diff --git a/.gitignore b/.gitignore
index e792a3f6e509811af884773c49c715bf525e7580..e541790a38735883601fc9b3b302cf2e9d940bae 100644
--- a/.gitignore
+++ b/.gitignore
@@ -200,8 +200,11 @@ celerybeat.pid
# SageMath parsed files
*.sage.py
+# we don't want dataset directories
**/test
**/train
+# but there is a unit test folder in main that we DO want to track
+!test
# user uplaod
/react_backend/child_networks
diff --git a/MetaAugment/autoaugment_learners/aa_learner.py b/MetaAugment/autoaugment_learners/aa_learner.py
index 561222a5fac35d5348f99da6a9fba31657afe133..48c05b95d5bef8de9c0405e949f1e6663e66e8ae 100644
--- a/MetaAugment/autoaugment_learners/aa_learner.py
+++ b/MetaAugment/autoaugment_learners/aa_learner.py
@@ -101,7 +101,7 @@ class aa_learner:
self.op_tensor_length = self.fun_num + p_bins + m_bins if discrete_p_m else self.fun_num +2
- def translate_operation_tensor(self, operation_tensor, return_log_prob=False, argmax=False):
+ def _translate_operation_tensor(self, operation_tensor, return_log_prob=False, argmax=False):
"""
takes in a tensor representing an operation and returns an actual operation which
is in the form of:
@@ -220,7 +220,7 @@ class aa_learner:
return operation
- def generate_new_policy(self):
+ def _generate_new_policy(self):
"""
Generate a new policy which can be fed into an AutoAugment object
by calling:
@@ -243,7 +243,7 @@ class aa_learner:
by calling: AutoAugment.subpolicies = policy
"""
- raise NotImplementedError('generate_new_policy not implemented in aa_learner')
+ raise NotImplementedError('_generate_new_policy not implemented in aa_learner')
def learn(self, train_dataset, test_dataset, child_network_architecture, iterations=15):
@@ -290,10 +290,10 @@ class aa_learner:
different policies
for _ in range(15):
- policy = self.generate_new_policy()
+ policy = self._generate_new_policy()
pprint(policy)
- reward = self.test_autoaugment_policy(policy,
+ reward = self._test_autoaugment_policy(policy,
child_network_architecture,
train_dataset,
test_dataset)
@@ -302,7 +302,7 @@ class aa_learner:
"""
- def test_autoaugment_policy(self,
+ def _test_autoaugment_policy(self,
policy,
child_network_architecture,
train_dataset,
@@ -329,7 +329,9 @@ class aa_learner:
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
if isinstance(child_network_architecture, types.FunctionType):
child_network = child_network_architecture()
elif isinstance(child_network_architecture, type):
@@ -394,10 +396,10 @@ class aa_learner:
# # This is dummy code
# # test out `n` random policies
# for _ in range(n):
- # policy = self.generate_new_policy()
+ # policy = self._generate_new_policy()
# pprint(policy)
- # reward, acc_list = self.test_autoaugment_policy(policy,
+ # reward, acc_list = self._test_autoaugment_policy(policy,
# child_network_architecture,
# train_dataset,
# test_dataset,
diff --git a/MetaAugment/autoaugment_learners/autoaugment.py b/MetaAugment/autoaugment_learners/autoaugment.py
index 5a8ecbcf6f0b8c6212a8c034a70d61476f4870f6..6baa28cd51c6a9c2a584aea0f3c772a42e55f92c 100644
--- a/MetaAugment/autoaugment_learners/autoaugment.py
+++ b/MetaAugment/autoaugment_learners/autoaugment.py
@@ -446,7 +446,7 @@ if __name__=='__main__':
- def test_autoaugment_policy(subpolicies, train_dataset, test_dataset):
+ def _test_autoaugment_policy(subpolicies, train_dataset, test_dataset):
aa_transform = AutoAugment()
aa_transform.subpolicies = subpolicies
@@ -470,8 +470,8 @@ if __name__=='__main__':
return best_acc, acc_log
- _, acc_log1 = test_autoaugment_policy(subpolicies1, train_dataset, test_dataset)
- _, acc_log2 = test_autoaugment_policy(subpolicies2, train_dataset, test_dataset)
+ _, acc_log1 = _test_autoaugment_policy(subpolicies1, train_dataset, test_dataset)
+ _, acc_log2 = _test_autoaugment_policy(subpolicies2, train_dataset, test_dataset)
