Update documentation #16

FredericWantiez · 2023-07-06T20:57:55Z

No description provided.

examples/smc.jl

src/SSMProblems.jl

FredericWantiez · 2023-07-24T21:00:29Z

573cab3 actually feels a bit weird, the model definition is now specialized to the concrete particle type:

function transition!!(rng::AbstractRNG, t::Int, particle::Particle{<:LinearSSM})
    if t == 1
        return Particle(LinearSSM(rand(rng, f0(t))))
    else
        return Particle(LinearSSM(rand(rng, f(t, particle.model.state))))
    end
end

where if we just dispatch on the model, like in 6394704, we can drop-in different particle estimators with different specialization of AbastractParticle{T}

yebai · 2023-07-28T15:16:11Z

function transition!!(rng::AbstractRNG, t::Int, particle::Particle{<:LinearSSM})
    if t == 1
        return Particle(LinearSSM(rand(rng, f0(t))))
    else
        return Particle(LinearSSM(rand(rng, f(t, particle.model.state))))
    end
end


function transition!!(rng::AbstractRNG, t::Int, m::LinearSSM, particle::Particle)
    if t == 1
        return Particle(rand(rng, f0(t)), m)
    else
        return Particle(rand(rng, f(t, particle.model.state)), m)
    end
end


struct Particle{T}
  z::T
end
Particle(z, m) = particleof(m)(z)

function particleof(::AbstractSSMProblem) 

particleof(::LinearSSM) = Vector{Float32}

yebai · 2023-08-03T11:19:29Z

@devmotion can you take a look at this PR? I'm very involved in the discussions so might be slightly biased.

yebai

Thanks, @FredericWantiez -- it looks pretty good to me. I left a lot of minor comments below. Most comments are typos-fixes or clarifications.

~~In addition, I think we should consider changing.~~

# Define the structure of the latent space
particleof(::LinearSSM) = Float64
dimension(::LinearSSM) = 2

into something like for clarity and convenience

SSMProblems.particle_eltype(::AbstractSSMModel) = typeof(0.) 
SSMProblems.particle_dimension(::LinearSSM) = 2

So the users only need optionally overload particle_eltype if needed.

EDIT: after introducing the AbstractStateSpaceModel type, I think the AbstractParticle type is redundant and should be removed for maximal simplicity. It is only used for dispatching, which is unnecessary due to the new model argument.

EDIT2: particle_eltype and particle_dimention can be removed too. Such information can be obtained by calling

particle = transition!!(rng, 1, m::LinearSSM, ) # step = 1
particle_eltype = eltype(particle)
particle_dimension = length(particle)

EDIT3: introduce a new transition!! function without the step argument, so it is clear it is for initialization.

src/SSMProblems.jl

docs/src/index.md

yebai · 2023-08-04T18:05:32Z

docs/src/index.md

+ ... # Sample from the transition density
+end
+
+function emission_logdensity(step, model::Model, particle::AbstractParticle) 


Suggested change

function emission_logdensity(step, model::Model, particle::AbstractParticle)

function emission_logdensity(step, model::Model, particle::AbstractParticle, observations)

docs/src/index.md

yebai · 2023-08-04T18:08:13Z

docs/src/index.md

+struct Particle{T<:AbstractStateSpaceModel} <: AbstractParticle{T} end 
+
+while !all(map(particle -> isdone(t, model, particles), particles)):
+ ancestors = resample(rng, logweigths)


We need to reset all log weights to zero after the resampling step.

Suggested change

ancestors = resample(rng, logweigths)

ancestors = resample(rng, logweigths)

logweights = zero(logweights)

What is resample, also a part of the interface? I think more explanations would make this example much clearer.

src/SSMProblems.jl

docs/src/index.md

src/SSMProblems.jl

docs/src/index.md

devmotion · 2023-08-04T22:15:59Z

docs/src/index.md

+- __Initialisation__: ``f_0(x)``
+- __Transition__: ``f(x)``
+- __Emission__: ``g(x)``


I think it can be a bit confusing here to use x which otherwise only denotes (internal) states. Maybe just define it in/via the model equations below and remove this list?

docs/src/index.md

devmotion · 2023-08-04T22:51:07Z

examples/smc.jl

 end
- return trace
+ return trace[1:(end - 1)] # Discard the root node


Just don't add this node in the first place? The indexing operation here actually creates a copy.

devmotion · 2023-08-04T22:51:33Z

examples/smc.jl

+function linearize(particle::Particle{T,V}) where {T,V}
+ trace = V[]
+ current = particle
+ while !isnothing(current)


In older Julia versions,

Suggested change

while !isnothing(current)

while current !== nothing

was more performant.

devmotion · 2023-08-04T22:52:12Z

examples/smc.jl

 end

-ParticleContainer = AbstractVector{<:Particle}
+function isdone(t, model::AbstractStateSpaceModel, particles::ParticleContainer)


I have my doubts about this function, as mentioned above.

devmotion · 2023-08-04T22:52:58Z

examples/smc.jl

 end

-ParticleContainer = AbstractVector{<:Particle}
+function isdone(t, model::AbstractStateSpaceModel, particles::ParticleContainer)
+ return all(map(particle -> isdone(t, model, particle), particles))


Without intermediate allocations1:

Suggested change

return all(map(particle -> isdone(t, model, particle), particles))

return all(particle -> isdone(t, model, particle), particles)

devmotion · 2023-08-04T22:55:31Z

examples/smc.jl

+ return logpdf(g(t, particle.state), observations[t])
+end
+
+isdone(t, ::LinearSSM, ::AbstractParticle) = t > T


