Training a HyperNetwork on MNIST and FashionMNIST

Package Imports

using Lux, ADTypes, ComponentArrays, LuxCUDA, MLDatasets, MLUtils, OneHotArrays, Optimisers,
      Printf, Random, Setfield, Statistics, Zygote

CUDA.allowscalar(false)

Loading Datasets

function load_dataset(::Type{dset}, n_train::Int, n_eval::Int, batchsize::Int) where {dset}
    imgs, labels = dset(:train)[1:n_train]
    x_train, y_train = reshape(imgs, 28, 28, 1, n_train), onehotbatch(labels, 0:9)

    imgs, labels = dset(:test)[1:n_eval]
    x_test, y_test = reshape(imgs, 28, 28, 1, n_eval), onehotbatch(labels, 0:9)

    return (
        DataLoader((x_train, y_train); batchsize=min(batchsize, n_train), shuffle=true),
        DataLoader((x_test, y_test); batchsize=min(batchsize, n_eval), shuffle=false))
end

function load_datasets(n_train=1024, n_eval=32, batchsize=256)
    return load_dataset.((MNIST, FashionMNIST), n_train, n_eval, batchsize)
end

load_datasets (generic function with 4 methods)

Implement a HyperNet Layer

function HyperNet(
        weight_generator::Lux.AbstractLuxLayer, core_network::Lux.AbstractLuxLayer)
    ca_axes = Lux.initialparameters(Random.default_rng(), core_network) |>
              ComponentArray |>
              getaxes
    return @compact(; ca_axes, weight_generator, core_network, dispatch=:HyperNet) do (x, y)
        # Generate the weights
        ps_new = ComponentArray(vec(weight_generator(x)), ca_axes)
        @return core_network(y, ps_new)
    end
end

HyperNet (generic function with 1 method)

Defining functions on the CompactLuxLayer requires some understanding of how the layer is structured, as such we don't recommend doing it unless you are familiar with the internals. In this case, we simply write it to ignore the initialization of the core_network parameters.

function Lux.initialparameters(rng::AbstractRNG, hn::CompactLuxLayer{:HyperNet})
    return (; weight_generator=Lux.initialparameters(rng, hn.layers.weight_generator),)
end

Create and Initialize the HyperNet

function create_model()
    # Doesn't need to be a MLP can have any Lux Layer
    core_network = Chain(FlattenLayer(), Dense(784, 256, relu), Dense(256, 10))
    weight_generator = Chain(Embedding(2 => 32), Dense(32, 64, relu),
        Dense(64, Lux.parameterlength(core_network)))

    model = HyperNet(weight_generator, core_network)
    return model
end

create_model (generic function with 1 method)

Define Utility Functions

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

function accuracy(model, ps, st, dataloader, data_idx)
    total_correct, total = 0, 0
    st = Lux.testmode(st)
    for (x, y) in dataloader
        target_class = onecold(y)
        predicted_class = onecold(first(model((data_idx, 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)

Training

function train()
    model = create_model()
    dataloaders = load_datasets()

    dev = gpu_device()
    rng = Xoshiro(0)
    ps, st = Lux.setup(rng, model) |> dev

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

    ### Lets train the model
    nepochs = 25
    for epoch in 1:nepochs, data_idx in 1:2
        train_dataloader, test_dataloader = dataloaders[data_idx] .|> dev

        stime = time()
        for (x, y) in train_dataloader
            (_, _, _, train_state) = Training.single_train_step!(
                AutoZygote(), loss, ((data_idx, x), y), train_state)
        end
        ttime = time() - stime

        train_acc = round(
            accuracy(model, train_state.parameters,
                train_state.states, train_dataloader, data_idx) * 100;
            digits=2)
        test_acc = round(
            accuracy(model, train_state.parameters,
                train_state.states, test_dataloader, data_idx) * 100;
            digits=2)

        data_name = data_idx == 1 ? "MNIST" : "FashionMNIST"

        @printf "[%3d/%3d] \t %12s \t Time %.5fs \t Training Accuracy: %.2f%% \t Test \
                 Accuracy: %.2f%%\n" epoch nepochs data_name ttime train_acc test_acc
    end

    println()

    test_acc_list = [0.0, 0.0]
    for data_idx in 1:2
        train_dataloader, test_dataloader = dataloaders[data_idx] .|> dev
        train_acc = round(
            accuracy(model, train_state.parameters,
                train_state.states, train_dataloader, data_idx) * 100;
            digits=2)
        test_acc = round(
            accuracy(model, train_state.parameters,
                train_state.states, test_dataloader, data_idx) * 100;
            digits=2)

        data_name = data_idx == 1 ? "MNIST" : "FashionMNIST"

