def __init__(self, ensemble=True, split=None, model_id=None):
"""
Constructor of COVID model evaluator class.
Arguments:
ensemble {str} -- Whether the model ensemble is used.
"""
self.root = os.path.join('/', *DIR_PATH.split('/')[:-1])
self.split = split
self.ensemble = ensemble
if model_id is None:
self.model_id = 'vgg_base'
elif model_id not in MODEL_FACTORY.keys():
raise ValueError(
f'Wrong model {model_id}. Options are:{MODEL_FACTORY.keys()}')
else:
self.model_id = model_id
if ensemble:
# retores 5 weight paths
self.weights_paths = [
os.path.join(self.root, 'trained_models', 'fold_' + str(fold),
"variables", "variables")
for fold in range(NUM_FOLDS)
]
else:
if split is None or split < 0 or split > 4:
raise ValueError(f'Provide split between 0 and 4, not {split}')
fold = split
self.weights_paths = [
os.path.join(self.root, 'trained_models', 'fold_' + str(fold),
"variables", "variables")
]
self.class_mappings = ['covid', 'pneunomia', 'regular']
self.models = [
MODEL_FACTORY[self.model_id]()
for _ in range(len(self.weights_paths))
]
# restore weights
try:
for model, path in zip(self.models, self.weights_paths):
model.load_weights(path)
except Exception:
raise Exception('Error in model restoring.')
print(f'Model restored. Class mappings are {self.class_mappings}')
# Initialize hyperparameters
DATA_DIR = args['data_dir']
MODEL_NAME = args['model_name']
FOLD = args['fold']
MODEL_DIR = os.path.join(args['model_dir'], MODEL_NAME, f'fold_{FOLD}')
LR = args['learning_rate']
EPOCHS = args['epochs']
BATCH_SIZE = args['batch_size']
MODEL_ID = args['model_id']
TRAINABLE_BASE_LAYERS = args['trainable_base_layers']
IMG_WIDTH, IMG_HEIGHT = args['img_width'], args['img_height']
LOG_SOFTMAX = args['log_softmax']
HIDDEN_SIZE = args['hidden_size']
# Check if model class exists
if MODEL_ID not in MODEL_FACTORY.keys():
raise ValueError(
f'Model {MODEL_ID} not implemented. Choose from {MODEL_FACTORY.keys()}'
)
if not os.path.exists(MODEL_DIR):
os.makedirs(MODEL_DIR)
print(f'Model parameters: {args}')
# grab the list of images in our dataset directory, then initialize
# the list of data (i.e., images) and class images
print('Loading images...')
imagePaths = list(paths.list_images(DATA_DIR))
data = []
labels = []
print(f'selected fold: {FOLD}')
def __init__(self,
weights_dir=None,
ensemble=True,
split=None,
model_id=None,
num_classes=3,
mc_dropout: bool = False,
test_augment: bool = False):
"""
Constructor of COVID model evaluator class.
Arguments:
ensemble {str} -- Whether the model ensemble is used.
num_classes: must be 3 or 4, how many classes the model was
trained on
"""
self.root = os.path.join('/', *DIR_PATH.split('/')[:-1])
if weights_dir is None:
weights_dir = os.path.join(self.root, "trained_models")
self.split = split
self.ensemble = ensemble
assert num_classes == 3 or num_classes == 4, "must be 3 or 4 classes"
if model_id is None:
self.model_id = 'vgg_base'
elif model_id not in MODEL_FACTORY.keys():
raise ValueError(
f'Wrong model {model_id}. Options are:{MODEL_FACTORY.keys()}')
else:
self.model_id = model_id
if ensemble:
# retores 5 weight paths
self.weights_paths = [
os.path.join(weights_dir, 'fold_' + str(fold), "best_weights",
"variables", "variables")
for fold in range(NUM_FOLDS)
]
else:
if split is None or split < 0 or split > 4:
raise ValueError(f'Provide split between 0 and 4, not {split}')
fold = split
self.weights_paths = [
os.path.join(weights_dir, 'fold_' + str(fold), "best_weights",
"variables", "variables")
]
self.class_mappings = CLASS_MAPPING[num_classes]
self.models = [
MODEL_FACTORY[self.model_id](num_classes=num_classes,
mc_dropout=mc_dropout)
for _ in range(len(self.weights_paths))
]
self.mc_dropout = mc_dropout
self.augmentor = ImageDataGenerator(rotation_range=10,
fill_mode='nearest',
horizontal_flip=True,
vertical_flip=True,
width_shift_range=0.1,
height_shift_range=0.1)
# restore weights
try:
for model, path in zip(self.models, self.weights_paths):
model.load_weights(path)
except Exception:
raise Exception('Error in model restoring.')
print(f'Model restored. Class mappings are {self.class_mappings}')