def __init__(self, inferrence_json_path, generator_obj):
self.inferrence_json_path = inferrence_json_path
self.generator_obj = generator_obj
local_json_loader = JsonLoader(inferrence_json_path)
local_json_loader.load_json()
self.json_data = local_json_loader.json_data
self.output_file = self.json_data["output_file"]
if "save_raw" in self.json_data.keys():
self.save_raw = self.json_data["save_raw"]
else:
self.save_raw = False
if "rescale" in self.json_data.keys():
self.rescale = self.json_data["rescale"]
else:
self.rescale = True
self.batch_size = self.generator_obj.batch_size
self.nb_datasets = len(self.generator_obj)
self.indiv_shape = self.generator_obj.get_output_size()
self.__load_model()
def __init__(self, inferrence_json_path, generator_obj):
self.inferrence_json_path = inferrence_json_path
self.generator_obj = generator_obj
local_json_loader = JsonLoader(inferrence_json_path)
local_json_loader.load_json()
self.json_data = local_json_loader.json_data
self.output_file = self.json_data["output_file"]
self.model_path = self.json_data["model_path"]
if "save_raw" in self.json_data.keys():
self.save_raw = self.json_data["save_raw"]
else:
self.save_raw = False
if "rescale" in self.json_data.keys():
self.rescale = self.json_data["rescale"]
else:
self.rescale = True
self.batch_size = self.generator_obj.batch_size
self.nb_datasets = len(self.generator_obj)
self.indiv_shape = self.generator_obj.get_output_size()
self.model = load_model(
self.model_path,
custom_objects={"annealed_loss": lc.loss_selector("annealed_loss")},
)
def unet_1024_search(path_json):
local_json_loader = JsonLoader(path_json)
local_json_loader.load_json()
json_data = local_json_loader.json_data
def local_network_function(input_img):
# encoder
local_input = input_img
for local_depth in range(json_data["network_depth"]):
local_conv = Conv2D(
2 ** local_depth * json_data["nb_features_scale"],
(3, 3),
activation="relu",
padding="same",
)(local_input)
local_output = MaxPooling2D(pool_size=(2, 2))(local_conv)
if json_data["unet"]:
if local_depth == 0:
u_net_shortcut = []
u_net_shortcut.append(local_conv)
local_input = local_output
# Deep CONV
deep_conv = Conv2D(
2 ** json_data["network_depth"] * json_data["nb_features_scale"],
(3, 3),
activation="relu",
padding="same",
)(local_input)
# decoder
local_input = deep_conv
for local_depth in range(json_data["network_depth"] - 1, -1, -1):
local_up = UpSampling2D((2, 2))(local_input)
if json_data["unet"]:
local_conc = Concatenate()(
[local_up, u_net_shortcut[local_depth]])
else:
local_conc = local_up
local_output = Conv2D(
2 ** local_depth * json_data["nb_features_scale"],
(3, 3),
activation="relu",
padding="same",
)(local_conc)
local_input = local_output
# output layer
final = Conv2D(1, (1, 1), activation=None,
padding="same")(local_output)
return final
return local_network_function
def __init__(self, inferrence_json_path, generator_obj):
self.inferrence_json_path = inferrence_json_path
self.generator_obj = generator_obj
local_json_loader = JsonLoader(inferrence_json_path)
local_json_loader.load_json()
self.json_data = local_json_loader.json_data
self.output_file = self.json_data["output_file"]
self.model_path = self.json_data["model_path"]
# This is used when output is a full volume to select only the center
# currently only set to true. Future implementation could make smarter
# scanning of the volume and leverage more
# than just the center pixel
if "single_voxel_output_single" in self.json_data.keys():
self.single_voxel_output_single = self.json_data[
"single_voxel_output_single"]
else:
self.single_voxel_output_single = True
self.model_path = self.json_data["model_path"]
self.model = load_model(self.model_path)
self.input_data_size = self.generator_obj.data_shape
def __init__(
self,
generator_obj,
test_generator_obj,
network_obj,
trainer_json_path,
auto_compile=True,
):
self.network_obj = network_obj
self.local_generator = generator_obj
self.local_test_generator = test_generator_obj
json_obj = JsonLoader(trainer_json_path)
