def load_weights(self, log_dir=None, type='latest', **kwargs):
if not log_dir:
log_dir = self.log_dir
if type == 'latest':
init_checkpoint_path = tfutils.get_latest_model_checkpoint_path(
log_dir, 'model.ckpt')
elif type == 'best_dice':
init_checkpoint_path = tfutils.get_latest_model_checkpoint_path(
log_dir, 'model_best_dice.ckpt')
elif type == 'best_loss':
init_checkpoint_path = tfutils.get_latest_model_checkpoint_path(
log_dir, 'model_best_loss.ckpt')
elif type == 'best_ged':
init_checkpoint_path = tfutils.get_latest_model_checkpoint_path(
log_dir, 'model_best_ged.ckpt')
elif type == 'iter':
assert 'iteration' in kwargs, "argument 'iteration' must be provided for type='iter'"
iteration = kwargs['iteration']
init_checkpoint_path = os.path.join(log_dir,
'model.ckpt-%d' % iteration)
else:
raise ValueError(
'Argument type=%s is unknown. type can be latest/iter.' % type)
self.saver.restore(self.sess, init_checkpoint_path)
def load_weights(self, log_dir=None, type='latest', **kwargs):
"""
Load weights into the model
:param log_dir: experiment directory into which all the checkpoints have been written
:param type: can be 'latest', 'best_wasserstein' (highest validation Wasserstein distance), or 'iter' (specific
iteration, requires passing the iteration argument with a valid step number from the checkpoint
files)
"""
if not log_dir:
log_dir = self.log_dir
if type == 'latest':
init_checkpoint_path = tf_utils.get_latest_model_checkpoint_path(
log_dir, 'model.ckpt')
elif type == 'best_wasserstein':
init_checkpoint_path = tf_utils.get_latest_model_checkpoint_path(
log_dir, 'model_best_wasserstein.ckpt')
elif type == 'iter':
assert 'iteration' in kwargs, "argument 'iteration' must be provided for type='iter'"
iteration = kwargs['iteration']
init_checkpoint_path = os.path.join(log_dir,
'model.ckpt-%d' % iteration)
else:
raise ValueError(
'Argument type=%s is unknown. type can be latest/best_wasserstein/iter.'
% type)
self.saver.restore(self.sess, init_checkpoint_path)
def _setup_log_dir_and_continue_mode(self):
# Default values
self.log_dir = os.path.join(sys_config.log_root, 'classifier', self.exp_config.experiment_name)
self.init_checkpoint_path = None
self.continue_run = False
self.init_step = 0
# If a checkpoint file already exists enable continue mode
if tf.gfile.Exists(self.log_dir):
init_checkpoint_path = tf_utils.get_latest_model_checkpoint_path(self.log_dir, 'model.ckpt')
if init_checkpoint_path is not False:
self.init_checkpoint_path = init_checkpoint_path
self.continue_run = True
self.init_step = int(self.init_checkpoint_path.split('/')[-1].split('-')[-1])
self.log_dir += '_cont'
logging.info('--------------------------- Continuing previous run --------------------------------')
logging.info('Checkpoint path: %s' % self.init_checkpoint_path)
logging.info('Latest step was: %d' % self.init_step)
logging.info('------------------------------------------------------------------------------------')
tf.gfile.MakeDirs(self.log_dir)
self.summary_writer = tf.summary.FileWriter(self.log_dir, self.sess.graph)
# Copy experiment config file to log_dir for future reference
shutil.copy(self.exp_config.__file__, self.log_dir)
