mmdetection-yolox

mmdetection_yolox

训练流程

一. 注册机制

注册机制：registry可以看成是一个类映射到一个字符串的映射。

from mmcv.cnn import MODELS as MMCV_MODELS
from mmcv.utils import Registry
#1. 创建注册表（表里填写映射）
MODELS = Registry('models', parent=MMCV_MODELS)
#2.将类模块注册到注册表中（字符串和类之间映射）'字符串Converter1' 类
@MODELS.register_module()
class Converter1(object):
    pass
#3.模块注册成功，通过configs使用这个转换器。得到实例化
#配置文件中的type：待实例化的类名；后面是初始化参数
converter_cfg = dict(type='Converter1', a=a_value, b=b_value)
converter = MODELS.build(converter_cfg)

二.训练代码流程【main.py 和 mmdet/apis/train.py】

from mmdet.datasets import build_dataset
from mmdet.models import build_detector
from mmdet.apis import train_detector
from mmcv import Config

#1.初始化配置文件
cfg=Config.fromfile('configs/123/yolox.py')
#2.初始化模型
model=build_detector(cfg.model) #得到的是yolox实例化类
#3.初始化数据集
datasets=[build_dataset(cfg.data.train)] #得到的datasets是没有经过pipeline处理的
train_detector(model,datasets,cfg,distributed=False,validate=True)#分布式=F

from mmdet.datasets import (build_dataloader, build_dataset）
from mmcv.parallel import MMDataParallel
from mmcv.runner import (DistSamplerSeedHook, EpochBasedRunner,
                         Fp16OptimizerHook, OptimizerHook, build_optimizer,
                         build_runner, get_dist_info)
def train_detector(model,
                   dataset,
                   cfg,
                   distributed=False,
                   validate=False,
                   timestamp=None,
                   meta=None):
    #4.初始化logger
    logger = get_root_logger(log_level=cfg.log_level)
    #5.初始化数据迭代器
    dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
    data_loaders = [
        build_dataloader(
            ds,
            cfg.data.samples_per_gpu,
            cfg.data.workers_per_gpu,
            # `num_gpus` will be ignored if distributed
            num_gpus=len(cfg.gpu_ids),#1
            dist=distributed, #Fasele
            seed=cfg.seed,#0
            runner_type=runner_type,
            persistent_workers=cfg.data.get('persistent_workers', False))
        #dataset是一个MultiImageMixDataset对象
        #每用for拿出一个就是，调用MultiImageMixDataset对象里的_getitem__函数
        #拿出来的事经过12条pipeline处理的数据
        for ds in dataset
    ]

    #6.model放在gpu上（distributed=False）
    model = MMDataParallel(
            model.cuda(cfg.gpu_ids[0]), #0
            device_ids=cfg.gpu_ids) #0
    #7.初始化optimizer
    optimizer = build_optimizer(model, cfg.optimizer)
    #8.初始化runner
    runner_type = 'EpochBasedRunner' #cfg中的就是这个
    runner = build_runner(
        cfg.runner,
        default_args=dict(
            model=model,
            optimizer=optimizer,
            work_dir=cfg.work_dir,
            logger=logger,
            meta=meta))#None
    #日志相关 an ugly workaround to make .log and .log.json filenames the same
    runner.timestamp = timestamp#nonne
    #fp16 setting 混合精度训练 | fp16 用于神经网络训练和预测   
    fp16_cfg = cfg.get('fp16', None)#有就拿出来，没有就是none
    if fp16_cfg is not None:#cdg中的是none
        optimizer_config = Fp16OptimizerHook(
            **cfg.optimizer_config, **fp16_cfg, distributed=distributed)
    elif distributed and 'type' not in cfg.optimizer_config:
        optimizer_config = OptimizerHook(**cfg.optimizer_config)
    else:##########################################只执行这里
        optimizer_config = cfg.optimizer_config
    # 9. 注册 hooks
    runner.register_training_hooks(
        cfg.lr_config,
        optimizer_config,
        cfg.checkpoint_config,
        cfg.log_config,
        cfg.get('momentum_config', None),
        custom_hooks_config=cfg.get('custom_hooks', None))

    # 10.register eval hooks
    if validate:
        # Support batch_size > 1 in validation
        val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
        if val_samples_per_gpu > 1:
            # Replace 'ImageToTensor' to 'DefaultFormatBundle'
            cfg.data.val.pipeline = replace_ImageToTensor(
                cfg.data.val.pipeline)
        val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
        val_dataloader = build_dataloader(
            val_dataset,
            samples_per_gpu=val_samples_per_gpu,
            workers_per_gpu=cfg.data.workers_per_gpu,
            dist=distributed,
            shuffle=False)
        eval_cfg = cfg.get('evaluation', {})
        eval_cfg['by_epoch'] = cfg.runner['type'] != 'IterBasedRunner'
        eval_hook = DistEvalHook if distributed else EvalHook
        # In this PR (https://github.com/open-mmlab/mmcv/pull/1193), the
        # priority of IterTimerHook has been modified from 'NORMAL' to 'LOW'.
        runner.register_hook(
            eval_hook(val_dataloader, **eval_cfg), priority='LOW')
    #11.加载权重
    if cfg.resume_from:
        runner.resume(cfg.resume_from)
    elif cfg.load_from:
        runner.load_checkpoint(cfg.load_from)
    #12.开始训练 
    runner.run(data_loaders, cfg.workflow)

