From 44c72191ff695d06f9a25e02b90f31885f05ba3d Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Fri, 24 Jul 2026 13:22:41 +0200
Subject: [PATCH 1/7] Learnable epsilon

---
 GNNlib/src/layers/conv.jl | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/GNNlib/src/layers/conv.jl b/GNNlib/src/layers/conv.jl
index ab40f13f2..6b7907f1d 100644
--- a/GNNlib/src/layers/conv.jl
+++ b/GNNlib/src/layers/conv.jl
@@ -252,7 +252,8 @@ function gin_conv(l, g::AbstractGNNGraph, x)
     xj, xi = expand_srcdst(g, x) 
  
     m = propagate(copy_xj, g, l.aggr, xj = xj)
-    return l.nn((1 .+ ofeltype(xi, l.ϵ)) .* xi .+ m)
+    eps = convert.(float(eltype(xi)), l.eps)
+    return l.nn((1 .+ eps) .* xi .+ m)
 end
 
 ####################### NNConv ######################################

From f86744fba9d714def02b28ad85a6515f242f1d77 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Fri, 24 Jul 2026 13:23:09 +0200
Subject: [PATCH 2/7] Learnable epsilon and flag train

---
 GNNLux/src/layers/conv.jl  | 36 ++++++++++++++++++++++++++++--------
 GNNLux/test/layers/conv.jl | 11 ++++++++++-
 2 files changed, 38 insertions(+), 9 deletions(-)

diff --git a/GNNLux/src/layers/conv.jl b/GNNLux/src/layers/conv.jl
index f67e54381..199494926 100644
--- a/GNNLux/src/layers/conv.jl
+++ b/GNNLux/src/layers/conv.jl
@@ -1287,7 +1287,7 @@ function Base.show(io::IO, l::GatedGraphConv)
 end
 
 @doc raw"""
-    GINConv(f, ϵ; aggr=+)
+    GINConv(f, [eps]; aggr=+, train_eps=false)
 
 Graph Isomorphism convolutional layer from paper [How Powerful are Graph Neural Networks?](https://arxiv.org/pdf/1810.00826.pdf).
 
@@ -1300,7 +1300,9 @@ where ``f_\Theta`` typically denotes a learnable function, e.g. a linear layer o
 # Arguments
 
 - `f`: A (possibly learnable) function acting on node features. 
-- `ϵ`: Weighting factor.
+- `eps`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
+- `train_eps`: If `true`, ``\epsilon`` is trainable. Default `false`.
+- `aggr`: Neighborhood aggregation function. Default `+`.
 
 # Examples:
 
@@ -1322,7 +1324,7 @@ x = randn(rng, Float32, in_channel, g.num_nodes)
 nn = Dense(in_channel, out_channel)
 
 # create layer
-l = GINConv(nn, 0.01f0, aggr = mean)
+l = GINConv(nn; eps = 0.01f0, train_eps = true, aggr = mean)
 
 # setup layer
 ps, st = LuxCore.setup(rng, l)
@@ -1333,15 +1335,31 @@ y, st = l(g, x, ps, st)       # size:  out_channel × num_nodes
 """
 @concrete struct GINConv <: GNNContainerLayer{(:nn,)}
     nn <: AbstractLuxLayer
-    ϵ <: Real
+    eps
     aggr
+    train_eps::Bool
 end
 
-GINConv(nn, ϵ; aggr = +) = GINConv(nn, ϵ, aggr)
+GINConv(nn, eps::Real; aggr = +, train_eps::Bool = false) =
+    GINConv(nn, train_eps ? [eps] : eps, aggr, train_eps)
+
+GINConv(nn, eps::Real, aggr) = GINConv(nn, eps; aggr)
+
+GINConv(nn; eps::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
+    GINConv(nn, eps; aggr, train_eps)
+
+function LuxCore.initialparameters(rng::AbstractRNG, l::GINConv)
+    nn = LuxCore.initialparameters(rng, l.nn)
+    return l.train_eps ? (; nn, eps = l.eps) : (; nn)
+end
+
+LuxCore.parameterlength(l::GINConv) = parameterlength(l.nn) + l.train_eps
+LuxCore.statelength(l::GINConv) = statelength(l.nn)
 
 function (l::GINConv)(g, x, ps, st)
     nn = StatefulLuxLayer{true}(l.nn, ps.nn, st.nn)
-    m = (; nn, l.ϵ, l.aggr)
+    eps = l.train_eps ? ps.eps : l.eps
+    m = (; nn, eps, l.aggr)
     y = GNNlib.gin_conv(m, g, x)
     stnew = (; nn = _getstate(nn))
     return y, stnew
@@ -1349,7 +1367,9 @@ end
 
 function Base.show(io::IO, l::GINConv)
     print(io, "GINConv($(l.nn)")
-    print(io, ", $(l.ϵ)")
+    print(io, ", eps=$(l.eps)")
+    l.train_eps && print(io, ", train_eps=true")
+    l.aggr == (+) || print(io, ", aggr=$(l.aggr)")
     print(io, ")")
 end
 
