fix zygote error (#579)

Author: CarloLucibello

GNNGraphs/src/GNNGraphs.jl

@@ -8,7 +8,7 @@ import NearestNeighbors
 import NNlib
 import StatsBase
 import KrylovKit
-using ChainRulesCore
+import ChainRulesCore as CRC
 using LinearAlgebra, Random, Statistics
 import MLUtils
 using MLUtils: getobs, numobs, ones_like, zeros_like, chunk, batch, rand_like

GNNGraphs/src/chainrules.jl

@@ -1,15 +1,16 @@
 # Taken from https://github.com/JuliaDiff/ChainRules.jl/pull/648
 # Remove when merged
 
-function ChainRulesCore.rrule(::Type{T}, ps::Pair...) where {T<:Dict}
+function CRC.rrule(::Type{T}, ps::Pair...) where {T<:Dict}
     ks = map(first, ps)
-    project_ks, project_vs = map(ProjectTo, ks), map(ProjectTo∘last, ps)
+    project_ks, project_vs = map(CRC.ProjectTo, ks), map(CRC.ProjectTo ∘ last, ps)
     function Dict_pullback(ȳ)
+        dy = CRC.unthunk(ȳ)
         dps = map(ks, project_ks, project_vs) do k, proj_k, proj_v
-            dk, dv = proj_k(getkey(ȳ, k, NoTangent())), proj_v(get(ȳ, k, NoTangent()))
-            Tangent{Pair{typeof(dk), typeof(dv)}}(first = dk, second = dv)
+            dk, dv = proj_k(getkey(dy, k, CRC.NoTangent())), proj_v(get(dy, k, CRC.NoTangent()))
+            CRC.Tangent{Pair{typeof(dk), typeof(dv)}}(first = dk, second = dv)
         end
-       return (NoTangent(), dps...)
+       return (CRC.NoTangent(), dps...)
     end
     return T(ps...), Dict_pullback
 end

GNNGraphs/src/convert.jl

@@ -78,7 +78,7 @@ function _findnz_idx(A)
     return s, t, nz
 end
 
-@non_differentiable _findnz_idx(A)
+CRC.@non_differentiable _findnz_idx(A)
 
 function to_coo(A::ADJMAT_T; dir = :out, num_nodes = nothing, weighted = true)
     s, t, nz = _findnz_idx(A)

GNNGraphs/src/gnnheterograph/gnnheterograph.jl

@@ -273,7 +273,7 @@ end
 
 # TODO this is not correct but Zygote cannot differentiate
 # through dictionary generation
-# @non_differentiable edge_type_subgraph(::Any...)
+# CRC.@non_differentiable edge_type_subgraph(::Any...)
 
 function _ntypes_from_edges(edge_ts::AbstractVector{<:EType})
     ntypes = Symbol[]
@@ -285,7 +285,7 @@ function _ntypes_from_edges(edge_ts::AbstractVector{<:EType})
     return ntypes
 end 
 
-@non_differentiable _ntypes_from_edges(::Any...)
+CRC.@non_differentiable _ntypes_from_edges(::Any...)
 
 function Base.getindex(g::GNNHeteroGraph, node_t::NType)
     return g.ndata[node_t]

GNNGraphs/src/query.jl

@@ -241,37 +241,46 @@ function Graphs.adjacency_matrix(g::GNNGraph{<:ADJMAT_T}, T::DataType = eltype(g
     return dir == :out ? A : A'
 end
 
-function ChainRulesCore.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
+function CRC.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
             dir = :out, weighted = true) where {G <: GNNGraph{<:ADJMAT_T}}
     A = adjacency_matrix(g, T; dir, weighted)
     if !weighted
         function adjacency_matrix_pullback_noweight(Δ)
-            return (NoTangent(), ZeroTangent(), NoTangent())  
+            return (CRC.NoTangent(), CRC.ZeroTangent(), CRC.NoTangent())  
         end
         return A, adjacency_matrix_pullback_noweight
     else
         function adjacency_matrix_pullback_weighted(Δ)
-            dg = Tangent{G}(; graph = Δ .* binarize(A))
-            return (NoTangent(), dg, NoTangent())  
+            dy = CRC.unthunk(Δ)
+            dg = CRC.Tangent{G}(; graph = dy .* binarize(dy))
+            return (CRC.NoTangent(), dg, CRC.NoTangent())  
         end
         return A, adjacency_matrix_pullback_weighted
     end
 end
 