plt.plot(acc_log1, label='subpolicies1')
plt.plot(acc_log2, label='subpolicies2')
diff --git a/MetaAugment/autoaugment_learners/evo_learner.py b/MetaAugment/autoaugment_learners/evo_learner.py
index 34cc2d44555423475914a1ba2528cfddb71aad57..6bf682c1595b0b731c4a68bfc8f619f953b31d29 100644
--- a/MetaAugment/autoaugment_learners/evo_learner.py
+++ b/MetaAugment/autoaugment_learners/evo_learner.py
@@ -1,12 +1,12 @@
import torch
-torch.manual_seed(0)
import torch.nn as nn
import pygad
import pygad.torchga as torchga
-import copy
+import torchvision
import torch
from MetaAugment.autoaugment_learners.aa_learner import aa_learner
+import MetaAugment.controller_networks as cont_n
class evo_learner(aa_learner):
@@ -14,7 +14,7 @@ class evo_learner(aa_learner):
def __init__(self,
# search space settings
sp_num=5,
- p_bins=10,
+ p_bins=11,
m_bins=10,
discrete_p_m=False,
exclude_method=[],
@@ -27,7 +27,7 @@ class evo_learner(aa_learner):
# evolutionary learner specific settings
num_solutions=5,
num_parents_mating=3,
- controller=None
+ controller=cont_n.evo_controller
):
super().__init__(
@@ -43,14 +43,19 @@ class evo_learner(aa_learner):
exclude_method=exclude_method
)
+ # evolutionary algorithm settings
+ self.controller = controller(
+ fun_num=self.fun_num,
+ p_bins=self.p_bins,
+ m_bins=self.m_bins,
+ sub_num_pol=self.sp_num
+ )
self.num_solutions = num_solutions
- self.controller = controller
self.torch_ga = torchga.TorchGA(model=self.controller, num_solutions=num_solutions)
self.num_parents_mating = num_parents_mating
self.initial_population = self.torch_ga.population_weights
- self.p_bins = p_bins
- self.sub_num_pol = sp_num
- self.m_bins = m_bins
+
+ # store our logs
self.policy_dict = {}
self.policy_result = []
@@ -58,6 +63,7 @@ class evo_learner(aa_learner):
assert num_solutions > num_parents_mating, 'Number of solutions must be larger than the number of parents mating!'
+
def get_full_policy(self, x):
"""
Generates the full policy (self.num_sub_pol subpolicies). Network architecture requires
@@ -77,7 +83,7 @@ class evo_learner(aa_learner):
section = self.fun_num + self.p_bins + self.m_bins
y = self.controller.forward(x)
full_policy = []
- for pol in range(self.sub_num_pol):
+ for pol in range(self.sp_num):
int_pol = []
for _ in range(2):
idx_ret = torch.argmax(y[:, (pol * section):(pol*section) + self.fun_num].mean(dim = 0))
@@ -167,10 +173,10 @@ class evo_learner(aa_learner):
prob2 += torch.sigmoid(y[idx, section+self.fun_num]).item()
if mag1 is not None:
# mag1 += min(max(0, (y[idx, self.auto_aug_agent.fun_num+1]).item()), 8)
- mag1 += 10 * torch.sigmoid(y[idx, self.fun_num+1]).item()
+ mag1 += min(9, 10 * torch.sigmoid(y[idx, self.fun_num+1]).item())
if mag2 is not None:
# mag2 += min(max(0, y[idx, section+self.auto_aug_agent.fun_num+1].item()), 8)
- mag2 += 10 * torch.sigmoid(y[idx, self.fun_num+1]).item()
+ mag2 += min(9, 10 * torch.sigmoid(y[idx, self.fun_num+1]).item())
counter += 1
@@ -240,7 +246,7 @@ class evo_learner(aa_learner):
self.policy_dict[trans1][trans2].append(new_set)
return False
else:
- self.policy_dict[trans1][trans2] = [new_set]
+ self.policy_dict[trans1] = {trans2: [new_set]}
if trans2 in self.policy_dict:
if trans1 in self.policy_dict[trans2]:
for test_pol in self.policy_dict[trans2][trans1]:
@@ -249,7 +255,7 @@ class evo_learner(aa_learner):
self.policy_dict[trans2][trans1].append(new_set)
return False
else:
- self.policy_dict[trans2][trans1] = [new_set]
+ self.policy_dict[trans2] = {trans1: [new_set]}
def set_up_instance(self, train_dataset, test_dataset, child_network_architecture):
@@ -277,6 +283,7 @@ class evo_learner(aa_learner):
weights_vector=solution)
self.controller.load_state_dict(model_weights_dict)
+ train_dataset.transform = torchvision.transforms.ToTensor()
