def __init__(self,
model_dir=None,
initial_iter=0,
tensorboard_every_n=0,
**_unused):
self.tensorboard_every_n = tensorboard_every_n
# creating new summary subfolder if it's not finetuning
self.summary_dir = get_latest_subfolder(
os.path.join(model_dir, 'logs'), create_new=initial_iter == 0)
self.writer_train = None
self.writer_valid = None
GRAPH_CREATED.connect(self.init_writer)
ITER_STARTED.connect(self.read_tensorboard_op)
ITER_FINISHED.connect(self.write_tensorboard)
def initialise_application(self, workflow_param, data_param):
"""
This function receives all parameters from user config file,
create an instance of application.
:param workflow_param: a dictionary of user parameters,
keys correspond to sections in the config file
:param data_param: a dictionary of input image parameters,
keys correspond to data properties to be used by image_reader
:return:
"""
try:
system_param = workflow_param.get('SYSTEM', None)
net_param = workflow_param.get('NETWORK', None)
train_param = workflow_param.get('TRAINING', None)
infer_param = workflow_param.get('INFERENCE', None)
app_param = workflow_param.get('CUSTOM', None)
except AttributeError:
tf.logging.fatal('parameters should be dictionaries')
raise
assert os.path.exists(system_param.model_dir), \
'Model folder not exists {}'.format(system_param.model_dir)
self.is_training = (system_param.action == "train")
# hardware-related parameters
self.num_threads = max(system_param.num_threads, 1) \
if self.is_training else 1
self.num_gpus = system_param.num_gpus \
if self.is_training else min(system_param.num_gpus, 1)
set_cuda_device(system_param.cuda_devices)
# set output TF model folders
self.model_dir = touch_folder(
os.path.join(system_param.model_dir, 'models'))
self.session_prefix = os.path.join(self.model_dir, FILE_PREFIX)
if self.is_training:
assert train_param, 'training parameters not specified'
summary_root = os.path.join(system_param.model_dir, 'logs')
self.summary_dir = get_latest_subfolder(
summary_root, create_new=train_param.starting_iter == 0)
# training iterations-related parameters
self.initial_iter = train_param.starting_iter
self.final_iter = train_param.max_iter
self.save_every_n = train_param.save_every_n
self.tensorboard_every_n = train_param.tensorboard_every_n
self.max_checkpoints = train_param.max_checkpoints
self.gradients_collector = GradientsCollector(
n_devices=max(self.num_gpus, 1))
action_param = train_param
else:
assert infer_param, 'inference parameters not specified'
self.initial_iter = infer_param.inference_iter
action_param = infer_param
self.outputs_collector = OutputsCollector(
n_devices=max(self.num_gpus, 1))
# create an application instance
assert app_param, 'application specific param. not specified'
app_module = ApplicationDriver._create_app(app_param.name)
self.app = app_module(net_param, action_param, self.is_training)
# initialise data input
self.app.initialise_dataset_loader(data_param, app_param)
# pylint: disable=not-context-manager
with self.graph.as_default(), tf.name_scope('Sampler'):
self.app.initialise_sampler()
def initialise_application(self, workflow_param, data_param):
"""
This function receives all parameters from user config file,
create an instance of application.
:param workflow_param: a dictionary of user parameters,
keys correspond to sections in the config file
:param data_param: a dictionary of input image parameters,
keys correspond to data properties to be used by image_reader
:return:
"""
try:
system_param = workflow_param.get('SYSTEM', None)
net_param = workflow_param.get('NETWORK', None)
train_param = workflow_param.get('TRAINING', None)
infer_param = workflow_param.get('INFERENCE', None)
app_param = workflow_param.get('CUSTOM', None)
except AttributeError:
tf.logging.fatal('parameters should be dictionaries')
raise
assert os.path.exists(system_param.model_dir), \
'Model folder not exists {}'.format(system_param.model_dir)
self.is_training = (system_param.action == "train")
# hardware-related parameters
self.num_threads = max(system_param.num_threads, 1) \
if self.is_training else 1
self.num_gpus = system_param.num_gpus \
if self.is_training else min(system_param.num_gpus, 1)
set_cuda_device(system_param.cuda_devices)
# set output TF model folders
self.model_dir = touch_folder(
os.path.join(system_param.model_dir, 'models'))
self.session_prefix = os.path.join(self.model_dir, FILE_PREFIX)
# set training params.
if self.is_training:
assert train_param, 'training parameters not specified'
summary_root = os.path.join(system_param.model_dir, 'logs')
self.summary_dir = get_latest_subfolder(
summary_root,
create_new=train_param.starting_iter == 0)
self.initial_iter = train_param.starting_iter
self.final_iter = max(train_param.max_iter, self.initial_iter)
self.save_every_n = train_param.save_every_n
self.tensorboard_every_n = train_param.tensorboard_every_n
self.max_checkpoints = \
max(train_param.max_checkpoints, self.max_checkpoints)
self.gradients_collector = GradientsCollector(
n_devices=max(self.num_gpus, 1))
self.validation_every_n = train_param.validation_every_n
if self.validation_every_n > 0:
self.validation_max_iter = max(self.validation_max_iter,
train_param.validation_max_iter)
action_param = train_param
else: # set inference params.