1. mmdet/models/builder.py/build_dtector(cfg.model)

cfg=Config.fromfile('configs/123/yolox.py')
model=build_detector(cfg.model)#根据配置文件信息得到实例化模型yolox

1）.创建注册表 MODELS

from mmcv.cnn import MODELS as MMCV_MODELS
from mmcv.utils import Registry

MODELS = Registry('models', parent=MMCV_MODELS)
DETECTORS = MODELS 

2）.通过在模块上方添加 @DETECTORS.register_module()语句，注册模块（在创建模块时将实现的模块注册到注册表中）

@DETECTORS.register_module()
class YOLOX(SingleStageDetector)

3）.通过build字句，根据cfg:model=dict(type=YOLOX)将YOLOX这个类实例化

model=def build_detector(
#cfg=Config.fromfile('configs/123/yolox.py').model
cfg, #这里的cf只是总配置文件中的model字典
train_cfg=None, 
test_cfg=None):
    return DETECTORS.build(
        cfg, default_args=dict(train_cfg=train_cfg, test_cfg=test_cfg))

2.mmdet/datasets/builder.py/build_dataset(cfg.data.train)

datasets=[build_dataset(cfg.data.train)]

1).传来的参数：cfg.data.train=train_dataset这个配置字典

传来的配置字典里有：数据集的类型，数据集的地址，数据处理10+2

train_dataset = dict(
    type='MultiImageMixDataset',
    dataset=dict(
        type=dataset_type,
        classes = ('human body','ball','circle cage',
        'square cage','tyre','metal bucket','cube','cylinder'),
        ann_file='data/sna/annotations/train.json',
        img_prefix='data/sna/train/',
        #2个数据处理
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='LoadAnnotations', with_bbox=True)
        ],
        filter_empty_gt=False,
    ),
    pipeline=train_pipeline)#10个数据处理

2）解读函数 build_dataset(cfg.data.train, default_args=None)

args:传来的配置字典里有：数据集的类型，数据集的地址，数据处理10+2
return:最后得到的dataset 是 MultiImageMixDataset(**cp_cfg) 一个对象

from mmcv.utils import Registr
DATASETS = Registry('dataset')
PIPELINES = Registry('pipeline')
#传过来的cfg是train_dataset这个字典
def build_dataset(cfg, default_args=None):
    from .dataset_wrappers import  MultiImageMixDataset
    # 函数isinstance()可以判断一个变量的类型
    # 当拿到变量 cfg  时，可以使用 isinstance 判断类型：
    # isinstance(cfg,(list, tuple))的结果是False
    # cfg 不是(list, tuple)类型 ，而cfg是个字典！！
    if isinstance(cfg, (list, tuple)):
        dataset = ConcatDataset([build_dataset(c, default_args) for c in cfg])
    # cfg字典中的键type；值MultiImageMixDataset（多图像混合数据集）
    elif cfg['type'] == 'MultiImageMixDataset':
        cp_cfg = copy.deepcopy(cfg)#cp_cfg长得和train_dataset一样
        # build_dataset(cp_cfg['dataset'])应该是build_from_cfgs)
        # cp_cfg['dataset']里面是一个COCOdataset对象
        cp_cfg['dataset'] = build_dataset(cp_cfg['dataset'])
        cp_cfg.pop('type')#去掉这个键值对
        #cp_cfg拿到了datasets，pipeline12个
        dataset = MultiImageMixDataset(**cp_cfg)
    elif isinstance(cfg.get('ann_file'), (list, tuple)):
        dataset = _concat_dataset(cfg, default_args)
    else:#返回一个COCOdataset对象
        dataset = build_from_cfg(cfg, DATASETS, default_args)
    return dataset

cocodataset类位于mmdet/datasets/coco.py。继承了customdataset类

customdataset类位于mmdet/datasets/custom.py

cp_cfg字典

这个字典里有两个键值对：dataset，pipeline。
cp_cfg = dict(
    dataset=是COCOdataset对象
    pipeline=train_pipeline)

train_pipeline：10个处理

train_pipeline = [
   1. dict(type='Mosaic', img_scale=img_scale, pad_val=114.0),
   2. dict(
        type='RandomAffine',
        scaling_ratio_range=(0.1, 2),
        border=(-img_scale[0] // 2, -img_scale[1] // 2)),
   3. dict(
        type='MixUp',
        img_scale=img_scale,
        ratio_range=(0.8, 1.6),
        pad_val=114.0),
   4. dict(type='YOLOXHSVRandomAug'),
   5. dict(type='RandomFlip', flip_ratio=0.5),
    # According to the official implementation, multi-scale
    # training is not considered here but in the
    # 'mmdet/models/detectors/yolox.py'.
   6. dict(type='Resize', img_scale=img_scale, keep_ratio=True),
   7. dict(
        type='Pad',
        pad_to_square=True,
        # If the image is three-channel, the pad value needs
        # to be set separately for each channel.
        pad_val=dict(img=(114.0, 114.0, 114.0))),
   8. dict(type='FilterAnnotations', min_gt_bbox_wh=(1, 1), keep_empty=False),
   9. dict(type='DefaultFormatBundle'),
  10. dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

最终得到的结果dataset = MultiImageMixDataset(**cp_cfg)

此类位于mmdet/datasets/dataset_wrappers.py中

@DATASETS.register_module()
class MultiImageMixDataset:
    """#多图像数据集包装器wrapper.