@@ -1879,4 +1899,4 @@ function (l::SAGEConv)(g, x, ps, st)
     m = (; ps.weight, bias = _getbias(ps), 
           l.σ, l.aggr)
     return GNNlib.sage_conv(m, g, x), st
-end
\ No newline at end of file
+end
diff --git a/GNNLux/test/layers/conv.jl b/GNNLux/test/layers/conv.jl
index 5bfb01825..959a0f79a 100644
--- a/GNNLux/test/layers/conv.jl
+++ b/GNNLux/test/layers/conv.jl
@@ -92,8 +92,17 @@
 
     @testset "GINConv" begin
         nn = Chain(Dense(in_dims => out_dims, tanh), Dense(out_dims => out_dims))
-        l = GINConv(nn, 0.5)
+        l = GINConv(nn; eps = 0.5)
         test_lux_layer(rng, l, g, x, sizey=(out_dims,g.num_nodes), container=true)
+        ps = LuxCore.initialparameters(rng, l)
+        @test !hasproperty(ps, :eps)
+
+        l = GINConv(nn; train_eps = true)
+        test_lux_layer(rng, l, g, x, sizey=(out_dims,g.num_nodes), container=true)
+        ps = LuxCore.initialparameters(rng, l)
+        @test ps.eps == [0.0f0]
+
+        @test GINConv(nn, 0.5).eps == 0.5
     end
 
     @testset "MEGNetConv" begin

From 1a3e313001ec32ada5a6ef3b25ac23bb23786df8 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Fri, 24 Jul 2026 13:23:28 +0200
Subject: [PATCH 3/7] Learnable epsilon and flag

---
 GraphNeuralNetworks/src/layers/conv.jl  | 27 +++++++++++++++++--------
 GraphNeuralNetworks/test/layers/conv.jl | 18 ++++++++++++++---
 2 files changed, 34 insertions(+), 11 deletions(-)

diff --git a/GraphNeuralNetworks/src/layers/conv.jl b/GraphNeuralNetworks/src/layers/conv.jl
index 10a340e90..4bd6cde76 100644
--- a/GraphNeuralNetworks/src/layers/conv.jl
+++ b/GraphNeuralNetworks/src/layers/conv.jl
@@ -590,7 +590,7 @@ function Base.show(io::IO, l::EdgeConv)
 end
 
 @doc raw"""
-    GINConv(f, ϵ; aggr=+)
+    GINConv(f, [eps]; aggr=+, train_eps=false)
 
 Graph Isomorphism convolutional layer from paper [How Powerful are Graph Neural Networks?](https://arxiv.org/pdf/1810.00826.pdf).
 
@@ -603,7 +603,9 @@ where ``f_\Theta`` typically denotes a learnable function, e.g. a linear layer o
 # Arguments
 
 - `f`: A (possibly learnable) function acting on node features.
-- `ϵ`: Weighting factor.
+- `eps`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
+- `train_eps`: If `true`, ``\epsilon`` is trainable. Default `false`.
+- `aggr`: Neighborhood aggregation function. Default `+`.
 
 # Examples:
 
@@ -619,28 +621,37 @@ g = GNNGraph(s, t)
 nn = Dense(in_channel, out_channel)
 
