def main(config_file: str):
config = cfg.load(config_file, cfg.Configuration)
# set up directories and logging
model_dir, result_dir = fs.prepare_directories(
config_file, cfg.Configuration, lambda: fs.get_directory_name(config))
config.model_dir = model_dir
config.result_dir = result_dir
print(config)
# set seed before model instantiation
print('Set seed to {}'.format(config.seed))
seed.set_seed(config.seed, config.cudnn_determinism)
# load train and valid subjects from split file (also test but it is unused)
subjects_train, subjects_valid, subjects_test = split.load_split(
config.split_file)
print('Train subjects:', subjects_train)
print('Valid subjects:', subjects_valid)
# set up data handling
data_handler = hdlr.MRFDataHandler(
config,
subjects_train,
subjects_valid,
subjects_test,
False,
padding_size=mdl.get_padding_size(config))
# extract a sample for model initialization
data_handler.dataset.set_extractor(data_handler.extractor_train)
sample = data_handler.dataset[0]
with tf.Session() as sess:
model = mdl.get_model(config)(sess, sample, config)
print('Number of parameters:', model.get_number_parameters())
logger = log.TensorFlowLogger(config.model_dir, sess,
model.epoch_summaries(),
model.batch_summaries(),
model.visualization_summaries())
trainer = train.AssemblingTesterTensorFlow(
data_handler, logger, config, model,
sess) # use this class to test the pipeline
#trainer = train.MRFTrainer(data_handler, logger, config, model, sess)
tf.get_default_graph().finalize(
) # to ensure that no ops are added during training, which would lead to
# a growing graph
trainer.train()
logger.close()
def main(model_dir: str, result_dir: str, do_occlusion: bool):
if not os.path.isdir(model_dir):
raise RuntimeError('Model dir "{}" does not exist'.format(model_dir))
config = cfg.load(
glob.glob(os.path.join(model_dir, '*config*.json'))[0],
cfg.Configuration)
split_file = glob.glob(os.path.join(model_dir, '*split*.json'))[0]
os.makedirs(result_dir, exist_ok=True)
# load train, valid, and test subjects from split file
subjects_train, subjects_valid, subjects_test = split.load_split(
split_file)
print('Test subjects:', subjects_test)
# set up data handling
data_handler = hdlr.MRFDataHandler(
config,
subjects_train,
subjects_valid,
subjects_test,
False,
padding_size=mdl.get_padding_size(config))
# extract a sample for model initialization
data_handler.dataset.set_extractor(data_handler.extractor_train)
sample = data_handler.dataset[0]
with tf.Session() as sess:
model = mdl.get_model(config)(sess, sample, config)
tester = test.MRFTensorFlowTester(data_handler, model, model_dir,
result_dir, config.maps, sess)
tester.load(os.path.join(model_dir, config.best_model_file_name))
print('Predict...')
tester.predict()
if do_occlusion:
occlude(tester, result_dir, sample[pymia_def.KEY_IMAGES].shape[-2])
def prepare_directories(config_file, config_cls,
directory_name_fn) -> (str, str):
"""Prepares the directories for an experiment.
Args:
config_file: The config file path.
config_cls: The config file class.
directory_name_fn: Lambda to a function that returns a the name for the directories to create, i.e. a str.
Returns:
A tuple with the paths to the created model and result directories.
"""
config = cfg.load(config_file, config_cls)
# create required directories
suffix = directory_name_fn()
model_dir = os.path.join(config.model_dir, suffix)
result_dir = os.path.join(config.result_dir, suffix)
# create directories
os.makedirs(model_dir, exist_ok=True)
os.makedirs(result_dir, exist_ok=True)
# copy config file
shutil.copyfile(config_file, os.path.join(result_dir, 'config.json'))
shutil.copyfile(config_file, os.path.join(model_dir, 'config.json'))
# copy split file
if os.path.exists(config.split_file):
shutil.copyfile(
config.split_file,
os.path.join(result_dir, os.path.basename(config.split_file)))
shutil.copyfile(
config.split_file,
os.path.join(model_dir, os.path.basename(config.split_file)))
return model_dir, result_dir