Examples¶
The maintained notebooks progress from a small synthetic regression problem to complete computer-vision, reinforcement-learning, and experiment-tracking workflows. Each notebook includes an installation cell, reports the installed torch-batteries version, and can be opened directly in Google Colab.
| Notebook | What it demonstrates | Expected workload | Links |
|---|---|---|---|
| Function Fitting with MLP | Training, validation, testing, prediction, metrics, and result inspection on synthetic data | CPU-friendly; no downloads; normally under a minute |   |
| Iris Classification with MLP | Hugging Face dataset loading, deterministic DataPack splits, implicit loaders, and held-out classification | CPU-friendly; tiny download; normally a few seconds | |
| Image Classification with CNN | MNIST DataPack lifecycle, implicit loaders, early stopping, and model checkpoints | Downloads MNIST; accelerator recommended | |
| Learning Rate Sweep with Early Stopping | Comparing learning rates, eager early stopping, and offline W&B experiment tracking | Downloads MNIST; trains five models; accelerator recommended | |
| FashionMNIST Diffusion | Gradient accumulation, mixed precision, clipping, scheduling, and streaming generation | Downloads FashionMNIST; CUDA or MPS strongly recommended | |
| CartPole Reinforcement Learning | DQN training, optimization events, stateful metrics, scheduling, and predict_iter() |
CPU-compatible; no dataset or credentials | |
| CIFAR-10 Transfer Learning | Resumable callbacks, metric types, top-k checkpoints, save/load, resume modes, and structured prediction | Downloads CIFAR-10 and ResNet18 weights; accelerator recommended | |
Where to start¶
Start with Function Fitting with MLP for the shortest direct-DataLoader introduction. Continue with Iris Classification with MLP for a compact DataPack workflow, then Image Classification with CNN for DataPack, callbacks, and checkpoints. Choose a specialized notebook according to the feature you want to explore.
The notebooks install their own example dependencies when run in a fresh notebook environment. Some datasets and pretrained weights are downloaded on first use; each notebook describes its prerequisites and expected hardware before running any training code.