 # create layer
-l = GINConv(nn, 0.01f0, aggr = mean)
+l = GINConv(nn; eps = 0.01f0, train_eps = true, aggr = mean)
 
 # forward pass
 y = l(g, x)
 ```
 """
-struct GINConv{R <: Real, NN, A} <: GNNLayer
+struct GINConv{E, NN, A} <: GNNLayer
     nn::NN
-    ϵ::R
+    eps::E
     aggr::A
+    train_eps::Bool
 end
 
 Flux.@layer :expand GINConv
-Flux.trainable(l::GINConv) = (nn = l.nn,)
+Flux.trainable(l::GINConv) = l.train_eps ? (; l.nn, l.eps) : (; l.nn)
 
-GINConv(nn, ϵ; aggr = +) = GINConv(nn, ϵ, aggr)
+GINConv(nn, eps::Real; aggr = +, train_eps::Bool = false) =
+    GINConv(nn, train_eps ? [eps] : eps, aggr, train_eps)
+
+GINConv(nn, eps::Real, aggr) = GINConv(nn, eps; aggr)
+
+GINConv(nn; eps::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
+    GINConv(nn, eps; aggr, train_eps)
 
 (l::GINConv)(g, x) = GNNlib.gin_conv(l, g, x)
 
 function Base.show(io::IO, l::GINConv)
     print(io, "GINConv($(l.nn)")
-    print(io, ", $(l.ϵ)")
+    print(io, ", eps=$(l.eps)")
+    l.train_eps && print(io, ", train_eps=true")
+    l.aggr == (+) || print(io, ", aggr=$(l.aggr)")
     print(io, ")")
 end
 
diff --git a/GraphNeuralNetworks/test/layers/conv.jl b/GraphNeuralNetworks/test/layers/conv.jl
index 7585b817d..fafd68fcd 100644
--- a/GraphNeuralNetworks/test/layers/conv.jl
+++ b/GraphNeuralNetworks/test/layers/conv.jl
@@ -278,19 +278,31 @@ end
     using .TestModule
     nn = Dense(D_IN, D_OUT)
 
-    l = GINConv(nn, 0.01, aggr = mean)
+    l = GINConv(nn; eps = 0.01, aggr = mean)
     for g in TEST_GRAPHS
         @test size(l(g, g.x)) == (D_OUT, g.num_nodes)
         test_gradients(l, g, g.x, rtol = RTOL_HIGH)
     end
+    @test keys(Flux.trainable(l)) == (:nn,)
 
-    @test !in(:eps, Flux.trainable(l))
+    l = GINConv(Dense(D_IN, D_OUT); train_eps = true, aggr = mean)
+    @test keys(Flux.trainable(l)) == (:nn, :eps)
+    @test l.eps == [0.0f0]
+    for g in TEST_GRAPHS
+        @test size(l(g, g.x)) == (D_OUT, g.num_nodes)
+    end
+    g = first(TEST_GRAPHS)
+    grad = Flux.gradient(l -> mean(l(g, g.x)), l)[1]
+    @test grad.eps isa AbstractVector
+    @test length(grad.eps) == 1
+
+    @test GINConv(nn, 0.5).eps == 0.5
 end
 
 @testitem "GINConv GPU" setup=[TolSnippet, TestModule] tags=[:gpu] begin
     using .TestModule
     nn = Dense(D_IN, D_OUT)
-    l = GINConv(nn, 0.01, aggr = mean)
+    l = GINConv(nn; train_eps = true, aggr = mean)
     for g in TEST_GRAPHS
         g.graph isa AbstractSparseMatrix && continue
         @test size(l(g, g.x)) == (D_OUT, g.num_nodes)

From c7b2b15edc286760995491b1e44b9b3591f4b318 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Fri, 24 Jul 2026 14:01:15 +0200
Subject: [PATCH 4/7] Change eps to unicode char

---
 GNNLux/src/layers/conv.jl               | 26 ++++++++++++-------------
 GNNLux/test/layers/conv.jl              |  8 ++++----
 GNNlib/src/layers/conv.jl               |  4 ++--
 GraphNeuralNetworks/src/layers/conv.jl  | 22 ++++++++++-----------
 GraphNeuralNetworks/test/layers/conv.jl | 12 ++++++------
 5 files changed, 36 insertions(+), 36 deletions(-)

diff --git a/GNNLux/src/layers/conv.jl b/GNNLux/src/layers/conv.jl
index 199494926..73a78f223 100644
--- a/GNNLux/src/layers/conv.jl
+++ b/GNNLux/src/layers/conv.jl
@@ -1287,7 +1287,7 @@ function Base.show(io::IO, l::GatedGraphConv)
 end
 
 @doc raw"""
-    GINConv(f, [eps]; aggr=+, train_eps=false)
+    GINConv(f, [ϵ]; aggr=+, train_eps=false)
 
 Graph Isomorphism convolutional layer from paper [How Powerful are Graph Neural Networks?](https://arxiv.org/pdf/1810.00826.pdf).
 
@@ -1300,7 +1300,7 @@ where ``f_\Theta`` typically denotes a learnable function, e.g. a linear layer o
 # Arguments
 
 - `f`: A (possibly learnable) function acting on node features. 
-- `eps`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
+- `ϵ`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
 - `train_eps`: If `true`, ``\epsilon`` is trainable. Default `false`.
 - `aggr`: Neighborhood aggregation function. Default `+`.
 
@@ -1324,7 +1324,7 @@ x = randn(rng, Float32, in_channel, g.num_nodes)
 nn = Dense(in_channel, out_channel)
 
 # create layer
-l = GINConv(nn; eps = 0.01f0, train_eps = true, aggr = mean)
