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.

First parameter sweep

Convert an existing Keras training function into a two-candidate Talos scan. The model code remains yours; Talos selects parameter combinations, records the results and retains model assets.

Get help and report issues

Use the documentation to isolate a failure, then use the Talos issue tracker for a reproducible bug or feature request. This page identifies the current repository routes; older community services are not required to use Talos.

Custom reducers

Pass a Python callable to Scan(reduction_method=…) to remove pending candidates using your own decision rule. The reducer receives the live Scan context and runs between trials; completed results remain recorded.

Bounded AutoML sweep

Run a two-trial binary Iris search using Talos AutoScan, AutoParams and built-in model architectures. The complete program combines the steps below.

AutoML: complete code

Run a two-trial binary Iris search using built-in Talos models. This page is the standalone companion to the walkthrough.

Keras sequence-generator sweep

Train a small convolutional model on real handwritten digits from scikit-learn, using a bounded offline dataset and a Talos sequence generator. The complete program combines the steps below.

Generator: complete code

Train two digit classifiers using a Talos sequence generator. This page is the standalone companion to the walkthrough.

Multiple-input Keras sweep

Train a functional Keras model that receives two aligned arrays, each containing a different pair of Iris features. The same pattern applies to larger multi-input models. The complete program combines the steps below.

Multiple inputs: complete code

Train a Keras model with two aligned feature arrays. This page is the standalone companion to the walkthrough.

Multiple-output Keras sweep

This example predicts two real outcomes from the Wisconsin Breast Cancer dataset: diagnosis and measured mean radius. The same multi-output pattern can model experiment outcomes, as in the original Telco Churn illustration of using hyperparameter optimization data to optimize hyperparameter optimization.

Multiple outputs: complete code

Train one Keras model with diagnosis and radius outputs. This page is the standalone companion to the walkthrough.

Native PyTorch sweep

Train two PyTorch networks while retaining the established Talos callback interface. The recipe records epoch metrics explicitly and provides a reconstruction factory for native Torch archives. The complete program guards training so archive restoration can import the model without starting another sweep.

PyTorch: complete code

Train two native PyTorch networks and retain their metrics and model factories. This page is the standalone companion to the walkthrough.

Typical Keras sweep

Train two small Iris classifiers while varying their hidden-layer activation. This is the starting recipe for adapting an existing Keras model; the complete program combines the same steps.

Typical model: complete code

Train two small Iris classifiers with different activations. This page is the standalone companion to the walkthrough.

Gamify

Use Scan(reduction_method="gamify") to inspect and change parameter-value statuses through a local JSON file between trials. This preserves the original human-in-the-loop search pattern: a researcher can remove pending values after observing completed results.

Local strategy

Use Scan(reductionmethod="localstrategy") to control pending work and live Scan settings between trials. local_strategy is similar to custom reducers with the difference that the optimization strategy can be changed during the experiment, as the function resides on the local machine.

Talos 2 migration

Talos 2 uses one search queue, executor and artifact format for Scan, native SFDs and the CLI. The core manages parameter experiments; data acquisition, model architecture and training remain caller-owned.

Optimization strategies

Talos supports several common optimization strategies:

Devices and parallel workers

Talos supports scenarios where on a single system one or more GPUs are handling one or more simultaneous jobs. GPU execution is handled by the selected TensorFlow, Keras or Torch backend. Install a backend build compatible with your device and drivers; see Backends. The examples are safe on CPU-only systems, where no GPU is configured.

Probabilistic reduction

Probabilistic reducers drop pending parameter combinations using completed model metrics. They formalize the research step of inspecting past results and deciding which values to stop trying. They can reduce completed work; whether that improves the experiment depends on the metric, sample size and search space.

SFD and manifest CLI

An SFD defines params(), prep(data, roundparams) and model(prepared, roundparams) in ordinary Python. Training remains your code. Model returns accept (history, model) or a dictionary containing numeric metrics plus model, history, optional predictions, backend and factory.

Model-search workflow

The goal of a deep learning experiment is to find one or more model candidates that meet a given performance expectation. Talos provides an API for both semi-automated and fully automated workflows. This guide connects the established path from an experiment idea to a model archive.

Support

Use the route matching the request: