uf

experiments/Manifest.toml

@@ -2,7 +2,7 @@
 
 julia_version = "1.9.3"
 manifest_format = "2.0"
-project_hash = "610dd01ba3ad744c63925263e3388bfa9abf58ab"
+project_hash = "f97f8534bd2363738071587ac7055fab77fc36f4"
 
 [[deps.ARFFFiles]]
 deps = ["CategoricalArrays", "Dates", "Parsers", "Tables"]

experiments/Project.toml

@@ -17,6 +17,7 @@ MLJEnsembles = "50ed68f4-41fd-4504-931a-ed422449fee0"
 MLJFlux = "094fc8d1-fd35-5302-93ea-dabda2abf845"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 MPI = "da04e1cc-30fd-572f-bb4f-1f8673147195"
+Metalhead = "dbeba491-748d-5e0e-a39e-b530a07fa0cc"
 Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
 TidierData = "fe2206b3-d496-4ee9-a338-6a095c4ece80"
 TidierPlots = "337ecbd1-5042-4e2a-ae6f-ca776f97570a"

experiments/models/additional_models.jl

@@ -48,3 +48,34 @@ Builds a LeNet-like convolutional neural network.
 """
 lenet5(builder=LeNetBuilder(5, 6, 16); kwargs...) = NeuralNetworkClassifier(builder=builder; acceleration=CUDALibs(), kwargs...)
 
+"""
+    ResNetBuilder
+
+MLJFlux builder for a ResNet.
+"""
+mutable struct ResNetBuilder end
+
+"""
+    MLJFlux.build(b::LeNetBuilder, rng, n_in, n_out)
+
+Overloads the MLJFlux build function for a LeNet-like convolutional neural network.
+"""
+function MLJFlux.build(b::ResNetBuilder, rng, n_in, n_out)
+    _n_in = Int(sqrt(n_in))
+    front = Metalhead.ResNet(18; pretrain=true, inchannels=1)
+    d = Flux.outputsize(front, (_n_in, _n_in, 1, 1)) |> first
+    back = Flux.Chain(
+        Dense(d, 120, relu),
+        Dense(120, 84, relu),
+        Dense(84, n_out),
+    )
+    chain = Flux.Chain(ECCCo.ToConv(_n_in), front, back)
+    return chain
+end
+
+"""
+    lenet5(builder=LeNetBuilder(5, 6, 16); kwargs...)
+
+Builds a LeNet-like convolutional neural network.
+"""
+resnet18(builder=ResNetBuilder(); kwargs...) = NeuralNetworkClassifier(builder=builder; acceleration=CUDALibs(), kwargs...)

experiments/setup_env.jl

@@ -18,6 +18,7 @@ using Flux
 using JointEnergyModels
 using LazyArtifacts
 using Logging
+using Metalhead
 using MLJBase: multiclass_f1score, accuracy, multiclass_precision, table, machine, fit!
 using MLJEnsembles
 using MLJFlux