+l = GINConv(nn; ϵ = 0.01f0, train_eps = true, aggr = mean)
 
 # setup layer
 ps, st = LuxCore.setup(rng, l)
@@ -1335,22 +1335,22 @@ y, st = l(g, x, ps, st)       # size:  out_channel × num_nodes
 """
 @concrete struct GINConv <: GNNContainerLayer{(:nn,)}
     nn <: AbstractLuxLayer
-    eps
+    ϵ
     aggr
     train_eps::Bool
 end
 
-GINConv(nn, eps::Real; aggr = +, train_eps::Bool = false) =
-    GINConv(nn, train_eps ? [eps] : eps, aggr, train_eps)
+GINConv(nn, ϵ::Real; aggr = +, train_eps::Bool = false) =
+    GINConv(nn, train_eps ? [ϵ] : ϵ, aggr, train_eps)
 
-GINConv(nn, eps::Real, aggr) = GINConv(nn, eps; aggr)
+GINConv(nn, ϵ::Real, aggr) = GINConv(nn, ϵ; aggr)
 
-GINConv(nn; eps::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
-    GINConv(nn, eps; aggr, train_eps)
+GINConv(nn; ϵ::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
+    GINConv(nn, ϵ; aggr, train_eps)
 
 function LuxCore.initialparameters(rng::AbstractRNG, l::GINConv)
     nn = LuxCore.initialparameters(rng, l.nn)
-    return l.train_eps ? (; nn, eps = l.eps) : (; nn)
+    return l.train_eps ? (; nn, ϵ = l.ϵ) : (; nn)
 end
 
 LuxCore.parameterlength(l::GINConv) = parameterlength(l.nn) + l.train_eps
@@ -1358,8 +1358,8 @@ LuxCore.statelength(l::GINConv) = statelength(l.nn)
 
 function (l::GINConv)(g, x, ps, st)
     nn = StatefulLuxLayer{true}(l.nn, ps.nn, st.nn)
-    eps = l.train_eps ? ps.eps : l.eps
-    m = (; nn, eps, l.aggr)
+    ϵ = l.train_eps ? ps.ϵ : l.ϵ
+    m = (; nn, ϵ, l.aggr)
     y = GNNlib.gin_conv(m, g, x)
     stnew = (; nn = _getstate(nn))
     return y, stnew
@@ -1367,7 +1367,7 @@ end
 
 function Base.show(io::IO, l::GINConv)
     print(io, "GINConv($(l.nn)")
-    print(io, ", eps=$(l.eps)")
+    print(io, ", ϵ=$(l.ϵ)")
     l.train_eps && print(io, ", train_eps=true")
     l.aggr == (+) || print(io, ", aggr=$(l.aggr)")
     print(io, ")")
diff --git a/GNNLux/test/layers/conv.jl b/GNNLux/test/layers/conv.jl
index 959a0f79a..d32c2772e 100644
--- a/GNNLux/test/layers/conv.jl
+++ b/GNNLux/test/layers/conv.jl
@@ -92,17 +92,17 @@
 
     @testset "GINConv" begin
         nn = Chain(Dense(in_dims => out_dims, tanh), Dense(out_dims => out_dims))
-        l = GINConv(nn; eps = 0.5)
+        l = GINConv(nn; ϵ = 0.5)
         test_lux_layer(rng, l, g, x, sizey=(out_dims,g.num_nodes), container=true)
         ps = LuxCore.initialparameters(rng, l)
-        @test !hasproperty(ps, :eps)
+        @test !hasproperty(ps, :ϵ)
 
         l = GINConv(nn; train_eps = true)
         test_lux_layer(rng, l, g, x, sizey=(out_dims,g.num_nodes), container=true)
         ps = LuxCore.initialparameters(rng, l)
-        @test ps.eps == [0.0f0]
+        @test ps.ϵ == [0.0f0]
 
-        @test GINConv(nn, 0.5).eps == 0.5
+        @test GINConv(nn, 0.5).ϵ == 0.5
     end
 
     @testset "MEGNetConv" begin
diff --git a/GNNlib/src/layers/conv.jl b/GNNlib/src/layers/conv.jl
index 6b7907f1d..685f3c6e4 100644
--- a/GNNlib/src/layers/conv.jl
+++ b/GNNlib/src/layers/conv.jl
@@ -252,8 +252,8 @@ function gin_conv(l, g::AbstractGNNGraph, x)
     xj, xi = expand_srcdst(g, x) 
  
     m = propagate(copy_xj, g, l.aggr, xj = xj)
-    eps = convert.(float(eltype(xi)), l.eps)
-    return l.nn((1 .+ eps) .* xi .+ m)
+    ϵ = convert.(float(eltype(xi)), l.ϵ)
+    return l.nn((1 .+ ϵ) .* xi .+ m)
 end
 
 ####################### NNConv ######################################
diff --git a/GraphNeuralNetworks/src/layers/conv.jl b/GraphNeuralNetworks/src/layers/conv.jl
index 4bd6cde76..b105aa93b 100644
--- a/GraphNeuralNetworks/src/layers/conv.jl
+++ b/GraphNeuralNetworks/src/layers/conv.jl
@@ -590,7 +590,7 @@ function Base.show(io::IO, l::EdgeConv)
 end
 
 @doc raw"""
-    GINConv(f, [eps]; aggr=+, train_eps=false)
+    GINConv(f, [ϵ]; aggr=+, train_eps=false)
 
 Graph Isomorphism convolutional layer from paper [How Powerful are Graph Neural Networks?](https://arxiv.org/pdf/1810.00826.pdf).
 
@@ -603,7 +603,7 @@ where ``f_\Theta`` typically denotes a learnable function, e.g. a linear layer o
 # Arguments
 
 - `f`: A (possibly learnable) function acting on node features.
-- `eps`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
