Skip to main content

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.

Install a Keras or TensorFlow backend before using it. Its default backend is tensorflow. Currently there are five supported architectures:

  • conv1d
  • lstm
  • bidirectional_lstm
  • simplernn
  • dense

AutoModel creates an input model for Scan(). Optimized for being used together with AutoParams() and expects one or more of the above architectures to be included in params dictionary, for example:

import talos
from sklearn.datasets import load_iris
x, y = load_iris(return_X_y=True)
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['network'] = ['dense', 'conv1d', 'lstm']
input_model = talos.autom8.AutoModel(task='multi_class', experiment_name='iris_architectures').model
scan_object = talos.Scan(x.astype('float32'), y, params=p, model=input_model,
                         experiment_name='iris_architectures', round_limit=3, seed=17)
assert len(scan_object.data) == 3

Arguments​

ArgumentDefaultDescription
taskrequiredbinary, multi_label, multi_class, or continuous for runnable presets. The constructor also accepts None with a metric list, subject to the output-layer boundary below.
experiment_namerequiredName shared with Scan(), used by the epoch-log callback.
metricNoneKeras metric names or objects in a list when task=None.
backend'tensorflow'TensorFlow/tf.keras or standalone 'keras'.

Setting task controls metric selection and output-layer construction. Choose a supported prediction task for a runnable preset.

These architecture presets build Keras models. For native PyTorch training, supply a Torch callback or use the Torch SFD template.

Callback result and boundaries​

The constructor signature is AutoModel(task, experiment_name, metric=None, backend='tensorflow'). Its .model has the callback signature model(x_train, y_train, x_val, y_val, params) and returns (history, fitted_model). The example produces three result rows and an epoch log for each trial in the run directory.

Supply one permutation's scalar values to the callback, including the architecture, width, hidden-layer shape, dropout, optimizer class, normalized learning rate, loss, batch size, epochs and output activation. AutoParams supplies the expected names. Non-dense presets reshape two-dimensional inputs to (rows, features, 1) internally; prediction inputs must match the fitted architecture's input shape.

Classification tasks use Talos F1 and accuracy metrics; continuous uses MAE and accuracy. When task=None, metric must be a list and the constructor also adds accuracy, but the current generated callback has no task-independent output-layer rule: fitting that preset raises ValueError for an unknown model task. Use a caller-owned callback for a custom task or metric contract. Native Torch backends raise ValueError; framework shape, loss and parameter errors propagate from the callback. This helper is for architecture exploration and does not validate the scientific suitability of those presets for a dataset.

Bound the candidate space with AutoParams and Scan, or use SFD and CLI for an explicit model implementation.