self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=self.batch_size)
for idx, (test_x, label_x) in enumerate(self.train_loader):
@@ -298,11 +305,11 @@ class evo_learner(aa_learner):
if len(self.policy_result) > self.sp_num:
self.policy_result = sorted(self.policy_result, key=lambda x: x[1], reverse=True)
self.policy_result = self.policy_result[:self.sp_num]
- print("Appended policy: ", self.policy_result)
+ print("appended policy: ", self.policy_result)
if fit_val > self.history_best[self.gen_count]:
- print("Best policy: ", full_policy)
+ print("best policy: ", full_policy)
self.history_best[self.gen_count] = fit_val
self.best_model = model_weights_dict
@@ -335,4 +342,3 @@ class evo_learner(aa_learner):
mutation_percent_genes = 0.1,
fitness_func=fitness_func,
on_generation = on_generation)
-
diff --git a/MetaAugment/autoaugment_learners/gru_learner.py b/MetaAugment/autoaugment_learners/gru_learner.py
index 5c15a4a41b086982aa543bda451c89bfa7eecba9..7ca8088cfd6291f068d1b0eb16c1effa89cfa427 100644
--- a/MetaAugment/autoaugment_learners/gru_learner.py
+++ b/MetaAugment/autoaugment_learners/gru_learner.py
@@ -85,7 +85,7 @@ class gru_learner(aa_learner):
self.softmax = torch.nn.Softmax(dim=0)
- def generate_new_policy(self):
+ def _generate_new_policy(self):
"""
The GRU controller pops out a new policy.
@@ -101,7 +101,7 @@ class gru_learner(aa_learner):
choice of function, prob, and mag seperately, so that the
resulting tensor's values sums up to 3.
- Then we input each tensor into self.translate_operation_tensor
+ Then we input each tensor into self._translate_operation_tensor
with parameter (return_log_prob=True), which outputs a tuple
in the form of ('img_function_name', prob, mag) and a float
representing the log probability that we chose the chosen
@@ -150,8 +150,8 @@ class gru_learner(aa_learner):
op2 = softmaxed_vectors[2*subpolicy_idx+1]
# translate both vectors
- op1, log_prob1 = self.translate_operation_tensor(op1, return_log_prob=True)
- op2, log_prob2 = self.translate_operation_tensor(op2, return_log_prob=True)
+ op1, log_prob1 = self._translate_operation_tensor(op1, return_log_prob=True)
+ op2, log_prob2 = self._translate_operation_tensor(op2, return_log_prob=True)
new_policy.append((op1,op2))
log_prob += (log_prob1+log_prob2)
@@ -177,10 +177,10 @@ class gru_learner(aa_learner):
for k in range(self.cont_mb_size):
# log_prob is $\sum_{t=1}^T log(P(a_t|a_{(t-1):1};\theta_c))$, used in PPO
- policy, log_prob = self.generate_new_policy()
+ policy, log_prob = self._generate_new_policy()
pprint(policy)
- reward = self.test_autoaugment_policy(policy,
+ reward = self._test_autoaugment_policy(policy,
child_network_architecture,
train_dataset,
test_dataset)
diff --git a/MetaAugment/autoaugment_learners/randomsearch_learner.py b/MetaAugment/autoaugment_learners/randomsearch_learner.py
index 2c35fb80ab15f7b2c51dfdcfbfcff942a6a70032..25fbd1125e5aaf5c88cb3fc2931f7612dc4d831b 100644
--- a/MetaAugment/autoaugment_learners/randomsearch_learner.py
+++ b/MetaAugment/autoaugment_learners/randomsearch_learner.py
@@ -46,7 +46,7 @@ class randomsearch_learner(aa_learner):
)
- def generate_new_discrete_operation(self):
+ def _generate_new_discrete_operation(self):
"""
generate a new random operation in the form of a tensor of dimension:
(fun_num + 11 + 10)
@@ -74,7 +74,7 @@ class randomsearch_learner(aa_learner):
return torch.cat([fun_t, prob_t, mag_t])
- def generate_new_continuous_operation(self):
+ def _generate_new_continuous_operation(self):
"""
Returns operation_tensor, which is a tensor representation of a random operation with
dimension:
@@ -101,7 +101,7 @@ class randomsearch_learner(aa_learner):
return fun_p_m
- def generate_new_policy(self):
+ def _generate_new_policy(self):
"""
Generates a new policy, with the elements chosen at random
(unifom random distribution).