-function ChainRulesCore.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
+function CRC.rrule(::typeof(adjacency_matrix), g::G, T::DataType; 
             dir = :out, weighted = true) where {G <: GNNGraph{<:COO_T}}
     A = adjacency_matrix(g, T; dir, weighted)
     w = get_edge_weight(g)
     if !weighted || w === nothing
         function adjacency_matrix_pullback_noweight(Δ)
-            return (NoTangent(), ZeroTangent(), NoTangent())  
+            return (CRC.NoTangent(), CRC.ZeroTangent(), CRC.NoTangent())  
         end
         return A, adjacency_matrix_pullback_noweight
     else
         function adjacency_matrix_pullback_weighted(Δ)
+            dy = CRC.unthunk(Δ)
             s, t = edge_index(g)
-            dg = Tangent{G}(; graph = (NoTangent(), NoTangent(), NNlib.gather(Δ, s, t)))
-            return (NoTangent(), dg, NoTangent())  
+            @show dy s t
+            #TODO using CRC.@thunk gives an error
+            #TODO use gather when https://github.com/FluxML/NNlib.jl/issues/625 is fixed
+            dw = zeros_like(w)
+            idx = CartesianIndex.(s, t) #TODO remove when https://github.com/FluxML/NNlib.jl/issues/626 is fixed
+            NNlib.gather!(dw, dy, idx)
+            @show dw
+            dg = CRC.Tangent{G}(; graph = (CRC.NoTangent(), CRC.NoTangent(), dw))
+            return (CRC.NoTangent(), dg, CRC.NoTangent())  
         end
         return A, adjacency_matrix_pullback_weighted
     end
@@ -378,34 +387,35 @@ function _degree(A::AbstractMatrix, T::Type, dir::Symbol, edge_weight::Bool, num
            vec(sum(A, dims = 1)) .+ vec(sum(A, dims = 2))
 end
 
-function ChainRulesCore.rrule(::typeof(_degree), graph, T, dir, edge_weight::Nothing, num_nodes)
+function CRC.rrule(::typeof(_degree), graph, T, dir, edge_weight::Nothing, num_nodes)
     degs = _degree(graph, T, dir, edge_weight, num_nodes)
     function _degree_pullback(Δ)
-        return (NoTangent(), NoTangent(), NoTangent(), NoTangent(), NoTangent(), NoTangent())
+        return ntuple(i -> (CRC.NoTangent(),), 6)
     end
     return degs, _degree_pullback
 end
 
-function ChainRulesCore.rrule(::typeof(_degree), A::ADJMAT_T, T, dir, edge_weight::Bool, num_nodes)
+function CRC.rrule(::typeof(_degree), A::ADJMAT_T, T, dir, edge_weight::Bool, num_nodes)
     degs = _degree(A, T, dir, edge_weight, num_nodes)
     if edge_weight === false
         function _degree_pullback_noweights(Δ)
-            return (NoTangent(), NoTangent(), NoTangent(), NoTangent(), NoTangent(), NoTangent())
+            return ntuple(i -> (CRC.NoTangent(),), 6)
         end
         return degs, _degree_pullback_noweights
     else
         function _degree_pullback_weights(Δ)
+            dy = CRC.unthunk(Δ)
             # We propagate the gradient only to the non-zero elements
             # of the adjacency matrix.
             bA = binarize(A)
             if dir == :in
-                dA = bA .* Δ'
+                dA = bA .* dy'
             elseif dir == :out
-                dA = Δ .* bA
+                dA = dy .* bA
             else # dir == :both
-                dA = Δ .* bA + Δ' .* bA
+                dA = dy .* bA + dy' .* bA
             end
-            return (NoTangent(), dA, NoTangent(), NoTangent(), NoTangent(), NoTangent())
+            return (CRC.NoTangent(), dA, CRC.NoTangent(), CRC.NoTangent(), CRC.NoTangent(), CRC.NoTangent())
         end
         return degs, _degree_pullback_weights
     end
@@ -452,7 +462,7 @@ function normalized_adjacency(g::GNNGraph, T::DataType = Float32;
         A = A + I
     end
     degs = vec(sum(A; dims = 2))
-    ChainRulesCore.ignore_derivatives() do
+    CRC.ignore_derivatives() do
         @assert all(!iszero, degs) "Graph contains isolated nodes, cannot compute `normalized_adjacency`."
     end
     inv_sqrtD = Diagonal(inv.(sqrt.(degs)))
@@ -609,12 +619,12 @@ function laplacian_lambda_max(g::GNNGraph, T::DataType = Float32;
     end
 end
 
