Guides
These guides take a model from its first parameter sweep to recorded results, controlled search and reusable model assets. They preserve the established Talos callback workflow alongside the SFD and CLI path.
Before you begin
Use a supported Python and backend installation, a writable experiment directory and a working training function. Each model recipe states its own dataset, framework and execution order. The bundled datasets used below are offline fixtures; a short recipe verifies an interface rather than the performance of a research model.
First sweep and migration
- First parameter sweep: compare a Keras model before and after adding Talos.
- Workflow: decide when to revise, evaluate or archive a search.
- Migration: keep current model callbacks and adopt corrected Talos 2 contracts.
- SFD and CLI: run file-based experiments, inspect artifacts and resume checkpoints.
Model recipes
| Model shape or task | Walkthrough | Complete runnable code |
|---|---|---|
| Typical Keras classification | Iris sweep | Complete Iris example |
| Multiple inputs | Aligned input arrays | Complete multiple-input example |
| Multiple outputs | Diagnosis and radius targets | Complete multiple-output example |
| Sequence generator | Handwritten digits | Complete generator example |
| Bounded AutoML | Binary Iris search | Complete AutoML example |
| Native PyTorch | Training history and factories | Complete PyTorch example |
The walkthroughs explain each step. Their complete-code companions are the corresponding standalone programs; use the walkthrough as the authority for prerequisites, data boundaries and interpretation.
Operate the search
- Choose a grid or sampled search.
- Configure probabilistic reduction or a custom reducer when the completed metrics support that decision.
- Use a local strategy or Gamify to change pending work during a run.
- Configure devices and worker shards for the environment where the callback executes.
A completed trial produces result rows and run artifacts. Reducers affect pending trials; device helpers configure a framework and do not launch concurrent workers. Each operational guide names the controls and limits it owns.
When a workflow fails
Check the backend contract before changing the search policy. Check Scan for callback arguments, data alignment and returned objects; check SFD and CLI for checkpoint identity and recovery errors. A reproducible bug report belongs in support.
Read next
Run the first sweep, or choose the recipe matching your current model. After training, continue with Analyze, Evaluate and Deploy.
Use the citation guide to identify the software and recorded experiment in a research paper.