Fix tests

Author: avik-pal

Project.toml

@@ -1,7 +1,7 @@
 name = "DeepEquilibriumNetworks"
 uuid = "6748aba7-0e9b-415e-a410-ae3cc0ecb334"
 authors = ["Avik Pal <avikpal@mit.edu>"]
-version = "1.3.0"
+version = "1.4.0"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
@@ -36,7 +36,7 @@ NonlinearSolve = "2"
 OrdinaryDiffEq = "6"
 Reexport = "1"
 SciMLBase = "2"
-SciMLSensitivity = "7"
+SciMLSensitivity = "7.43"
 Setfield = "1"
 Static = "0.6, 0.7, 0.8"
 SteadyStateDiffEq = "1.16"

README.md

@@ -24,30 +24,31 @@ Pkg.add("DeepEquilibriumNetworks")
 ## Quickstart
 
 ```julia
-import DeepEquilibriumNetworks as DEQs
-import Lux
-import Random
-import Zygote
+using DeepEquilibriumNetworks, Lux, Random, Zygote
+# using LuxCUDA, LuxAMDGPU ## Install and Load for GPU Support
 
 seed = 0
 rng = Random.default_rng()
 Random.seed!(rng, seed)
 
-model = Lux.Chain(Lux.Dense(2, 2),
-    DEQs.DeepEquilibriumNetwork(Lux.Parallel(+,
-            Lux.Dense(2, 2; use_bias=false),
-            Lux.Dense(2, 2; use_bias=false)),
-        DEQs.ContinuousDEQSolver(;
-            abstol=0.1f0,
-            reltol=0.1f0,
-            abstol_termination=0.1f0,
-            reltol_termination=0.1f0)))
-
-ps, st = gpu.(Lux.setup(rng, model))
-x = gpu(rand(rng, Float32, 2, 1))
-y = gpu(rand(rng, Float32, 2, 1))
-
-gs = Zygote.gradient(p -> sum(abs2, model(x, p, st)[1][1] .- y), ps)[1]
+model = Chain(Dense(2 => 2),
+    DeepEquilibriumNetwork(Parallel(+,
+            Dense(2 => 2; use_bias=false),
+            Dense(2 => 2; use_bias=false)),
+        ContinuousDEQSolver(; abstol=0.1f0, reltol=0.1f0, abstol_termination=0.1f0,
+            reltol_termination=0.1f0);
+        save_everystep=true))
+
+gdev = gpu_device()
+cdev = cpu_device()
+
+ps, st = Lux.setup(rng, model) |> gdev
+x = rand(rng, Float32, 2, 1) |> gdev
+y = rand(rng, Float32, 2, 1) |> gdev
+
+model(x, ps, st)
+
+gs = only(Zygote.gradient(p -> sum(abs2, first(first(model(x, p, st))) .- y), ps))
 ```
 
 ## Citation

docs/src/index.md

@@ -2,7 +2,7 @@
 
 DeepEquilibriumNetworks.jl is a framework built on top of
 [DifferentialEquations.jl](https://docs.sciml.ai/DiffEqDocs/stable/) and
-[Lux.jl](https://docs.sciml.ai/Lux/stable/), enabling the efficient training and inference for
+[Lux.jl](https://lux.csail.mit.edu/), enabling the efficient training and inference for
 Deep Equilibrium Networks (Infinitely Deep Neural Networks).
 
