MNIST Classification with SimpleChains

SimpleChains.jl is an excellent framework for training small neural networks. In this tutorial we will demonstrate how to use the same API as Lux.jl to train a model using SimpleChains.jl. We will use the tutorial from SimpleChains.jl as a reference.

Package Imports

using Lux, MLUtils, Optimisers, Zygote, OneHotArrays, Random, Statistics, Printf, Reactant
using MLDatasets: MNIST
using SimpleChains: SimpleChains

Reactant.set_default_backend("cpu")

Precompiling Lux...
   9506.8 ms  ✓ Lux
  1 dependency successfully precompiled in 10 seconds. 104 already precompiled.
Precompiling LuxMLUtilsExt...
   2090.0 ms  ✓ Lux → LuxMLUtilsExt
  1 dependency successfully precompiled in 2 seconds. 164 already precompiled.
Precompiling LuxZygoteExt...
   2519.9 ms  ✓ Lux → LuxZygoteExt
  1 dependency successfully precompiled in 3 seconds. 143 already precompiled.
Precompiling Reactant...
  93936.8 ms  ✓ Enzyme
   6799.6 ms  ✓ Enzyme → EnzymeGPUArraysCoreExt
  87129.8 ms  ✓ Reactant
  3 dependencies successfully precompiled in 188 seconds. 77 already precompiled.
Precompiling LuxEnzymeExt...
   6880.7 ms  ✓ Enzyme → EnzymeSpecialFunctionsExt
   6720.1 ms  ✓ Enzyme → EnzymeLogExpFunctionsExt
  14677.4 ms  ✓ Enzyme → EnzymeStaticArraysExt
  14858.8 ms  ✓ Enzyme → EnzymeChainRulesCoreExt
   7600.3 ms  ✓ Lux → LuxEnzymeExt
  5 dependencies successfully precompiled in 22 seconds. 145 already precompiled.
Precompiling OptimisersReactantExt...
  16579.6 ms  ✓ Reactant → ReactantStatisticsExt
  18982.8 ms  ✓ Optimisers → OptimisersReactantExt
  2 dependencies successfully precompiled in 19 seconds. 88 already precompiled.
Precompiling LuxCoreReactantExt...
  16683.5 ms  ✓ LuxCore → LuxCoreReactantExt
  1 dependency successfully precompiled in 17 seconds. 85 already precompiled.
Precompiling MLDataDevicesReactantExt...
  16920.8 ms  ✓ MLDataDevices → MLDataDevicesReactantExt
  1 dependency successfully precompiled in 17 seconds. 82 already precompiled.
Precompiling LuxLibReactantExt...
  17230.5 ms  ✓ Reactant → ReactantSpecialFunctionsExt
  17285.3 ms  ✓ Reactant → ReactantKernelAbstractionsExt
  17350.2 ms  ✓ LuxLib → LuxLibReactantExt
  16585.9 ms  ✓ Reactant → ReactantArrayInterfaceExt
  4 dependencies successfully precompiled in 34 seconds. 158 already precompiled.
Precompiling WeightInitializersReactantExt...
  16727.3 ms  ✓ WeightInitializers → WeightInitializersReactantExt
  1 dependency successfully precompiled in 17 seconds. 96 already precompiled.
Precompiling ReactantAbstractFFTsExt...
  16652.4 ms  ✓ Reactant → ReactantAbstractFFTsExt
  1 dependency successfully precompiled in 17 seconds. 82 already precompiled.
Precompiling ReactantOneHotArraysExt...
  16995.6 ms  ✓ Reactant → ReactantOneHotArraysExt
  1 dependency successfully precompiled in 17 seconds. 104 already precompiled.
Precompiling ReactantNNlibExt...
  19518.6 ms  ✓ Reactant → ReactantNNlibExt
  1 dependency successfully precompiled in 20 seconds. 103 already precompiled.
Precompiling LuxReactantExt...
  11941.4 ms  ✓ Lux → LuxReactantExt
  1 dependency successfully precompiled in 12 seconds. 180 already precompiled.
Precompiling ReactantOffsetArraysExt...
  16965.4 ms  ✓ Reactant → ReactantOffsetArraysExt
  1 dependency successfully precompiled in 17 seconds. 82 already precompiled.
Precompiling LuxSimpleChainsExt...
   1840.0 ms  ✓ Lux → LuxSimpleChainsExt
  1 dependency successfully precompiled in 2 seconds. 122 already precompiled.
2025-07-14 00:09:05.891939: I external/xla/xla/service/service.cc:153] XLA service 0xc3dc490 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
2025-07-14 00:09:05.891973: I external/xla/xla/service/service.cc:161]   StreamExecutor device (0): Quadro RTX 5000, Compute Capability 7.5
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1752451745.892822 2724621 se_gpu_pjrt_client.cc:1370] Using BFC allocator.
I0000 00:00:1752451745.892907 2724621 gpu_helpers.cc:136] XLA backend allocating 12528893952 bytes on device 0 for BFCAllocator.
I0000 00:00:1752451745.892943 2724621 gpu_helpers.cc:177] XLA backend will use up to 4176297984 bytes on device 0 for CollectiveBFCAllocator.
I0000 00:00:1752451745.905474 2724621 cuda_dnn.cc:471] Loaded cuDNN version 90800

