Mix-In Methods

The file named work_flow_mixin.py presented as a class, which contains some commonly methods of training, testing and so on. For easily management methods of engines, they are departed independent.

Solver

Here stores helpful components during training phase or testing phase, such as optimizer, learning rate scheduler and so on. New optimizer and scheduler can be put in here. Different from modules, implementation of these components isn’t necessary to register as a module. They can be called by engine module once their names are assigned in config file.

Base Engine

Module Engine inherits from BaseModule and WorkFlowMixIn (work_flow_mixin.py), which is the kernel of all modules. Normally for different algorithms, only the method of training and testing needed to be newly implemented, so for convenient these kind of methods were moved into class WorkFlowMixIn. Its API is shown below:

Note

For customization of engine class, it is also necessary to register it to Registry for use.

class Engine(BaseModule, WorkFlowMixIn, metaclass=ABCMeta):

    '''
    Module takes in charge of organize modules and process.

    Inputs: algorithm (dict): algorithm configuration.
            data (dict): dataset configuration.
            cfg (dict): other configurations. Such as phases, optimizer name
                        and so on.
    '''

    def __init__(self,
                 algorithm=None,
                 data=None,
                 cfg=None,
                 **kwargs):

        self.init_engine()
        self.set_logger()

    def init_engine(self):
        '''
        initialize the engine with its input argument.
        '''

    def set_logger(self):
        '''
        set up logger.
        '''

    def build_models(self):
        '''
        building algorithm, schedule them as a property of engine.
        '''

    def build_dataset(self):
        '''
        building dataset, schedule it as a property of engine.
        '''

    def build_optimizer(self):
        '''
        build optimizer when training.
        '''

    def resume(self, filename, resume_optimizer=True):
        '''
        can be called in main.py, useful for training resume.
        '''

    def cur_lr(self):
        '''
        return current learning rate.
        '''