 ## Installation
@@ -17,30 +17,30 @@ Pkg.add("DeepEquilibriumNetworks")
 ## Quick-start
 
 ```julia
-import DeepEquilibriumNetworks as DEQs
-import Lux
-import Random
-import Zygote
+using DeepEquilibriumNetworks, Lux, Random, Zygote
+# using LuxCUDA, LuxAMDGPU ## Install and Load for GPU Support
 
 seed = 0
 rng = Random.default_rng()
 Random.seed!(rng, seed)
+model = Chain(Dense(2 => 2),
+    DeepEquilibriumNetwork(Parallel(+,
+            Dense(2 => 2; use_bias=false),
+            Dense(2 => 2; use_bias=false)),
+        ContinuousDEQSolver(; abstol=0.1f0, reltol=0.1f0, abstol_termination=0.1f0,
+            reltol_termination=0.1f0);
+        save_everystep=true))
 
-model = Lux.Chain(Lux.Dense(2, 2),
-    DEQs.DeepEquilibriumNetwork(Lux.Parallel(+,
-            Lux.Dense(2, 2; use_bias=false),
-            Lux.Dense(2, 2; use_bias=false)),
-        DEQs.ContinuousDEQSolver(;
-            abstol=0.1f0,
-            reltol=0.1f0,
-            abstol_termination=0.1f0,
-            reltol_termination=0.1f0)))
-
-ps, st = gpu.(Lux.setup(rng, model))
-x = gpu(rand(rng, Float32, 2, 1))
-y = gpu(rand(rng, Float32, 2, 1))
-
-gs = Zygote.gradient(p -> sum(abs2, model(x, p, st)[1][1] .- y), ps)[1]
+gdev = gpu_device()
+cdev = cpu_device()
+
+ps, st = Lux.setup(rng, model) |> gdev
+x = rand(rng, Float32, 2, 1) |> gdev
+y = rand(rng, Float32, 2, 1) |> gdev
+
+model(x, ps, st)
+
+gs = only(Zygote.gradient(p -> sum(abs2, first(first(model(x, p, st))) .- y), ps))
 ```
 
 ## Citation

src/DeepEquilibriumNetworks.jl

@@ -16,6 +16,7 @@ import ChainRulesCore as CRC
 import ConcreteStructs: @concrete
 
 const DEQs = DeepEquilibriumNetworks
+const ∂∅ = CRC.NoTangent()
 
 ## FIXME: Uses of nothing was removed in Lux 0.5 with a deprecation. It was not updated
 ##        here
@@ -33,6 +34,13 @@ include("layers/evaluate.jl")
 
 include("chainrules.jl")
 
+# Start of Weird Patches
+# Honestly no clue why this is needed! -- probably a whacky fix which shouldn't be ever
+# needed.
+ZygoteRules.gradtuple1(::NamedTuple{()}) = (nothing, nothing, nothing, nothing, nothing)
+ZygoteRules.gradtuple1(x::NamedTuple) = collect(values(x))
+# End of Weird Patches
+
 # Useful Shorthand
 export DEQs
 

src/chainrules.jl

@@ -1,25 +1,27 @@
+__backing(Δ::CRC.Tangent) = __backing(CRC.backing(Δ))
+__backing(Δ::Tuple) = __backing.(Δ)
+__backing(Δ::NamedTuple{F}) where {F} = NamedTuple{F}(__backing(values(Δ)))
+__backing(Δ) = Δ
+
 function CRC.rrule(::Type{<:DeepEquilibriumSolution}, z_star::T, u0::T, residual::T,
     jacobian_loss::R, nfe::Int) where {T, R <: AbstractFloat}
     function deep_equilibrium_solution_pullback(dsol)
-        return (CRC.NoTangent(), dsol.z_star, dsol.u0, dsol.residual, dsol.jacobian_loss,
-            dsol.nfe)
+        return (∂∅, dsol.z_star, dsol.u0, dsol.residual, dsol.jacobian_loss, dsol.nfe)
     end
     return (DeepEquilibriumSolution(z_star, u0, residual, jacobian_loss, nfe),
         deep_equilibrium_solution_pullback)
 end
 
 function _safe_getfield(x::NamedTuple{fields}, field) where {fields}
-    return field ∈ fields ? getfield(x, field) : CRC.NoTangent()
+    return field ∈ fields ? getfield(x, field) : ∂∅
 end
 
 function CRC.rrule(::Type{T}, args...) where {T <: NamedTuple}
     y = T(args...)
     function nt_pullback(dy)
         fields = fieldnames(T)
-        if dy isa CRC.Tangent
-            dy = CRC.backing(dy)
-        end
-        return (CRC.NoTangent(), _safe_getfield.((dy,), fields)...)
+        dy isa CRC.Tangent && (dy = CRC.backing(dy))
+        return (∂∅, _safe_getfield.((dy,), fields)...)
     end
     return y, nt_pullback
 end
@@ -28,20 +30,20 @@ function CRC.rrule(::typeof(Setfield.set), obj, l::Setfield.PropertyLens{field},
     val) where {field}
     res = Setfield.set(obj, l, val)
     function setfield_pullback(Δres)
-        if Δres isa CRC.Tangent
-            Δres = CRC.backing(Δres)
-        end
-        Δobj = Setfield.set(obj, l, CRC.NoTangent())
-        return (CRC.NoTangent(), Δobj, CRC.NoTangent(), getfield(Δres, field))
+        Δres = __backing(Δres)
+        Δobj = Setfield.set(obj, l, ∂∅)
+        return (∂∅, Δobj, ∂∅, getfield(Δres, field))
     end
     return res, setfield_pullback
 end
 
