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.

Install TensorFlow for the default callback, or choose an installed Keras backend through start(backend=...). The caller owns dataset acquisition and scientific train/validation/test separation.

Constructing AutoScan stores configuration. Calling its start() method runs the experiment synchronously and returns the Scan object.

import talos
from sklearn.datasets import load_iris
x, y = load_iris(return_X_y=True)
auto = talos.autom8.AutoScan(task='multi_class', experiment_name='iris_autoscan', max_param_values=2)
p = talos.autom8.AutoParams(task='multi_class', network=False, resample_params=1).params
p.update({'epochs': [1], 'first_neuron': [8], 'hidden_layers': [0],
          'batch_size': [16], 'dropout': [0.], 'losses': ['sparse_categorical_crossentropy'],
          'kernel_initializer': ['glorot_uniform'], 'activation': ['relu'],
          'shapes': ['brick'], 'lr': [1.]})
p['activation'] = ['relu', 'elu']
scan_object = auto.start(x.astype('float32'), y, params=p, round_limit=2, seed=17)
assert len(scan_object.data) == 2

NOTE: auto.start() accepts all Scan() arguments.

Arguments

Argument	Default	Description
task	required	binary, multi_label, multi_class, continuous, or None with a caller-supplied model.
experiment_name	required	Name used by the resulting Scan and default epoch logs.
max_param_values	None	Additional per-parameter resampling limit when parameters are generated automatically.

Set task according to the prediction problem. For custom metrics or task=None, supply both an explicit parameter dictionary and a caller-owned training callback through auto.start(params=..., model=...); the automatic presets require a supported task.

Start the experiment

The signature is AutoScan(task, experiment_name, max_param_values=None) followed by start(x, y, **kwargs). Start accepts:

Argument	Input	Description
x	array or list of arrays	prediction features
y	array or list of arrays	prediction outcome variable
kwargs	arguments	any Scan() argument can be passed into AutoScan.start()

Defaults and boundaries

An explicit params dictionary bypasses AutoParams and max_param_values. Otherwise AutoParams first applies its default four-values-per-parameter resampling, then max_param_values may reduce those lists further. An explicit model bypasses AutoModel. Remaining keyword arguments are forwarded to Scan; experiment_name comes from the AutoScan constructor.

The example completes two trials, returns .data with two result rows, and writes the same local artifacts as Scan. With generated presets, include round_limit or another search limit before starting. Native Torch training needs a supplied model callback because the default AutoModel rejects native Torch architecture generation.

Invalid presets, missing model parameters, framework errors and Scan input errors propagate. This helper chooses presets; it does not infer a task from labels or select a scientific evaluation protocol.

Read next

Use Scan for all execution arguments, AutoParams to narrow the space, or AutoPredict to compare fitted candidates on held-out data.