        @printf "[FINAL] \t %12s \t Training Accuracy: %.2f%% \t Test Accuracy: \
                 %.2f%%\n" data_name train_acc test_acc
        test_acc_list[data_idx] = test_acc
    end
    return test_acc_list
end

test_acc_list = train()

[  1/ 25] 	        MNIST 	 Time 70.33614s 	 Training Accuracy: 24.80% 	 Test Accuracy: 25.00%
[  1/ 25] 	 FashionMNIST 	 Time 0.02729s 	 Training Accuracy: 30.66% 	 Test Accuracy: 21.88%
[  2/ 25] 	        MNIST 	 Time 0.02900s 	 Training Accuracy: 50.68% 	 Test Accuracy: 34.38%
[  2/ 25] 	 FashionMNIST 	 Time 0.02873s 	 Training Accuracy: 55.47% 	 Test Accuracy: 46.88%
[  3/ 25] 	        MNIST 	 Time 0.02848s 	 Training Accuracy: 59.86% 	 Test Accuracy: 62.50%
[  3/ 25] 	 FashionMNIST 	 Time 0.02878s 	 Training Accuracy: 64.26% 	 Test Accuracy: 56.25%
[  4/ 25] 	        MNIST 	 Time 0.02857s 	 Training Accuracy: 67.58% 	 Test Accuracy: 59.38%
[  4/ 25] 	 FashionMNIST 	 Time 0.02825s 	 Training Accuracy: 70.12% 	 Test Accuracy: 56.25%
[  5/ 25] 	        MNIST 	 Time 0.02198s 	 Training Accuracy: 77.15% 	 Test Accuracy: 62.50%
[  5/ 25] 	 FashionMNIST 	 Time 0.05481s 	 Training Accuracy: 75.68% 	 Test Accuracy: 65.62%
[  6/ 25] 	        MNIST 	 Time 0.02101s 	 Training Accuracy: 79.79% 	 Test Accuracy: 65.62%
[  6/ 25] 	 FashionMNIST 	 Time 0.02097s 	 Training Accuracy: 74.80% 	 Test Accuracy: 65.62%
[  7/ 25] 	        MNIST 	 Time 0.02040s 	 Training Accuracy: 84.77% 	 Test Accuracy: 68.75%
[  7/ 25] 	 FashionMNIST 	 Time 0.02082s 	 Training Accuracy: 76.95% 	 Test Accuracy: 71.88%
[  8/ 25] 	        MNIST 	 Time 0.02092s 	 Training Accuracy: 86.72% 	 Test Accuracy: 71.88%
[  8/ 25] 	 FashionMNIST 	 Time 0.02103s 	 Training Accuracy: 81.84% 	 Test Accuracy: 68.75%
[  9/ 25] 	        MNIST 	 Time 0.02125s 	 Training Accuracy: 88.77% 	 Test Accuracy: 68.75%
[  9/ 25] 	 FashionMNIST 	 Time 0.02035s 	 Training Accuracy: 79.10% 	 Test Accuracy: 59.38%
[ 10/ 25] 	        MNIST 	 Time 0.03415s 	 Training Accuracy: 91.02% 	 Test Accuracy: 71.88%
[ 10/ 25] 	 FashionMNIST 	 Time 0.02111s 	 Training Accuracy: 80.66% 	 Test Accuracy: 65.62%
[ 11/ 25] 	        MNIST 	 Time 0.02066s 	 Training Accuracy: 93.26% 	 Test Accuracy: 68.75%
[ 11/ 25] 	 FashionMNIST 	 Time 0.02097s 	 Training Accuracy: 82.23% 	 Test Accuracy: 71.88%
[ 12/ 25] 	        MNIST 	 Time 0.02039s 	 Training Accuracy: 93.46% 	 Test Accuracy: 68.75%
[ 12/ 25] 	 FashionMNIST 	 Time 0.02096s 	 Training Accuracy: 83.79% 	 Test Accuracy: 68.75%
[ 13/ 25] 	        MNIST 	 Time 0.02109s 	 Training Accuracy: 94.73% 	 Test Accuracy: 62.50%
[ 13/ 25] 	 FashionMNIST 	 Time 0.02091s 	 Training Accuracy: 86.82% 	 Test Accuracy: 75.00%
[ 14/ 25] 	        MNIST 	 Time 0.02070s 	 Training Accuracy: 96.68% 	 Test Accuracy: 65.62%
[ 14/ 25] 	 FashionMNIST 	 Time 0.02682s 	 Training Accuracy: 86.04% 	 Test Accuracy: 71.88%
[ 15/ 25] 	        MNIST 	 Time 0.01987s 	 Training Accuracy: 97.66% 	 Test Accuracy: 68.75%
[ 15/ 25] 	 FashionMNIST 	 Time 0.02206s 	 Training Accuracy: 88.87% 	 Test Accuracy: 62.50%
[ 16/ 25] 	        MNIST 	 Time 0.02040s 	 Training Accuracy: 98.34% 	 Test Accuracy: 68.75%
[ 16/ 25] 	 FashionMNIST 	 Time 0.02056s 	 Training Accuracy: 89.75% 	 Test Accuracy: 71.88%
[ 17/ 25] 	        MNIST 	 Time 0.02081s 	 Training Accuracy: 98.83% 	 Test Accuracy: 68.75%
[ 17/ 25] 	 FashionMNIST 	 Time 0.02094s 	 Training Accuracy: 90.14% 	 Test Accuracy: 75.00%
[ 18/ 25] 	        MNIST 	 Time 0.02048s 	 Training Accuracy: 99.02% 	 Test Accuracy: 68.75%
[ 18/ 25] 	 FashionMNIST 	 Time 0.02077s 	 Training Accuracy: 88.09% 	 Test Accuracy: 68.75%
[ 19/ 25] 	        MNIST 	 Time 0.02040s 	 Training Accuracy: 99.22% 	 Test Accuracy: 62.50%
[ 19/ 25] 	 FashionMNIST 	 Time 0.02096s 	 Training Accuracy: 89.75% 	 Test Accuracy: 65.62%
[ 20/ 25] 	        MNIST 	 Time 0.02105s 	 Training Accuracy: 99.02% 	 Test Accuracy: 65.62%
[ 20/ 25] 	 FashionMNIST 	 Time 0.02067s 	 Training Accuracy: 89.06% 	 Test Accuracy: 62.50%
[ 21/ 25] 	        MNIST 	 Time 0.02052s 	 Training Accuracy: 99.12% 	 Test Accuracy: 62.50%
[ 21/ 25] 	 FashionMNIST 	 Time 0.02091s 	 Training Accuracy: 88.96% 	 Test Accuracy: 62.50%
[ 22/ 25] 	        MNIST 	 Time 0.02045s 	 Training Accuracy: 99.71% 	 Test Accuracy: 68.75%
[ 22/ 25] 	 FashionMNIST 	 Time 0.02075s 	 Training Accuracy: 92.38% 	 Test Accuracy: 68.75%
[ 23/ 25] 	        MNIST 	 Time 0.02851s 	 Training Accuracy: 99.22% 	 Test Accuracy: 65.62%
[ 23/ 25] 	 FashionMNIST 	 Time 0.02059s 	 Training Accuracy: 92.29% 	 Test Accuracy: 65.62%
[ 24/ 25] 	        MNIST 	 Time 0.02039s 	 Training Accuracy: 98.54% 	 Test Accuracy: 68.75%
[ 24/ 25] 	 FashionMNIST 	 Time 0.02092s 	 Training Accuracy: 88.38% 	 Test Accuracy: 65.62%
[ 25/ 25] 	        MNIST 	 Time 0.02041s 	 Training Accuracy: 99.22% 	 Test Accuracy: 65.62%
[ 25/ 25] 	 FashionMNIST 	 Time 0.02087s 	 Training Accuracy: 89.84% 	 Test Accuracy: 68.75%