+- `ϵ`: Initial value of the weighting factor ``\epsilon``. Default `0.0f0`.
 - `train_eps`: If `true`, ``\epsilon`` is trainable. Default `false`.
 - `aggr`: Neighborhood aggregation function. Default `+`.
 
@@ -621,7 +621,7 @@ g = GNNGraph(s, t)
 nn = Dense(in_channel, out_channel)
 
 # create layer
-l = GINConv(nn; eps = 0.01f0, train_eps = true, aggr = mean)
+l = GINConv(nn; ϵ = 0.01f0, train_eps = true, aggr = mean)
 
 # forward pass
 y = l(g, x)
@@ -629,27 +629,27 @@ y = l(g, x)
 """
 struct GINConv{E, NN, A} <: GNNLayer
     nn::NN
-    eps::E
+    ϵ::E
     aggr::A
     train_eps::Bool
 end
 
 Flux.@layer :expand GINConv
-Flux.trainable(l::GINConv) = l.train_eps ? (; l.nn, l.eps) : (; l.nn)
+Flux.trainable(l::GINConv) = l.train_eps ? (; l.nn, l.ϵ) : (; l.nn)
 
-GINConv(nn, eps::Real; aggr = +, train_eps::Bool = false) =
-    GINConv(nn, train_eps ? [eps] : eps, aggr, train_eps)
+GINConv(nn, ϵ::Real; aggr = +, train_eps::Bool = false) =
+    GINConv(nn, train_eps ? [ϵ] : ϵ, aggr, train_eps)
 
-GINConv(nn, eps::Real, aggr) = GINConv(nn, eps; aggr)
+GINConv(nn, ϵ::Real, aggr) = GINConv(nn, ϵ; aggr)
 
-GINConv(nn; eps::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
-    GINConv(nn, eps; aggr, train_eps)
+GINConv(nn; ϵ::Real = 0.0f0, aggr = +, train_eps::Bool = false) =
+    GINConv(nn, ϵ; aggr, train_eps)
 
 (l::GINConv)(g, x) = GNNlib.gin_conv(l, g, x)
 
 function Base.show(io::IO, l::GINConv)
     print(io, "GINConv($(l.nn)")
-    print(io, ", eps=$(l.eps)")
+    print(io, ", ϵ=$(l.ϵ)")
     l.train_eps && print(io, ", train_eps=true")
     l.aggr == (+) || print(io, ", aggr=$(l.aggr)")
     print(io, ")")
diff --git a/GraphNeuralNetworks/test/layers/conv.jl b/GraphNeuralNetworks/test/layers/conv.jl
index fafd68fcd..7cecaf143 100644
--- a/GraphNeuralNetworks/test/layers/conv.jl
+++ b/GraphNeuralNetworks/test/layers/conv.jl
@@ -278,7 +278,7 @@ end
     using .TestModule
     nn = Dense(D_IN, D_OUT)
 
-    l = GINConv(nn; eps = 0.01, aggr = mean)
+    l = GINConv(nn; ϵ = 0.01, aggr = mean)
     for g in TEST_GRAPHS
         @test size(l(g, g.x)) == (D_OUT, g.num_nodes)
         test_gradients(l, g, g.x, rtol = RTOL_HIGH)
@@ -286,17 +286,17 @@ end
     @test keys(Flux.trainable(l)) == (:nn,)
 
     l = GINConv(Dense(D_IN, D_OUT); train_eps = true, aggr = mean)
-    @test keys(Flux.trainable(l)) == (:nn, :eps)
-    @test l.eps == [0.0f0]
+    @test keys(Flux.trainable(l)) == (:nn, :ϵ)
+    @test l.ϵ == [0.0f0]
     for g in TEST_GRAPHS
         @test size(l(g, g.x)) == (D_OUT, g.num_nodes)
     end
     g = first(TEST_GRAPHS)
     grad = Flux.gradient(l -> mean(l(g, g.x)), l)[1]
-    @test grad.eps isa AbstractVector
-    @test length(grad.eps) == 1
+    @test grad.ϵ isa AbstractVector
+    @test length(grad.ϵ) == 1
 
-    @test GINConv(nn, 0.5).eps == 0.5
+    @test GINConv(nn, 0.5).ϵ == 0.5
 end
 
 @testitem "GINConv GPU" setup=[TolSnippet, TestModule] tags=[:gpu] begin

From 65f5d5c7b09e87ad0820f9454b8207cd1816ea29 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Fri, 24 Jul 2026 19:40:50 +0200
Subject: [PATCH 5/7] Fix gpu problem

---
 GNNlib/src/layers/conv.jl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/GNNlib/src/layers/conv.jl b/GNNlib/src/layers/conv.jl
index 685f3c6e4..14171e69d 100644
--- a/GNNlib/src/layers/conv.jl
+++ b/GNNlib/src/layers/conv.jl
@@ -252,7 +252,7 @@ function gin_conv(l, g::AbstractGNNGraph, x)
     xj, xi = expand_srcdst(g, x) 
  
     m = propagate(copy_xj, g, l.aggr, xj = xj)
-    ϵ = convert.(float(eltype(xi)), l.ϵ)
+    ϵ = l.ϵ isa Real ? ofeltype(xi, l.ϵ) : l.ϵ
     return l.nn((1 .+ ϵ) .* xi .+ m)
 end
 

From 1d99de0f78d49282610b596bde0932e897aecfa0 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Wed, 30 Sep 2026 15:09:19 +0200
Subject: [PATCH 6/7] Address GIN epsilon review feedback

---
 CHANGELOG.md                            | 18 +++++++++++++++---
 GNNLux/Project.toml                     |  4 ++--
 GNNLux/src/layers/conv.jl               |  2 +-
 GNNLux/test/layers/conv.jl              |  4 ++--
 GNNlib/Project.toml                     |  2 +-
 GraphNeuralNetworks/Project.toml        |  4 ++--
 GraphNeuralNetworks/src/layers/conv.jl  |  2 +-