Loading MNIST

function loadmnist(batchsize, train_split)
    # Load MNIST
    N = parse(Bool, get(ENV, "CI", "false")) ? 1500 : nothing
    dataset = MNIST(; split=:train)
    if N !== nothing
        imgs = dataset.features[:, :, 1:N]
        labels_raw = dataset.targets[1:N]
    else
        imgs = dataset.features
        labels_raw = dataset.targets
    end

    # Process images into (H, W, C, BS) batches
    x_data = Float32.(reshape(imgs, size(imgs, 1), size(imgs, 2), 1, size(imgs, 3)))
    y_data = onehotbatch(labels_raw, 0:9)
    (x_train, y_train), (x_test, y_test) = splitobs((x_data, y_data); at=train_split)

    return (
        # Use DataLoader to automatically minibatch and shuffle the data
        DataLoader(collect.((x_train, y_train)); batchsize, shuffle=true, partial=false),
        # Don't shuffle the test data
        DataLoader(collect.((x_test, y_test)); batchsize, shuffle=false, partial=false),
    )
end

loadmnist (generic function with 1 method)

Define the Model

lux_model = Chain(
    Conv((5, 5), 1 => 6, relu),
    MaxPool((2, 2)),
    Conv((5, 5), 6 => 16, relu),
    MaxPool((2, 2)),
    FlattenLayer(3),
    Chain(Dense(256 => 128, relu), Dense(128 => 84, relu), Dense(84 => 10)),
)

Chain(
    layer_1 = Conv((5, 5), 1 => 6, relu),  # 156 parameters
    layer_2 = MaxPool((2, 2)),
    layer_3 = Conv((5, 5), 6 => 16, relu),  # 2_416 parameters
    layer_4 = MaxPool((2, 2)),
    layer_5 = Lux.FlattenLayer{Static.StaticInt{3}}(static(3)),
    layer_6 = Chain(
        layer_1 = Dense(256 => 128, relu),  # 32_896 parameters
        layer_2 = Dense(128 => 84, relu),  # 10_836 parameters
        layer_3 = Dense(84 => 10),      # 850 parameters
    ),
)         # Total: 47_154 parameters,
          #        plus 0 states.

We now need to convert the lux_model to SimpleChains.jl. We need to do this by defining the ToSimpleChainsAdaptor and providing the input dimensions.

adaptor = ToSimpleChainsAdaptor((28, 28, 1))
simple_chains_model = adaptor(lux_model)

SimpleChainsLayer(
    Chain(
        layer_1 = Conv((5, 5), 1 => 6, relu),  # 156 parameters
        layer_2 = MaxPool((2, 2)),
        layer_3 = Conv((5, 5), 6 => 16, relu),  # 2_416 parameters
        layer_4 = MaxPool((2, 2)),
        layer_5 = Lux.FlattenLayer{Static.StaticInt{3}}(static(3)),
        layer_6 = Chain(
            layer_1 = Dense(256 => 128, relu),  # 32_896 parameters
            layer_2 = Dense(128 => 84, relu),  # 10_836 parameters
            layer_3 = Dense(84 => 10),  # 850 parameters
        ),
    ),
)         # Total: 47_154 parameters,
          #        plus 0 states.

Helper Functions

const lossfn = CrossEntropyLoss(; logits=Val(true))

function accuracy(model, ps, st, dataloader)
    total_correct, total = 0, 0
    st = Lux.testmode(st)
    for (x, y) in dataloader
        target_class = onecold(Array(y))
        predicted_class = onecold(Array(first(model(x, ps, st))))
        total_correct += sum(target_class .== predicted_class)
        total += length(target_class)
    end
    return total_correct / total
end

accuracy (generic function with 1 method)

Define the Training Loop

function train(model, dev=cpu_device(); rng=Random.default_rng(), kwargs...)
    train_dataloader, test_dataloader = dev(loadmnist(128, 0.9))
    ps, st = dev(Lux.setup(rng, model))

    vjp = dev isa ReactantDevice ? AutoEnzyme() : AutoZygote()

    train_state = Training.TrainState(model, ps, st, Adam(3.0f-4))

    if dev isa ReactantDevice
        x_ra = first(test_dataloader)[1]
        model_compiled = Reactant.with_config(;
            dot_general_precision=PrecisionConfig.HIGH,
            convolution_precision=PrecisionConfig.HIGH,
        ) do
            @compile model(x_ra, ps, Lux.testmode(st))
        end
    else
        model_compiled = model
    end