    适用于多幅图像的训练，如混合数据增强 mosaic and mixup. 
    Args:
        dataset (:obj:`CustomDataset`): 要混合的数据集（因为custom是所有类的父类）
        pipeline (Sequence[dict]): 数据预处理操作（一个操作是一个字典）
        dynamic_scale (tuple[int], optional): 图片动态放缩的比例. Default to None. .
        skip_type_keys (list[str], optional): 要跳过的pipeline. Default to None.
    """

    def __init__(self,
                 dataset,#键值对1：dataset 对应COCODataset对象
                 pipeline,#键值对2：pipeline=train_pipeline（这是一个列表，里面有10个字典）
                 dynamic_scale=None,
                 skip_type_keys=None):
        self._skip_type_keys = skip_type_keys#none
        self.pipeline = [] #10个pipeine的处理模块
        self.pipeline_types = [] #10个pipeline的type
        for transform in pipeline:#每次拿出一个字典
        #一共有10个字典
            if isinstance(transform, dict):
                self.pipeline_types.append(transform['type'])
                #transform：配置字典。PIPELINES：注册表
                #实例化类对象transform
                transform = build_from_cfg(transform, PIPELINES)#实例化这个处理模块
                self.pipeline.append(transform)
       '''
    pipeline_types 里面有12个字符串
    self.pipeline_types=['Mosaic','RandomAffine','MixUp','YOLOXHSVRandomAug',
    'RandomFlip', 'Resize','Pad','FilterAnnotations',
    'DefaultFormatBundle', 'Collect']
    
    transform 里面有12个类对象
    self.pipeline=[Mosaic,RandomAffine,MixUp,YOLOXHSVRandomAug,
    RandomFlip,Resize,Pad,FilterAnnotations,
    DefaultFormatBundle, Collect]
       '''
        self.dataset = dataset#cocodataset对象
        self.CLASSES = dataset.CLASSES 
        #就是调用COCOdataset对象中的len方法：结果=annotations/train.json中的数据条数
        #目前不知道num_samples 是图片个数还是图片里的bbox的个数
        #len(dataset)=dataset.__len__()，现在知道了应该是图片的索引
        self.num_samples = len(dataset)

   #这个是python的magic方法。例如对一个对象dataset= MultiImageMixDataset(**cp_cfg)
   #dataset_item = dataset[idx]
   #等价于dataset =  dataset.__getitem__(idx)
       def __getitem__(self, idx):
       # 获取当前某一张的图片信息 
       #self.dataset[idx]是调用cocodataset中的_.__getitem__.
       #1.custome中的.__getitem__
       #2.custom中的prepare_train_img(self, idx)
       #3.cocodataset中的load_annotations(self, ann_file):
       #4.cocodataset中的 get_ann_info(self, idx):
       #5.custom pre_pipeline(results)
       #####self.dataset[idx]得到的事经过load1load2这两个pipeline处理的第idx这张图片
       #####return self.pipeline(results)：最终得到的是一个results字典。
       #self.dataset[idx]就是 那个根据 img的idx 最终得到的 results字典。
       #6.这个results传进去，再经过10个pipeline处理，得到最终的results字典
       results = copy.deepcopy(self.dataset[idx])#custom里得到的那个result字典
       # 遍历 transorm,其中可以包括 mosaic 、mixup、flip 等各种  transform 
        for (transform, transform_type) in zip(self.pipeline, self.pipeline_types):
            # 考虑到某些训练阶段需要动态关闭掉部分数据增强，故引入   _skip_type_keys       
            if self._skip_type_keys is not None and \
                    transform_type in self._skip_type_keys:
                continue
            ##如果transform 中含有 get_indexes 方法，就先走这一步。
            if hasattr(transform, 'get_indexes'):
                indexes = transform.get_indexes(self.dataset)
                if not isinstance(indexes, collections.abc.Sequence):
                    indexes = [indexes]
                # 得到混合图片信息   
                mix_results = [
                    copy.deepcopy(self.dataset[index]) for index in indexes
                ]
                results['mix_results'] = mix_results
       #这个results传进去，再经过10个pipeline处理，得到最终的results字典
            results = transform(results)#执行transfor的call函数

            if 'mix_results' in results:
                results.pop('mix_results')

        return results

transform处理

所有的 pipeline处理都针对的是一张图片，最终得到的datasets对象，里面的__getitem__方法是获得了一张图片的信息。得到一张图片的信息就是一个results字典。
先进行俩pipeline操作得到原始的ressults
Mosaic,RandomAffine,MixUp,YOLOXHSVRandomAug, RandomFlip,Resize,Pad,FilterAnnotations,
DefaultFormatBundle, Collect
经过这些pipeline处理就是对results的键值对进行增删
最终得到的results字典里面只有4个键值对：img img_meta gtbox gtlabel
如果调用datasets中的_getitem_方法，最后的得到的是collect的输出

collect_results=
dict('img':DC(tensor,stacked=True,cpu_only=False),#DC是datacontainer的缩写
     'img_metas':DC(dict('flip':bool,
                       'ori_shape':tuple,
                       'img_shape':tuple,
                       'pad_shape':tuple,
                       'scale_factor':float or ndarray(4,),
                       'img_norm_cfg':dict('mean':ndarray,'std':ndarray)),stacked=False,cpu_only=True),
      'gt_bboxes':DC(tensor,stacked=Fasle,cpu_only=False),
      'gt_labels':DC(tensor,stacked=False,cpu_only=False)
    )