 GraphNeuralNetworks/test/layers/conv.jl |  4 ++--
 8 files changed, 26 insertions(+), 14 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2ae90818b..6dcc6b4e9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,6 +10,11 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and the packages adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 Entries link to the pull request that introduced them.
 
+## GNNLux.jl — Unreleased (towards 0.3.0)
+
+**Added**
+- `GINConv` now supports a trainable `ϵ` parameter through the `train_eps` keyword ([#697]).
+
 ## GNNGraphs.jl — Unreleased (towards 1.6.0)
 
 **Added**
@@ -19,7 +24,7 @@ Entries link to the pull request that introduced them.
 - Added an `EnzymeCore` package extension marking `scaled_laplacian` inactive for Enzyme, matching its existing `@non_differentiable` ChainRules declaration. Enzyme previously differentiated the Krylov eigensolve inside it and failed; this unblocks `ChebConv`, `GConvGRUCell` and `GConvLSTMCell` ([#706]).
 - Enzyme can now differentiate through the adjacency-matrix → COO graph conversion: a new internal keyword-free helper `_to_coo_graph` avoids the union-typed keyword handling of the `GNNGraph(g; graph_type)` constructor and of `to_coo`, which Enzyme's type analysis cannot compile. Together with upstream fixes in Enzyme ≥ 0.13.197 this removes the `EnzymeInternalError` crash on `:dense`/`:sparse` graphs ([#703]).
 
-## GNNLux.jl — Unreleased (towards 0.2.0)
+## GNNLux.jl 0.2.0 — 2026-07-22
 
 **Added**
 - Added pooling layers (`GlobalPool`, `GlobalAttentionPool`, `TopKPool`) ([#576]).
@@ -33,12 +38,18 @@ Entries link to the pull request that introduced them.
 **Removed**
 - Removed the `A3TGCN` layer, mirroring its removal from the Flux frontend ([#696]).
 
-## GNNlib.jl — Unreleased (towards 1.4.1)
+## GNNlib.jl — Unreleased (towards 1.5.0)
+
+**Added**
+- Extended `gin_conv` to accept a scalar-shaped array for `ϵ`, enabling frontends to expose it as a trainable parameter ([#697]).
 
 **Fixed**
 - Fixed `Enzyme.gradient` failing when differentiating `GCNConv`, `SGConv` and `TAGConv` on `:dense`/`:sparse` adjacency graphs: their adjacency-matrix fallbacks now convert via the Enzyme-differentiable `GNNGraphs._to_coo_graph` instead of the keyword `GNNGraph` constructor. Requires GNNGraphs ≥ 1.5.2 and, for Enzyme, Enzyme ≥ 0.13.197 ([#703]).
 
-## GraphNeuralNetworks.jl — Unreleased (towards 1.1.1)
+## GraphNeuralNetworks.jl — Unreleased (towards 1.2.0)
+
+**Added**
+- `GINConv` now supports a trainable `ϵ` parameter through the `train_eps` keyword ([#697]).
 
 **Changed**
 - `GNNRecurrence` now hands each cell an indexed time slice instead of an `eachslice` view, matching the Lux frontend. Enzyme's type analysis fails on `SubArray` cell inputs, so this unblocks `DCGRU`, `EvolveGCNO`, `GConvGRU` and `GConvLSTM` under Enzyme ([#707]).
@@ -234,6 +245,7 @@ Lux implementations of the graph convolutional, pooling, and temporal layers
 [#623]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/issues/623
 [#695]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/695
 [#696]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/696
+[#697]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/697
 [#703]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/703
 [#704]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/704
 [#707]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/707
diff --git a/GNNLux/Project.toml b/GNNLux/Project.toml
index 0fb3b2e05..86b566822 100644
--- a/GNNLux/Project.toml
+++ b/GNNLux/Project.toml
@@ -1,6 +1,6 @@
 name = "GNNLux"
 uuid = "e8545f4d-a905-48ac-a8c4-ca114b98986d"
-version = "0.2.0"
+version = "0.3.0-DEV"
 authors = ["Carlo Lucibello and contributors"]
 
 [workspace]
@@ -26,7 +26,7 @@ GNNlib = {path = "../GNNlib"}
 [compat]
 ConcreteStructs = "0.2.3"
 GNNGraphs = "1.4"
-GNNlib = "1.4"
+GNNlib = "1.5"
 Lux = "1"
 LuxCore = "1"
 MLUtils = "0.4.4"
diff --git a/GNNLux/src/layers/conv.jl b/GNNLux/src/layers/conv.jl
index 73a78f223..387a57caf 100644
--- a/GNNLux/src/layers/conv.jl
+++ b/GNNLux/src/layers/conv.jl
@@ -1341,7 +1341,7 @@ y, st = l(g, x, ps, st)       # size:  out_channel × num_nodes
