Hierarchy of Configuration File

All the modules used in an algorithm can be easily set in the configuration file with yaml format.

Below a simple algorithm is taken as example for presenting the hierarchy of configuration file. Commonly the whole yaml file can be separated to 6 main parts: engine, algorithm, train_cfg, test_cfg, data and runtime. For special parts like engine, algorithm and data, they must include key of type, which indicates the name of “engine class” for use. Other parameters for building module can be followed after the type key with format of key: value. Especially, In the part of runtime, workflow phases are assigned here, for instance, setting of it work_flow: [['train', 2], ['val', 1]] means the engine runs every one validation phase after every two training epoch.

Tip

The names of workflow phases must correspond to the names of instance methods implemented inside engines. For example, as for task of Multi Targets Tracking, ‘test_emb’ (testing ReID network) or ‘test_det’ (testing Detector) can be added to the work_flow without modifying another setting of configuration file only if they are attributes of engine instance. More details are elaborated in chapter of engine.

# YAML Configuration File

engine:
    type: <Class Name of Engine>
algorithm:
    type: <Class Name of Algorithm>
    ...
    backbone:
        type: <Class Name of Backbone Module>
        ...
    head:
        type: <Class Name of Head Module>
        ...
    ...
train_cfg:
test_cfg:
train_pipeline:
val_pipeline:
test_pipeline:
data:
    type: <Class Name of Data Module>
    ...
runtime:

Feature of Every Part

Parts	Features
Engine	Preparing and Executing work flows (e.g., training, testing and so on)
Algorithm	Combination of sub modules (e.g., backbone, head and so on)
Train_cfg	Configuration of some special training setting
Test_cfg	Configuration of some special testing setting
Data	Building dataset as well as data flow pipeline
Runtime	Setting up optimizer, lr scheduler, logger and workflow phase

Management of Modules

Property

A base module takes in charge of the management of all modules, it’s to say that all modules (data module, algorithm module or engine module and so on) inherit from the base module, like the image show below. This design is convenient to improve the communication of independent modules from top to bottom.

Organization

The importation and scheduling of modules is qualified by Registry Mechanics. A new module can be easily applied by two lines code:

 from dolphin.utils import Registers, base

 @Registers.backbone.register
 class ResNet(base.BaseModelModule):
     def __init__():
         ...