def setUp(self):
self.config = get_config('config.yaml', 'validate')
converter_config = get_config('config.yaml', 'prepare')
self.converter = Converter(converter_config)
if not self.config['cuda']:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
self.sess, self.age, self.gender, self.images_pl = load_network(
self.config['pretrained_model_folder_or_file'])
def __init__(self, config):
# parameters
self._models_config = get_config(config, 'models')
self._train_config = get_config(config, 'train')
self._dataset_config = get_config(
config, 'datasets')[self._train_config['dataset']]
if not self._train_config['balance_dataset']:
self._dataset_config['balance'] = None
lr_method = self._train_config['learning_rate']
lr_config = get_config(config, 'learning_rates')
self._lr_config = lr_config[
'linear'] if lr_method == 'test_lr' else lr_config[lr_method]
self.learning_rate_manager = LearningRateManager(
lr_method, self._lr_config)
self.model = models[self._train_config['model']](
**self._models_config[self._train_config['model']])
self.num_epochs = self._train_config['epochs']
self.train_size = 0
self.test_size = None
self.batch_size = self._train_config['batch_size']
self.save_frequency = self._train_config['save_frequency']
self.val_frequency = self._train_config['val_frequency']
self.mode = self._train_config['mode']
self.model_path = self._train_config['model_path']
self.experiment_folder = self.get_experiment_folder(self.mode)
# operations
self.global_step = self.model.global_step
self.train_mode = tf.placeholder(tf.bool)
self.init_op = None
self.train_op = None
self.reset_global_step_op = None
self.train_summary = None
self.train_init_op = None
self.test_summary = None
self.test_init_op = None
self.images = tf.placeholder(tf.float32, shape=[None, 256, 256, 3])
self.age_labels = tf.placeholder(tf.int32)
self.gender_labels = tf.placeholder(tf.int32)
self.test_lr = list(
) if self.learning_rate_manager.method_name == 'test_lr' else None
def validate(config):
validation_config = get_config(config, 'dataset_validation')
dataset_config = get_config(config,
'datasets')[validation_config['dataset']]
batch_size = validation_config['batch']
num_epochs = validation_config['epochs']
next_data_element, train_init_op, train_size = init_data_loader(
batch_size,
dataset_config['full_desc_path'],
dataset_config['images_path'],
dataset_config['balance'],
epochs=num_epochs)
print('dataset_size: ', train_size)
print('train_size // batch_size', train_size // batch_size)
num_batches = train_size // batch_size + \
(train_size % batch_size != 0)
with tf.Graph().as_default() and tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('num_epochs*num_batches', num_epochs * num_batches)
sess.run(train_init_op)
for batch_idx in range(num_epochs * num_batches):
train_images, train_age_labels, train_gender_labels, file_paths = sess.run(
next_data_element)
print(f'batch_inx: {batch_idx}, file_paths_len: {len(file_paths)}')
import argparse
from functools import partial
from age_gender.utils.config_parser import get_config
from age_gender.utils.converter import ConverterManager
from age_gender.utils.splitter import train_test_split_dataset
if __name__ == "__main__":
config = get_config('config.yaml', 'prepare')
parser = argparse.ArgumentParser()
choices = {'convert_dataset': ConverterManager(config).run,
'split_dataset': partial(train_test_split_dataset, config)}
parser.add_argument('command', type=str, choices=choices.keys(), help='dataset preparation command')
args = parser.parse_args()
choices[args.command]()
json.dump(positive, Path(config['results_path']).joinpath('positive.json').open(mode='w'))
def load_network(config):
pretrained_model_folder_or_file = config['pretrained_model_folder_or_file']
sess = tf.Session()
images_pl = tf.placeholder(tf.float32, shape=[None, 256, 256 ,3], name='input_image')
images_norm = tf.map_fn(lambda frame: tf.image.per_image_standardization(frame), images_pl)
model = InceptionResnetV1(phase_train=False, is_training=False)
variables_to_restore, age_logits, gender_logits = model.inference(images_norm)
gender = tf.argmax(tf.nn.softmax(gender_logits), 1)
age_ = tf.cast(tf.constant([i for i in range(0, 101)]), tf.float32)
age = tf.reduce_sum(tf.multiply(tf.nn.softmax(age_logits), age_), axis=1)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
saver = ModelSaver(variables_to_restore)
saver.restore_model(sess, pretrained_model_folder_or_file)
return sess, age, gender, images_pl
if __name__ == '__main__':
config = get_config('config.yaml','mining')
converter_config = get_config('config.yaml','prepare')
converter = Converter(converter_config)
if not config['cuda']:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
sess, age, gender, images_pl = load_network(config)
Miner(config, converter, sess).run()
if name != 'lr':
metrics_deque[name].append(metric)
metric = np.mean(metrics_deque[name])
test_lr_chunk[name] = float(metric)
summary = summary_pb2.Summary.Value(tag=f'{mode}/{name}',
simple_value=metric)
summaries_list.append(summary)
if test_lr is not None:
test_lr_chunk['lr'] = float(
metrics_and_errors['lr']
) # необходимо значение lr взятое из train стадии
test_lr.append(test_lr_chunk)
summaries = summary_pb2.Summary(value=summaries_list)
return summaries
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--config",
type=str,
default="config.yaml",
help="config")
args = parser.parse_args()
config = get_config(args.config)
if not config['train']['cuda']:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
if config['train']['mode'] not in ['start', 'continue', 'test']:
raise ValueError('Invalid mode!')
ModelManager(config).train()
def init_data_loader(self, dataset_path):
dataset_json = json.load(Path(dataset_path).open())
data_folder = os.path.dirname(dataset_path)
if Path(self._config['results_path']).is_file():
prev_result = json.load(open(self._config['results_path'], 'r'))
for name in self.results_names:
self.results[name] += prev_result[name]
processed_files = [
fn[fn.find(data_folder) + len(data_folder) + 1:]
for fn in prev_result['file_name']
]
dataset_json = list(
filter(lambda it: it['file_name'] not in processed_files,
dataset_json))
loader = DataLoader(dataset_json, data_folder)
dataset = loader.create_dataset(perform_shuffle=False,
batch_size=self.batch_size)
iterator = tf.data.Iterator.from_structure(dataset.output_types,
dataset.output_shapes)
next_data_element = iterator.get_next()
data_loader_init_op = iterator.make_initializer(dataset)
return next_data_element, data_loader_init_op, loader.dataset_len()
if __name__ == '__main__':
config = get_config('config.yaml', 'inference')
print(config)
if not config['cuda']:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
ModelManager(config).inference()
validation_config = get_config(config, 'dataset_validation')
dataset_config = get_config(config,
'datasets')[validation_config['dataset']]
batch_size = validation_config['batch']
num_epochs = validation_config['epochs']
next_data_element, train_init_op, train_size = init_data_loader(
batch_size,
dataset_config['full_desc_path'],
dataset_config['images_path'],
dataset_config['balance'],
epochs=num_epochs)
print('dataset_size: ', train_size)
print('train_size // batch_size', train_size // batch_size)
num_batches = train_size // batch_size + \
(train_size % batch_size != 0)
with tf.Graph().as_default() and tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('num_epochs*num_batches', num_epochs * num_batches)
sess.run(train_init_op)
for batch_idx in range(num_epochs * num_batches):
train_images, train_age_labels, train_gender_labels, file_paths = sess.run(
next_data_element)
print(f'batch_inx: {batch_idx}, file_paths_len: {len(file_paths)}')
if __name__ == '__main__':
config = get_config('config.yaml')
if not config['dataset_validation']['cuda']:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
validate(config)
