Merge pull request #7 from MurrellGroup/elements_and_batching

Adding some States->Elements and back

Author: murrellb

Project.toml

@@ -5,6 +5,7 @@ version = "0.1.2"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
 ForwardBackward = "e879419d-bb0f-4252-adee-d266c51ac92d"
 LogExpFunctions = "2ab3a3ac-af41-5b50-aa03-7779005ae688"
 Manifolds = "1cead3c2-87b3-11e9-0ccd-23c62b72b94e"
@@ -14,16 +15,11 @@ StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 
 [compat]
 Adapt = "4.1.1"
+FillArrays = "1.13.0"
 ForwardBackward = "0.1.0"
 LogExpFunctions = "0.3.29"
 Manifolds = "0.10.12"
 NNlib = "0.9.27"
 OneHotArrays = "0.2.6"
 StatsBase = "0.34.4"
 julia = "1.9"
-
-[extras]
-Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-
-[targets]
-test = ["Test"]

src/Flowfusion.jl

@@ -17,9 +17,6 @@ Later:
 =#
 
 
-
-
-
 module Flowfusion
 
 using ForwardBackward, OneHotArrays, Adapt, Manifolds, NNlib, LogExpFunctions
@@ -31,6 +28,8 @@ include("loss.jl")
 include("processes.jl")
 include("doob.jl")
 
+include("batching.jl")
+
 export
     #Processes not in ForwardBackward.jl
     InterpolatingDiscreteFlow,

src/batching.jl

@@ -0,0 +1,43 @@
+#This will be in Flowfusion:
+element(state, seqindex) = selectdim(state, ndims(state), seqindex:seqindex)
+element(state, seqindex, batchindex) = element(selectdim(state, ndims(state), batchindex), seqindex)
+element(S::MaskedState, inds...) = element(S.S, inds...)
+element(S::ContinuousState, inds...) = ContinuousState(element(S.state, inds...))
+element(S::ManifoldState, inds...) = ManifoldState(S.M, element(S.state, inds...))
+element(S::DiscreteState, inds...) = DiscreteState(S.K, element(S.state, inds...))
+element(S::Tuple{Vararg{Flowfusion.UState}}, inds...) = element.(S, inds...)
+
+
+#Create a "zero" state appropriate for the type. Tricky for manifolds, but we just want rotations working for now I think.
+zerostate(element::T, expandsize...) where T <: ContinuousState = T(similar(tensor(element), size(tensor(element))[1:end-1]..., expandsize...) .= 0)
+zerostate(element::DiscreteState{<:AbstractArray{<:Signed}}, expandsize...) = DiscreteState(element.K,similar(tensor(element), size(tensor(element))[1:end-1]..., expandsize...) .= element.K)
+zerostate(element::DiscreteState, expandsize...) = Flowfusion.onehot(DiscreteState(element.K,zeros(Int,expandsize...) .= element.K))
+function zerostate(element::T, expandsize...) where T <: Union{ManifoldState{<:Rotations},ManifoldState{<:SpecialOrthogonal}}
+    newtensor = similar(tensor(element), size(tensor(element))[1:end-1]..., expandsize...) .= 0
+    for i in 1:manifold_dimension(element.M)
+        selectdim(selectdim(newtensor, 1,i),1,i) .= 1
+    end
+    return ManifoldState(element.M, eachslice(newtensor, dims=ntuple(i -> 2+i, length(expandsize))))
+end
+#Pls test this general version with other manifolds? Not sure this will handle the various underlying representations
+function zerostate(element::ManifoldState, expandsize...)
+    newtensor = similar(tensor(element), size(tensor(element))[1:end-1]..., expandsize...) .= 0
+    return ManifoldState(element.M, eachslice(newtensor, dims=ntuple(i -> 2+i, length(expandsize))))
+end
+
+#In general, these will be different lengths, so we use an array of arrays as input.
+#Doesn't work for onehot states yet.
+function regroup(elarray::AbstractArray{<:AbstractArray})
+    example_tuple = elarray[1][1]
+    maxlen = maximum(length.(elarray))
+    b = length(elarray)
+    newstates = [zerostate(example_tuple[i],maxlen,b) for i in 1:length(example_tuple)]
+    for i in 1:b
+        for j in 1:length(elarray[i])
+            for k in 1:length(example_tuple)
+                element(tensor(newstates[k]),j,i) .= tensor(elarray[i][j][k])
+            end
+        end
+    end
+    return Tuple(newstates)
+end

test/Project.toml

@@ -0,0 +1,4 @@
+[deps]
+ForwardBackward = "e879419d-bb0f-4252-adee-d266c51ac92d"
+Manifolds = "1cead3c2-87b3-11e9-0ccd-23c62b72b94e"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

test/runtests.jl

@@ -44,7 +44,7 @@ using ForwardBackward
 
     @testset "Bridge, step" begin
 
-        siz = (5,6,7)
+        siz = (5,6)
         XC() = ContinuousState(randn(5, siz...))
         XD() = DiscreteState(5, rand(1:5, siz...))
         MT = Torus(2)