new docs

autoaug/autoaugment_learners/AaLearner.py

@@ -367,8 +367,6 @@ class AaLearner:
             accuracy (float): best accuracy reached in any
         """
 
-
-
         # we create an instance of the child network that we're going
         # to train. The method of creation depends on the type of 
         # input we got for child_network_architecture

docs/source/autoaug/aa_learners/autoaug.autoaugment_learners.AaLearner.rst

-:mod:`autoaug.autoaugment_learners`.AaLearner 
-==============================================
-
-.. currentmodule:: autoaug.autoaugment_learners
-
-.. autoclass:: AaLearner
-    :members:
\ No newline at end of file

docs/source/autoaug/aa_learners/autoaug.autoaugment_learners.EvoLearner.rst

-:mod:`autoaug.autoaugment_learners`.EvoLearner 
-==============================================
-
-.. currentmodule:: autoaug.autoaugment_learners
-
-.. autoclass:: EvoLearner
-    :members:
\ No newline at end of file

docs/source/autoaug/aa_learners/autoaug.autoaugment_learners.GruLearner.rst

-:mod:`autoaug.autoaugment_learners`.GruLearner 
-==============================================
-
-.. currentmodule:: autoaug.autoaugment_learners
-
-.. autoclass:: GruLearner
-    :members:
\ No newline at end of file

docs/source/autoaug/aa_learners/autoaug.autoaugment_learners.RsLearner.rst

-:mod:`autoaug.autoaugment_learners`.RsLearner 
-==============================================
-
-.. currentmodule:: autoaug.autoaugment_learners
-
-.. autoclass:: RsLearner
-    :members:
\ No newline at end of file

docs/source/autoaug/aa_learners/autoaug.autoaugment_learners.UcbLearner.rst

-:mod:`autoaug.autoaugment_learners`.UcbLearner 
-==============================================
-
-.. currentmodule:: autoaug.autoaugment_learners
-
-.. autoclass:: UcbLearner
-    :members:
\ No newline at end of file

docs/source/autoaug/autoaugment_learners.rst

-AutoAugment learners
---------------------
-
-.. toctree::
-   :maxdepth: 3
-   :caption: autoaugment_learners
-
-   aa_learners/autoaug.autoaugment_learners.AaLearner
-   aa_learners/autoaug.autoaugment_learners.EvoLearner
-   aa_learners/autoaug.autoaugment_learners.GruLearner
-   aa_learners/autoaug.autoaugment_learners.RsLearner
-   aa_learners/autoaug.autoaugment_learners.UcbLearner
\ No newline at end of file

docs/source/index.rst

@@ -5,32 +5,14 @@
 
 .. toctree::
    :maxdepth: 4
-   :caption: Usage (developers might find this useful):
+   :caption: How-to Guides:
 
-   usage/tutorial_for_team
-   usage/autoaugment_helperclass
+   howto/howto_main
 
 
 
 .. toctree::
    :maxdepth: 4
-   :caption: API Reference:
+   :caption: Explanations:
 
-   autoaug/autoaugment_learners
-   autoaug/aa_learners/autoaug.autoaugment_learners.AaLearner
-   autoaug/aa_learners/autoaug.autoaugment_learners.EvoLearner
-   autoaug/aa_learners/autoaug.autoaugment_learners.GruLearner
-   autoaug/aa_learners/autoaug.autoaugment_learners.RsLearner
-   autoaug/aa_learners/autoaug.autoaugment_learners.UcbLearner
-
-
-
-.. 
-   I've commented this out for now
-
-   Indices and tables
-   ==================
-
-   * :ref:`genindex`
-   * :ref:`modindex`
-   * :ref:`search`
+   explanations/autoaugment_learners

docs/source/usage/autoaugment_helperclass.rst

-AutoAugment object
-------------------
-
-######################################################################################################
-How to use a ``AutoAugment`` object to apply AutoAugment policies to ``Datasets`` objects
-######################################################################################################
-
-This is a page dedicated to demonstrating functionalities of :class:`AutoAugment`, which
-we use as a helper class to help us apply AutoAugment policies to datasets.
-
-This is a tutorial (in the sense describe in https://documentation.divio.com/structure/).
-
-For an example of how the material is used in our library, see the source code of
-:meth:`AaLearner._test_autoaugment_policy <autoaug.autoaugment_learners.AaLearner>`.
-
-Let's say we have a policy within the search space specified by the original 
-AutoAugment paper:
-
-.. code-block::
-
-    my_policy = [
-                        (("Invert", 0.8, None), ("Contrast", 0.2, 6)),
-                        (("Rotate", 0.7, 2), ("Invert", 0.8, None)),
-                        (("Sharpness", 0.8, 1), ("Sharpness", 0.9, 3)),
-                        (("ShearY", 0.5, 8), ("Invert", 0.7, None)),
-                        (("AutoContrast", 0.5, None), ("Equalize", 0.9, None))
-                        ]
-
-And that we also have a dataset that we want to apply this policy to:
-
-.. code-block::
-
-    train_dataset = datasets.MNIST(root='./datasets/mnist/train', train=True)
-    test_dataset = datasets.MNIST(root='./datasets/mnist/test', train=False,
-                                transform=torchvision.transforms.ToTensor())
-
-The ``train_dataset`` object will have an attribute ``.transform`` with the 
-default value ``None``.
-The ``.transform`` attribute takes a function which takes an image as an input 
-and returns a transformed image.
-
-We need a function which will apply the ``my_policy`` and we use 
-an ``AutoAugment`` for this job.
-
-.. code-block::
-    :caption: Creating an ``AutoAugment`` object and imbueing it with ``my_policy``.
-
-    aa_transform = AutoAugment()
-    aa_transform.subpolicies = my_policy
-    train_transform = transforms.Compose([
-                                            aa_transform,
-                                            transforms.ToTensor()
-                                        ])
-
-We can use ``train_transform`` as an image function:
-
-.. code-block::
-    :caption: This function call will return an augmented image
-
-    augmented_image = train_transform(original_image)
-
-We usually apply an image function to a ``Dataset`` like this:
-
-.. code-block::
-
-    train_dataset = datasets.MNIST(root='./datasets/mnist/train', train=True, transform=my_function)
-
-However, in our library we often have to apply a image function *after* the ``Dataset`` 
-object was already created. (For example, a ``Dataset`` object is created and trained on
-multiple times using different policies).
-In this case, we alter the ``.transform`` attribute:
-
-.. code-block::
-
-    train_dataset.transform = train_transform
-
-Now if we can create a ``DataLoader`` object from ``train_dataset``, it will automatically
-apply ``my_policy``.
\ No newline at end of file

docs/source/usage/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``.
-
-
-.. code-block::
-
-    import autoaug.autoaugment_learners as aal
-    import autoaug.child_networks as cn
-
-    import torchvision.datasets as datasets
-    import torchvision
-
-
-
-Defining the problem setting:
-
-.. code-block::
-
-    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)
-
-.. code-block::
-
-    # 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.
-
-.. code-block::
-    
-    print(aa_agent.history)
\ No newline at end of file