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.

This page covers framework selection and normalized callback results. Installation owns dependency extras, Scan owns the five-argument Python interface, and SFD and CLI owns structured experiments. Dataset acquisition and framework-specific training loops remain caller responsibilities.

TensorFlow

Install talos[tensorflow] on Python 3.11 or newer (TensorFlow 2.20+, Keras 3.15+). Existing five-argument callbacks returning (history, model) work unchanged. The separate legacy-tensorflow lane preserves TensorFlow 2.14 / Keras 2.14 applications on Python 3.10–3.11. This compatibility lane retains known upstream advisories; modern extras are the default for new work.

Keras

On Python 3.11 or newer, install talos[keras,tensorflow], or talos[keras,torch] with KERAS_BACKEND=torch. Set the backend before importing Keras. Native .keras artifacts retain trained weights; pass custom_objects for custom layers/metrics when required.

PyTorch

Install talos[torch] (Torch 2.13+) on Python 3.10 or newer. Return (history, model) or a structured SFD result. The documented historical return of a network carrying .history plus its parameter iterator is normalized to the trained network.

For portable restoration, provide an importable factory and constructor configuration, as demonstrated in the Torch SFD. Native artifacts use state_dict; nested class instances need an explicit restoration factory. Prediction preserves the network's existing training mode around inference.

See migration for archive and resume behavior.

Callback and adapter interfaces

Import the user-facing entry points from talos. A Scan callback accepts (x_train, y_train, x_val, y_val, params) and returns (history, fitted_model). History may be a framework History object or a dictionary mapping metric names to sequences of numeric scalar values. Each trial's summary uses the final value of each nonempty sequence; .round_history retains every supplied epoch value.

SFD model functions may instead return a dictionary with metrics, history, model, backend and model_factory fields. Metric-only experiments are valid, but predicting or deploying requires a trained model. Invalid histories, non-string metric names and nonnumeric metric values raise TypeError before the trial is committed.

For custom integrations, talos.backends.backend_for(model=None, backend=None) returns the inferred or named adapter, and normalise_result(output, backend=None, model_factory=None) validates callback output. Supported framework hints are keras, tensorflow and torch, with tf/tf.keras and pytorch aliases. A hint can be supplied through Scan's backend keyword or an SFD's backend attribute. Unsupported hints raise ValueError.

Framework	Prediction	Native artifact and restoration
Standalone Keras	Calls model.predict() with verbose=0 by default.	.keras with an integrity digest; loads with compile=False. Custom objects may be supplied explicitly.
TensorFlow/tf.keras	Same Keras prediction contract.	.keras, plus historical JSON/H5 compatibility in Restore.
Native Torch	Uses model.predict() if provided; otherwise calls the module under evaluation mode and torch.no_grad().	CPU-copied .pt state dictionary plus constructor configuration and factory identity.

Torch arrays are moved to the model's device and converted to float32; existing tensors retain their dtype. List/tuple inputs are positional module inputs and dictionaries are keyword inputs. Standard module outputs are converted back to NumPy, including nested structures. The adapter restores the original training flag after successful inference. A custom .predict() method owns its own inference behavior.

For Torch portability, use an importable class or function with a constructor configuration dictionary (talos_config), or pass model_factory=(factory, configuration). Generic containers such as torch.nn.Sequential need an explicit factory. Model and source checksum failures stop restoration; compatible installed framework versions are still required on the destination.

Minimal adapter check

talos.backends.backend_for(backend='torch').name returns 'torch' without fitting a model or importing Torch. This checks adapter dispatch only; the native Torch walkthrough verifies a complete training callback and restored predictions.

Read next

Start with the bounded Scan example, use the native Torch walkthrough, or migrate an existing model to an SFD.