tutorial_for_team.rst

aa_learner object and its children

This is a page dedicated to demonstrating functionalities of :class:`aa_learner`.

This is a how-to guide (in the sense describe in https://documentation.divio.com/structure/).

How to use the aa_learner class to find an optimal policy for a dataset-child_network pair

This section can also be read as a .py file in ./tutorials/how_use_aalearner.py.

import MetaAugment.autoaugment_learners as aal
import MetaAugment.child_networks as cn

import torchvision.datasets as datasets
import torchvision

Defining the problem setting:

train_dataset = datasets.MNIST(root='./MetaAugment/datasets/mnist/train',
                                train=True, download=True, transform=None)
test_dataset = datasets.MNIST(root='./MetaAugment/datasets/mnist/test',
                        train=False, download=True, transform=torchvision.transforms.ToTensor())
child_network = cn.lenet

Note

It is important not to type

child_network = cn.lenet()

We need the child_network variable to be a type object, not a nn.Module object because the child_network will be called multiple times to initialize a nn.Module of its architecture multiple times. This is because every time we need to evaluate a different policy, we need to train another new network of the same architecture.

Using the random search learner to evaluate randomly generated policies: (You can use any other learner in place of random search learner as well)

# aa_agent = aal.gru_learner()
# aa_agent = aal.evo_learner()
# aa_agent = aal.ucb_learner()
# aa_agent = aal.ac_learner()
aa_agent = aal.randomsearch_learner()
aa_agent.learn(train_dataset, test_dataset, child_network)

Viewing the results:

.history is a list containing all the policies tested and the respective accuracies obtained when trained using them.

print(aa_agent.history)