 end
 
 GINConv(nn, ϵ::Real; aggr = +, train_eps::Bool = false) =
-    GINConv(nn, train_eps ? [ϵ] : ϵ, aggr, train_eps)
+    GINConv(nn, train_eps ? Float32[ϵ] : ϵ, aggr, train_eps)
 
 GINConv(nn, ϵ::Real, aggr) = GINConv(nn, ϵ; aggr)
 
diff --git a/GNNLux/test/layers/conv.jl b/GNNLux/test/layers/conv.jl
index d32c2772e..f472b95b2 100644
--- a/GNNLux/test/layers/conv.jl
+++ b/GNNLux/test/layers/conv.jl
@@ -97,10 +97,10 @@
         ps = LuxCore.initialparameters(rng, l)
         @test !hasproperty(ps, :ϵ)
 
-        l = GINConv(nn; train_eps = true)
+        l = GINConv(nn; ϵ = 0.5, train_eps = true)
         test_lux_layer(rng, l, g, x, sizey=(out_dims,g.num_nodes), container=true)
         ps = LuxCore.initialparameters(rng, l)
-        @test ps.ϵ == [0.0f0]
+        @test ps.ϵ == Float32[0.5]
 
         @test GINConv(nn, 0.5).ϵ == 0.5
     end
diff --git a/GNNlib/Project.toml b/GNNlib/Project.toml
index 4c2e00a75..26f51d6fa 100644
--- a/GNNlib/Project.toml
+++ b/GNNlib/Project.toml
@@ -1,7 +1,7 @@
 name = "GNNlib"
 uuid = "a6a84749-d869-43f8-aacc-be26a1996e48"
 authors = ["Carlo Lucibello and contributors"]
-version = "1.4.1-DEV"
+version = "1.5.0-DEV"
 
 [workspace]
 projects = ["test", "docs", "../GNNGraphs"]
diff --git a/GraphNeuralNetworks/Project.toml b/GraphNeuralNetworks/Project.toml
index ce79f146f..eff1c2b24 100644
--- a/GraphNeuralNetworks/Project.toml
+++ b/GraphNeuralNetworks/Project.toml
@@ -1,7 +1,7 @@
 name = "GraphNeuralNetworks"
 uuid = "cffab07f-9bc2-4db1-8861-388f63bf7694"
 authors = ["Carlo Lucibello and contributors"]
-version = "1.1.1-DEV"
+version = "1.2.0-DEV"
 
 [workspace]
 projects = ["test", "docs"]
@@ -25,7 +25,7 @@ ChainRulesCore = "1"
 ConcreteStructs = "0.2.3"
 Flux = "0.16.0"
 GNNGraphs = "1.4"
-GNNlib = "1.4"
+GNNlib = "1.5"
 LinearAlgebra = "1"
 MLUtils = "0.4"
 MacroTools = "0.5"
diff --git a/GraphNeuralNetworks/src/layers/conv.jl b/GraphNeuralNetworks/src/layers/conv.jl
index b105aa93b..569ba3559 100644
--- a/GraphNeuralNetworks/src/layers/conv.jl
+++ b/GraphNeuralNetworks/src/layers/conv.jl
@@ -638,7 +638,7 @@ Flux.@layer :expand GINConv
 Flux.trainable(l::GINConv) = l.train_eps ? (; l.nn, l.ϵ) : (; l.nn)
 
 GINConv(nn, ϵ::Real; aggr = +, train_eps::Bool = false) =
-    GINConv(nn, train_eps ? [ϵ] : ϵ, aggr, train_eps)
+    GINConv(nn, train_eps ? Float32[ϵ] : ϵ, aggr, train_eps)
 
 GINConv(nn, ϵ::Real, aggr) = GINConv(nn, ϵ; aggr)
 
diff --git a/GraphNeuralNetworks/test/layers/conv.jl b/GraphNeuralNetworks/test/layers/conv.jl
index 7cecaf143..bc0525e2a 100644
--- a/GraphNeuralNetworks/test/layers/conv.jl
+++ b/GraphNeuralNetworks/test/layers/conv.jl
@@ -285,9 +285,9 @@ end
     end
     @test keys(Flux.trainable(l)) == (:nn,)
 
-    l = GINConv(Dense(D_IN, D_OUT); train_eps = true, aggr = mean)
+    l = GINConv(Dense(D_IN, D_OUT); ϵ = 0.5, train_eps = true, aggr = mean)
     @test keys(Flux.trainable(l)) == (:nn, :ϵ)
-    @test l.ϵ == [0.0f0]
+    @test l.ϵ == Float32[0.5]
     for g in TEST_GRAPHS
         @test size(l(g, g.x)) == (D_OUT, g.num_nodes)
     end

From 1b51c53d48dcd675fd25b6856b2e3a8facab3480 Mon Sep 17 00:00:00 2001
From: Aurora Rossi <aurora98rossi@gmail.com>
Date: Wed, 30 Sep 2026 15:16:52 +0200
Subject: [PATCH 7/7] Fix

---
 CHANGELOG.md                     | 18 +++---------------
 GNNLux/Project.toml              |  4 ++--
 GNNlib/Project.toml              |  2 +-
 GraphNeuralNetworks/Project.toml |  4 ++--
 4 files changed, 8 insertions(+), 20 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6dcc6b4e9..2ae90818b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,11 +10,6 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and the packages adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 Entries link to the pull request that introduced them.
 
-## GNNLux.jl — Unreleased (towards 0.3.0)
-
-**Added**
-- `GINConv` now supports a trainable `ϵ` parameter through the `train_eps` keyword ([#697]).
-
 ## GNNGraphs.jl — Unreleased (towards 1.6.0)
 
 **Added**
@@ -24,7 +19,7 @@ Entries link to the pull request that introduced them.
