def create_dataset(args):
''' Create train and val dataset
'''
# Data loading code
if args.task == 'syn_t1':
dataset = MRISynDataset
elif args.task == 'seg_oct':
dataset = OCTSegDataset
else:
raise NotImplementedError
train_dataset = dataset(args, train=True, augment=True)
val_dataset_list = []
if args.vimg_path:
val_dataset = dataset(args, train=False, augment=False)
# return a list of val_dataset, each contains one subject
if args.__dict__.get('sub_name',False):
if os.path.isfile(args.sub_name):
with open(args.sub_name) as f:
dataname = f.read().splitlines()
datalist = val_dataset.datalist
labellist = val_dataset.labellist
for name in dataname:
val_dataset.datalist = sorted([_ for _ in datalist if name in str(_)])
val_dataset.labellist = sorted([_ for _ in labellist if name in str(_)])
val_dataset_list.append(deepcopy(val_dataset))
else:
val_dataset_list = [val_dataset]
elif args.split and not args.__dict__.get('sub_name',False):
train_dataset, val_dataset = split_data(dataset=train_dataset,
split=args.split,
switch=args.test or args.evaluate)
train_dataset = train_dataset
val_dataset = val_dataset
val_dataset_list = [val_dataset]
else:
# no validation used
val_dataset = dataset(args, train=False)
val_dataset_list = [val_dataset]
logger.info('Found {} training samples'.format(len(train_dataset)))
logger.info('Found {} validation subjects with total {} samples'.format(len(val_dataset_list),len(val_dataset)))
train_loader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = [DataLoader(dataset=_,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True) for _ in val_dataset_list]
return train_loader, val_loader
def run():
# declare global variables
global config
global config_path
# add config file to experiment
experiment.add_artifact(config_path)
if config['dataset'].get('link', True):
dataset_config_path = f'../configs/datasets/{ config["dataset"]["link"] }'
experiment.add_artifact(dataset_config_path)
config['dataset'].update(json.load(open(dataset_config_path)))
# dataset specific variables
folds = config['dataset']['split']
data_directory = config['dataset']['path']
# split dataset into k folds
split_dirs = split_data(folds, data_directory, data_name)
# total results dictionary
results = {
'f1': [],
'rec': [],
'acc': [],
'mcc': [],
'prec': [],
'spec': []
}
# iterate over each dataset split
for split_index in range(len(split_dirs)):
# print current validation split index
print(f'start validating on split {split_index}')
# restart keras session
K.clear_session()
# prepare dataset by distributing the k splits
# into training and validation sets
training_directory, validation_directory = prepare_dataset(
split_dirs, split_index, len(split_dirs))
if config['dataset'].get('validation_extension', False):
extension_path = config['dataset']['validation_extension_path']
for class_extension in os.listdir(extension_path):
class_path = os.path.join(extension_path, class_extension)
target_path = os.path.join(validation_directory,
class_extension)
for filename in os.listdir(class_path):
shutil.copy(os.path.join(class_path, filename),
os.path.join(target_path, filename))
# print training directories for sanity
print(f'training on {training_directory}')
print(f'validation on {validation_directory}')
# load model from model file or build it using a build file.
if config['model'].get('load_model', False):
model = load_model(config['model']['model_splits'][split_index])
else:
model_builder_path = config['model']['build_file']
model_builder = importlib.import_module(
f'models.{model_builder_path}')
model = model_builder.build(config)
# train model and get last weigths
if config['model'].get('train', True):
print("Start training...")
model = train(model, config, experiment, training_directory,
validation_directory, f'split_{split_index}')
evaluate(model, config, experiment, validation_directory,
f'split_{split_index}')
# if fine tune, train model again on config link found in config
if config.get('fine_tuning', {}).get(
'link', False) and config['model'].get('train', True):
print("Start fine tuning...")
# load config link from config
fine_tuning_config_name = config['fine_tuning']['link']
fine_tuning_config_path = f'../configs/links/{fine_tuning_config_name}'
fine_tuning_config = json.load(open(fine_tuning_config_path))
if fine_tuning_config['dataset'].get('link', True):
dataset_config_path = f'../configs/datasets/{fine_tuning_config["dataset"]["link"]}'
experiment.add_artifact(dataset_config_path)
fine_tuning_config['dataset'].update(
json.load(open(dataset_config_path)))
# add link config to experiment
experiment.add_artifact(fine_tuning_config_path)
# train using new config
model = train(model, fine_tuning_config, experiment,
training_directory, validation_directory,
f'fine_split_{split_index}')
# evaluate train model and get metrics
print("Start evaluation...")
split_results = evaluate(model, config, experiment,
validation_directory, f'split_{split_index}')
# merge split results with total results
for key in split_results:
results[key].append(split_results[key])
print(key, results[key])
# log results
log_cross_validation_results(full_kfold_summary_file_path, results,
experiment_name, folds)
log_to_results_comparison(results, experiment_name, folds)
experiment.add_artifact(full_kfold_summary_file_path)
experiment.add_artifact(all_results_file_path)