@@ -115,10 +115,10 @@ class randomsearch_learner(aa_learner):
for i in range(2):
# if our agent uses discrete representations of probability and magnitude
if self.discrete_p_m:
- new_op = self.generate_new_discrete_operation()
+ new_op = self._generate_new_discrete_operation()
else:
- new_op = self.generate_new_continuous_operation()
- new_op = self.translate_operation_tensor(new_op)
+ new_op = self._generate_new_continuous_operation()
+ new_op = self._translate_operation_tensor(new_op)
ops.append(new_op)
new_subpolicy = tuple(ops)
@@ -135,10 +135,10 @@ class randomsearch_learner(aa_learner):
iterations=15):
# test out `iterations` number of random policies
for _ in range(iterations):
- policy = self.generate_new_policy()
+ policy = self._generate_new_policy()
pprint(policy)
- reward = self.test_autoaugment_policy(policy,
+ reward = self._test_autoaugment_policy(policy,
child_network_architecture,
train_dataset,
test_dataset)
diff --git a/MetaAugment/autoaugment_learners/ucb_learner.py b/MetaAugment/autoaugment_learners/ucb_learner.py
index fdf735bea1916897ee5e28a27bd67c10b5751581..6ed010fc6668ba2577815b6e65382ceef7241a74 100644
--- a/MetaAugment/autoaugment_learners/ucb_learner.py
+++ b/MetaAugment/autoaugment_learners/ucb_learner.py
@@ -47,7 +47,7 @@ class ucb_learner(randomsearch_learner):
# attributes used in the UCB1 algorithm
self.num_policies = num_policies
- self.policies = [self.generate_new_policy() for _ in range(num_policies)]
+ self.policies = [self._generate_new_policy() for _ in range(num_policies)]
self.avg_accs = [None]*self.num_policies
self.best_avg_accs = []
@@ -67,7 +67,7 @@ class ucb_learner(randomsearch_learner):
and add to our list of policies
"""
- self.policies += [self.generate_new_policy() for _ in range(n)]
+ self.policies += [self._generate_new_policy() for _ in range(n)]
# all the below need to be lengthened to store information for the
# new policies
@@ -96,7 +96,7 @@ class ucb_learner(randomsearch_learner):
# test that one
this_policy_idx = self.avg_accs.index(None)
this_policy = self.policies[this_policy_idx]
- acc = self.test_autoaugment_policy(
+ acc = self._test_autoaugment_policy(
this_policy,
child_network_architecture,
train_dataset,
@@ -111,7 +111,7 @@ class ucb_learner(randomsearch_learner):
# one with the best q_plus_cnt value
this_policy_idx = np.argmax(self.q_plus_cnt)
this_policy = self.policies[this_policy_idx]
- acc = self.test_autoaugment_policy(
+ acc = self._test_autoaugment_policy(
this_policy,
child_network_architecture,
train_dataset,
diff --git a/benchmark/scripts/util_04_22.py b/benchmark/scripts/util_04_22.py
index 62c0456af78549bfaa0599a1a61c40a7eb78e806..8d7aa6b18f30b4a07e2531b04ba13c6ace9e0dbf 100644
--- a/benchmark/scripts/util_04_22.py
+++ b/benchmark/scripts/util_04_22.py
@@ -111,7 +111,7 @@ def rerun_best_policy(
print(f'{_}/{repeat_num}')
temp_agent = aal.aa_learner(**config)
accs.append(
- temp_agent.test_autoaugment_policy(megapol,
+ temp_agent._test_autoaugment_policy(megapol,
child_network_architecture,
train_dataset,
test_dataset,
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/aa_learners.rst b/docs/source/MetaAugment_library/autoaugment_learners/aa_learners.rst
index 4fc99146806c3b71f88e9173a9751d7edf354c27..db9d7b9fb0a1cade9b28ac93e8e8b6c6dab8493b 100644
--- a/docs/source/MetaAugment_library/autoaugment_learners/aa_learners.rst