# the following line is to be backward compatible in case
# new parameter logics are added.
json_obj.set_default("apply_learning_decay", 0)
json_data = json_obj.json_data
self.output_dir = json_data["output_dir"]
self.run_uid = json_data["run_uid"]
self.model_string = json_data["model_string"]
self.batch_size = json_data["batch_size"]
self.steps_per_epoch = json_data["steps_per_epoch"]
self.loss_type = json_data["loss"]
self.nb_gpus = json_data["nb_gpus"]
self.period_save = json_data["period_save"]
self.learning_rate = json_data["learning_rate"]
if 'checkpoints_dir' in json_data.keys():
self.checkpoints_dir = json_data["checkpoints_dir"]
else:
self.checkpoints_dir = self.output_dir
if "use_multiprocessing" in json_data.keys():
self.use_multiprocessing = json_data["use_multiprocessing"]
else:
self.use_multiprocessing = True
if "caching_validation" in json_data.keys():
self.caching_validation = json_data["caching_validation"]
else:
self.caching_validation = True
self.output_model_file_path = os.path.join(
self.output_dir,
self.run_uid + "_" + self.model_string + "_model.h5"
)
if "nb_workers" in json_data.keys():
self.workers = json_data["nb_workers"]
else:
self.workers = 16
# These parameters are related to setting up the
# behavior of learning rates
self.apply_learning_decay = json_data["apply_learning_decay"]
if self.apply_learning_decay == 1:
self.initial_learning_rate = json_data["initial_learning_rate"]
self.epochs_drop = json_data["epochs_drop"]
self.nb_times_through_data = json_data["nb_times_through_data"]
# Generator has to be initialized first to provide
# input size of network
self.initialize_generator()
if self.nb_gpus > 1:
mirrored_strategy = tensorflow.distribute.MirroredStrategy()
with mirrored_strategy.scope():
if auto_compile:
self.initialize_network()
self.initialize_callbacks()
self.initialize_loss()
self.initialize_optimizer()
if auto_compile:
self.compile()
else:
if auto_compile:
self.initialize_network()
self.initialize_callbacks()
self.initialize_loss()
self.initialize_optimizer()
if auto_compile:
self.compile()
def __init__(self, json_path):
local_json_loader = JsonLoader(json_path)
local_json_loader.load_json()
self.json_data = local_json_loader.json_data
self.local_mean = 1
self.local_std = 1
def __init__(
self,
generator_obj,
test_generator_obj,
network_obj,
trainer_json_path,
auto_compile=True,
):
self.network_obj = network_obj
self.local_generator = generator_obj
self.local_test_generator = test_generator_obj
json_obj = JsonLoader(trainer_json_path)
# the following line is to be backward compatible in case
# new parameter logics are added.
json_obj.set_default("apply_learning_decay", 0)
json_data = json_obj.json_data
self.output_dir = json_data["output_dir"]
self.run_uid = json_data["run_uid"]
self.model_string = json_data["model_string"]
self.batch_size = json_data["batch_size"]
self.steps_per_epoch = json_data["steps_per_epoch"]
self.loss_type = json_data["loss"]
self.nb_gpus = json_data["nb_gpus"]
self.period_save = json_data["period_save"]
self.learning_rate = json_data["learning_rate"]
if "nb_workers" in json_data.keys():
self.workers = json_data["nb_workers"]
else:
self.workers = 16
# These parameters are related to setting up the
# behavior of learning rates
self.apply_learning_decay = json_data["apply_learning_decay"]
if self.apply_learning_decay == 1:
self.initial_learning_rate = json_data["initial_learning_rate"]
self.epochs_drop = json_data["epochs_drop"]
self.nb_times_through_data = json_data["nb_times_through_data"]
# Generator has to be initialized first to provide
# input size of network
#self.initialize_generator()
# modified by sk 2020/11/20
# "nb_max_epoch" aims to reduce to total number of training
# "nb_max_epoch" should not exceed "else" condition
if "nb_max_epoch" in json_data.keys():
self.epochs = json_data["nb_max_epoch"]
else:
self.epochs = self.nb_times_through_data * int(
np.floor(len(self.local_generator) / self.steps_per_epoch)
)
if self.nb_gpus > 1:
mirrored_strategy = tensorflow.distribute.MirroredStrategy()
with mirrored_strategy.scope():
self.initialize_network()
self.initialize_callbacks()
self.initialize_loss()
self.initialize_optimizer()
if auto_compile:
self.compile()
else:
if auto_compile:
self.initialize_network()
self.initialize_callbacks()
self.initialize_loss()
self.initialize_optimizer()
if auto_compile:
self.compile()