6）mmcv.utils.build_from_cfg函数定义位于mmcv/mmcv/utils/registry.py中

def build_from_cfg(cfg, registry, default_args=None):
    """从配置字典中构建模块.
    Args:
        cfg (dict): Config dict. 至少包含 "type".这个键值对（模块参数可不用初始但要指定名称）
        registry (:obj:`Registry`): 这个registry 来搜寻 type模块在哪
        default_args (dict, optional): 默认的初始化参数
    Returns:
        object: 返回实例化对象
    """
    # 1.判断输入的参数形式是否正确
    if not isinstance(cfg, dict):
        raise TypeError(f'cfg must be a dict, but got {type(cfg)}')
    if 'type' not in cfg:
        raise KeyError(
            f'the cfg dict must contain the key "type", but got {cfg}')
    if not isinstance(registry, Registry):
        raise TypeError('registry must be an mmcv.Registry object, '
                        f'but got {type(registry)}')
    if not (isinstance(default_args, dict) or default_args is None):
        raise TypeError('default_args must be a dict or None, '
                        f'but got {type(default_args)}')

    # 2.将obj_cls设置为config的type对应的class。
    args = cfg.copy() # dict.copy()用来返回一个字典的浅复制。里面东西=cfg这个传入的字典
    obj_type = args.pop('type') # 从args中删除type这个键值对，obj_type等于type对应的值。
    #3.obj_cls就是得到了类名
    if is_str(obj_type):
    #registry.get(obj_type)是提取名为obj_type的class赋给obj_cls。
        obj_cls = registry.get(obj_type) 
        if obj_cls is None:
            raise KeyError(
                f'{obj_type} is not in the {registry.name} registry')
    elif inspect.isclass(obj_type): # 检查是否为类
        obj_cls = obj_type
    else:
        raise TypeError(
            f'type must be a str or valid type, but got {type(obj_type)}')

    if default_args is not None:
        for name, value in default_args.items():
    # dict.setdefault()：如果字典中包含有给定键，则返回该键对应的值，否则返回为该键设置的值。
            args.setdefault(name, value) 
          
    return obj_cls(**args)#返回对象（且里面有配置字典所给定的初始值）

3. mmdet/datasets/builder.py/build_dataloader(一堆参数)

1）传来的参数

    #最终得到的data_loaders是一个列表，里面是一堆dataloader
    #[]之进行一次循环，里面只有一个元素，就是一个dataloader对象
    data_loaders = [
        build_dataloader(
            ds,
            cfg.data.samples_per_gpu,#8 batchsize
            cfg.data.workers_per_gpu,#4 线程
            # `num_gpus` will be ignored if distributed
            num_gpus=len(cfg.gpu_ids),#1
            dist=distributed,#False:非分布式训练
            seed=cfg.seed,#0
            runner_type=runner_type,#EpochBasedRunner
            #好像不重要，应该是false。关于pytorch版本的东西
            persistent_workers=cfg.data.get('persistent_workers', False))
        for ds in dataset #只进行一次循环
    ]   #dataset是一个MultiImageMixDataset对象
        #每用for拿出一个就是，调用MultiImageMixDataset对象里的_getitem__函数
        #拿出来的是经过12条pipeline处理的某一张图片的数据
        #但是这里的dataset是一个列表，列表里只有一个元素，就是muli对象
        #所以for循环只执行一次

2）解读函数build_dataloader

1.先利用参数构造一个sampler采样器（作为DataLoader的参数）

mmdet.datasets.samplers.group_sampler

sampler = GroupSampler(dataset,samples_per_gpu) if shuffle else None

2.用一下collate函数（作为DataLoader的参数）

mmcv.parallel.collate
torch.utils.data.dataloader.efault_collate

collate_fn=partial(collate, samples_per_gpu=samples_per_gpu),#这是一个函数

3.data_loader 是 DataLoader 对象

from torch.utils.data import DataLoader
函数 build_dataloader 返回值return data_loader
传来的参数：

  data_loader = DataLoader(
        dataset,#还是最开始传入的dataset：一张图片的results字典
        batch_size=batch_size,#8
        sampler=sampler,#前面获得的sampler对象，，每次返回一个整数下标索引
        num_workers=num_workers,#4
        batch_sampler=batch_sampler,#none
        #（collate这个函数，8）。如果用到collate_fn（x）：x和8做collate的操作。collate:
        collate_fn=partial(collate, samples_per_gpu=samples_per_gpu),#这是一个函数
        pin_memory=False,
        worker_init_fn=init_fn,#none
        **kwargs)