[FINAL] 	        MNIST 	 Training Accuracy: 99.80% 	 Test Accuracy: 65.62%
[FINAL] 	 FashionMNIST 	 Training Accuracy: 89.84% 	 Test Accuracy: 68.75%

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 1.10.5
Commit 6f3fdf7b362 (2024-08-27 14:19 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
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (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

CUDA runtime 12.5, artifact installation
CUDA driver 12.5
NVIDIA driver 555.42.6

CUDA libraries: 
- CUBLAS: 12.5.3
- CURAND: 10.3.6
- CUFFT: 11.2.3
- CUSOLVER: 11.6.3
- CUSPARSE: 12.5.1
- CUPTI: 2024.2.1 (API 23.0.0)
- NVML: 12.0.0+555.42.6

Julia packages: 
- CUDA: 5.4.3
- CUDA_Driver_jll: 0.9.2+0
- CUDA_Runtime_jll: 0.14.1+0

Toolchain:
- Julia: 1.10.5
- LLVM: 15.0.7

Environment:
- JULIA_CUDA_HARD_MEMORY_LIMIT: 100%

1 device:
  0: NVIDIA A100-PCIE-40GB MIG 1g.5gb (sm_80, 2.170 GiB / 4.750 GiB available)

---

This page was generated using Literate.jl.