niftynet.application.base_application module

Interface of NiftyNet application

class SingletonApplication

Bases: type

Maintaining a global application instance.

classmethod clear()

Remove the instance. :return:

class BaseApplication

Bases: object

BaseApplication represents an interface.

Each application type_str should support to use the standard training and inference driver.

REQUIRED_CONFIG_SECTION = None

SUPPORTED_PHASES = set([u'evaluation', u'inference', u'training'])

is_validation = None

readers = None

sampler = None

net = None

optimiser = None

gradient_op = None

output_decoder = None

outputs_collector = None

gradients_collector = None

total_loss = None

patience = None

performance_history = []

mode = None

initialise_dataset_loader(data_param=None, task_param=None, data_partitioner=None)

this function initialise self.readers

Parameters:

• data_param – input modality specifications
• task_param – contains task keywords for grouping data_param
• data_partitioner – specifies train/valid/infer splitting if needed

Returns:

initialise_sampler()

Samplers take self.reader as input and generates sequences of ImageWindow that will be fed to the networks

This function sets self.sampler.

initialise_network()

This function create an instance of network and sets self.net

Returns:None

connect_data_and_network(outputs_collector=None, gradients_collector=None)

Adding sampler output tensor and network tensors to the graph.

Parameters:

• outputs_collector –
• gradients_collector –

Returns:

interpret_output(batch_output)

Implement output interpretations, e.g., save to hard drive cache output windows.

Parameters:batch_output – outputs by running the tf graph Returns:True indicates the driver should continue the loop False indicates the drive should stop

add_inferred_output_like(data_param, task_param, name)

This function adds entries to parameter objects to enable the evaluation action to automatically read in the output of a previous inference run if inference is not explicitly specified.

This can be used in an application if there is a data section entry in the configuration file that matches the inference output. In supervised learning, the reference data section would often match the inference output and could be used here. Otherwise, a template data section could be used.

Parameters:

• data_param –
• task_param –
• name – name of input parameter to copy parameters from

Returns:

modified data_param and task_param

set_iteration_update(iteration_message)

At each iteration application_driver calls:

output = tf.session.run(variables_to_eval, feed_dict=data_dict)

to evaluate TF graph elements, where variables_to_eval and data_dict are retrieved from iteration_message.ops_to_run and iteration_message.data_feed_dict

(In addition to the variables collected by self.output_collector).

The output of tf.session.run(…) will be stored at iteration_message.current_iter_output, and can be accessed from engine.handler_network_output.OutputInterpreter.

override this function for more complex operations (such as learning rate decay) according to iteration_message.current_iter.

get_sampler()

Get samplers of the application

Returns:niftynet.engine.sampler_* instances

add_validation_flag()

Add a TF placeholder for switching between train/valid graphs, this function sets self.is_validation. self.is_validation can be used by applications.

Returns:

action

A string indicating the action in train/inference/evaluation

Returns:

is_training

boolean value indicating if the phase is training

Type:return

is_inference

boolean value indicating if the phase is inference

Type:return

is_evaluation

boolean value indicating if the action is evaluation

Type:return

static stop()

remove application instance if there’s any.

Returns: