Skip to main content

Talos capabilities

Key features​

Talos organizes model search around a training callback, a parameter dictionary and reusable result objects. The Python workflow retains the model code you already use; SFD and CLI execution add a file-based experiment definition.

  • perform a hyperparameter scan
  • perform a model architecture search
  • analyze and visualize results
  • evaluate results to find best model candidates
  • make predictions with selected models
  • archive supported models and run assets in zip files
  • restore models in a compatible runtime
  • create AutoML pipelines

Propose a feature in the issue tracker.

How to use​

Before running these examples, install the backend used by the callback and execute the linked Scan setup. The working directory must be writable because scans and archives create files.

Talos provides commands for conducting and analyzing experiments, evaluating candidates and making predictions. Archive commands transfer supported model assets between compatible systems; they do not start an application server.

The primary commands return a class object specific to the command, with various properties. These are outlined in the corresponding sections of the documentation.

Analyze(), Evaluate(), Predict(), and Deploy() accept the class object resulting from Scan() as input; Restore() accepts an archive path.

Conduct an experiment​

Examples below use the held-out Iris setup in Scan → Minimal Example. Run that setup first; it defines scan_object, p, input_model, x, y, x_val, y_val, x_test, and y_test.

scan_object = talos.Scan(x, y, p, input_model, 'overview',
                         x_val=x_val, y_val=y_val, seed=17, disable_progress_bar=True)

Analyze experiment results​

talos.Analyze(scan_object)

Prepare held-out evaluation​

talos.Evaluate(scan_object)

Prepare predictions​

talos.Predict(scan_object)

Create an archive​

talos.Deploy(scan_object, model_name='deployed_package', metric='val_loss', asc=True)

Restore the archive​

talos.Restore('deployed_package.zip')

In addition to the primary commands, various utilities can be accessed through talos.utils, datasets, parameters, and models from talos.templates, and AutoML features through talos.autom8.

Creating an experiment​

An experiment needs three inputs:

  • a hyperparameter dictionary
  • a working Keras, tf.keras or Torch training callback
  • a Talos experiment configuration

1. Prepare the input model​

In order to prepare a Keras model for a Talos experiment, you simply replace parameters you want to include in the scan, with references to the parameter dictionary.

2. Define the parameter space​

In a regular Python dictionary, you declare the hyperparameters and the boundaries you want to include in the experiment.

3. Configure the experiment​

To start the experiment, you input the parameter dictionary and the Keras model into Talos with the option for Grid, Random, or Probabilistic optimization strategy.

4. Inspect the completed trials​

The completed scan exposes trial metrics and selected parameter values in scan_object.data; the run directory retains CSV results and recovery artifacts. Use those observations to decide whether to revise the search or evaluate a selected candidate on untouched data.

Outputs and failure boundaries​

Analyze, Evaluate and Predict construct helpers around the completed run; constructing them does not itself run held-out evaluation or prediction. Use their reference methods for those operations. Deploy writes deployed_package.zip; Restore loads that archive using the matching framework and, when required, an importable model factory.

Data loading, preprocessing, training and scientific interpretation remain caller responsibilities. A missing metric, incompatible model input shape or unavailable archive factory needs correction at that boundary; changing the search policy will not repair it. See Scan, Backends and Restore.

Run the first parameter sweep, follow the workflow, or move an existing callback to SFD and CLI.