A simple PyTorch CNN trained on the CIFAR-10 dataset. Comes in a few different flavors:
models/train_basic.py—simple no-frills training script.models/train.py—training script with optional parameters, useful for e.g. hyperparameter search. Also incorporates TensorBoard support.models/distributed_train.py—training script with Horovod support, to demonstrate Spell's distributed training feature.models/wandb_train.py—training script leveraging Spell's Weights & Biases integration.servers/serve.py—example model server script compatible with any of the models in this repo.
Refer to the notebooks/ folder for implementation notes.
See also the Spell quickstart.