Addition of step_warmup (#117)

* added step_warmup which is can be overloaded when convenient * added step_warmup to docs * Update src/interface.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * introduce new kwarg `num_warmup` to `sample` which uses `step_warmup` * updated docs * allow combination of discard_initial and num_warmup * added docstring for mcmcsample * Apply suggestions from code review Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: David Widmann <[email protected]> * Update src/sample.jl Co-authored-by: David Widmann <[email protected]> * removed docstring and deferred description of keyword arguments to the docs * Update src/sample.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update src/sample.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * added num_warmup to common keyword arguments docs * also allow step_warmup for the initial step * simplify logic for discarding fffinitial samples * Apply suggestions from code review Co-authored-by: David Widmann <[email protected]> * also report progress for the discarded samples * Apply suggestions from code review Co-authored-by: David Widmann <[email protected]> * move progress-report to end of for-loop for discard samples * move step_warmup to the inner while loops too * Update src/sample.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Apply suggestions from code review Co-authored-by: David Widmann <[email protected]> * reverted to for-loop * Update src/sample.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * added accidentanly removed comment * Update src/sample.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * fixed formatting * fix typo * Apply suggestions from code review Co-authored-by: David Widmann <[email protected]> * Added testing of warmup steps * Added checks as @devmotion requested * Removed unintended change in previous commit * Bumped patch version * Bump minor version instead of patch version since this is a new feature --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: David Widmann <[email protected]>
TuringLang · Oct 4, 2024 · a18c12f · a18c12f · torfjelde · Oct 4, 2024
1 parent 8431b31
commit a18c12f
Show file tree

Hide file tree

Showing 7 changed files with 193 additions and 34 deletions.
diff --git a/Project.toml b/Project.toml
@@ -3,7 +3,7 @@ uuid = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
 keywords = ["markov chain monte carlo", "probabilistic programming"]
 license = "MIT"
 desc = "A lightweight interface for common MCMC methods."
-version = "5.3.0"
+version = "5.4.0"
 
 [deps]
 BangBang = "198e06fe-97b7-11e9-32a5-e1d131e6ad66"

diff --git a/docs/src/api.md b/docs/src/api.md
@@ -71,9 +71,13 @@ Common keyword arguments for regular and parallel sampling are:
 - `progress` (default: `AbstractMCMC.PROGRESS[]` which is `true` initially):  toggles progress logging
 - `chain_type` (default: `Any`): determines the type of the returned chain
 - `callback` (default: `nothing`): if `callback !== nothing`, then
-  `callback(rng, model, sampler, sample, state, iteration)` is called after every sampling step,
-  where `sample` is the most recent sample of the Markov chain and `state` and `iteration` are the current state and iteration of the sampler
-- `discard_initial` (default: `0`): number of initial samples that are discarded
+  `callback(rng, model, sampler, sample, iteration)` is called after every sampling step,
+  where `sample` is the most recent sample of the Markov chain and `iteration` is the current iteration
+- `num_warmup` (default: `0`): number of "warm-up" steps to take before the first "regular" step, 
+   i.e. number of times to call [`AbstractMCMC.step_warmup`](@ref) before the first call to 
+   [`AbstractMCMC.step`](@ref).
+- `discard_initial` (default: `num_warmup`): number of initial samples that are discarded. Note that
+  if `discard_initial < num_warmup`, warm-up samples will also be included in the resulting samples.
 - `thinning` (default: `1`): factor by which to thin samples.
 - `initial_state` (default: `nothing`): if `initial_state !== nothing`, the first call to [`AbstractMCMC.step`](@ref)
   is passed `initial_state` as the `state` argument.