This is an example of a function that closes over a non-const global. Often this leads to very bad performance.

yebai · 2023-09-08T14:32:44Z

src/SSMProblems.jl

 abstract type AbstractParticleCache end

 """
- transition!!(rng, step, particle[, cache])
+ transition!!(rng, model[, timestep, state, cache])


timestep can be made optional for homogenous SSMs.

yebai · 2023-09-08T14:39:13Z

examples/smc.jl

-# Particle Filter implementation
-struct Particle{T} # Here we just need a tree
- parent::Union{Particle,Nothing}
+# Particle Filter 


Do we want to keep this code block in the module?

yebai · 2023-09-08T14:46:24Z

src/SSMProblems.jl


 Simulate the particle for the next time step from the forward dynamics.
 """
 function transition!! end

 """
- transition_logdensity(step, particle, x[, cache])
+ transition_logdensity(model, timestep, prev_state, next_state[, cache])


Suggested change

transition_logdensity(model, timestep, prev_state, next_state[, cache])

transition_logdensity(model, timestep, prev_state, current_state[, cache])

src/SSMProblems.jl

Co-authored-by: David Widmann <devmotion@users.noreply.github.com>

yebai · 2023-09-13T21:18:13Z

src/utils/particles.jl

@@ -0,0 +1,32 @@
+# Concrete Particles as LinkedLists


@FredericWantiez can we wrap this file into a sub module, e.g. Utils.

Created one Utils, not sure that's the best name

It follows my lazy naming method. Please feel free to change it if you find a better name.

FredericWantiez changed the title ~~Fred/doc~~ Update documentation Jul 6, 2023

FredericWantiez added 3 commits July 6, 2023 22:14

Add details to doc

837847e

Fix source

3648f4c

Typo

f7033db

FredericWantiez force-pushed the fred/doc branch from 433d2c1 to f7033db Compare July 6, 2023 21:14

Update SSM Interface

673e670

github-actions bot reviewed Jul 12, 2023

View reviewed changes

examples/smc.jl Outdated Show resolved Hide resolved

examples/smc.jl Outdated Show resolved Hide resolved

FredericWantiez added 2 commits July 19, 2023 22:41

Fix linearize bug

8732544

Interface

6394704

github-actions bot reviewed Jul 20, 2023

View reviewed changes

src/SSMProblems.jl Outdated Show resolved Hide resolved

FredericWantiez added 2 commits July 20, 2023 21:05

Format

f7bdf6a

Trying things

573cab3

yebai mentioned this pull request Jul 28, 2023

Change name of AbstractParticle #14

Merged

FredericWantiez added 5 commits July 29, 2023 18:50

Fix transition!!

71c74cd

Format

d603f56

Fix doc

a3af900

Helper

d40c277

Format

8f64bf4

yebai requested changes Aug 4, 2023

View reviewed changes

yebai reviewed Aug 4, 2023

View reviewed changes

src/SSMProblems.jl Outdated Show resolved Hide resolved

yebai reviewed Aug 4, 2023

View reviewed changes

docs/src/index.md Outdated Show resolved Hide resolved

yebai reviewed Aug 4, 2023

View reviewed changes

src/SSMProblems.jl Outdated Show resolved Hide resolved

devmotion reviewed Aug 4, 2023

View reviewed changes

Forget about particles

593d305

yebai reviewed Sep 8, 2023

View reviewed changes

FredericWantiez added 2 commits September 13, 2023 21:51

Optional timestep

0a21829

Add utils

18ab51e

Apply suggestions from code review

5bcb4f1

Co-authored-by: David Widmann <devmotion@users.noreply.github.com>

yebai reviewed Sep 13, 2023

View reviewed changes

Utils module

09f1687

yebai merged commit 4153da3 into main Sep 13, 2023
3 checks passed

yebai deleted the fred/doc branch September 13, 2023 22:30

yebai mentioned this pull request Sep 13, 2023

A bit confused regarding the interface #9

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Update documentation #16

Update documentation #16

FredericWantiez commented Jul 6, 2023

FredericWantiez commented Jul 24, 2023

yebai commented Jul 28, 2023

yebai commented Aug 3, 2023

yebai left a comment •

edited

Loading

yebai Aug 4, 2023

yebai Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

devmotion Aug 4, 2023

yebai Sep 8, 2023

yebai Sep 8, 2023

yebai Sep 8, 2023

yebai Sep 13, 2023

FredericWantiez Sep 13, 2023

yebai Sep 13, 2023

	function emission_logdensity(step, model::Model, particle::AbstractParticle)
	function emission_logdensity(step, model::Model, particle::AbstractParticle, observations)

	ancestors = resample(rng, logweigths)
	ancestors = resample(rng, logweigths)
	logweights = zero(logweights)

	return all(map(particle -> isdone(t, model, particle), particles))
	return all(particle -> isdone(t, model, particle), particles)

	transition_logdensity(model, timestep, prev_state, next_state[, cache])
	transition_logdensity(model, timestep, prev_state, current_state[, cache])

Update documentation #16

Update documentation #16

Conversation

FredericWantiez commented Jul 6, 2023

FredericWantiez commented Jul 24, 2023

yebai commented Jul 28, 2023

yebai commented Aug 3, 2023

yebai left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yebai left a comment •

edited

Loading