DataLoader类对象

class DataLoader(object):
    __initialized = False
 
    def __init__(self, 
    dataset, #可能是一张图片的信息，应该是所有图片的信息阿
    batch_size=1, #8
    #（default）是否随即组成每个batch，
    #因为下面定义了sampler，控制 Dataset 输出的数据顺序，
    #可能是就是规定了一种特定的组成batch的方法。
    #就算是组成batch既不是随即的，也不是按顺序的，是有一定规律的
    shuffle=False, 
    sampler=None, #数据采样器 从数据集中随机取样，，每次返回一个整数下标索引
    batch_sampler=None,#none
    num_workers=0, #0（default）
    collate_fn=default_collate,#collate函数 合并样本列表以形成小批量
    pin_memory=False, #（default）
    drop_last=False, #（default）如果datasizze不能整除8，是否把最后一个batchdrop掉
    timeout=0, #（default）
    worker_init_fn=None):#none
    """
    Data loader. Combines a dataset and a sampler, and provides
    single- or multi-process iterators over the dataset.
    Arguments:
        dataset (Dataset): dataset from which to load the data.
        batch_size (int, optional): how many samples per batch to load
            (default: 1).
        shuffle (bool, optional): set to ``True`` to have the data reshuffled
            at every epoch (default: False).
        sampler (Sampler, optional): defines the strategy to draw samples from
            the dataset. If specified, ``shuffle`` must be False.
            定义从中提取样本的策略数据集。如果指定，“shuffle”必须为False。
        batch_sampler (Sampler, optional): like sampler, but returns a batch of
            indices at a time. Mutually exclusive with batch_size, shuffle,
            sampler, and drop_last.
        num_workers (int, optional): how many subprocesses to use for data
            loading. 0 means that the data will be loaded in the main process.
            (default: 0)
        collate_fn (callable, optional): merges a list of samples to form a mini-batch."""
        
        self.dataset = dataset
        self.batch_size = batch_size
        self.num_workers = num_workers
        self.collate_fn = collate_fn
        self.pin_memory = pin_memory
        self.drop_last = drop_last
        self.timeout = timeout
        self.worker_init_fn = worker_init_fn
        self.sampler = sampler
        self.__initialized = True
 
        if batch_sampler is None:
            batch_sampler = BatchSampler(sampler, batch_size, drop_last)
        self.batch_sampler = batch_sampler
 
    def __setattr__(self, attr, val):
        if self.__initialized and attr in ('batch_size', 'sampler', 'drop_last'):
            raise ValueError('{} attribute should not be set after {} is '
                             'initialized'.format(attr, self.__class__.__name__))
 
        super(DataLoader, self).__setattr__(attr, val)
 
    def __iter__(self):
        return _DataLoaderIter(self)
 
    def __len__(self):
        return len(self.batch_sampler)

4.mmcv/parallel/MMDataParallel（单gpu版本的model上层封装）

model = MMDataParallel(
            model.cuda(cfg.gpu_ids[0]), #0
            device_ids=cfg.gpu_ids) #0

mmcv.parallel .MMDataParallel
model是MMDataParallel类对象。

from itertools import chain。
from torch.nn.parallel import DataParallel
from .scatter_gather import scatter_kwargs

class MMDataParallel(DataParallel): # 继承于pytorch.DataParallel
    def __init__(self, *args, dim=0, **kwargs):
        # 构造函数和pytorch的DataParallel一致
        super(MMDataParallel, self).__init__(*args, dim=dim, **kwargs)
        self.dim = dim

    def forward(self, *inputs, **kwargs):
        # 在api/test.py和api/inference.py中调用model(return_loss=False, rescale=True, **data_batch)所以实际上，这里参数只使用了kwargs，inputs为空tuple()
        return super().forward(*inputs, **kwargs) 
        # pytorch的forward，实现了把数据平均分发到各个 GPU 上
        #每个 GPU 实际的数据量为batch/gpu_num

    def scatter(self, inputs, kwargs, device_ids): # 非常重要######
        return scatter_kwargs(inputs, kwargs, device_ids, dim=self.dim) 

    def train_step(self, *inputs, **kwargs):
        # 参数就使用了inputs，为(data_loader[i])
        # kwargs为空字典{}
        # MMDataParallel只支持单GPU
        assert len(self.device_ids) == 1,
        for t in chain(self.module.parameters(), self.module.buffers()): 
            # 遍历parameter和buffer，parameter记录需要更新的参数，buffer记录不需要更新的参数
            # 是判断模型是不是在主GPU上
            if t.device != self.src_device_obj:
                raise RuntimeError(
                    'module must have its parameters and buffers '
                    f'on device {self.src_device_obj} (device_ids[0]) but '
                    f'found one of them on device: {t.device}')

        inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)# 非常重要######
        return self.module.train_step(*inputs[0], **kwargs[0])# 调用真正的module

    def val_step(self, *inputs, **kwargs):
        # 参数就使用了inputs，为(data_loader[i])
        # kwargs为空字典{}
        # MMDataParallel只支持单GPU，MMDistributedDataParallel才支持多GPU
        assert len(self.device_ids) == 1
        for t in chain(self.module.parameters(), self.module.buffers()):
            if t.device != self.src_device_obj:
                raise RuntimeError(
                    'module must have its parameters and buffers '
                    f'on device {self.src_device_obj} (device_ids[0]) but '
                    f'found one of them on device: {t.device}')

        inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids)# 非常重要######
        return self.module.val_step(*inputs[0], **kwargs[0]) # 调用真正的module