assert infer_param, 'inference parameters not specified'
self.initial_iter = infer_param.inference_iter
action_param = infer_param
self.outputs_collector = OutputsCollector(
n_devices=max(self.num_gpus, 1))
# create an application instance
assert app_param, 'application specific param. not specified'
app_module = ApplicationDriver._create_app(app_param.name)
self.app = app_module(net_param, action_param, system_param.action)
# initialise data input
data_partitioner = ImageSetsPartitioner()
# clear the cached file lists
data_partitioner.reset()
do_new_partition = \
self.is_training and self.initial_iter == 0 and \
(not os.path.isfile(system_param.dataset_split_file)) and \
(train_param.exclude_fraction_for_validation > 0 or
train_param.exclude_fraction_for_inference > 0)
data_fractions = None
if do_new_partition:
assert train_param.exclude_fraction_for_validation > 0 or \
self.validation_every_n <= 0, \
'validation_every_n is set to {}, ' \
'but train/validation splitting not available,\nplease ' \
'check "exclude_fraction_for_validation" in the config ' \
'file (current config value: {}).'.format(
self.validation_every_n,
train_param.exclude_fraction_for_validation)
data_fractions = (train_param.exclude_fraction_for_validation,
train_param.exclude_fraction_for_inference)
if data_param:
data_partitioner.initialise(
data_param=data_param,
new_partition=do_new_partition,
ratios=data_fractions,
data_split_file=system_param.dataset_split_file)
if data_param and self.is_training and self.validation_every_n > 0:
assert data_partitioner.has_validation, \
'validation_every_n is set to {}, ' \
'but train/validation splitting not available.\nPlease ' \
'check dataset partition list {} ' \
'(remove file to generate a new dataset partition). ' \
'Or set validation_every_n to -1.'.format(
self.validation_every_n, system_param.dataset_split_file)
# initialise readers
self.app.initialise_dataset_loader(
data_param, app_param, data_partitioner)
self._data_partitioner = data_partitioner
# pylint: disable=not-context-manager
with self.graph.as_default(), tf.name_scope('Sampler'):
self.app.initialise_sampler()
def initialise_application(self, workflow_param, data_param):
"""
This function receives all parameters from user config file,
create an instance of application.
:param workflow_param: a dictionary of user parameters,
keys correspond to sections in the config file
:param data_param: a dictionary of input image parameters,
keys correspond to data properties to be used by image_reader
:return:
"""
try:
system_param = workflow_param.get('SYSTEM', None)
net_param = workflow_param.get('NETWORK', None)
train_param = workflow_param.get('TRAINING', None)
infer_param = workflow_param.get('INFERENCE', None)
app_param = workflow_param.get('CUSTOM', None)
except AttributeError:
tf.logging.fatal('parameters should be dictionaries')
raise
assert os.path.exists(system_param.model_dir), \
'Model folder not exists {}'.format(system_param.model_dir)
self.is_training = (system_param.action == "train")
# hardware-related parameters
self.num_threads = max(system_param.num_threads, 1) \
if self.is_training else 1
self.num_gpus = system_param.num_gpus \
if self.is_training else min(system_param.num_gpus, 1)
set_cuda_device(system_param.cuda_devices)
# set output TF model folders
self.model_dir = touch_folder(
os.path.join(system_param.model_dir, 'models'))
self.session_prefix = os.path.join(self.model_dir, FILE_PREFIX)
if self.is_training:
assert train_param, 'training parameters not specified'
summary_root = os.path.join(system_param.model_dir, 'logs')
self.summary_dir = get_latest_subfolder(
summary_root, create_new=train_param.starting_iter == 0)
self.initial_iter = train_param.starting_iter
self.final_iter = max(train_param.max_iter, self.initial_iter)
self.save_every_n = train_param.save_every_n
self.tensorboard_every_n = train_param.tensorboard_every_n
self.max_checkpoints = \
max(train_param.max_checkpoints, self.max_checkpoints)
self.gradients_collector = GradientsCollector(
n_devices=max(self.num_gpus, 1))
self.validation_every_n = train_param.validation_every_n
if self.validation_every_n > 0:
self.validation_max_iter = max(self.validation_max_iter,
train_param.validation_max_iter)
action_param = train_param
else:
assert infer_param, 'inference parameters not specified'
self.initial_iter = infer_param.inference_iter
action_param = infer_param
self.outputs_collector = OutputsCollector(
n_devices=max(self.num_gpus, 1))
# create an application instance
assert app_param, 'application specific param. not specified'
app_module = ApplicationDriver._create_app(app_param.name)
self.app = app_module(net_param, action_param, self.is_training)
# initialise data input
data_partitioner = ImageSetsPartitioner()
# clear the cached file lists
data_partitioner.reset()
do_new_partition = \
self.is_training and self.initial_iter == 0 and \
(not os.path.isfile(system_param.dataset_split_file)) and \
(train_param.exclude_fraction_for_validation > 0 or
train_param.exclude_fraction_for_inference > 0)
data_fractions = None
if do_new_partition:
assert train_param.exclude_fraction_for_validation > 0 or \
self.validation_every_n <= 0, \
'validation_every_n is set to {}, ' \
'but train/validation splitting not available,\nplease ' \
'check "exclude_fraction_for_validation" in the config ' \
'file (current config value: {}).'.format(
self.validation_every_n,
train_param.exclude_fraction_for_validation)
data_fractions = (train_param.exclude_fraction_for_validation,
train_param.exclude_fraction_for_inference)
if data_param:
data_partitioner.initialise(
data_param=data_param,
new_partition=do_new_partition,
ratios=data_fractions,
data_split_file=system_param.dataset_split_file)
if data_param and self.is_training and self.validation_every_n > 0:
assert data_partitioner.has_validation, \
'validation_every_n is set to {}, ' \
'but train/validation splitting not available.\nPlease ' \
'check dataset partition list {} ' \
'(remove file to generate a new dataset partition). ' \
'Or set validation_every_n to -1.'.format(
self.validation_every_n, system_param.dataset_split_file)
# initialise readers
self.app.initialise_dataset_loader(data_param, app_param,
data_partitioner)
self._data_partitioner = data_partitioner
# pylint: disable=not-context-manager
with self.graph.as_default(), tf.name_scope('Sampler'):
self.app.initialise_sampler()