-@non_differentiable edge_index(x...)
-@non_differentiable adjacency_list(x...)
-@non_differentiable graph_indicator(x...)
-@non_differentiable has_multi_edges(x...)
-@non_differentiable Graphs.has_self_loops(x...)
-@non_differentiable is_bidirected(x...)
-@non_differentiable normalized_adjacency(x...) # TODO remove this in the future
-@non_differentiable normalized_laplacian(x...) # TODO remove this in the future
-@non_differentiable scaled_laplacian(x...) # TODO remove this in the future
+CRC.@non_differentiable edge_index(x...)
+CRC.@non_differentiable adjacency_list(x...)
+CRC.@non_differentiable graph_indicator(x...)
+CRC.@non_differentiable has_multi_edges(x...)
+CRC.@non_differentiable Graphs.has_self_loops(x...)
+CRC.@non_differentiable is_bidirected(x...)
+CRC.@non_differentiable normalized_adjacency(x...) # TODO remove this in the future
+CRC.@non_differentiable normalized_laplacian(x...) # TODO remove this in the future
+CRC.@non_differentiable scaled_laplacian(x...) # TODO remove this in the future

GNNGraphs/src/transform.jl

@@ -808,8 +808,8 @@ function _unbatch_edgemasks(s, t, num_graphs, cumnum_nodes)
     return edgemasks
 end
 
-@non_differentiable _unbatch_nodemasks(::Any...)
-@non_differentiable _unbatch_edgemasks(::Any...)
+CRC.@non_differentiable _unbatch_nodemasks(::Any...)
+CRC.@non_differentiable _unbatch_edgemasks(::Any...)
 
 """
     getgraph(g::GNNGraph, i; nmap=false)
@@ -998,10 +998,10 @@ dense_zeros_like(x, sz = size(x)) = dense_zeros_like(x, eltype(x), sz)
 # """
 ci2t(ci::AbstractVector{<:CartesianIndex}, dims) = ntuple(i -> map(x -> x[i], ci), dims)
 
-@non_differentiable negative_sample(x...)
-@non_differentiable add_self_loops(x...)     # TODO this is wrong, since g carries feature arrays, needs rrule
-@non_differentiable remove_self_loops(x...)  # TODO this is wrong, since g carries feature arrays, needs rrule
-@non_differentiable dense_zeros_like(x...)
+CRC.@non_differentiable negative_sample(x...)
+CRC.@non_differentiable add_self_loops(x...)     # TODO this is wrong, since g carries feature arrays, needs rrule
+CRC.@non_differentiable remove_self_loops(x...)  # TODO this is wrong, since g carries feature arrays, needs rrule
+CRC.@non_differentiable dense_zeros_like(x...)
 
 """
     ppr_diffusion(g::GNNGraph{<:COO_T}, alpha =0.85f0) -> GNNGraph

GNNGraphs/src/utils.jl

@@ -292,9 +292,9 @@ end
 
 binarize(x) = map(>(0), x)
 
-@non_differentiable binarize(x...)
-@non_differentiable edge_encoding(x...)
-@non_differentiable edge_decoding(x...)
+CRC.@non_differentiable binarize(x...)
+CRC.@non_differentiable edge_encoding(x...)
+CRC.@non_differentiable edge_decoding(x...)
 
 ### PRINTING #####
 
@@ -330,11 +330,11 @@ function dims2string(d)
     join(map(string, d), '×')
 end
 
-@non_differentiable normalize_graphdata(::NamedTuple{(), Tuple{}})
-@non_differentiable normalize_graphdata(::Nothing)
+CRC.@non_differentiable normalize_graphdata(::NamedTuple{(), Tuple{}})
+CRC.@non_differentiable normalize_graphdata(::Nothing)
 
 iscuarray(x::AbstractArray) = false 
-@non_differentiable iscuarray(::Any)
+CRC.@non_differentiable iscuarray(::Any)
 
 
 @doc raw"""

GNNGraphs/test/Project.toml

@@ -1,5 +1,4 @@
 [deps]
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 FiniteDifferences = "26cc04aa-876d-5657-8c51-4c34ba976000"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 GNNGraphs = "aed8fd31-079b-4b5a-b342-a13352159b8c"
@@ -9,6 +8,7 @@ LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
+NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
@@ -20,7 +20,6 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 TestItems = "1c621080-faea-4a02-84b6-bbd5e436b8fe"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
-cuDNN = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 
 [compat]
 GPUArraysCore = "0.1"

GraphNeuralNetworks/examples/Project.toml

@@ -1,19 +1,16 @@
 [deps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
-DiffEqFlux = "aae7a2af-3d4f-5e19-a356-7da93b79d9d0"
 DifferentialEquations = "0c46a032-eb83-5123-abaf-570d42b7fbaa"
 Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
 GraphNeuralNetworks = "cffab07f-9bc2-4db1-8861-388f63bf7694"
 Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd"
-NNlibCUDA = "a00861dc-f156-4864-bf3c-e6376f28a68d"
 
 [compat]
-DiffEqFlux = "2"
-Flux = "0.13"
-GraphNeuralNetworks = "0.6"
+Flux = "0.16"
+GraphNeuralNetworks = "1"
 Graphs = "1"
 MLDatasets = "0.7"
-julia = "1.9"
+julia = "1.10"