Talos documentation
Talos runs parameter sweeps for Keras, TensorFlow and PyTorch models. Keep the established Python Scan interface or describe an experiment in a single-file definition (SFD) and run it through Python or the CLI. Both paths use the same experiment core and recorded run artifacts.
Start with your task
| I want to… | Start here |
|---|---|
| Run my first sweep | First parameter sweep, then the typical example |
| Keep an existing Talos model working | Migration, Scan, and backend contracts |
| Define a reproducible experiment for Python or the CLI | SFD and CLI |
| Use PyTorch, multiple inputs, multiple outputs or a generator | Model recipes |
| Reduce or control a running search | Optimization strategies, local strategy, and Gamify |
| Compare, evaluate or reuse model candidates | Workflow, Analyze, Evaluate, and Predict |
| Package or restore a trained model | Deploy and Restore |
| Maintain Talos or its documentation | Developer home and maintenance verification |
From model to recorded result
- Choose a supported backend environment. The core owns sweep execution; your framework owns model construction, training and device behavior.
- Supply data and a working training callback to Scan, or place the parameter, preparation and model functions in an SFD. Data loading and scientific preprocessing remain experiment code.
- Declare the candidate values and search policy. The parameter domain selects pending combinations; configured reducers or live controls may remove later candidates.
- The experiment runner prepares and trains each selected candidate. Backend adapters normalize its history and retain the trained model according to the backend contract.
- Inspect the first observable result in
scan.dataor the native run result, then open the run directory’sresults.csv. SFD and CLI describes manifests, source snapshots, round records, checkpoints and resume validation. - Analyze tuning results, evaluate selected candidates on held-out data, and predict or archive a chosen model. Talos ends at recorded experiments and reusable model assets; application serving and scientific conclusions remain yours.
The workflow guide connects these steps and explains when to repeat the search.
Documentation map
| Section | Canonical responsibility | Entry point |
|---|---|---|
| Overview | Product boundary, capabilities, workflow and direction | Capabilities and development priorities |
| Guides | Complete reader jobs, model recipes and operating a sweep | Guides |
| Reference | Public interfaces, defaults, return values and error boundaries | Reference |
| Developer | Contributions, verification, documentation and maintenance | Developer |
| Packages | Source ownership, public entry points and optional dependencies | Talos package |
Scope and responsibility
Talos owns parameter selection, experiment execution, result recording, reducer and intervention controls, checkpoint validation and supported model archive contracts. It accepts caller-provided arrays and preparation functions; it does not supply a finance data pipeline, decide whether a metric is scientifically suitable, establish causal validity or deploy an application server.
A seed controls supported sampling and recorded RNG state. Hardware, framework kernels, external services and opaque data streams can impose further reproducibility limits. Review backend compatibility, recovery contracts and maintenance evidence for the claims proven in each environment.
Read next
Start with the first sweep for the Python interface, SFD and CLI for a file-based experiment, or migration for an existing Talos project.
Cite an experiment
Use the citation guide to cite the software version and retain manifest and run identities with research artifacts.