diff --git a/docs/src/design.md b/docs/src/design.md
@@ -63,6 +63,15 @@ the sampling step of the inference method.
 AbstractMCMC.step
 ```
 
+If one also has some special handling of the warmup-stage of sampling, then this can be specified by overloading
+
+```@docs
+AbstractMCMC.step_warmup
+```
+
+which will be used for the first `num_warmup` iterations, as specified as a keyword argument to [`AbstractMCMC.sample`](@ref). 
+Note that this is optional; by default it simply calls [`AbstractMCMC.step`](@ref) from above.
+
 ## Collecting samples
 
 !!! note

diff --git a/src/interface.jl b/src/interface.jl
@@ -73,6 +73,23 @@ current `state` of the sampler.
 """
 function step end
 
+"""
+    step_warmup(rng, model, sampler[, state; kwargs...])
+
+Return a 2-tuple of the next sample and the next state of the MCMC `sampler` for `model`.
+
+When sampling using [`sample`](@ref), this takes the place of [`AbstractMCMC.step`](@ref) in the first
+`num_warmup` number of iterations, as specified by the `num_warmup` keyword to [`sample`](@ref).
+This is useful if the sampler has an initial "warmup"-stage that is different from the
+standard iteration.
+
+By default, this simply calls [`AbstractMCMC.step`](@ref).
+"""
+step_warmup(rng, model, sampler; kwargs...) = step(rng, model, sampler; kwargs...)
+function step_warmup(rng, model, sampler, state; kwargs...)
+    return step(rng, model, sampler, state; kwargs...)
+end
+
 """
     samples(sample, model, sampler[, N; kwargs...])
 

diff --git a/src/sample.jl b/src/sample.jl
@@ -43,6 +43,11 @@ isdone(rng, model, sampler, samples, state, iteration; kwargs...)
 ```
 where `state` and `iteration` are the current state and iteration of the sampler, respectively.
 It should return `true` when sampling should end, and `false` otherwise.
+
+# Keyword arguments
+
+See https://turinglang.org/AbstractMCMC.jl/dev/api/#Common-keyword-arguments for common keyword
+arguments.
 """
 function StatsBase.sample(
     rng::Random.AbstractRNG,
@@ -80,6 +85,11 @@ end
 
 Sample `nchains` Monte Carlo Markov chains from the `model` with the `sampler` in parallel
 using the `parallel` algorithm, and combine them into a single chain.
+
+# Keyword arguments
+
+See https://turinglang.org/AbstractMCMC.jl/dev/api/#Common-keyword-arguments for common keyword
+arguments.
 """
 function StatsBase.sample(
     rng::Random.AbstractRNG,
@@ -94,7 +104,6 @@ function StatsBase.sample(
 end
 
 # Default implementations of regular and parallel sampling.
-
 function mcmcsample(
     rng::Random.AbstractRNG,
     model::AbstractModel,
@@ -103,15 +112,28 @@ function mcmcsample(
     progress=PROGRESS[],
     progressname="Sampling",
     callback=nothing,
-    discard_initial=0,
+    num_warmup::Int=0,
+    discard_initial::Int=num_warmup,
     thinning=1,
     chain_type::Type=Any,
     initial_state=nothing,
     kwargs...,
 )
     # Check the number of requested samples.
     N > 0 || error("the number of samples must be ≥ 1")
+    discard_initial >= 0 ||
+        throw(ArgumentError("number of discarded samples must be non-negative"))
+    num_warmup >= 0 ||
+        throw(ArgumentError("number of warm-up samples must be non-negative"))
     Ntotal = thinning * (N - 1) + discard_initial + 1
+    Ntotal >= num_warmup || throw(
+        ArgumentError("number of warm-up samples exceeds the total number of samples")
+    )
+
+    # Determine how many samples to drop from `num_warmup` and the
+    # main sampling process before we start saving samples.
+    discard_from_warmup = min(num_warmup, discard_initial)
+    keep_from_warmup = num_warmup - discard_from_warmup
 
     # Start the timer
     start = time()
@@ -126,22 +148,41 @@ function mcmcsample(
         end
 
         # Obtain the initial sample and state.
-        sample, state = if initial_state === nothing
-            step(rng, model, sampler; kwargs...)
+        sample, state = if num_warmup > 0
+            if initial_state === nothing
+                step_warmup(rng, model, sampler; kwargs...)
+            else
+                step_warmup(rng, model, sampler, initial_state; kwargs...)
+            end
         else
-            step(rng, model, sampler, initial_state; kwargs...)
+            if initial_state === nothing
+                step(rng, model, sampler; kwargs...)
+            else
+                step(rng, model, sampler, initial_state; kwargs...)
+            end
+        end
+
+        # Update the progress bar.
+        itotal = 1
+        if progress && itotal >= next_update
+            ProgressLogging.@logprogress itotal / Ntotal
+            next_update = itotal + threshold
         end
 
