autoaug.autoaugment_learners.UcbLearner

class autoaug.autoaugment_learners.UcbLearner(num_sub_policies=5, p_bins=11, m_bins=10, exclude_method=[], batch_size=8, toy_size=1, learning_rate=0.1, max_epochs=inf, early_stop_num=30, num_policies=100)

Uses the UCB1 algorithm originally developed for multi-armed bandit problems. Recommended when

• Optimal hyperparameters for training the CNN (such as batch size and learning rate) has not been found yet.

• Using a toy dataset or a toy CNN

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 11.

• m_bins (int, optional) – number of bins we divide the magnitude space. Defaults to 10.

• 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-2.

• 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_policies (int, optional) – Number of policies we want to serach over. Defaults to 100.

history

list of policies that has been input into self._test_autoaugment_policy as well as their respective obtained accuracies

Type

list

augmentation_space

list of image functions that the user has chosen to include in the search space.

Type

list

policies

A list of policies which we are currently searching over.

Type

list

avg_accs

A list where the nth element indicates the average accuracy obtained by the nth policy.

Type

list

Notes

As opposed the the other learners, this searches over a subset of the entire search space (specified in the AutoAugment paper). The size of the subset is initialized to be self.num_policies. But we can increase it by running self.make_more_policies(). For example, we initialize the learner with self.num_policies=7, run self.learn(iterations=20) to learn about the seven policies we have in our self.policies. Then run self.make_more_policies(n=5) to add 5 more policies to self.policies. Then we can run self.learn(iterations=20) to continue the UCB1 algorithm with the extended search space.

References

Peter Auer, et al.

“Finite-time Analysis of the Multiarmed Bandit Problem” https://homes.di.unimi.it/~cesabian/Pubblicazioni/ml-02.pdf

get_mega_policy(number_policies=5)

Produces a mega policy, based on the n best subpolicies (evo learner)/policies (other learners)

Parameters

• int (number_policies ->) – Number of (sub)policies to be included in the mega

• policy –

Returns

megapolicy -> [subpolicy, subpolicy, …]

get_n_best_policies(number_policies=5)

returns the n best policies

Parameters

number_policies (int) – Number of (sub)policies to return

Returns

list of best n policies

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

continue the UCB algorithm for iterations number of turns

make_more_policies(n)

generates n more random policies and adds it to self.policies

Parameters

n (int) – how many more policies to we want to randomly generate and add to our list of policies