diff --git a/autoaug/autoaugment_learners/AaLearner.py b/autoaug/autoaugment_learners/AaLearner.py
index f14a107ebc5ac73c6c6fe1e9202936c652ab50d8..c0566d4edba29079f70e1d3e0334393894796a03 100644
--- a/autoaug/autoaugment_learners/AaLearner.py
+++ b/autoaug/autoaugment_learners/AaLearner.py
@@ -13,32 +13,51 @@ import types
class AaLearner:
- """
- The parent class for all AaLearner's
-
- Attributes:
- op_tensor_length (int): what is the dimension of the tensor that represents
- each 'operation' (which is made up of fun_name, prob,
- and mag).
-
- See Also
- --------
+ """The parent class for all AaLearner's
+ Contains utility methods that child AaLearner's use.
- Notes
- -----
+ Args:
+ sp_num (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.
- References
- ----------
-
+ m_bins (int, optional): number of bins we divide the magnitude space.
+ Defaults to 10.
+
+ discrete_p_m (bool, optional):
+ Whether or not the agent should represent probability and
+ magnitude as discrete variables as the out put of the
+ controller (A controller can be a neural network, genetic
+ algorithm, etc.). Defaults to False
- Examples
- --------
+ 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.
+
+ exclude_method (list, optional): list of names(:type:str) of image operations
+ the user wants to exclude from the search space. Defaults to [].
+
+
+ Attributes:
+ history (list): list of policies that has been input into
+ self._test_autoaugment_policy as well as their respective obtained
+ accuracies
+
+ augmentation_space (list): list of image functions that the user has chosen to
+ include in the search space.
"""
+
def __init__(self,
# parameters that define the search space
sp_num=5,
@@ -53,29 +72,7 @@ class AaLearner:
early_stop_num=20,
exclude_method = [],
):
- """
- Args:
- sp_num (int, optional): number of subpolicies per policy. Defaults to 5.
- fun_num (int, optional): number of image functions in our search space.
- Defaults to 14.
- p_bins (int, optional): number of bins we divide the interval [0,1] for
- probabilities. Defaults to 11.
- m_bins (int, optional): number of bins we divide the magnitude space.
- Defaults to 10.
- discrete_p_m (bool, optional):
- Whether or not the agent should represent probability and
- magnitude as discrete variables as the out put of the
- controller (A controller can be a neural network, genetic
- algorithm, etc.). Defaults to False
-
- 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.
- """
+
# related to defining the search space
self.sp_num = sp_num
self.p_bins = p_bins
@@ -298,31 +295,36 @@ class AaLearner:
2. <see how good that policy is>
3. <save how good the policy is in a list/dictionary and
(if applicable,) update the controller (e.g. RL agent)>
-
- Args:
- train_dataset (torchvision.dataset.vision.VisionDataset)
- test_dataset (torchvision.dataset.vision.VisionDataset)
- child_network_architecture (Union[function, nn.Module]):
- NOTE This can be both, for example,
- MyNetworkArchitecture
- and
- MyNetworkArchitecture()
- iterations (int): how many different policies do you want to test
- Returns:
- none
-
-
- If child_network_architecture is a <function>, then we make an
- instance of it. If this is a <nn.Module>, we make a copy.deepcopy
+
+ If ``child_network_architecture`` is a ``<function>``, then we make an
+ instance of it. If this is a ``<nn.Module>``, we make a ``copy.deepcopy``
of it. We make a copy of it because we we want to keep an untrained
(initialized but not trained) version of the child network
architecture, because we need to train it multiple times
- for each policy. Keeping child_network_architecture as a `function` is
- potentially better than keeping it as a nn.Module because every
+ for each policy. Keeping ``child_network_architecture`` as a ``<function>`` is
+ potentially better than keeping it as a ``<nn.Module>`` because every
time we make a new instance, the weights are differently initialized
which means that our results will be less biased
(https://en.wikipedia.org/wiki/Bias_(statistics)).
-
+
+
+ Args:
+ train_dataset (torchvision.dataset.vision.VisionDataset):
+ test_dataset (torchvision.dataset.vision.VisionDataset):
+ child_network_architecture (Union[function, nn.Module]):
+ This can be both, for example,
+
+ ``LeNet``
+
+ and
+
+ ``LeNet()``
+
+ iterations (int): how many different policies do you want to test
+
+ Returns:
+ none
+
Example code:
@@ -332,7 +334,7 @@ class AaLearner:
for _ in range(15):
policy = self._generate_new_policy()
- pprint(policy)
+ print(policy)
reward = self._test_autoaugment_policy(policy,
child_network_architecture,
train_dataset,
@@ -449,11 +451,11 @@ class AaLearner:
Args:
- number_policies -> int: Number of (sub)policies to be included in the mega
+ number_policies (int): Number of (sub)policies to be included in the mega
policy
Returns:
- megapolicy -> [subpolicy, subpolicy, ...]
+ megapolicy ([subpolicy, subpolicy, ...])
"""
number_policies = max(number_policies, len(self.history))
diff --git a/autoaug/autoaugment_learners/GruLearner.py b/autoaug/autoaugment_learners/GruLearner.py
index 6db3ba5cc5f75b81947ebe52b9df32bcd67194c2..c2ef521e58c4bb3f473fe4ca33a76830deaebeb5 100644
--- a/autoaug/autoaugment_learners/GruLearner.py
+++ b/autoaug/autoaugment_learners/GruLearner.py
@@ -24,21 +24,64 @@ class GruLearner(AaLearner):
(https://arxiv.org/abs/1412.3555), which is why we substituted
the LSTM for the GRU.
-
- See Also
- --------
+ Args:
+ sp_num (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.
- Notes
- -----
+ m_bins (int, optional): number of bins we divide the magnitude space.