里面最重要的方法scatter，是对dataloader输出的数据进行解封装，变成pytorch可以处理的格式，放到模型上处理，得到的处理结果就是模型的输入。self.module.train_step。

# num_gpus就是分布式训练时的gpu数量，默认为1，数组中只有一个dict元素
tuple(num_gpus * 
        dict('img': tensor(Batch,C,H,W) ,
             'img_metas': list[Batch*dict('flip','ori_shape'……)],
             'gt_bboxes': list[Batch*tensor],
             'gt_labels': list[Batch*tensor]
            )
      )

5.mmcv/runner-build_optimizer

# model是MMDataParallel对象，对原始model进行了封装
optimizer = build_optimizer(model, cfg.optimizer)

由配置文件构造模块

cfg.optimizer = dict(
    type='SGD',
    lr=0.01/8,
    momentum=0.9,
    weight_decay=5e-4,
    nesterov=True,
    paramwise_cfg=dict(norm_decay_mult=0., bias_decay_mult=0.))
    #关闭 norm 和 bias 的weight decay。
optimizer_config = dict(grad_clip=None)

函数解析 build_optimizer(model, cfg.optimizer)。最后就是得到一个sgd模块(model)

def build_optimizer(model, cfg):
    optimizer_cfg = copy.deepcopy(cfg)#先复制一份配置文件
    #pop先赋值，再去掉
    constructor_type = optimizer_cfg.pop('constructor',
                                         'DefaultOptimizerConstructor')
    paramwise_cfg = optimizer_cfg.pop('paramwise_cfg', None)
    # 得到的 optim_constructor 是通过配置文件中的type构造的sgd模块把应该是
    optim_constructor = build_optimizer_constructor(
        dict(
            type=constructor_type,
            optimizer_cfg=optimizer_cfg,
            paramwise_cfg=paramwise_cfg))
    #优化器我就不看了
    optimizer = optim_constructor(model)
    return optimizer

'''
cfg=dict(
            type=constructor_type,
            optimizer_cfg=optimizer_cfg,
            paramwise_cfg=paramwise_cfg)
'''
def build_optimizer_constructor(cfg):
    return build_from_cfg(cfg, OPTIMIZER_BUILDERS)

6.mmcv/runner/builder/build_runner

 runner_type = 'EpochBasedRunner' #cfg中的就是这个
 runner = build_runner(
        cfg.runner,
        default_args=dict(
        model=model,
        optimizer=optimizer,
        work_dir=cfg.work_dir,
        logger=logger,
        meta=meta))#None
       

由配置文件构造模块

cfg.runner = dict(type='EpochBasedRunner', max_epochs=300)

函数解析 build_runner(model, cfg.optimizer)。最后就是得到一个EpochBasedRunner模块(model)
（mmcv.runner.epoch_based_runner.py）继承了BaseRunner（mmcv.runner.base_runner.py）

def build_runner(cfg, default_args=None):
    runner_cfg = copy.deepcopy(cfg)
    constructor_type = runner_cfg.pop('constructor', 'DefaultRunnerConstructor')
    #得到的 runner_constructor 是通过配置文件中的type构造的sgd模块把应该是
    runner_constructor = build_runner_constructor(
        dict(
            type=constructor_type,
            runner_cfg=runner_cfg,
            default_args=default_args))
    runner = runner_constructor()
    return runner

def build_runner_constructor(cfg):
    return RUNNER_BUILDERS.build(cfg)

7.runner.register_training_hooks()训练hook

runner.register_training_hooks(
        cfg.lr_config,
        optimizer_config,
        cfg.checkpoint_config,
        cfg.log_config,
        cfg.get('momentum_config', None),#配置文件里没有，这里是拿出来的是none
        custom_hooks_config=cfg.get('custom_hooks', None))

register_training_hooks接受所有的配置文件（一堆字典）然后发给不同函数

    def register_training_hooks(self,
                                lr_config,##
                                optimizer_config=None,##
                                checkpoint_config=None,##
                                log_config=None,##
                                momentum_config=None,#none
                                timer_config=dict(type='IterTimerHook'),#就是默认
                                custom_hooks_config=None):##
        """Register default and custom hooks for training.

        Default and custom hooks include:

        +----------------------+-------------------------+
        | Hooks                | Priority   优先级 字符和数字的对应关系          |
        +======================+=========================+
        | LrUpdaterHook        | VERY_HIGH (10)          |
        +----------------------+-------------------------+
        | MomentumUpdaterHook  | HIGH (30)               |
        +----------------------+-------------------------+
        | OptimizerStepperHook | ABOVE_NORMAL (40)       |
        +----------------------+-------------------------+
        | CheckpointSaverHook  | NORMAL (50)             |
        +----------------------+-------------------------+
        | IterTimerHook        | LOW (70)                |
        +----------------------+-------------------------+
        | LoggerHook(s)        | VERY_LOW (90)           |
        +----------------------+-------------------------+
        | CustomHook(s)        | defaults to NORMAL (50) |
        +----------------------+-------------------------+

        If custom hooks have same priority with default hooks, custom hooks
        will be triggered after default hooks.先触发的fault再触发custom
        """
        self.register_lr_hook(lr_config)##
        self.register_momentum_hook(momentum_config)##none
        self.register_optimizer_hook(optimizer_config)##
        self.register_checkpoint_hook(checkpoint_config)##
        self.register_timer_hook(timer_config)##默认
        self.register_logger_hooks(log_config)##
        self.register_custom_hooks(custom_hooks_config)##

1）register_lr_hook 【YOLOXLrUpdaterHook】10

YOLOXLrUpdaterHook在mmdet/core/hook/yolox_lrupdater_hook.py训练的学习率策略

'''
lr_config = dict(
    _delete_=True,
    policy='YOLOX',最后变成YOLOXLrUpdaterHook类型
    warmup='exp',
    by_epoch=False,
    warmup_by_epoch=True,
    warmup_ratio=1,
    warmup_iters=5,  # 5 epoch
    num_last_epochs=num_last_epochs,
    min_lr_ratio=0.05
    step=[8, 11])
'''
def register_lr_hook(self, lr_config):
        elif isinstance(lr_config, dict):
            policy_type = lr_config.pop('policy')#=yolox
            # If the type of policy is all in lower case, e.g., 'cyclic',
            # then its first letter will be capitalized, e.g., to be 'Cyclic'.
            # This is for the convenient usage of Lr updater.
            # Since this is not applicable for `
            # CosineAnnealingLrUpdater`,
            # the string will not be changed if it contains capital letters.
            #这一堆注释是为了改名的，没有意义
            if policy_type == policy_type.lower():
                policy_type = policy_type.title()
            #hook_type=YOLOXLrUpdaterHook
            hook_type = policy_type + 'LrUpdaterHook'
            lr_config['type'] = hook_type
            #根据配置文件 lr_config 构建一个LrUpdaterHook对象
            hook = mmcv.build_from_cfg(lr_config, HOOKS)
        #将hook插入hooks队列中，并指定优先级
        self.register_hook(hook, priority='VERY_HIGH==10')

2）register_momentum_hook 【none】

'''
none
'''
    def register_momentum_hook(self, momentum_config):
        if momentum_config is None:
            return

3）register_optimizer_hook 【OptimizerHook】40

'''
optimizer_config = dict(grad_clip=None)
'''
    def register_optimizer_hook(self, optimizer_config):
        #这个hook是OptimizerHook
        if isinstance(optimizer_config, dict):
            optimizer_config.setdefault('type', 'OptimizerHook')
            hook = mmcv.build_from_cfg(optimizer_config, HOOKS)
        #将这个hook插入到hook序列中
        self.register_hook(hook, priority='ABOVE_NORMAL')

4）register_checkpoint_hook 【CheckpointHook】50

'''
checkpoint_config = dict(interval=10)
'''
def register_checkpoint_hook(self, checkpoint_config):
        if isinstance(checkpoint_config, dict):
            checkpoint_config.setdefault('type', 'CheckpointHook')
            hook = mmcv.build_from_cfg(checkpoint_config, HOOKS)
        self.register_hook(hook, priority='NORMAL')

5）register_timer_hook 【IterTimerHook】70

'''
timer_config=dict(type='IterTimerHook')
'''
    def register_timer_hook(self, timer_config):
        if isinstance(timer_config, dict):
            timer_config_ = copy.deepcopy(timer_config)
            hook = mmcv.build_from_cfg(timer_config_, HOOKS)
        else:
            hook = timer_config
        self.register_hook(hook, priority='LOW')

6）register_logger_hooks 【TextLoggerHook】90

'''
 log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook'),
    ])
'''
 def register_logger_hooks(self, log_config):

        log_interval = log_config['interval'] #50
        for info in log_config['hooks']:
            logger_hook = mmcv.build_from_cfg(
                info, HOOKS, default_args=dict(interval=log_interval))
            self.register_hook(logger_hook, priority='VERY_LOW')

7) register_custom_hooks 【YOLOX Sy Exp】50

'''
custom_hooks = [
    dict(
        type='YOLOXModeSwitchHook',
        num_last_epochs=num_last_epochs,
        priority=48),
    dict(
        type='SyncNormHook',
        num_last_epochs=num_last_epochs,
        interval=interval,
        priority=48),
    dict(
        type='ExpMomentumEMAHook',
        resume_from=resume_from,
        momentum=0.0001,
        priority=49)
] '''
    def register_custom_hooks(self, custom_config):
        for item in custom_config:
            if isinstance(item, dict):
                self.register_hook_from_cfg(item)
            else:
                self.register_hook(item, priority='NORMAL')

8.runner.load_checkpoint(cfg.load_from)

