def __init__(self, images="cifar10", labels=None, config=None):
"""Initializes DeepAugment object
Does following steps:
1. load and preprocess data
2. create child model
3. create controller
4. create notebook (for recording trainings)
5. do initial training
6. create objective function
7. evaluate objective function without augmentation
Args:
images (numpy.array/str): array with shape (n_images, dim1, dim2 , channel_size), or a string with name of keras-dataset (cifar10, fashion_mnsit)
labels (numpy.array): labels of images, array with shape (n_images) where each element is an integer from 0 to number of classes
config (dict): dictionary of configurations, for updating the default config which is:
{
"model": "basiccnn",
"method": "bayesian_optimization",
"train_set_size": 2000,
"val_set_size": 100,
"opt_samples": 3,
"opt_last_n_epochs": 3,
"opt_initial_points": 10,
"child_epochs": 50,
"child_first_train_epochs": 0,
"child_batch_size": 64,
"pre_aug_weights_path": "pre_aug_weights.h5",
"logging": logging,
"notebook_path": f"{EXPERIMENT_FOLDER_PATH}/notebook.csv"
}
"""
self.config = DEFAULT_CONFIG
if config!=None: self.config.update(config)
self.iterated = 0 # keep tracks how many times optimizer iterated
self._load_and_preprocess_data(images, labels)
# define main objects
self.child_model = ChildCNN(self.input_shape, self.num_classes, self.config)
self.controller = Controller(self.config)
self.notebook = Notebook(self.config)
if self.config["child_first_train_epochs"] > 0:
self._do_initial_training()
self.child_model.save_pre_aug_weights()
self.objective_func = Objective(
self.data, self.child_model, self.notebook, self.config
)
#self._evaluate_objective_func_without_augmentation()
self.trial_hyperparams = None
self.trial_no = 0
def run_model(dataset_name, num_classes, epochs, batch_size, policies_path):
data, input_shape = DataOp.load(dataset_name)
data = DataOp.preprocess_normal(data)
wrn_28_10 = ChildCNN(
model_name="wrn_28_10",
input_shape=input_shape,
batch_size=batch_size,
num_classes=num_classes,
pre_augmentation_weights_path="initial_model_weights.h5",
logging=logging,
)
if policies_path == "dont_augment":
policy_str = "non_augmented"
else:
policy_str = "augmented"
csv_logger = CSVLogger(
f"{EXPERIMENT_FOLDER_PATH}/wrn_28_10_training_on_{dataset_name}_{policy_str}.csv"
)
if policies_path == "dont_augment":
history = wrn_28_10.fit_normal(data,
epochs=epochs,
csv_logger=csv_logger)
print(f"Reached validation accuracy is {history['val_acc'][-1]}")
else:
datagen = deepaugment_image_generator(
data["X_train"],
data["y_train"],
policies_path,
batch_size=batch_size,
augment_chance=0.8,
)
print("fitting the model")
history = wrn_28_10.fit_with_generator(
datagen,
data["X_val"],
data["y_val"],
train_data_size=len(data["X_train"]),
epochs=epochs,
csv_logger=csv_logger,
)
print(f"Reached validation accuracy is {history['val_acc'][-1]}")
class DeepAugment:
"""Initiliazes commponents of DeepAugment (e.g. Controller, Child-model, Notebook) and optimizes image augmentation hyperparameters
"""
@logger(logfile_dir=EXPERIMENT_FOLDER_PATH)
def __init__(self, images="cifar10", labels=None, config=None):
"""Initializes DeepAugment object
Does following steps:
1. load and preprocess data
2. create child model
3. create controller
4. create notebook (for recording trainings)
5. do initial training
6. create objective function
7. evaluate objective function without augmentation
Args:
images (numpy.array/str): array with shape (n_images, dim1, dim2 , channel_size), or a string with name of keras-dataset (cifar10, fashion_mnsit)
labels (numpy.array): labels of images, array with shape (n_images) where each element is an integer from 0 to number of classes
config (dict): dictionary of configurations, for updating the default config which is:
{
"model": "basiccnn",
"method": "bayesian_optimization",
"train_set_size": 2000,
"opt_samples": 3,
"opt_last_n_epochs": 3,
"opt_initial_points": 10,
"child_epochs": 50,
"child_first_train_epochs": 0,
"child_batch_size": 64,
"pre_aug_weights_path": "pre_aug_weights.h5",
"logging": logging,
"notebook_path": f"{EXPERIMENT_FOLDER_PATH}/notebook.csv"
}
"""
self.config = DEFAULT_CONFIG
if config!=None: self.config.update(config)
self.iterated = 0 # keep tracks how many times optimizer iterated
self._load_and_preprocess_data(images, labels)
# define main objects
self.child_model = ChildCNN(self.input_shape, self.num_classes, self.config)
self.controller = Controller(self.config)
self.notebook = Notebook(self.config)
if self.config["child_first_train_epochs"] > 0:
self._do_initial_training()
self.child_model.save_pre_aug_weights()
self.objective_func = Objective(
self.data, self.child_model, self.notebook, self.config
)
self._evaluate_objective_func_without_augmentation()
