autoaug.autoaugment_learners.EvoLearner

class autoaug.autoaugment_learners.EvoLearner(num_sub_policies=5, p_bins=1, m_bins=1, exclude_method=[], learning_rate=0.1, max_epochs=inf, early_stop_num=20, batch_size=8, toy_size=1, num_solutions=5, num_parents_mating=3, controller=<class 'autoaug.controller_networks.EvoController.EvoController'>)

Evolutionary Strategy learner

This learner generates neural networks that predict optimal augmentation policies. Hence, there is no backpropagation or gradient descent. Instead, training is done by randomly changing weights of the ‘parent’ networks, where parents are determined by their ability to produce policies that increase the accuracy of the child network.

Parameters

• num_sub_policies (int, optional) – number of subpolicies per policy. Defaults to 5.

• p_bins (int, optional) – number of bins we divide the interval [0,1] for probabilities. e.g. (0.0, 0.1, … 1.0) Defaults to 1.

• m_bins (int, optional) – number of bins we divide the magnitude space. Defaults to 1.

• exclude_method (list, optional) – list of names(:type:str) of image operations the user wants to exclude from the search space. Defaults to [].

• batch_size (int, optional) – child_network training parameter. Defaults to 32.

• toy_size (int, optional) – child_network training parameter. ratio of original dataset used in toy dataset. Defaults to 0.1.

• learning_rate (float, optional) – child_network training parameter. Defaults to 1e-1.

• max_epochs (Union[int, float], optional) – child_network training parameter. Defaults to float(‘inf’).

• early_stop_num (int, optional) – child_network training parameter. Defaults to 20.

• num_solutions (int, optional) – Number of offspring spawned at each generation of the algorithm. Default 5

• num_parents_mating (int, optional) – Number of networks chosen as parents for the next generation of networks Defaults to 3

• controller (nn.Module, optional) – Controller network for the evolutionary algorithm. Defaults to cont_n.EvoController

Notes

The Evolutionary algorithm runs in generations, and so batches of child networks are trained at specific time intervals.

Examples

from autoaug.autoaugment_learners.EvlLearner import EvoLearner evo_learner = EvoLearner()

learn(train_dataset, test_dataset, child_network_architecture, iterations=15, return_weights=False)

Runs the GA instance and returns the model weights as a dictionary

Parameters

bool (return_weights ->) – Determines if the weight of the GA network should be returned

Returns

• If return_weights – Network weights -> dict

• Else – Solution -> Best GA instance solution

  Solution fitness -> float

  Solution_idx -> int