vars_vals[i]
args = prepare_arg_values(node_args, model.evaluation_env, loop_vars)
if !is_stochastic
value = Base.invokelatest(node_function; args...)
BangBang.@set!! model.evaluation_env = setindex!!(
Expand Down Expand Up @@ -318,11 +362,11 @@ function AbstractPPL.condition(
new_parameters = setdiff(model.parameters, var_group)

sorted_blanket_with_vars = if sorted_nodes isa Nothing
sorted_nodes
model.eval_cache.sorted_nodes
else
filter(
vn -> vn in union(markov_blanket(model.g, new_parameters), new_parameters),
model.sorted_nodes,
model.eval_cache.sorted_nodes,
)
end

Expand All @@ -338,7 +382,9 @@ function AbstractPPL.condition(
end
end

new_model = BUGSModel(model, new_parameters, sorted_blanket_with_vars, evaluation_env)
new_model = BUGSModel(
model, g, new_parameters, sorted_blanket_with_vars, evaluation_env
)
return BangBang.setproperty!!(new_model, :g, g)
end

Expand All @@ -347,18 +393,19 @@ function AbstractPPL.decondition(model::BUGSModel, var_group::Vector{<:VarName})
base_model = model.base_model isa Nothing ? model : model.base_model

new_parameters = [
v for v in base_model.sorted_nodes if v in union(model.parameters, var_group)
v for
v in base_model.eval_cache.sorted_nodes if v in union(model.parameters, var_group)
] # keep the order

markov_blanket_with_vars = union(
markov_blanket(base_model.g, new_parameters), new_parameters
)
sorted_blanket_with_vars = filter(
vn -> vn in markov_blanket_with_vars, base_model.sorted_nodes
vn -> vn in markov_blanket_with_vars, base_model.eval_cache.sorted_nodes
)

new_model = BUGSModel(
model, new_parameters, sorted_blanket_with_vars, base_model.evaluation_env
model, model.g, new_parameters, sorted_blanket_with_vars, base_model.evaluation_env
)
evaluate_env, _ = evaluate!!(new_model)
return BangBang.setproperty!!(new_model, :evaluation_env, evaluate_env)
Expand All @@ -374,12 +421,15 @@ function check_var_group(var_group::Vector{<:VarName}, model::BUGSModel)
end

function AbstractPPL.evaluate!!(rng::Random.AbstractRNG, model::BUGSModel)
(; evaluation_env, g, sorted_nodes) = model
(; evaluation_env, g) = model
vi = deepcopy(evaluation_env)
logp = 0.0
for vn in sorted_nodes
(; is_stochastic, node_function, node_args, loop_vars) = g[vn]
args = prepare_arg_values(Val(node_args), evaluation_env, loop_vars)
for (i, vn) in enumerate(model.eval_cache.sorted_nodes)
is_stochastic = model.eval_cache.is_stochastic_vals[i]
node_function = model.eval_cache.node_function_vals[i]
node_args = model.eval_cache.node_args_vals[i]
loop_vars = model.eval_cache.loop_vars_vals[i]
args = prepare_arg_values(node_args, evaluation_env, loop_vars)
if !is_stochastic
value = node_function(; args...)
evaluation_env = setindex!!(evaluation_env, value, vn)
Expand All @@ -394,11 +444,14 @@ function AbstractPPL.evaluate!!(rng::Random.AbstractRNG, model::BUGSModel)
end

function AbstractPPL.evaluate!!(model::BUGSModel)
(; sorted_nodes, g, evaluation_env) = model
logp = 0.0
for vn in sorted_nodes
(; is_stochastic, node_function, node_args, loop_vars) = g[vn]
args = prepare_arg_values(Val(node_args), evaluation_env, loop_vars)
evaluation_env = deepcopy(model.evaluation_env)
for (i, vn) in enumerate(model.eval_cache.sorted_nodes)
is_stochastic = model.eval_cache.is_stochastic_vals[i]
node_function = model.eval_cache.node_function_vals[i]
node_args = model.eval_cache.node_args_vals[i]
loop_vars = model.eval_cache.loop_vars_vals[i]
args = prepare_arg_values(node_args, evaluation_env, loop_vars)
if !is_stochastic
value = node_function(; args...)
evaluation_env = setindex!!(evaluation_env, value, vn)
Expand Down Expand Up @@ -428,13 +481,16 @@ function AbstractPPL.evaluate!!(model::BUGSModel, flattened_values::AbstractVect
model.untransformed_var_lengths
end

g = model.g
evaluation_env = deepcopy(model.evaluation_env)
current_idx = 1
logp = 0.0
for vn in model.sorted_nodes
(; is_stochastic, is_observed, node_function, node_args, loop_vars) = g[vn]
args = prepare_arg_values(Val(node_args), evaluation_env, loop_vars)
for (i, vn) in enumerate(model.eval_cache.sorted_nodes)
is_stochastic = model.eval_cache.is_stochastic_vals[i]
is_observed = model.eval_cache.is_observed_vals[i]
node_function = model.eval_cache.node_function_vals[i]
node_args = model.eval_cache.node_args_vals[i]
loop_vars = model.eval_cache.loop_vars_vals[i]
args = prepare_arg_values(node_args, evaluation_env, loop_vars)
if !is_stochastic
value = node_function(; args...)
evaluation_env = BangBang.setindex!!(evaluation_env, value, vn)
Expand Down
4 changes: 2 additions & 2 deletions test/graphs.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,11 +47,11 @@ c = @varname c
cond_model = AbstractPPL.condition(model, setdiff(model.parameters, [c]))
# tests for MarkovBlanketBUGSModel constructor
@test cond_model.parameters == [c]
@test Set(Symbol.(cond_model.sorted_nodes)) == Set([:l, :a, :b, :f, :c])
@test Set(Symbol.(cond_model.eval_cache.sorted_nodes)) == Set([:l, :a, :b, :f, :c])

decond_model = AbstractPPL.decondition(cond_model, [a, l])
@test Set(Symbol.(decond_model.parameters)) == Set([:a, :c, :l])
@test Set(Symbol.(decond_model.sorted_nodes)) ==
@test Set(Symbol.(decond_model.eval_cache.sorted_nodes)) ==
Set([:l, :b, :f, :a, :d, :e, :c, :h, :g, :i])

c_value = 4.0
Expand Down

2 comments on commit b466e8e

@sunxd3
Copy link
Member Author

@sunxd3 sunxd3 commented on b466e8e Nov 2, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@JuliaRegistrator register

@JuliaRegistrator
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Registration pull request created: JuliaRegistries/General/118543

Tip: Release Notes

Did you know you can add release notes too? Just add markdown formatted text underneath the comment after the text
"Release notes:" and it will be added to the registry PR, and if TagBot is installed it will also be added to the
release that TagBot creates. i.e.

@JuliaRegistrator register

Release notes:

## Breaking changes

- blah

To add them here just re-invoke and the PR will be updated.

Tagging

After the above pull request is merged, it is recommended that a tag is created on this repository for the registered package version.

This will be done automatically if the Julia TagBot GitHub Action is installed, or can be done manually through the github interface, or via:

git tag -a v0.6.3 -m "<description of version>" b466e8e1587f826205bd2583694a25f362d183b9
git push origin v0.6.3

Please sign in to comment.