         # Discard initial samples.
-        for i in 1:discard_initial
-            # Update the progress bar.
-            if progress && i >= next_update
-                ProgressLogging.@logprogress i / Ntotal
-                next_update = i + threshold
+        for j in 1:discard_initial
+            # Obtain the next sample and state.
+            sample, state = if j ≤ num_warmup
+                step_warmup(rng, model, sampler, state; kwargs...)
+            else
+                step(rng, model, sampler, state; kwargs...)
             end
 
-            # Obtain the next sample and state.
-            sample, state = step(rng, model, sampler, state; kwargs...)
+            # Update the progress bar.
+            if progress && (itotal += 1) >= next_update
+                ProgressLogging.@logprogress itotal / Ntotal
+                next_update = itotal + threshold
+            end
         end
 
         # Run callback.
@@ -151,19 +192,16 @@ function mcmcsample(
         samples = AbstractMCMC.samples(sample, model, sampler, N; kwargs...)
         samples = save!!(samples, sample, 1, model, sampler, N; kwargs...)
 
-        # Update the progress bar.
-        itotal = 1 + discard_initial
-        if progress && itotal >= next_update
-            ProgressLogging.@logprogress itotal / Ntotal
-            next_update = itotal + threshold
-        end
-
         # Step through the sampler.
         for i in 2:N
             # Discard thinned samples.
             for _ in 1:(thinning - 1)
                 # Obtain the next sample and state.
-                sample, state = step(rng, model, sampler, state; kwargs...)
+                sample, state = if i ≤ keep_from_warmup
+                    step_warmup(rng, model, sampler, state; kwargs...)
+                else
+                    step(rng, model, sampler, state; kwargs...)
+                end
 
                 # Update progress bar.
                 if progress && (itotal += 1) >= next_update
@@ -173,7 +211,11 @@ function mcmcsample(
             end
 
             # Obtain the next sample and state.
-            sample, state = step(rng, model, sampler, state; kwargs...)
+            sample, state = if i ≤ keep_from_warmup
+                step_warmup(rng, model, sampler, state; kwargs...)
+            else
+                step(rng, model, sampler, state; kwargs...)
+            end
 
             # Run callback.
             callback === nothing ||
@@ -217,28 +259,51 @@ function mcmcsample(
     progress=PROGRESS[],
     progressname="Convergence sampling",
     callback=nothing,
-    discard_initial=0,
+    num_warmup=0,
+    discard_initial=num_warmup,
     thinning=1,
     initial_state=nothing,
     kwargs...,
 )
+    # Check the number of requested samples.
+    discard_initial >= 0 ||
+        throw(ArgumentError("number of discarded samples must be non-negative"))
+    num_warmup >= 0 ||
+        throw(ArgumentError("number of warm-up samples must be non-negative"))
+
+    # Determine how many samples to drop from `num_warmup` and the
+    # main sampling process before we start saving samples.
+    discard_from_warmup = min(num_warmup, discard_initial)
+    keep_from_warmup = num_warmup - discard_from_warmup
 
     # Start the timer
     start = time()
     local state
 
     @ifwithprogresslogger progress name = progressname begin
         # Obtain the initial sample and state.
-        sample, state = if initial_state === nothing
-            step(rng, model, sampler; kwargs...)
+        sample, state = if num_warmup > 0
+            if initial_state === nothing
+                step_warmup(rng, model, sampler; kwargs...)
+            else
+                step_warmup(rng, model, sampler, initial_state; kwargs...)
+            end
         else
-            step(rng, model, sampler, state; kwargs...)
+            if initial_state === nothing
+                step(rng, model, sampler; kwargs...)
+            else
+                step(rng, model, sampler, initial_state; kwargs...)
+            end
         end
 
         # Discard initial samples.
-        for _ in 1:discard_initial
+        for j in 1:discard_initial
             # Obtain the next sample and state.
-            sample, state = step(rng, model, sampler, state; kwargs...)
+            sample, state = if j ≤ num_warmup
+                step_warmup(rng, model, sampler, state; kwargs...)
+            else
+                step(rng, model, sampler, state; kwargs...)
+            end
         end
 