    ### Lets train the model
    nepochs = 10
    tr_acc, te_acc = 0.0, 0.0
    for epoch in 1:nepochs
        stime = time()
        for (x, y) in train_dataloader
            _, _, _, train_state = Training.single_train_step!(
                vjp, lossfn, (x, y), train_state
            )
        end
        ttime = time() - stime

        tr_acc =
            accuracy(
                model_compiled, train_state.parameters, train_state.states, train_dataloader
            ) * 100
        te_acc =
            accuracy(
                model_compiled, train_state.parameters, train_state.states, test_dataloader
            ) * 100

        @printf "[%2d/%2d] \t Time %.2fs \t Training Accuracy: %.2f%% \t Test Accuracy: \
                 %.2f%%\n" epoch nepochs ttime tr_acc te_acc
    end

    return tr_acc, te_acc
end

train (generic function with 2 methods)

Finally Training the Model

First we will train the Lux model

tr_acc, te_acc = train(lux_model, reactant_device())

[ 1/10] 	 Time 393.58s 	 Training Accuracy: 20.47% 	 Test Accuracy: 19.53%
[ 2/10] 	 Time 0.11s 	 Training Accuracy: 35.78% 	 Test Accuracy: 35.16%
[ 3/10] 	 Time 0.10s 	 Training Accuracy: 47.81% 	 Test Accuracy: 43.75%
[ 4/10] 	 Time 0.18s 	 Training Accuracy: 57.66% 	 Test Accuracy: 51.56%
[ 5/10] 	 Time 0.10s 	 Training Accuracy: 67.11% 	 Test Accuracy: 57.03%
[ 6/10] 	 Time 0.10s 	 Training Accuracy: 72.73% 	 Test Accuracy: 66.41%
[ 7/10] 	 Time 0.10s 	 Training Accuracy: 76.02% 	 Test Accuracy: 72.66%
[ 8/10] 	 Time 0.10s 	 Training Accuracy: 80.16% 	 Test Accuracy: 77.34%
[ 9/10] 	 Time 0.10s 	 Training Accuracy: 81.88% 	 Test Accuracy: 77.34%
[10/10] 	 Time 0.13s 	 Training Accuracy: 83.91% 	 Test Accuracy: 78.91%

Now we will train the SimpleChains model

tr_acc, te_acc = train(simple_chains_model)

[ 1/10] 	 Time 866.75s 	 Training Accuracy: 31.02% 	 Test Accuracy: 25.00%
[ 2/10] 	 Time 12.10s 	 Training Accuracy: 48.91% 	 Test Accuracy: 44.53%
[ 3/10] 	 Time 12.10s 	 Training Accuracy: 61.72% 	 Test Accuracy: 58.59%
[ 4/10] 	 Time 12.16s 	 Training Accuracy: 67.34% 	 Test Accuracy: 55.47%
[ 5/10] 	 Time 12.10s 	 Training Accuracy: 71.64% 	 Test Accuracy: 63.28%
[ 6/10] 	 Time 12.13s 	 Training Accuracy: 76.72% 	 Test Accuracy: 70.31%
[ 7/10] 	 Time 12.12s 	 Training Accuracy: 81.02% 	 Test Accuracy: 74.22%
[ 8/10] 	 Time 12.13s 	 Training Accuracy: 82.81% 	 Test Accuracy: 75.00%
[ 9/10] 	 Time 12.12s 	 Training Accuracy: 84.69% 	 Test Accuracy: 78.91%
[10/10] 	 Time 12.19s 	 Training Accuracy: 86.80% 	 Test Accuracy: 82.81%

On my local machine we see a 3-4x speedup when using SimpleChains.jl. The conditions of the server this documentation is being built on is not ideal for CPU benchmarking hence, the speedup may not be as significant and even there might be regressions.

Appendix

using InteractiveUtils
InteractiveUtils.versioninfo()

if @isdefined(MLDataDevices)
    if @isdefined(CUDA) && MLDataDevices.functional(CUDADevice)
        println()
        CUDA.versioninfo()
    end

    if @isdefined(AMDGPU) && MLDataDevices.functional(AMDGPUDevice)
        println()
        AMDGPU.versioninfo()
    end
end

Julia Version 1.11.6
Commit 9615af0f269 (2025-07-09 12:58 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 48 × AMD EPYC 7402 24-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver2)
Threads: 48 default, 0 interactive, 24 GC (on 2 virtual cores)
Environment:
  JULIA_CPU_THREADS = 2
  LD_LIBRARY_PATH = /usr/local/nvidia/lib:/usr/local/nvidia/lib64
  JULIA_PKG_SERVER = 
  JULIA_NUM_THREADS = 48
  JULIA_CUDA_HARD_MEMORY_LIMIT = 100%
  JULIA_PKG_PRECOMPILE_AUTO = 0
  JULIA_DEBUG = Literate
  JULIA_DEPOT_PATH = /root/.cache/julia-buildkite-plugin/depots/01872db4-8c79-43af-ab7d-12abac4f24f6

---

This page was generated using Literate.jl.