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")

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1758300376.952413 1096581 service.cc:158] XLA service 0x34d2d430 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
I0000 00:00:1758300376.952498 1096581 service.cc:166]   StreamExecutor device (0): Quadro RTX 5000, Compute Capability 7.5
I0000 00:00:1758300376.952518 1096581 service.cc:166]   StreamExecutor device (1): Quadro RTX 5000, Compute Capability 7.5
I0000 00:00:1758300376.958115 1096581 se_gpu_pjrt_client.cc:1338] Using BFC allocator.
I0000 00:00:1758300376.958326 1096581 gpu_helpers.cc:136] XLA backend allocating 12526534656 bytes on device 0 for BFCAllocator.
I0000 00:00:1758300376.958567 1096581 gpu_helpers.cc:136] XLA backend allocating 12526534656 bytes on device 1 for BFCAllocator.
I0000 00:00:1758300376.958589 1096581 gpu_helpers.cc:177] XLA backend will use up to 4175511552 bytes on device 0 for CollectiveBFCAllocator.
I0000 00:00:1758300376.958608 1096581 gpu_helpers.cc:177] XLA backend will use up to 4175511552 bytes on device 1 for CollectiveBFCAllocator.
I0000 00:00:1758300376.973485 1096581 cuda_dnn.cc:463] Loaded cuDNN version 91200

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

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

Finally Training the Model

First we will train the Lux model

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

[ 1/10] 	 Time 412.42s 	 Training Accuracy: 7.50% 	 Test Accuracy: 11.72%
[ 2/10] 	 Time 0.09s 	 Training Accuracy: 23.83% 	 Test Accuracy: 20.31%
[ 3/10] 	 Time 0.08s 	 Training Accuracy: 36.72% 	 Test Accuracy: 32.81%
[ 4/10] 	 Time 0.09s 	 Training Accuracy: 49.45% 	 Test Accuracy: 45.31%
[ 5/10] 	 Time 0.09s 	 Training Accuracy: 60.08% 	 Test Accuracy: 51.56%
[ 6/10] 	 Time 0.09s 	 Training Accuracy: 65.62% 	 Test Accuracy: 54.69%
[ 7/10] 	 Time 0.15s 	 Training Accuracy: 70.86% 	 Test Accuracy: 63.28%
[ 8/10] 	 Time 0.08s 	 Training Accuracy: 73.98% 	 Test Accuracy: 64.06%
[ 9/10] 	 Time 0.09s 	 Training Accuracy: 76.33% 	 Test Accuracy: 71.09%
[10/10] 	 Time 0.09s 	 Training Accuracy: 79.53% 	 Test Accuracy: 73.44%

Now we will train the SimpleChains model

tr_acc, te_acc = train(simple_chains_model)

[ 1/10] 	 Time 891.83s 	 Training Accuracy: 30.94% 	 Test Accuracy: 31.25%
[ 2/10] 	 Time 12.36s 	 Training Accuracy: 43.05% 	 Test Accuracy: 42.97%
[ 3/10] 	 Time 12.39s 	 Training Accuracy: 54.37% 	 Test Accuracy: 60.16%
[ 4/10] 	 Time 12.39s 	 Training Accuracy: 61.72% 	 Test Accuracy: 60.94%
[ 5/10] 	 Time 12.39s 	 Training Accuracy: 69.77% 	 Test Accuracy: 70.31%
[ 6/10] 	 Time 12.42s 	 Training Accuracy: 74.14% 	 Test Accuracy: 67.97%
[ 7/10] 	 Time 12.38s 	 Training Accuracy: 79.38% 	 Test Accuracy: 74.22%
[ 8/10] 	 Time 12.39s 	 Training Accuracy: 81.56% 	 Test Accuracy: 75.78%
[ 9/10] 	 Time 12.45s 	 Training Accuracy: 83.91% 	 Test Accuracy: 79.69%
[10/10] 	 Time 13.99s 	 Training Accuracy: 86.33% 	 Test Accuracy: 79.69%

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 1.11.7
Commit f2b3dbda30a (2025-09-08 12:10 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
  JULIA_DEPOT_PATH = /root/.cache/julia-buildkite-plugin/depots/01872db4-8c79-43af-ab7d-12abac4f24f6
  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

---

This page was generated using Literate.jl.