def set_model_name(self, model_name):
"""
Set Model Name in config file
Args:
model_name: Name of model
"""
pipeline = read_config(self.config_path)
pipeline.model.name = model_name
write_config(pipeline=pipeline, config_path=self.config_path)
def set_batch_size(self, batch_size):
"""
Set Training Batch Size in config file
Args:
batch_size: an Integer for the batch size used for training
"""
pipeline = read_config(self.config_path)
pipeline.train_config.batch_size = batch_size
write_config(pipeline=pipeline, config_path=self.config_path)
def set_train_epochs(self, train_epochs, steps_per_epoch):
"""
Convert Epochs to Steps and write to config file
Args:
train_epochs: Number of Epochs to train the model
steps_per_epoch: Number of training steps for one epoch
"""
pipeline = read_config(self.config_path)
pipeline.train_config.num_steps = train_epochs * steps_per_epoch
write_config(pipeline=pipeline, config_path=self.config_path)
def set_val_tfrecord(self, val_tfrecord_path):
"""
Set Validation Tfrecord Path in config file
Args:
val_tfrecord_path: Path of validation or test tfrecord
"""
pipeline = read_config(self.config_path)
pipeline.eval_input_reader[0].tf_record_input_reader.input_path[
0] = val_tfrecord_path
write_config(pipeline=pipeline, config_path=self.config_path)
def set_train_tfrecord(self, train_tfrecord_path):
"""
Set Train Tfrecord Path in config file
Args:
train_tfrecord_path: Path of training tfrecord
"""
pipeline = read_config(self.config_path)
pipeline.train_input_reader.tf_record_input_reader.input_path[
0] = train_tfrecord_path
write_config(pipeline=pipeline, config_path=self.config_path)
def set_labelmap(self, labelmap_path):
"""
Set Label Map Path in config file
Args:
labelmap_path: Path of 'label_map.pbtxt' file
"""
pipeline = read_config(self.config_path)
pipeline.train_input_reader.label_map_path = labelmap_path
pipeline.eval_input_reader[0].label_map_path = labelmap_path
write_config(pipeline=pipeline, config_path=self.config_path)
def set_num_classes(self, num_classes):
"""
Set Number of classes (look at labelmap.pbtxt) in config file
Args:
num_classes: Number of classes (without background class)
"""
pipeline = read_config(self.config_path)
if self.model_type == 'ssd':
pipeline.model.ssd.num_classes = num_classes
elif self.model_type == 'prob_two_stage':
pipeline.model.prob_two_stage.num_classes = num_classes
write_config(pipeline=pipeline, config_path=self.config_path)
def set_optimizer(self, optimizer_name, scheduler_name, first_decay_epochs,
train_epochs, steps_per_epoch, learning_rate):
"""
Set Optimizer in config file
Args:
optimizer_name: String Name of Optimizer: either 'adam' or 'sgd'
scheduler_name: String Name of Scheduler: either 'restart' for Cosine Decay with Restarts or
'decay' for standard Cosine Decay
first_decay_epochs: Number of epochs for Warmup Phase of schedulers
train_epochs: Number of total train epochs
steps_per_epoch: Number of training steps for one epoch
learning_rate: Base Learning Rate for Training
"""
pipeline = read_config(self.config_path)
if optimizer_name not in ['adam', 'sgd']:
raise ValueError(
'String optimizer_name must be one of: [adam, sgd]')
if optimizer_name == 'adam':
optimizer = pipeline.train_config.optimizer.adam_optimizer
pipeline.train_config.optimizer.adam_optimizer.epsilon = 1e-8
elif optimizer_name == 'sgd':
optimizer = pipeline.train_config.optimizer.momentum_optimizer
pipeline.train_config.optimizer.momentum_optimizer.momentum_optimizer_value = 0.9
if scheduler_name not in ['restart', 'decay']:
raise ValueError(
'String optimizer_name must be one of: [restart, decay]')
if scheduler_name == 'restart':
scheduler = optimizer.learning_rate.cosine_restart_learning_rate
scheduler.first_decay_steps = first_decay_epochs * steps_per_epoch
scheduler.initial_learning_rate = learning_rate
elif scheduler_name == 'decay':
scheduler = optimizer.learning_rate.cosine_decay_learning_rate
scheduler.learning_rate_base = learning_rate
scheduler.warmup_steps = first_decay_epochs * steps_per_epoch
scheduler.total_steps = train_epochs * steps_per_epoch
write_config(pipeline=pipeline, config_path=self.config_path)
def set_weight_information(self,
add_weight_as_input=False,
input_method='input',
add_weight_as_output_gpo=False,
add_weight_as_output_gesamt=False):
"""
Set Weight integration in config file
Args:
add_weight_as_input: True if to add Weight as additional Input to model
input_method: Type of method for weight as input: For 'ssd' the strings 'input-multiply' and 'fpn-multiply'
are valid. For 'prob_two_stage' the strings 'input-multiply', 'fpn-multiply', 'roi-multiply' and
'concat' are valid
add_weight_as_output_gpo: True if using weight per object as additional output in model
add_weight_as_output_gesamt: True if using whole weight of image as additional output in model
(only works for 'ssd')
"""
pipeline = read_config(self.config_path)
if self.model_type == 'ssd':
if input_method not in ['input-multiply', 'fpn-multiply']:
raise ValueError(
'String input_method must be one of: [input-multiply, fpn-multiply]'
)
pipeline.model.ssd.add_weight_as_input = add_weight_as_input
pipeline.model.ssd.input_method = input_method
pipeline.model.ssd.add_weight_as_output_gpo = add_weight_as_output_gpo
pipeline.model.ssd.add_weight_as_output_gesamt = add_weight_as_output_gesamt
elif self.model_type == 'prob_two_stage':
if input_method not in [
'input-multiply', 'fpn-multiply', 'roi-multiply', 'concat'
]:
raise ValueError(
'String input_method must be one of: [input-multiply, fpn-multiply, roi-multiply, concat]'
)
pipeline.model.prob_two_stage.add_weight_as_input = add_weight_as_input
pipeline.model.prob_two_stage.input_method = input_method
pipeline.model.prob_two_stage.add_weight_as_output_gpo = add_weight_as_output_gpo
write_config(pipeline=pipeline, config_path=self.config_path)