+++ b/docs/source/MetaAugment_library/autoaugment_learners/aa_learners.rst
@@ -6,5 +6,7 @@ AutoAugment learners
:toctree: generated
MetaAugment.autoaugment_learners.aa_learner
+ MetaAugment.autoaugment_learners.evo_learner
MetaAugment.autoaugment_learners.gru_learner
- MetaAugment.autoaugment_learners.randomsearch_learner
\ No newline at end of file
+ MetaAugment.autoaugment_learners.randomsearch_learner
+ MetaAugment.autoaugment_learners.ucb_learner
\ No newline at end of file
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.aa_learner.rst b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.aa_learner.rst
index 010241cbfcfc58968726e3f4c177ff7e4f70ad25..85be0d0150aa95092c10990d6d17c147c799b3af 100644
--- a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.aa_learner.rst
+++ b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.aa_learner.rst
@@ -14,11 +14,7 @@
.. autosummary::
~aa_learner.__init__
- ~aa_learner.demo_plot
- ~aa_learner.generate_new_policy
~aa_learner.learn
- ~aa_learner.test_autoaugment_policy
- ~aa_learner.translate_operation_tensor
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.evo_learner.rst b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.evo_learner.rst
new file mode 100644
index 0000000000000000000000000000000000000000..37f06b00f70a6e120fdd1f2dd30612da5bf7b7e6
--- /dev/null
+++ b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.evo_learner.rst
@@ -0,0 +1,27 @@
+MetaAugment.autoaugment\_learners.evo\_learner
+==============================================
+
+.. currentmodule:: MetaAugment.autoaugment_learners
+
+.. autoclass:: evo_learner
+
+
+ .. automethod:: __init__
+
+
+ .. rubric:: Methods
+
+ .. autosummary::
+
+ ~evo_learner.__init__
+ ~evo_learner.get_full_policy
+ ~evo_learner.get_single_policy_cov
+ ~evo_learner.in_pol_dict
+ ~evo_learner.learn
+ ~evo_learner.set_up_instance
+
+
+
+
+
+
\ No newline at end of file
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.gru_learner.rst b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.gru_learner.rst
index f5cbe4429c79106ddd6e14c8f54a37226f196a3a..23eb306c38dbc495a4a6c775cf850fe913172876 100644
--- a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.gru_learner.rst
+++ b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.gru_learner.rst
@@ -14,11 +14,7 @@
.. autosummary::
~gru_learner.__init__
- ~gru_learner.demo_plot
- ~gru_learner.generate_new_policy
~gru_learner.learn
- ~gru_learner.test_autoaugment_policy
- ~gru_learner.translate_operation_tensor
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.randomsearch_learner.rst b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.randomsearch_learner.rst
index 5bfb303d8fc1fcb8f8638c186b2e4ecc69be8861..72903e47cdaf1a51c7922127343848ada0a12675 100644
--- a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.randomsearch_learner.rst
+++ b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.randomsearch_learner.rst
@@ -14,13 +14,7 @@
.. autosummary::
~randomsearch_learner.__init__
- ~randomsearch_learner.demo_plot
- ~randomsearch_learner.generate_new_continuous_operation
- ~randomsearch_learner.generate_new_discrete_operation
- ~randomsearch_learner.generate_new_policy
~randomsearch_learner.learn
- ~randomsearch_learner.test_autoaugment_policy
- ~randomsearch_learner.translate_operation_tensor
diff --git a/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.ucb_learner.rst b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.ucb_learner.rst
