tutorial_for_team.rst

AaLearner object and its children

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

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

How to use the AaLearner 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 autoaug.autoaugment_learners as aal
import autoaug.child_networks as cn

import torchvision.datasets as datasets
import torchvision

Defining the problem setting:

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

Warning

In earlier versions, we had to write child_network_architecture=cn.LeNet and not child_network_architecture=cn.LeNet(). But now we can do both. Both types of objects can be input into AaLearner.learn().

More precisely, the child_network_architecture parameter has to be either as nn.Module, a function which returns a nn.Module, or a type which inherits nn.Module.

A downside (or maybe the upside??) of doing one of the latter two is that the same randomly initialized weights are used for every policy.

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.GruLearner()
# aa_agent = aal.EvoLearner()
# aa_agent = aal.UcbLearner()
# aa_agent = aal.ac_learner()
aa_agent = aal.RsLearner(
                                sp_num=7,
                                toy_size=0.01,
                                batch_size=4,
                                learning_rate=0.05,
                                max_epochs=float('inf'),
                                early_stop_num=35,
                                )
aa_agent.learn(train_dataset,
            test_dataset,
            child_network_architecture=child_network_architecture,
            iterations=15000)

You can set further hyperparameters when defining a AaLearner.

Also, depending on what learner you are using, there might be unique hyperparameters. For example, in the GRU learner you can tune the exploration parameter alpha.

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)