PyTorch lifecycle field guide

A compact lookup for the baseline lab.

Training batch

1. model.train() once per training phase.
2. Move model, features, and labels to one device.
3. optimizer.zero_grad().
4. logits = model(features).
5. loss = loss_fn(logits, labels).
6. loss.backward().
7. optimizer.step().

Evaluation batch

Use model.eval() and with torch.no_grad():. The first changes layer behavior; the second disables gradient recording.

Reproducible checkpoint

Save model and optimizer state dictionaries, epoch, seed, config, ordered labels, dataset checksum, preprocessing/weights identity, and best validation metric. Load with an explicit map_location. Recreate the model and optimizer before restoring state.

Device triage

Start with a CPU one-batch shape test. On GPU, record PyTorch/CUDA/device versions and reduce batch size before changing the model. Reproduction means matching inputs and configuration within a declared tolerance, not promising bitwise identity across hardware.