normalizer,
args.batch_size,
shuffle=True)
val_data_gen = utils.BatchGenDeepSupervision(val_data_loader,
discretizer,
normalizer,
args.batch_size,
shuffle=False)
else:
# Set number of batches in one epoch
train_nbatches = 2000
val_nbatches = 1000
if args.small_part:
train_nbatches = 40
val_nbatches = 40
train_data_gen = utils.BatchGen(train_reader, discretizer, normalizer,
args.batch_size, train_nbatches, True)
val_data_gen = utils.BatchGen(val_reader, discretizer, normalizer,
args.batch_size, val_nbatches, False)
if args.mode == 'train':
# Prepare training
path = os.path.join(
args.output_dir, 'keras_states/' + model.final_name +
'.chunk{epoch}.test{val_loss}.state')
metrics_callback = keras_utils.DecompensationMetrics(
train_data_gen=train_data_gen,
val_data_gen=val_data_gen,
deep_supervision=args.deep_supervision,
batch_size=args.batch_size,
# Load data and prepare generators
if args.deep_supervision:
train_data_gen = utils.BatchGenDeepSupervisoin(train_data_loader,
discretizer, normalizer,
args.batch_size)
val_data_gen = utils.BatchGenDeepSupervisoin(val_data_loader, discretizer,
normalizer, args.batch_size)
else:
# Set number of batches in one epoch
train_nbatches = 2000
val_nbatches = 1000
if (args.small_part):
train_nbatches = 40
val_nbatches = 40
train_data_gen = utils.BatchGen(train_reader, discretizer, normalizer,
args.batch_size, train_nbatches)
val_data_gen = utils.BatchGen(val_reader, discretizer, normalizer,
args.batch_size, val_nbatches)
#train_data_gen.steps = train_reader.get_number_of_examples() // args.batch_size
#val_data_gen.steps = val_reader.get_number_of_examples() // args.batch_size
if args.mode == 'train':
# Prepare training
path = 'keras_states/' + model.final_name + '.chunk{epoch}.test{val_loss}.state'
metrics_callback = keras_utils.MetricsBinaryFromGenerator(
train_data_gen, val_data_gen, args.batch_size, args.verbose)
# make sure save directory exists
dirname = os.path.dirname(path)
if not os.path.exists(dirname):
# Load data and prepare generators
if args.deep_supervision:
train_data_gen = utils.BatchGenDeepSupervisoin(train_data_loader,
discretizer, normalizer,
args.batch_size, True)
val_data_gen = utils.BatchGenDeepSupervisoin(val_data_loader, discretizer,
normalizer, args.batch_size,
False)
else:
# Set number of batches in one epoch
train_nbatches = 2000
val_nbatches = 1000
if args.small_part:
train_nbatches = 40
val_nbatches = 40
train_data_gen = utils.BatchGen(train_reader, discretizer, normalizer,
args.batch_size, train_nbatches, True)
val_data_gen = utils.BatchGen(val_reader, discretizer, normalizer,
args.batch_size, val_nbatches, False)
if args.mode == 'train':
# Prepare training
path = 'keras_states/' + model.final_name + '.chunk{epoch}.test{val_loss}.state'
metrics_callback = keras_utils.MetricsBinaryFromGenerator(
train_data_gen, val_data_gen, args.batch_size, args.verbose)
# make sure save directory exists
dirname = os.path.dirname(path)
if not os.path.exists(dirname):
os.makedirs(dirname)
saver = ModelCheckpoint(path, verbose=1, period=args.save_every)