new file mode 100644
index 0000000000000000000000000000000000000000..83f80f4893a8376490d67aeeb1943866cef3ffa3
--- /dev/null
+++ b/docs/source/MetaAugment_library/autoaugment_learners/generated/MetaAugment.autoaugment_learners.ucb_learner.rst
@@ -0,0 +1,24 @@
+MetaAugment.autoaugment\_learners.ucb\_learner
+==============================================
+
+.. currentmodule:: MetaAugment.autoaugment_learners
+
+.. autoclass:: ucb_learner
+
+
+ .. automethod:: __init__
+
+
+ .. rubric:: Methods
+
+ .. autosummary::
+
+ ~ucb_learner.__init__
+ ~ucb_learner.learn
+ ~ucb_learner.make_more_policies
+
+
+
+
+
+
\ No newline at end of file
diff --git a/docs/source/usage/autoaugment_helperclass.rst b/docs/source/usage/autoaugment_helperclass.rst
index 75f080fb94093d67b94c2fb27a19778453a6505f..cc361da4f66c41c937803225498bc20a7fdf5b0d 100644
--- a/docs/source/usage/autoaugment_helperclass.rst
+++ b/docs/source/usage/autoaugment_helperclass.rst
@@ -11,7 +11,7 @@ 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:`aa_learner.test_autoaugment_policy <MetaAugment.autoaugment_learners.aa_learner>`.
+:meth:`aa_learner._test_autoaugment_policy <MetaAugment.autoaugment_learners.aa_learner>`.
Let's say we have a policy within the search space specified by the original
AutoAugment paper:
diff --git a/temp_util/wapp_util.py b/temp_util/wapp_util.py
index cde572f7f1e0ba1590fb5685b7212f4d0b3b173a..bb10113311f77d32f3f0b996e76d71f16fa0d184 100644
--- a/temp_util/wapp_util.py
+++ b/temp_util/wapp_util.py
@@ -54,6 +54,7 @@ def parse_users_learner_spec(
p_bins=11,
m_bins=10,
discrete_p_m=True,
+ exclude_method=exclude_method,
# hyperparameters for when training the child_network
batch_size=batch_size,
toy_size=toy_size,
@@ -63,52 +64,55 @@ def parse_users_learner_spec(
# ucb_learner specific hyperparameter
num_policies=num_policies
)
- pprint(learner.policies)
-
- learner.learn(
- train_dataset=train_dataset,
- test_dataset=test_dataset,
- child_network_architecture=child_archi,
- iterations=5
- )
elif auto_aug_learner == 'Evolutionary Learner':
- network = cont_n.evo_controller(fun_num=num_funcs, p_bins=1, m_bins=1, sub_num_pol=1)
- child_network = cn.LeNet()
learner = aal.evo_learner(
- network=network,
- fun_num=num_funcs,
- p_bins=1,
- mag_bins=1,
- sub_num_pol=1,
- ds = ds,
- ds_name=ds_name,
- exclude_method=exclude_method,
- child_network=child_network
- )
+ # parameters that define the search space
+ sp_num=num_sub_policies,
+ p_bins=11,
+ m_bins=10,
+ discrete_p_m=True,
+ exclude_method=exclude_method,
+ # hyperparameters for when training the child_network
+ batch_size=batch_size,
+ toy_size=toy_size,
+ learning_rate=learning_rate,
+ max_epochs=max_epochs,
+ early_stop_num=early_stop_num,
+ )
learner.run_instance()
elif auto_aug_learner == 'Random Searcher':
agent = aal.randomsearch_learner(
- sp_num=num_sub_policies,
- batch_size=batch_size,
- learning_rate=learning_rate,
- toy_size=toy_size,
- max_epochs=max_epochs,
- early_stop_num=early_stop_num,
- )
- agent.learn(train_dataset,
- test_dataset,
- child_network_architecture=child_archi,
- iterations=iterations)
+ # parameters that define the search space
+ sp_num=num_sub_policies,
+ p_bins=11,
+ m_bins=10,
+ discrete_p_m=True,
+ exclude_method=exclude_method,
+ # hyperparameters for when training the child_network
+ batch_size=batch_size,
+ toy_size=toy_size,
