root_path = os.path.abspath('..')
if root_path not in sys.path:
sys.path.append(root_path)
import config
from faces_loader import load_faces, build_dataset
from ssd_utils import output_encoder
from ssd_utils.metrics import MeanAveragePrecision
from utils import import_by_name, MeanAveragePrecisionCallback
test_image_paths, test_bnd_boxes = load_faces(split='test')
test_data = build_dataset(test_image_paths, test_bnd_boxes,
image_size=config.IMAGE_SIZE,
batch_size=config.BATCH_SIZE)
meanAP_metric = MeanAveragePrecision()
results = {'architecture': [],
'train_samples': [],
'train_type': []}
for run in range(1, config.NUM_RUNS+1):
results['run_{}'.format(run)] = []
for architecture in config.ARCHITECTURES:
model_class = import_by_name('ssd_utils.networks.' + architecture)
model = model_class(num_classes=len(config.CLASSES))
anchors = model.get_anchors(image_shape=config.IMAGE_SIZE + (3,))
root_path = os.path.abspath(os.path.join('..'))
if root_path not in sys.path:
sys.path.append(root_path)
import config
from faces_loader import load_faces, build_dataset
from ssd_utils.ssd_loss import SSDLoss
from utils import import_by_name, train_test_split_tensors, MeanAveragePrecisionCallback
# Load train and validation data
train_image_paths, train_bnd_boxes = load_faces(split='train')
valid_image_paths, valid_bnd_boxes = load_faces(split='valid')
valid_data = build_dataset(valid_image_paths, valid_bnd_boxes,
image_size=config.IMAGE_SIZE,
batch_size=config.BATCH_SIZE)
for run in range(1, config.NUM_RUNS+1):
weights_dir = 'weights_{}'.format(run)
history_dir = 'history_{}'.format(run)
os.makedirs(weights_dir, exist_ok=True)
os.makedirs(history_dir, exist_ok=True)
for architecture in config.ARCHITECTURES:
model_class = import_by_name('ssd_utils.networks.' + architecture)
model_name = architecture.lower() + '_pretrained'
model_file = model_name + '.h5'
model_path = os.path.join(weights_dir, model_file)
sys.path.append(root_path)
import config_pretraining_ablation as config
from faces_loader import load_faces, build_dataset
from ssd_utils.ssd_loss import SSDLoss
from utils import import_by_name, train_test_split_tensors, MeanAveragePrecisionCallback
# Load train and validation data
train_image_paths, train_bnd_boxes = load_faces(split='train')
valid_image_paths, valid_bnd_boxes = load_faces(split='valid')
fake_image_paths, fake_bnd_boxes = load_faces(root=os.path.join(
'..', 'data', 'faces_fake'),
split='fake')
valid_data = build_dataset(valid_image_paths,
valid_bnd_boxes,
image_size=config.IMAGE_SIZE,
batch_size=config.BATCH_SIZE)
for run in range(1, config.NUM_RUNS + 1):
weights_dir = 'weights_pretraining_ablation_{}'.format(run)
history_dir = 'history_pretraining_ablation_{}'.format(run)
os.makedirs(weights_dir, exist_ok=True)
os.makedirs(history_dir, exist_ok=True)
for architecture in config.ARCHITECTURES:
model_class = import_by_name('ssd_utils.networks.' + architecture)
for prop_fake_samples in config.PROP_FAKE_SAMPLES:
num_fake_samples = int(prop_fake_samples * config.NUM_REAL_SAMPLES)
print('\n\nINFO: Considering {} real and {} fake samples'.format(