 - Added an `EnzymeCore` package extension marking `scaled_laplacian` inactive for Enzyme, matching its existing `@non_differentiable` ChainRules declaration. Enzyme previously differentiated the Krylov eigensolve inside it and failed; this unblocks `ChebConv`, `GConvGRUCell` and `GConvLSTMCell` ([#706]).
 - Enzyme can now differentiate through the adjacency-matrix → COO graph conversion: a new internal keyword-free helper `_to_coo_graph` avoids the union-typed keyword handling of the `GNNGraph(g; graph_type)` constructor and of `to_coo`, which Enzyme's type analysis cannot compile. Together with upstream fixes in Enzyme ≥ 0.13.197 this removes the `EnzymeInternalError` crash on `:dense`/`:sparse` graphs ([#703]).
 
-## GNNLux.jl 0.2.0 — 2026-07-22
+## GNNLux.jl — Unreleased (towards 0.2.0)
 
 **Added**
 - Added pooling layers (`GlobalPool`, `GlobalAttentionPool`, `TopKPool`) ([#576]).
@@ -38,18 +33,12 @@ Entries link to the pull request that introduced them.
 **Removed**
 - Removed the `A3TGCN` layer, mirroring its removal from the Flux frontend ([#696]).
 
-## GNNlib.jl — Unreleased (towards 1.5.0)
-
-**Added**
-- Extended `gin_conv` to accept a scalar-shaped array for `ϵ`, enabling frontends to expose it as a trainable parameter ([#697]).
+## GNNlib.jl — Unreleased (towards 1.4.1)
 
 **Fixed**
 - Fixed `Enzyme.gradient` failing when differentiating `GCNConv`, `SGConv` and `TAGConv` on `:dense`/`:sparse` adjacency graphs: their adjacency-matrix fallbacks now convert via the Enzyme-differentiable `GNNGraphs._to_coo_graph` instead of the keyword `GNNGraph` constructor. Requires GNNGraphs ≥ 1.5.2 and, for Enzyme, Enzyme ≥ 0.13.197 ([#703]).
 
-## GraphNeuralNetworks.jl — Unreleased (towards 1.2.0)
-
-**Added**
-- `GINConv` now supports a trainable `ϵ` parameter through the `train_eps` keyword ([#697]).
+## GraphNeuralNetworks.jl — Unreleased (towards 1.1.1)
 
 **Changed**
 - `GNNRecurrence` now hands each cell an indexed time slice instead of an `eachslice` view, matching the Lux frontend. Enzyme's type analysis fails on `SubArray` cell inputs, so this unblocks `DCGRU`, `EvolveGCNO`, `GConvGRU` and `GConvLSTM` under Enzyme ([#707]).
@@ -245,7 +234,6 @@ Lux implementations of the graph convolutional, pooling, and temporal layers
 [#623]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/issues/623
 [#695]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/695
 [#696]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/696
-[#697]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/697
 [#703]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/703
 [#704]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/704
 [#707]: https://github.com/JuliaGraphs/GraphNeuralNetworks.jl/pull/707
diff --git a/GNNLux/Project.toml b/GNNLux/Project.toml
index 86b566822..0fb3b2e05 100644
--- a/GNNLux/Project.toml
+++ b/GNNLux/Project.toml
@@ -1,6 +1,6 @@
 name = "GNNLux"
 uuid = "e8545f4d-a905-48ac-a8c4-ca114b98986d"
-version = "0.3.0-DEV"
+version = "0.2.0"
 authors = ["Carlo Lucibello and contributors"]
 
 [workspace]
@@ -26,7 +26,7 @@ GNNlib = {path = "../GNNlib"}
 [compat]
 ConcreteStructs = "0.2.3"
 GNNGraphs = "1.4"
-GNNlib = "1.5"
+GNNlib = "1.4"
 Lux = "1"
 LuxCore = "1"
 MLUtils = "0.4.4"
diff --git a/GNNlib/Project.toml b/GNNlib/Project.toml
index 26f51d6fa..4c2e00a75 100644
--- a/GNNlib/Project.toml
+++ b/GNNlib/Project.toml
@@ -1,7 +1,7 @@
 name = "GNNlib"
 uuid = "a6a84749-d869-43f8-aacc-be26a1996e48"
 authors = ["Carlo Lucibello and contributors"]
-version = "1.5.0-DEV"
+version = "1.4.1-DEV"
 
 [workspace]
 projects = ["test", "docs", "../GNNGraphs"]
diff --git a/GraphNeuralNetworks/Project.toml b/GraphNeuralNetworks/Project.toml
index eff1c2b24..ce79f146f 100644
--- a/GraphNeuralNetworks/Project.toml
+++ b/GraphNeuralNetworks/Project.toml
@@ -1,7 +1,7 @@
 name = "GraphNeuralNetworks"
 uuid = "cffab07f-9bc2-4db1-8861-388f63bf7694"
 authors = ["Carlo Lucibello and contributors"]
-version = "1.2.0-DEV"
+version = "1.1.1-DEV"
 
 [workspace]
 projects = ["test", "docs"]
@@ -25,7 +25,7 @@ ChainRulesCore = "1"
 ConcreteStructs = "0.2.3"
 Flux = "0.16.0"
 GNNGraphs = "1.4"
-GNNlib = "1.5"
+GNNlib = "1.4"
 LinearAlgebra = "1"
 MLUtils = "0.4"
 MacroTools = "0.5"