PyTorch · ~1 month

A PyTorch train/val loop on toy moons

Can I run a standard PyTorch train/eval loop on small data, keep a checkpoint, and write the metrics down?

Data: sklearn.datasets.make_moons — 1,000 samples, 20% noise, stratified 80/20 train/val.
Setup: MLP, hidden sizes 32 and 16, ReLU. Adam (lr=0.01), cross-entropy, batch 32, 50 epochs on CPU. Checkpoint by validation accuracy.
What PyTorch handled: Autograd via loss.backward()
torch.optim.Adam
DataLoader
nn.Sequential
Also in the repo: CONCEPTS.md on autograd and optimizers
Optional: python scripts/train_fastai.py --epochs 50
Numbers: Final trainloss 0.071 · accuracy 97.1%
Final validationloss 0.029 · accuracy 98.5%
Best val accuracy1.00 at epoch 41
Figures: 1,000 points, 20% noise. Not linearly separable.
Train vs validation loss and accuracy. Val beating train is a small-set quirk on this data.
Validation-set confusion matrix.
What I take from it: The loop and checkpointing work end to end. Val hitting 1.00 just means the moons are an easy shape to separate.
Reproduce: make reproduce # or: python scripts/train_torch.py --epochs 50