+ learning_rate=learning_rate,
+ max_epochs=max_epochs,
+ early_stop_num=early_stop_num,
+ )
elif auto_aug_learner == 'GRU Learner':
agent = aal.gru_learner(
- sp_num=num_sub_policies,
- batch_size=batch_size,
- learning_rate=learning_rate,
- toy_size=toy_size,
- max_epochs=max_epochs,
- early_stop_num=early_stop_num,
- )
- agent.learn(train_dataset,
- test_dataset,
- child_network_architecture=child_archi,
- iterations=iterations)
\ No newline at end of file
+ # parameters that define the search space
+ sp_num=num_sub_policies,
+ p_bins=11,
+ m_bins=10,
+ discrete_p_m=True,
+ exclude_method=exclude_method,
+ # hyperparameters for when training the child_network
+ batch_size=batch_size,
+ toy_size=toy_size,
+ learning_rate=learning_rate,
+ max_epochs=max_epochs,
+ early_stop_num=early_stop_num,
+ )
+
+
+ agent.learn(train_dataset,
+ test_dataset,
+ child_network_architecture=child_archi,
+ iterations=iterations)
\ No newline at end of file
diff --git a/test/MetaAugment/test_aa_learner.py b/test/MetaAugment/test_aa_learner.py
index b1524988939e9adac0b45285921b8d058f087887..64517cd090360a4f150d1d1e794df6eb08cd6c90 100644
--- a/test/MetaAugment/test_aa_learner.py
+++ b/test/MetaAugment/test_aa_learner.py
@@ -7,13 +7,13 @@ import torchvision.datasets as datasets
import random
-def test_translate_operation_tensor():
+def test__translate_operation_tensor():
"""
- See if aa_learner class's translate_operation_tensor works
+ See if aa_learner class's _translate_operation_tensor works
by feeding many (valid) inputs in it.
We make a lot of (fun_num+p_bins_m_bins,) size tensors, softmax
- them, and feed them through the translate_operation_tensor method
+ them, and feed them through the _translate_operation_tensor method
to see if it doesn't break
"""
@@ -44,7 +44,7 @@ def test_translate_operation_tensor():
mag_t = softmax(mag_t * alpha)
softmaxed_vector = torch.cat((fun_t, prob_t, mag_t))
- agent.translate_operation_tensor(softmaxed_vector)
+ agent._translate_operation_tensor(softmaxed_vector)
# discrete_p_m=False
@@ -73,10 +73,10 @@ def test_translate_operation_tensor():
softmaxed_vector = torch.cat((fun_t, prob_t, mag_t))
- agent.translate_operation_tensor(softmaxed_vector)
+ agent._translate_operation_tensor(softmaxed_vector)
-def test_test_autoaugment_policy():
+def test__test_autoaugment_policy():
agent = aal.aa_learner(
sp_num=5,
p_bins=11,
@@ -107,7 +107,7 @@ def test_test_autoaugment_policy():
train=False, download=True,
transform=torchvision.transforms.ToTensor())
- acc = agent.test_autoaugment_policy(
+ acc = agent._test_autoaugment_policy(
policy,
child_network_architecture,
train_dataset,
@@ -116,4 +116,41 @@ def test_test_autoaugment_policy():
)
assert isinstance(acc, float)
+
+
+def test_exclude_method():
+ """
+ we want to see if the exclude_methods
+ parameter is working properly in aa_learners
+ """
+
+ exclude_method = [
+ 'ShearX',
+ 'Color',
+ 'Brightness',
+ 'Contrast'
+ ]
+ agent = aal.gru_learner(
+ exclude_method=exclude_method
+ )
+ for _ in range(200):
+ new_pol, _ = agent._generate_new_policy()
+ print(new_pol)
+ for (op1, op2) in new_pol:
+ image_function_1 = op1[0]
+ image_function_2 = op2[0]
+ assert image_function_1 not in exclude_method
+ assert image_function_2 not in exclude_method
+
+ agent = aal.randomsearch_learner(
+ exclude_method=exclude_method
+ )
+ for _ in range(200):
+ new_pol= agent._generate_new_policy()
+ print(new_pol)