         # Run callback.
@@ -250,16 +315,23 @@ function mcmcsample(
 
         # Step through the sampler until stopping.
         i = 2
-
         while !isdone(rng, model, sampler, samples, state, i; progress=progress, kwargs...)
             # Discard thinned samples.
             for _ in 1:(thinning - 1)
                 # Obtain the next sample and state.
-                sample, state = step(rng, model, sampler, state; kwargs...)
+                sample, state = if i ≤ keep_from_warmup
+                    step_warmup(rng, model, sampler, state; kwargs...)
+                else
+                    step(rng, model, sampler, state; kwargs...)
+                end
             end
 
             # Obtain the next sample and state.
-            sample, state = step(rng, model, sampler, state; kwargs...)
+            sample, state = if i ≤ keep_from_warmup
+                step_warmup(rng, model, sampler, state; kwargs...)
+            else
+                step(rng, model, sampler, state; kwargs...)
+            end
 
             # Run callback.
             callback === nothing ||

diff --git a/test/sample.jl b/test/sample.jl
@@ -575,6 +575,45 @@
         @test all(chain[i].b == ref_chain[i + discard_initial].b for i in 1:N)
     end
 
+    @testset "Warm-up steps" begin
+        # Create a chain and discard initial samples.
+        Random.seed!(1234)
+        N = 100
+        num_warmup = 50
+
+        # Everything should be discarded here.
+        chain = sample(MyModel(), MySampler(), N; num_warmup=num_warmup)
+        @test length(chain) == N
+        @test !ismissing(chain[1].a)
+
+        # Repeat sampling without discarding initial samples.
+        # On Julia < 1.6 progress logging changes the global RNG and hence is enabled here.
+        # https://github.com/TuringLang/AbstractMCMC.jl/pull/102#issuecomment-1142253258
+        Random.seed!(1234)
+        ref_chain = sample(
+            MyModel(), MySampler(), N + num_warmup; progress=VERSION < v"1.6"
+        )
+        @test all(chain[i].a == ref_chain[i + num_warmup].a for i in 1:N)
+        @test all(chain[i].b == ref_chain[i + num_warmup].b for i in 1:N)
+
+        # Some other stuff.
+        Random.seed!(1234)
+        discard_initial = 10
+        chain_warmup = sample(
+            MyModel(),
+            MySampler(),
+            N;
+            num_warmup=num_warmup,
+            discard_initial=discard_initial,
+        )
+        @test length(chain_warmup) == N
+        @test all(chain_warmup[i].a == ref_chain[i + discard_initial].a for i in 1:N)
+        # Check that the first `num_warmup - discard_initial` samples are warmup samples.
+        @test all(
+            chain_warmup[i].is_warmup == (i <= num_warmup - discard_initial) for i in 1:N
+        )
+    end
+
     @testset "Thin chain by a factor of `thinning`" begin
         # Run a thinned chain with `N` samples thinned by factor of `thinning`.
         Random.seed!(100)

diff --git a/test/utils.jl b/test/utils.jl
@@ -3,8 +3,11 @@ struct MyModel <: AbstractMCMC.AbstractModel end
 struct MySample{A,B}
     a::A
     b::B
+    is_warmup::Bool
 end
 
+MySample(a, b) = MySample(a, b, false)
+
 struct MySampler <: AbstractMCMC.AbstractSampler end
 struct AnotherSampler <: AbstractMCMC.AbstractSampler end
 
@@ -16,6 +19,21 @@ end
 
 MyChain(a, b) = MyChain(a, b, NamedTuple())
 
+function AbstractMCMC.step_warmup(
+    rng::AbstractRNG,
+    model::MyModel,
+    sampler::MySampler,
+    state::Union{Nothing,Integer}=nothing;
+    loggers=false,
+    initial_params=nothing,
+    kwargs...,
+)
+    transition, state = AbstractMCMC.step(
+        rng, model, sampler, state; loggers, initial_params, kwargs...
+    )
+    return MySample(transition.a, transition.b, true), state
+end
+
 function AbstractMCMC.step(
     rng::AbstractRNG,
     model::MyModel,