mmselfsup调用逻辑

dataloader

IterBaseRunner

    + register_training_hooks()

    + run()    + train()    + val()    + resume()    + save_checkpoint()

optimizer

配置参数config（dict）model，dataset，schedules，runtime

utils/register.py

传统的训练流程

openmmlab-mmselfsup

装饰

BaseAlgorithm

+ train_step()   # 调用forward- forward()  # 调用forward_train

OPTIMIZER_BUILDERS = Registry('optimizer builder')

将训练过程抽象到底层mmcv，以便匹配多数算法。具体某种算法的实现在实例到model中实现

DATASETS = Registry('dataset')PIPELINES = Registry('pipeline')DATASOURCE = Registry('datasource')

model

构建modelbuild_algorithm()

runner

Runner

Config

+ fromfile(filename):return cfg

EpochBaseRunner

+ train()+ run()+ run_iter()   # 调用model中的train_step()进行推论

+ val()+ save_checkpoint()

while(epochs max_epoch)

dataset

datasets

构建datasetsbuild_dataset()

runner/optimizer/build.py

dataloader=build_dataloader(dataset)

model.train_step()

Hook

+ before_run(runner)+ before_train_epoch(runner)+ before_train_iter(runner)+ after_train_iter(runner)+ after_train_epoch(runner)+ before_val_epoch(runner)+ before_val_iter(runner)+ after_val_iter(runner)+ after_val_epoch(runner)+ after_run(runner)+ get_triggered_stages()

runner/hooks

将需要执行的Hook注册进cfg中配置的runner，比如“EpochBaseRunner”，最终会将hook按优先级加入到_hooks列表内，run的时候遍历调用_hooks

cnn

base.py

CustomHook

+ custom_function(runner)

train.py

runner.register_hook()

RUNNER_BUILDERS.build(cfg)

将注册的Register模块传入build_from_cfg()生成需要的对像

cfg=Config().fromfile()

cnn.bricks.registry

OPTIMIZERS = Registry('optimizer')

RUNNER = Registry('runner')

runner.register_training_hooks()

train()

train_model流程

MODELS

+ build_algorithm(cfg)+ build_backbone(cfg)+ build_head(cfg)+ build_neck(cfg)+ build_memory(cfg)

EvalHook

+ evaluation()

+ before_run(runner)+ before_train_epoch(runner)+ before_train_iter(runner)+ after_train_iter(runner)+ after_train_epoch(runner)

runner/build.py

run()

OptimizerHook

+ after_train_iter(runner)

mmcv

dataset=build_dataset(cfg.data)

CheckpointHook

+ before_run(runner)

+ after_train_iter(runner)+ after_train_epoch(runner)

config

MOCO(BaseAlgorithm)

+ forward_train()   # 调用forward- forward()

RUNNER_BUILDERS = Registry('runner builder')

Register

+ build() # 编译模块（接口）+ build_func() # 编译（外部函数）

+ register_module() # 注册模块

LrUpdaterHook

+ get_lr(runner)+ get_regular_lr(runner)+ get_warmup_lr(runner)

+ before_run(runner)+ after_train_iter(runner)+ after_train_epoch(runner)

mmselfsup/mmselfsup

BaseRunner

+ run()# 虚函数，在继承类中实现+ train()# 虚函数，在继承类中实现+ val()# 虚函数，在继承类中实现+ save_checkpoint()

- _hooks+ register_hook():return _hooks+ call_hook()+ resume()+ register_lr_hook(cfg)+ register_momentum_hook(cfg)+ register_optimizer_hook(cfg)+ register_checkpoint_hook(cfg)+ register_logger_hook(cfg)+ register_timer_hook(cfg)+ register_custom_hook(cfg)+ register_training_hook(cfg)

HOOKS = Registry('hook')

mmselfsup/tools

MMCV_ATTENTION=Register('attention')

logger

api/train.py