+ for (op1, op2) in new_pol:
+ image_function_1 = op1[0]
+ image_function_2 = op2[0]
+ assert image_function_1 not in exclude_method
+ assert image_function_2 not in exclude_method
\ No newline at end of file
diff --git a/test/MetaAugment/test_evo_learner.py b/test/MetaAugment/test_evo_learner.py
new file mode 100644
index 0000000000000000000000000000000000000000..b917fb3934555585c2b95a412580ab27f517c081
--- /dev/null
+++ b/test/MetaAugment/test_evo_learner.py
@@ -0,0 +1,44 @@
+import MetaAugment.autoaugment_learners as aal
+import MetaAugment.child_networks as cn
+import torchvision
+import torchvision.datasets as datasets
+from pprint import pprint
+
+def test_evo_learner():
+ child_network_architecture = cn.SimpleNet
+ train_dataset = datasets.FashionMNIST(root='./datasets/fashionmnist/train',
+ train=True, download=True, transform=None)
+ test_dataset = datasets.FashionMNIST(root='./datasets/fashionmnist/test',
+ train=False, download=True,
+ transform=torchvision.transforms.ToTensor())
+
+
+ learner = aal.evo_learner(
+ # parameters that define the search space
+ sp_num=5,
+ p_bins=11,
+ m_bins=10,
+ discrete_p_m=True,
+ exclude_method=['ShearX'],
+ # hyperparameters for when training the child_network
+ batch_size=8,
+ toy_size=0.0001,
+ learning_rate=1e-1,
+ max_epochs=float('inf'),
+ early_stop_num=30,
+ # evolutionary learner specific settings
+ num_solutions=3,
+ num_parents_mating=2,
+ )
+
+ # learn on the 3 policies we generated
+ learner.learn(
+ train_dataset=train_dataset,
+ test_dataset=test_dataset,
+ child_network_architecture=child_network_architecture,
+ iterations=2
+ )
+
+
+if __name__=="__main__":
+ test_evo_learner()
diff --git a/test/MetaAugment/test_gru_learner.py b/test/MetaAugment/test_gru_learner.py
index cd52b0e95f710c8cddf8a9afdbe67a86acb8fb07..b2ea8930df73959c63933804e72d28ce9bdeca84 100644
--- a/test/MetaAugment/test_gru_learner.py
+++ b/test/MetaAugment/test_gru_learner.py
@@ -6,9 +6,9 @@ import torchvision.datasets as datasets
import random
-def test_generate_new_policy():
+def test__generate_new_policy():
"""
- make sure gru_learner.generate_new_policy() is robust
+ make sure gru_learner._generate_new_policy() is robust
with respect to different values of sp_num, fun_num,
p_bins, and m_bins
"""
@@ -24,7 +24,7 @@ def test_generate_new_policy():
cont_mb_size=2
)
for _ in range(4):
- new_policy = agent.generate_new_policy()
+ new_policy = agent._generate_new_policy()
assert isinstance(new_policy[0], list), new_policy
diff --git a/test/MetaAugment/test_randomsearch_learner.py b/test/MetaAugment/test_randomsearch_learner.py
index 61e9f9cd8c86be854dcd8d42cc9ceddde8ada3bc..6c5a935071bfacc8c2e1c8d276732aa7585afa60 100644
--- a/test/MetaAugment/test_randomsearch_learner.py
+++ b/test/MetaAugment/test_randomsearch_learner.py
@@ -6,9 +6,9 @@ import torchvision.datasets as datasets
import random
-def test_generate_new_policy():
+def test__generate_new_policy():
"""
- make sure randomsearch_learner.generate_new_policy() is robust
+ make sure randomsearch_learner._generate_new_policy() is robust
with respect to different values of sp_num, fun_num,
p_bins, and m_bins
"""
@@ -27,7 +27,7 @@ def test_generate_new_policy():
discrete_p_m=discrete_p_m
)
for _ in range(4):
- new_policy = agent.generate_new_policy()
+ new_policy = agent._generate_new_policy()
assert isinstance(new_policy, list), new_policy
discrete_p_m = True