def optimize(self, iterations=300):
"""Optimize objective function hyperparameters using controller and child model
Args:
iterations (int): number of optimization iterations, which the child model will be run
Returns:
pandas.DataFrame: top policies (with highest expected accuracy increase)
"""
# iteratate optimizer
for trial_no in range(self.iterated + 1, self.iterated + iterations + 1):
trial_hyperparams = self.controller.ask()
print("trial:", trial_no, "\n", trial_hyperparams)
f_val = self.objective_func.evaluate(trial_no, trial_hyperparams)
self.controller.tell(trial_hyperparams, f_val)
self.iterated += iterations # update number of previous iterations
self.top_policies = self.notebook.get_top_policies(20)
self.notebook.output_top_policies()
print("\ntop policies are:\n", self.top_policies)
return self.top_policies
def image_generator_with_top_policies(self, images, labels, batch_size=None):
"""
Args:
images (numpy.array): array with shape (N,dim,dim,channek-size)
labels (numpy.array): array with shape (N), where each eleemnt is an integer from 0 to num_classes-1
batch_size (int): batch size of the generator on demand
Returns:
generator: generator for augmented images
"""
if batch_size is None:
batch_size = self.config["child_batch_size"]
top_policies_list = self.top_policies[
['A_aug1_type', 'A_aug1_magnitude', 'A_aug2_type', 'A_aug2_magnitude',
'B_aug1_type', 'B_aug1_magnitude', 'B_aug2_type', 'B_aug2_magnitude',
'C_aug1_type', 'C_aug1_magnitude', 'C_aug2_type', 'C_aug2_magnitude',
'D_aug1_type', 'D_aug1_magnitude', 'D_aug2_type', 'D_aug2_magnitude',
'E_aug1_type', 'E_aug1_magnitude', 'E_aug2_type', 'E_aug2_magnitude']
].to_dict(orient="records")
return deepaugment_image_generator(images, labels, top_policies_list, batch_size=batch_size)
def _load_and_preprocess_data(self, images, labels):
"""Loads and preprocesses data
Records `input_shape`, `data`, and `num_classes` into `self
Args:
images (numpy.array/str): array with shape (n_images, dim1, dim2 , channel_size), or a string with name of keras-dataset (cifar10, fashion_mnsit)
labels (numpy.array): labels of images, array with shape (n_images) where each element is an integer from 0 to number of classes
"""
if isinstance(images, str):
X, y, self.input_shape = DataOp.load(images)
else:
X, y = images, labels
self.input_shape = X.shape[1:]
self.data = DataOp.preprocess(X, y, self.config["train_set_size"])
self.num_classes = DataOp.find_num_classes(self.data)
def _do_initial_training(self):
"""Do the first training without augmentations
Training weights will be used as based to further child model trainings
"""
history = self.child_model.fit(
self.data, epochs=self.config["child_first_train_epochs"]
)
#BUG
#self.notebook.record(
# -1, ["first", 0.0, "first", 0.0, "first", 0.0, 0.0], 1, None, history
#)
def _evaluate_objective_func_without_augmentation(self):
"""Find out what would be the accuracy if augmentation are not applied
"""
no_aug_hyperparams = ["rotate", 0.0, "rotate", 0.0,
"rotate", 0.0, "rotate", 0.0,
"rotate", 0.0, "rotate", 0.0,
"rotate", 0.0, "rotate", 0.0,
"rotate", 0.0, "rotate", 0.0]
f_val = self.objective_func.evaluate(0, no_aug_hyperparams)
self.controller.tell(no_aug_hyperparams, f_val)
def run_full_model(images,
labels,
test_proportion=0.1,
model="wrn_28_10",
epochs=200,
batch_size=64,
policies_path="dont_augment"):
data = {}
data["X_train"], data["X_val"], data["y_train"], data[
"y_val"] = train_test_split(images,
labels,
test_size=test_proportion,
shuffle=True)
data = DataOp.preprocess_normal(data)
input_shape = data["X_train"][0].shape
num_classes = data["y_train"].shape[1]
cnn_config = {
"model": model,
"weights": "imagenet",
"input_shape": input_shape,
"child_batch_size": batch_size,
"pre_augmentation_weights_path": "initial_model_weights.h5",
"logging": logging
}
full_model = ChildCNN(input_shape=input_shape,
num_classes=num_classes,
config=cnn_config)
if policies_path == "dont_augment":
policy_str = "non_augmented"
else:
policy_str = "augmented"
csv_logger = CSVLogger(
f"{EXPERIMENT_FOLDER_PATH}/wrn_28_10_training_on_{policy_str}.csv")
if policies_path == "dont_augment":
history = full_model.fit_normal(data,
epochs=epochs,
csv_logger=csv_logger)
print(f"Reached validation accuracy is {history['val_acc'][-1]}")
else:
datagen = deepaugment_image_generator(
data["X_train"],
data["y_train"],
policies_path,
batch_size=batch_size,
augment_chance=0.8,
)
file = open('datagen.txt', 'w')
file.write(datagen)
file.close()
print("fitting the model")
history = full_model.fit_with_generator(
datagen,
data["X_val"],
data["y_val"],
train_data_size=len(data["X_train"]),
epochs=epochs,
csv_logger=csv_logger,
)
print(f"Reached validation accuracy is {history['val_acc'][-1]}")