diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index df6506e28a7767cd17b4c6e4ea15ebb3c4416d85..9d829da36400d8e3f5b33efa0197f101747a34b3 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -1,14 +1,17 @@
-build-job:
- stage: build
- script:
- - python3 -m venv venv
- - . venv/bin/activate
- - pip3 install wheel
- - pip3 install -r requirements.txt --no-cache-dir
+# build-job:
+# stage: build
+# script:
+# - python3 -m venv venv
+# - . venv/bin/activate
+# - pip3 install wheel
+# - pip3 install -r requirements.txt --no-cache-dir
test-job:
stage: test
script:
+ - python3 -m venv venv
+ - . venv/bin/activate
+ - pip3 install -r requirements.txt --no-cache-dir
- pytest
# deploy-job:
diff --git a/MetaAugment/UCB1_JC_py.py b/MetaAugment/UCB1_JC_py.py
index 8ba8c93d7cfd6fb5fad91499e5a46faad7c1d91a..27322463dc6b679751aac4b2483f0e12f4352cdb 100644
--- a/MetaAugment/UCB1_JC_py.py
+++ b/MetaAugment/UCB1_JC_py.py
@@ -102,7 +102,7 @@ def sample_sub_policy(policies, policy, num_sub_policies):
"""Sample policy, open and apply above transformations"""
-def run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, early_stop_num, iterations, IsLeNet, ds_name=None):
+def run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, early_stop_num, early_stop_flag, average_validation, iterations, IsLeNet, ds_name=None):
# get number of policies and sub-policies
num_policies = len(policies)
@@ -172,12 +172,13 @@ def run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, earl
train_loader, test_loader = create_toy(train_dataset, test_dataset, batch_size, toy_size)
# create model
+ device = 'cuda' if torch.cuda.is_available() else 'cpu'
if IsLeNet == "LeNet":
- model = LeNet(img_height, img_width, num_labels, img_channels)
+ model = LeNet(img_height, img_width, num_labels, img_channels).to(device) # added .to(device)
elif IsLeNet == "EasyNet":
- model = EasyNet(img_height, img_width, num_labels, img_channels)
+ model = EasyNet(img_height, img_width, num_labels, img_channels).to(device) # added .to(device)
elif IsLeNet == 'SimpleNet':
- model = SimpleNet(img_height, img_width, num_labels, img_channels)
+ model = SimpleNet(img_height, img_width, num_labels, img_channels).to(device) # added .to(device)
else:
model = pickle.load(open(f'datasets/childnetwork', "rb"))
@@ -185,8 +186,8 @@ def run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, earl
cost = nn.CrossEntropyLoss()
best_acc = train_child_network(model, train_loader, test_loader, sgd,
- cost, max_epochs, early_stop_num, logging=False,
- print_every_epoch=False)
+ cost, max_epochs, early_stop_num, early_stop_flag,
+ average_validation, logging=False, print_every_epoch=False)
# update q_values
if policy < num_policies:
@@ -222,6 +223,8 @@ if __name__=='__main__':
toy_size = 0.02 # total propeortion of training and test set we use
max_epochs = 100 # max number of epochs that is run if early stopping is not hit
early_stop_num = 10 # max number of worse validation scores before early stopping is triggered
+ early_stop_flag = True # implement early stopping or not
+ average_validation = [15,25] # if not implementing early stopping, what epochs are we averaging over
num_policies = 5 # fix number of policies
num_sub_policies = 5 # fix number of sub-policies in a policy
iterations = 100 # total iterations, should be more than the number of policies
@@ -230,7 +233,7 @@ if __name__=='__main__':
# generate random policies at start
policies = generate_policies(num_policies, num_sub_policies)
- q_values, best_q_values = run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, early_stop_num, iterations, IsLeNet)
+ q_values, best_q_values = run_UCB1(policies, batch_size, learning_rate, ds, toy_size, max_epochs, early_stop_num, early_stop_flag, average_validation, iterations, IsLeNet)
plt.plot(best_q_values)
diff --git a/MetaAugment/main.py b/MetaAugment/main.py
index b9642879b5896f8f874b230267a36384acd38291..61dec50cc98134d6931d721c79b2d53fd11c750f 100644
--- a/MetaAugment/main.py
+++ b/MetaAugment/main.py
@@ -41,6 +41,8 @@ def train_child_network(child_network,
cost,
max_epochs=2000,
early_stop_num=10,
+ early_stop_flag=True,
+ average_validation=[15,25],
logging=False,
print_every_epoch=True):
if torch.cuda.is_available():
@@ -94,17 +96,26 @@ def train_child_network(child_network,
# correct += torch.sum(_.numpy(), axis=-1)
_sum += _.shape[0]
- # update best validation accuracy if it was higher, otherwise increase early stop count
+
acc = correct / _sum
+ if average_validation[0] <= _epoch <= average_validation[1]:
+ total_val += acc
+
+ # update best validation accuracy if it was higher, otherwise increase early stop count
if acc > best_acc :
best_acc = acc
early_stop_cnt = 0
else:
early_stop_cnt += 1
- # exit if validation gets worse over 10 runs
- if early_stop_cnt >= early_stop_num:
+ # exit if validation gets worse over 10 runs and using early stopping
+ if early_stop_cnt >= early_stop_num and early_stop_flag:
+ break
+
+ # exit if using fixed epoch length
+ if _epoch >= average_validation[1] and not early_stop_flag:
+ best_acc = total_val / (average_validation[1] - average_validation[0] + 1)
break
if print_every_epoch:
diff --git a/requirements.txt b/requirements.txt
index 6892830fe80aef2d9340766deb85b60815f117f7..a19a95d5bc34cbf5c929cdc214e7ed249757ea16 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -2,8 +2,9 @@ torch
torchvision
numpy
Flask
+wheel
react
pytest
tqdm
pygad
-matplotlib
\ No newline at end of file
+matplotlib