Module 06 — Saving & Loading 💾
A trained model that only lives in memory is one crash away from gone. This module is how you make your work persistent, resumable, and shareable.
You just spent Module 05 teaching a model to be good — running epoch after epoch until its validation accuracy peaked. All of that learning lives in the model's weights, sitting in RAM. The moment your script ends or your machine reboots, it vanishes. Saving and loading is how you capture that hard-won knowledge to disk so you can reload it later for predictions, hand it to a teammate, deploy it to a server, or pick up training exactly where you left off. It's the unglamorous plumbing that turns a training experiment into a real, reusable asset.
PyTorch's approach is refreshingly simple once you learn the one golden rule: save the state_dict, not the whole model object. You met the state_dict back in Module 03 — it's just an ordered dictionary of every parameter's values. Saving that dictionary (rather than pickling the entire Python object) keeps your saved files portable and robust to code changes. From that single idea flow everything else: checkpoints that bundle the model and optimizer state so training can resume seamlessly, and device-aware loading so a model trained on a GPU can run on a laptop CPU. Recall the "which epoch was best?" question that ended Module 05 — this module gives you the tools to save exactly that model.
This module is split into three sub-modules — work through them in order.
📚 Sub-Modules
| # | Sub-Module | What you'll learn |
|---|---|---|
| 01 | state_dict Basics |
Saving/loading a state_dict with torch.save/torch.load — and why not to pickle the whole model |
| 02 | Checkpoints & Resuming | Bundling model + optimizer + epoch to pause and resume training seamlessly |
| 03 | Loading for Inference | map_location across devices, eval() discipline, and loading weights safely |
| 04 | Best Model & Early Stopping | Tracking the best validation score, saving that checkpoint, and stopping early |
🎯 By the end of this module, you'll be able to...
- Save and reload a model's weights the recommended, portable way.
- Create checkpoints that let you resume training from an exact point.
- Load a model onto any device for inference the safe way.
- Track and save the best-performing model and stop training early to avoid overfitting.
✅ Prerequisites
The state_dict and train()/eval() ideas from Module 03, and a trained model from the loop in Module 05.
⬅️ Prev module: 05 · The Training Loop · ➡️ Next module: 07 · Computer Vision