-function CRC.rrule(::typeof(_construct_problem), deq::AbstractDEQs, dudt, z, ps, x)
+function CRC.rrule(::typeof(_construct_problem), deq::AbstractDEQs, dudt, z,
+    ps::NamedTuple{F}, x) where {F}
     prob = _construct_problem(deq, dudt, z, ps, x)
     function ∇_construct_problem(Δ)
-        return (CRC.NoTangent(), CRC.NoTangent(), CRC.NoTangent(), Δ.u0,
-            (; model = Δ.p.ps), Δ.p.x)
+        Δ = __backing(Δ)
+        nograds = NamedTuple{F}(ntuple(i -> ∂∅, length(F)))
+        return (∂∅, ∂∅, ∂∅, Δ.u0, merge(nograds, (; model=Δ.p.ps)), Δ.p.x)
     end
     return prob, ∇_construct_problem
 end

src/layers/evaluate.jl

@@ -14,8 +14,7 @@ end
 @inline _postprocess_output(_, z_star) = z_star
 
 @inline function _construct_problem(::AbstractDEQs, dudt, z, ps, x)
-    return SteadyStateProblem(ODEFunction{false}(dudt), z,
-        NamedTuple{(:ps, :x)}((ps.model, x)))
+    return SteadyStateProblem(ODEFunction{false}(dudt), z, (; ps=ps.model, x))
 end
 
 @inline _fix_solution_output(_, x) = x
@@ -48,7 +47,9 @@ function (deq::AbstractDEQs)(x::AbstractArray, ps, st::NamedTuple, ::Val{false})
 
     prob = _construct_problem(deq, dudt, z, ps, x)
     sol = solve(prob, deq.solver; deq.sensealg, deq.kwargs...)
-    z_star = sol.u
+    _z_star = sol.u
+    # Handle Neural ODEs
+    z_star = _z_star isa Vector{<:AbstractArray} ? last(_z_star) : _z_star
 
     if _jacobian_regularization(deq)
         rng = Lux.replicate(st.rng)

src/layers/mdeq.jl

@@ -282,8 +282,7 @@ end
 """
     MultiScaleNeuralODE(main_layers::Tuple, mapping_layers::Matrix,
                         post_fuse_layer::Union{Nothing,Tuple}, solver, scales;
-                        sensealg=InterpolatingAdjoint(; autojacvec=ZygoteVJP()),
-                        kwargs...)
+                        sensealg=GaussAdjoint(; autojacvec=ZygoteVJP()), kwargs...)
 
 Multiscale Neural ODE with Input Injection.
 
@@ -334,7 +333,7 @@ See also: [`DeepEquilibriumNetwork`](@ref), [`SkipDeepEquilibriumNetwork`](@ref)
 """
 function MultiScaleNeuralODE(main_layers::Tuple, mapping_layers::Matrix,
     post_fuse_layer::Union{Nothing, Tuple}, solver, scales::NTuple{N, NTuple{L, Int64}};
-    sensealg=InterpolatingAdjoint(; autojacvec=ZygoteVJP()), kwargs...) where {N, L}
+    sensealg=GaussAdjoint(; autojacvec=ZygoteVJP()), kwargs...) where {N, L}
     l1 = Parallel(nothing, main_layers...)
     l2 = BranchLayer(Parallel.(+, map(x -> tuple(x...), eachrow(mapping_layers))...)...)
 
@@ -357,7 +356,7 @@ function _get_initial_condition(deq::MultiScaleNeuralODE, x, ps, st)
 end
 
 @inline function _construct_problem(::MultiScaleNeuralODE, dudt, z, ps, x)
-    return ODEProblem(ODEFunction{false}(dudt), z, (0.0f0, 1.0f0), ps.model)
+    return ODEProblem(ODEFunction{false}(dudt), z, (0.0f0, 1.0f0), (; ps=ps.model, x))
 end
 
 @inline _fix_solution_output(::MultiScaleNeuralODE, x) = x[end]