Reference
Talos exposes parameter sweeps through the Python Scan pattern and through SFD/CLI experiments. This section describes public callables, defaults, result fields and failure boundaries. Guides own end-to-end jobs; Developer owns maintenance and documentation proof.
Analyze
talos.Analyze reads experiment results for metric summaries, parameter comparisons and plots. talos.Reporting remains an alias for the same class. Import it from talos; training and model selection belong to Scan and Predict.
AutoModel
talos.autom8.AutoModel creates a five-argument training callback for Scan. Import it through talos.autom8. The helper builds and fits Keras architectures when Scan invokes .model; constructing AutoModel does not start a scan.
AutoParams
talos.autom8.AutoParams generates a parameter dictionary for Scan and provides methods for changing individual candidate lists. Import it through talos.autom8; it constructs parameters without training a model.
AutoPredict
talos.autom8.AutoPredict selects candidates from a completed Scan, scores them on held-out observations, and predicts with the winning trained model. Import it through talos.autom8. It is a function returning the same Scan object after adding evaluation and prediction fields.
AutoScan
talos.autom8.AutoScan combines AutoParams, AutoModel and Scan for early architecture exploration. Import it through talos.autom8. Default parameter generation covers common hyperparameters, network shapes, sizes and architectures; a small explicit parameter dictionary makes the first run bounded.
Backends
Talos runs standalone Keras, TensorFlow/tf.keras and native PyTorch callbacks through the same parameter-sweep, result and artifact interfaces. The core does not import a framework until training, prediction, serialization or a framework-specific helper requires it.
Deploy
talos.Deploy selects a trained model from Scan or a RunResult and writes a local ZIP archive for Restore. Import it from talos. Construction performs packaging immediately; it does not publish a model or start a service.
Energy draw callback
talos.callbacks.PowerDraw records sampled GPU power draw in watts at epoch begin and end. Import it from talos.callbacks; talos.utils.powerdrawappend adds the summary to a training History object.
Evaluate
talos.Evaluate scores a trained model from a completed Scan or RunResult on caller-supplied held-out observations. Import it from talos. Evaluate(scan_object) stores the run; its .evaluate() method selects one fitted model and scores subsets without retraining.
Generator
talos.utils.generator yields repeating NumPy batches; talos.utils.SequenceGenerator provides finite, indexed batches for Keras model.fit(). Import them from talos.utils. They batch caller-supplied arrays without acquiring data, splitting validation sets or changing a parameter sweep. Custom framework generators can be used in a training callback as usual.
Hidden layers and shapes
talos.model.hidden_layers appends Keras Dense and Dropout layers to a model being built inside a Scan callback. Import it from talos.model or talos.utils. It makes the number and width of hidden layers sweep parameters; it does not train the model or add the output layer.
Installation
Install Talos into an isolated environment, then choose the framework extra your model uses. The package exports the Python API and the talos command; optional frameworks are loaded only when required. This page owns installation choices rather than training or CLI configuration.
Learning-rate normalizer
talos.model.lr_normalizer converts a shared learning-rate scale to optimizer-specific values using fixed divisors. Import it from talos.model.normalizers, talos.model, or talos.utils. Invoke it explicitly when constructing an optimizer; adding an lr candidate to Scan does not apply normalization automatically.
Metrics
talos.utils.metrics provides lazy Keras-compatible training metrics. The same module is available as talos.metrics.keras_metrics. This page covers metrics supplied to model.compile(); Evaluate separately defines held-out F1 and MAE scoring after a sweep.
Monitoring
Talos exposes round progress, printed parameters, per-epoch CSV logs and Keras training plots. Scan owns round-level progress; callbacks observe the caller's training loop. This page covers talos.callbacks.TrainingPlot and ExperimentLog, plus Scan's output flags.
Predict
talos.Predict selects a trained model from Scan or a RunResult and performs inference on caller-supplied features. Import it from talos. Construction stores the run; prediction happens when .predict() or .predict_classes() is called.
Restore
talos.Restore reads a trusted Deploy ZIP archive and restores the selected fitted model, result table and experiment assets. Import it from talos. Construction extracts the archive and loads the model immediately; it does not run training.
Scan
talos.Scan configures and runs a parameter-sweep experiment synchronously through a five-argument training callback. Import it from talos. Hyperparameter candidates belong in params; execution, sampling, reduction and persistence options belong in Scan arguments.
Templates
talos.templates exposes dataset acquisition helpers, preset parameter dictionaries, model callbacks and complete Scan pipelines. Import it through talos.templates. These existing Python templates support education, testing and development; framework-specific SFD templates separately support CLI experiments.
Security
Supported versions
Third-party ownership
Talos declares Python runtime and optional framework dependencies in pyproject.toml, CI toolchains in requirements/ci/, and the documentation toolchain in docs-site/package.json and its lockfile. Review dependency licenses when adding or materially upgrading packages.