AutoParams

talos.autom8.AutoParams generates a parameter dictionary for Scan and provides methods for changing individual candidate lists. Import it through talos.autom8; it constructs parameters without training a model.

Automatic optimizer generation imports the chosen framework backend. Install TensorFlow for the default backend='tensorflow', or select backend='keras' or backend='torch' with the corresponding extra. Native Torch optimizer lists still require a caller-owned Torch model; AutoModel builds Keras architectures.

Create a parameter dictionary

import talos
p = talos.autom8.AutoParams().params
assert 'optimizer' in p and 'network' in p

NOTE: The above example yields a very large permutation space so configure Scan() accordingly with fraction_limit.

Keep the helper object

param_object = talos.autom8.AutoParams()
assert isinstance(param_object.params, dict)

Methods on param_object change individual candidate lists. For example:

Modify a parameter

param_object.batch_size(min_size=20, max_size=100, steps=10)
assert param_object.params['batch_size'] == list(range(20, 100, 10))

Now the modified params dictionary can be accessed through param_object.params

Extend an existing dictionary

params_dict = talos.autom8.AutoParams(p.copy(), task='multi_label').params
assert params_dict['losses'] == ['categorical_crossentropy']

Declare task explicitly when generating presets for a prediction problem other than 'binary' (binary, multi_label, multi_class, or continuous).

Arguments

Argument	Default	Description
params	None	Create a dictionary, or start from the supplied dictionary.
task	'binary'	binary, multi_class, multi_label, or continuous; selects preset losses and output activations.
replace	True	Overwrite existing keys when parameter methods add values; False fills only missing keys.
auto	True	Generate all available parameter presets.
network	True	Generate architecture candidates; False sets network=['dense'].
resample_params	4	Keep at most this many values per parameter, or False to retain all values.
backend	'tensorflow'	Framework from which automatic optimizer classes are imported.

Parameter methods

The params property returns the parameter dictionary which can be used as an input to Scan().

The resample_params method accepts n and keeps at most that many candidate values for each parameter.

The remaining methods manipulate individual parameters in the dictionary.

activations For controlling the corresponding parameter in the parameters dictionary.

batch_size For controlling the corresponding parameter in the parameters dictionary.

dropout For controlling the corresponding parameter in the parameters dictionary.

epochs For controlling the corresponding parameter in the parameters dictionary.

kernel_initializers For controlling the corresponding parameter in the parameters dictionary.

last_activations For controlling the corresponding parameter in the parameters dictionary.

layers For controlling the corresponding parameter (i.e. hidden_layers) in the parameters dictionary.

losses For controlling the corresponding parameter in the parameters dictionary.

lr For controlling the corresponding parameter in the parameters dictionary.

networks For controlling the Talos present network architectures (dense, lstm, bidirectional_lstm, conv1d, and simplernn). NOTE: the use of preset networks requires the use of the input model from AutoModel() for Scan().

neurons For controlling the corresponding parameter (i.e. first_neuron) in the parameters dictionary.

optimizers For controlling the corresponding parameter in the parameters dictionary.

shapes For controlling the Talos preset network shapes (brick, funnel, and triangle).

shapes_slope For controlling the shape parameter with a floating point value to set the slope of the network from input layer to output layer.

Mutation and sampling

The signature is AutoParams(params=None, task='binary', replace=True, auto=True, network=True, resample_params=4, backend='tensorflow'). Methods modify .params and return None; .params is the dictionary passed to Scan. Automatic additions operate on a supplied dictionary before resampling creates a new dictionary, so pass a copy when preserving the original input matters.

resample_params(n) requires a positive integer and chooses up to n evenly spaced positions from each candidate list, including its endpoints. It does not sample trial combinations or assign probabilities. Integer range methods use Python's exclusive upper bound; float range methods use NumPy's exclusive upper bound.

Method	Default candidate control
layers(min_layers=0, max_layers=6, steps=1)	Hidden-layer count range.
dropout(min_dropout=0, max_dropout=.85, steps=.1)	Dropout fractions rounded to two decimal places.
neurons(min_neuron=8, max_neuron=None, steps=None)	Preset powers of two, or an integer range when maximum and step are supplied together.
batch_size(min_size=8, max_size=None, steps=None)	Preset batch sizes, or a supplied integer range.
epochs(min_epochs=50, max_epochs=None, steps=None)	Preset epoch counts, or a supplied integer range.
shapes_slope(min_slope=0, max_slope=.6, steps=.1)	Fractional contraction slopes.
shapes, optimizers, activations, losses, kernel_initializers, lr, networks, last_activations	Pass a candidate list, or keep the method's 'auto' preset.

Unsupported task names fail preset lookup. Empty or malformed candidate lists are not repaired by this helper. Even four values per parameter yield a large Cartesian space: use Scan's limits or a deliberately small dictionary before training.

Read next

Use AutoModel for the matching architecture callback, AutoScan to combine the presets, or optimization strategies to choose search limits.