Decompose forward function into initialize, predict, update (#105)

* Decompose forward function into initialize, predict, update

* Fixes

* Remove unused predict

---------

Co-authored-by: Guillaume Dalle <[email protected]>

src/HiddenMarkovModels.jl

@@ -38,6 +38,7 @@ include("utils/lightdiagnormal.jl")
 include("utils/lightcategorical.jl")
 include("utils/limits.jl")
 
+include("inference/predict.jl")
 include("inference/forward.jl")
 include("inference/viterbi.jl")
 include("inference/forward_backward.jl")

src/inference/forward.jl

@@ -63,6 +63,27 @@ function initialize_forward(
     return ForwardStorage(α, logL, B, c)
 end
 
+function _forward_digest_observation!(
+    current_state_marginals::AbstractVector{<:Real},
+    current_obs_likelihoods::AbstractVector{<:Real},
+    hmm::AbstractHMM,
+    obs,
+    control,
+)
+    a, b = current_state_marginals, current_obs_likelihoods
+
+    obs_logdensities!(b, hmm, obs, control)
+    logm = maximum(b)
+    b .= exp.(b .- logm)
+
+    a .*= b
+    c = inv(sum(a))
+    lmul!(c, a)
+
+    logL = -log(c) + logm
+    return c, logL
+end
+
 function _forward!(
     storage::ForwardOrForwardBackwardStorage,
     hmm::AbstractHMM,
@@ -73,36 +94,19 @@ function _forward!(
 )
     (; α, B, c, logL) = storage
     t1, t2 = seq_limits(seq_ends, k)
-
-    # Initialization
-    Bₜ₁ = view(B, :, t1)
-    obs_logdensities!(Bₜ₁, hmm, obs_seq[t1], control_seq[t1])
-    logm = maximum(Bₜ₁)
-    Bₜ₁ .= exp.(Bₜ₁ .- logm)
-
-    init = initialization(hmm)
-    αₜ₁ = view(α, :, t1)
-    αₜ₁ .= init .* Bₜ₁
-    c[t1] = inv(sum(αₜ₁))
-    lmul!(c[t1], αₜ₁)
-
-    logL[k] = -log(c[t1]) + logm
-
-    # Loop
-    for t in t1:(t2 - 1)
-        Bₜ₊₁ = view(B, :, t + 1)
-        obs_logdensities!(Bₜ₊₁, hmm, obs_seq[t + 1], control_seq[t + 1])
-        logm = maximum(Bₜ₊₁)
-        Bₜ₊₁ .= exp.(Bₜ₊₁ .- logm)
-
-        trans = transition_matrix(hmm, control_seq[t])
-        αₜ, αₜ₊₁ = view(α, :, t), view(α, :, t + 1)
-        mul!(αₜ₊₁, transpose(trans), αₜ)
-        αₜ₊₁ .*= Bₜ₊₁
-        c[t + 1] = inv(sum(αₜ₊₁))
-        lmul!(c[t + 1], αₜ₊₁)
-
-        logL[k] += -log(c[t + 1]) + logm
+    logL[k] = zero(eltype(logL))
+    for t in t1:t2
+        αₜ = view(α, :, t)
+        Bₜ = view(B, :, t)
+        if t == t1
+            copyto!(αₜ, initialization(hmm))
+        else
+            αₜ₋₁ = view(α, :, t - 1)
+            predict_next_state!(αₜ, hmm, αₜ₋₁, control_seq[t - 1])
+        end
+        cₜ, logLₜ = _forward_digest_observation!(αₜ, Bₜ, hmm, obs_seq[t], control_seq[t])
+        c[t] = cₜ
+        logL[k] += logLₜ
     end
 
     @argcheck isfinite(logL[k])

src/inference/predict.jl

@@ -0,0 +1,10 @@
+function predict_next_state!(
+    next_state_marginals::AbstractVector{<:Real},
+    hmm::AbstractHMM,
+    current_state_marginals::AbstractVector{<:Real},
+    control=nothing,
+)
+    trans = transition_matrix(hmm, control)
+    mul!(next_state_marginals, transpose(trans), current_state_marginals)
+    return next_state_marginals
+end

src/types/abstract_hmm.jl

@@ -72,6 +72,9 @@ log_initialization(hmm::AbstractHMM) = elementwise_log(initialization(hmm))
     transition_matrix(hmm, control)
 
 Return the matrix of state transition probabilities for `hmm` (possibly when `control` is applied).
+
+!!! note
+    When processing sequences, the control at time `t` influences the transition from time `t` to `t+1` (and not from time `t-1` to `t`).
 """
 function transition_matrix end
 
@@ -82,6 +85,9 @@ function transition_matrix end
 Return the matrix of state transition log-probabilities for `hmm` (possibly when `control` is applied).
 
 Falls back on `transition_matrix`.
+
+!!! note
+    When processing sequences, the control at time `t` influences the transition from time `t` to `t+1` (and not from time `t-1` to `t`).
 """
 function log_transition_matrix(hmm::AbstractHMM, control)
     return elementwise_log(transition_matrix(hmm, control))