Tutorial

Fashion-MNIST classification

The current implicit-autograd classifier over the complete Fashion-MNIST training and held-out test splits.

oa::MatrixImplicit autogradCheckpoint manager

Contract	Value
Dataset	60,000 train and 10,000 held-out test images
Model	784 → Linear(128, ReLU) → Linear(10)
Training	5 epochs, batch 64, AdamW lr=0.001
Pass criteria	Loss decreases; test accuracy >70%; checkpoint accuracy within 0.5 points

Model

1class MnistClassifier : public oa::Module {
2public:
3    MnistClassifier() {
4        fc1_ = oa::makeShared<oa::Linear>(784, 128);
5        fc1_->setActivation(oa::Activation::Relu);
6        fc2_ = oa::makeShared<oa::Linear>(128, 10);
7        registerModule("fc1", fc1_);
8        registerModule("fc2", fc2_);
9    }
10
11    oa::Matrix forward(const oa::Matrix& input) override {
12        auto normalized = oa::FnMatrix::scale(input, 1.0F / 255.0F);
13        return fc2_->forward(fc1_->forward(normalized));
14    }
15};
16

Training

The batch ring is sized from MaxAsyncSubmissions(), allowing GPU work to overlap CPU sampling while preserving matrix lifetimes.

1while (not training.loop.isDone()) {
2    oa::Matrix batchX;
3    oa::Matrix batchY;
4    if (not trainLoader.nextBatch(batchX, batchY)) {
5        trainLoader.reset();
6        trainLoader.nextBatch(batchX, batchY);
7    }
8
9    optimizer->zeroGrad();
10    oa::GradientTape tape;
11    auto logits = model->forward(batchX);
12    auto loss = oa::FnLoss::crossEntropy(logits, batchY);
13    tape.backward(loss);
14    training.loop.next(loss);
15}
16training.loop.finish();
17

Real held-out evaluation

Accuracy is computed over all 10,000 test images, not the final training minibatch.

Capability-aware precision

Weight initialization follows the active OA weight dtype; execution remains selected by device capability.

Measured training loop

oa::ItTraining owns progress, wall/GPU timing, optimizer completion, and summary metrics.

Artifact verification

oa::CheckpointManager saves model and AdamW state, reloads the best checkpoint, and re-runs evaluation.

Build and run

cmake --build build/release --target TutorialMnistClassifierAg -j
OA_MNIST_DATA=/path/to/FashionMNIST/raw   ./bin/release/sdk/tutorials/ml/tuMnistClassifierAg

C++ source Python source