+ Defaults to 10.
+ discrete_p_m (bool, optional):
+ Whether or not the agent should represent probability and
+ magnitude as discrete variables as the out put of the
+ controller (A controller can be a neural network, genetic
+ algorithm, etc.). Defaults to False
- References
- ----------
+ 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.
+
+ exclude_method (list, optional): list of names(:type:str) of image operations
+ the user wants to exclude from the search space. Defaults to [].
+
+ alpha (float, optional): Exploration parameter. It is multiplied to
+ operation tensors before they're softmaxed.
+ The lower this value, the more smoothed the output
+ of the softmaxed will be, hence more exploration.
+ Defaults to 0.2.
+
+ cont_mb_size (int, optional): Controller Minibatch Size. How many
+ policies do we test in order to calculate the
+ PPO(proximal policy update) gradient to update
+ the controller. Defaults to
+ Attributes:
+ history (list): list of policies that has been input into
+ self._test_autoaugment_policy as well as their respective obtained
+ accuracies
+
+ augmentation_space (list): list of image functions that the user has chosen to
+ include in the search space.
- Examples
- --------
+ References
+ ----------
+ Ekin D. Cubuk, et al.
+ "AutoAugment: Learning Augmentation Policies from Data"
+ arXiv:1805.09501
+ Junyoung Chung, et al.
+ "Empirical Evaluation of Gated Recurrent Neural
+ Networks on Sequence Modeling"
+ https://arxiv.org/abs/1412.3555
@@ -61,18 +104,7 @@ class GruLearner(AaLearner):
alpha=0.2,
cont_mb_size=4,
cont_lr=0.03):
- """
- Args:
- alpha (float, optional): Exploration parameter. It is multiplied to
- operation tensors before they're softmaxed.
- The lower this value, the more smoothed the output
- of the softmaxed will be, hence more exploration.
- Defaults to 0.2.
- cont_mb_size (int, optional): Controller Minibatch Size. How many
- policies do we test in order to calculate the
- PPO(proximal policy update) gradient to update
- the controller. Defaults to
- """
+
if discrete_p_m==True:
print('Warning: Incompatible discrete_p_m=True input into GruLearner. \
discrete_p_m=False will be used')
diff --git a/autoaug/autoaugment_learners/RsLearner.py b/autoaug/autoaugment_learners/RsLearner.py
index da0f7695ec94f7387c4d2dcdb42b7c5b0951b957..e1562a431099b6166fb9ef3b63416df6106530de 100644
--- a/autoaug/autoaugment_learners/RsLearner.py
+++ b/autoaug/autoaugment_learners/RsLearner.py
@@ -13,26 +13,60 @@ import pickle
class RsLearner(AaLearner):
- """
+ """This agent tests out randomly generated augmentation policies.
+
Tests randomly sampled policies from the search space specified by the AutoAugment
- paper. Acts as a baseline for other AaLearner's.
+ paper. Random search has been shown to be a hard baseline to beat for many other
+ hyper-parameter optimization tasks (see References below).
+ Hence, this learner acts as a difficult baseline for other AaLearner's.
-
- See Also
- --------
+ Args:
+ sp_num (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.
- Notes
- -----
+ m_bins (int, optional): number of bins we divide the magnitude space.
+ Defaults to 10.
+ discrete_p_m (bool, optional):
+ Whether or not the agent should represent probability and
+ magnitude as discrete variables as the out put of the
+ controller (A controller can be a neural network, genetic
+ algorithm, etc.). Defaults to False
- References
- ----------
+ 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.
+
+ exclude_method (list, optional): list of names(:type:str) of image operations
+ the user wants to exclude from the search space. Defaults to [].
+ Attributes:
+ history (list): list of policies that has been input into
+ self._test_autoaugment_policy as well as their respective obtained
+ accuracies
- Examples
- --------
+ augmentation_space (list): list of image functions that the user has chosen to
+ include in the search space.
+ References
+ ----------
+ Ekin D. Cubuk, et al.
+ "AutoAugment: Learning Augmentation Policies from Data"
+ arXiv:1805.09501
+
+ Bergstra James, Yoshua Bengio
+ "Random Search for Hyper-Parameter Optimization"
+ https://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf
"""
def __init__(self,
diff --git a/autoaug/autoaugment_learners/UcbLearner.py b/autoaug/autoaugment_learners/UcbLearner.py
index ded7d67a87dbb9582e3753a520b9d15d6de39ff1..7e3e707780bec802c254fe08aab563faacfb612c 100644
--- a/autoaug/autoaugment_learners/UcbLearner.py
+++ b/autoaug/autoaugment_learners/UcbLearner.py
@@ -15,24 +15,71 @@ class UcbLearner(RsLearner):
- Using a toy dataset or a toy CNN
+ Args:
+ sp_num (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.
+
+ discrete_p_m (bool, optional):
+ Whether or not the agent should represent probability and
+ magnitude as discrete variables as the out put of the
+ controller (A controller can be a neural network, genetic
+ algorithm, etc.). Defaults to False
+
+ 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.
+
+ exclude_method (list, optional): list of names(:type:str) of image operations
+ the user wants to exclude from the search space. Defaults to [].
+ num_policies (int, optional): Number of policies we want to serach over.
+ Defaults to 100.
- See Also
- --------
+ Attributes:
+ history (list): list of policies that has been input into
+ self._test_autoaugment_policy as well as their respective obtained
+ accuracies
+ augmentation_space (list): list of image functions that the user has chosen to
+ include in the search space.
- Notes
- -----
+ policies (list): A list of policies which we are currently searching over.
+ avg_accs (list): A list where the nth element indicates the average accuracy
+ obtained by the nth policy.
- References
- ----------
-
- Examples
- --------
+ 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
"""
def __init__(self,