'''
配置文件：
resume_from = None
load_from = 'checkpoints/yolox_l_8x8_300e_coco_20211126_140236-d3bd2b23.pth'
'''
 if cfg.resume_from:#none
        runner.resume(cfg.resume_from)
    elif cfg.load_from:#加载预训练权重文件
        runner.load_checkpoint(cfg.load_from)

9.runner.run()

'''
data_loaders=上面获得的dataloaders
cfg.workflow=[('train',1)]只是表示跑一个epoch是一个循环
'''
runner.run(data_loaders, cfg.workflow)

1）runner.run(data_loaders, workflow）

def run(self, data_loaders, workflow, max_epochs=None, **kwargs):
        """Start running！！！！！！！！

        Args:
            data_loaders (list[:obj:`DataLoader`]): Dataloaders for training and validation.
            workflow (list[tuple]): A list of (phase, epochs) to specify the
                running order and epochs. E.g, [('train', 2), ('val', 1)] means
                running 2 epochs for training and 1 epoch for validation,
                iteratively.
        """
       #data_loaders的长度=1，里面只有一个对象。
       assert len(data_loaders) == len(workflow)
       #0._max_epochs
       self._max_epochs = max_epochs#300
       #1.计算所有的iter（300* len(data_loaders[i])）
       #data_loaders[i]应该是个字典的，它的len是什么
       #data_loaders[i]是个字典，len（字典）=有几个键值对？？？？？？
       self._max_iters = self._max_epochs * len(data_loaders[i])#i=0
       #2.self.work_dir=cfg.work_dir='work_dir_yoloxl'
       work_dir = self.work_dir
       #一堆log操作
       #2.运行run前的hook
       #call__hook函数里面是 getattr，就是执行名字为：before_run 的hook
       self.call_hook('before_run')
        while self.epoch < self._max_epochs:#self.epoch <300
                mode=train
                epochs =1
                #保证mode是个字符串=train
                if isinstance(mode, str): 
                   #epoch_runner被赋予了一个方法名的字符串=runner.train()这个方法
                    epoch_runner = getattr(self, mode)
                for _ in range(epochs):#epochs=1
                    if mode == 'train' and self.epoch >= self._max_epochs:
                        break
                    #运行runner.train()这个方法
                    ###############################
                    epoch_runner(data_loaders[i], **kwargs)

        time.sleep(1)  # wait for some hooks like loggers to finish
        self.call_hook('after_run')

2）runner.train(dataloder)

def train(self, data_loader):
        self.model.train() ############
        self.mode = 'train'#mode赋值
        self.data_loader = data_loader #data_loaders[i]是一个batch的字典
        self._max_iters = self._max_epochs * len(self.data_loader)
        self.call_hook('before_train_epoch')#call_hook函数：执行那些操作
        time.sleep(2)  # Prevent possible deadlock during epoch transition
        for i, data_batch in enumerate(self.data_loader):
        ##每次拿出的data_batch就是一个小批量的数据，进行一次梯度下降
            self._inner_iter = i
            self.call_hook('before_train_iter')
	            self.run_iter(data_batch, train_mode=True)##############
            self.call_hook('after_train_iter')
            self._iter += 1

        self.call_hook('after_train_epoch')
        self._epoch += 1

3）runner.run_iter(data_batch, train_mode=True)

def run_iter(self, data_batch, train_mode):
        elif train_mode:
            outputs = self.model.train_step(data_batch, self.optimizer)
        self.outputs = outputs

4）YOLOX,SingleStageDetector,Base

1.model=yolox在mmdet/models/detectors/yolox.py
2.SingleStageDetector在mmdet/models/detectors/single_stage.py
3.BaseDetector在mmdet/models/detectors/base.py
yolox继承SingleStageDetector继承BaseDetector

1调用BaseDetector中的train_step。2.train_step再调用BaseDetector的foward。3.forward再调用被yolox重写的forward_train。4.forward_train里面又调用了single_stage中的forward_stage。5.得到的结果就是run_iter中得到的outputs

10.validate

1）创建测试集

2）创建数据迭代器

3）runner.register注册hook

    if validate:#true
        # Support batch_size > 1 in validation
        # val_samples_per_gpu=1
        val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
        if val_samples_per_gpu > 1:
            # Replace 'ImageToTensor' to 'DefaultFormatBundle'
            cfg.data.val.pipeline = replace_ImageToTensor(
                cfg.data.val.pipeline)
        #创建测试集
        '''
        val=dict(
        type=dataset_type,
        classes=('human body', 'ball', 'circle cage', 'square cage', 'tyre', 'metal bucket', 'cube', 'cylinder'),
        ann_file='data/sna/annotations/val.json',
        img_prefix='data/sna/val/',
        pipeline=test_pipeline),
     
        test_pipeline = [
        dict(type='LoadImageFromFile'),
        dict(
        type='MultiScaleFlipAug',
        img_scale=img_scale,
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(
                type='Pad',
                pad_to_square=True,
                pad_val=dict(img=(114.0, 114.0, 114.0))),
            dict(type='DefaultFormatBundle'),
            dict(type='Collect', keys=['img'])
        ])]
        '''
        val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
        #创建数据迭代器
        val_dataloader = build_dataloader(
            val_dataset,
            samples_per_gpu=val_samples_per_gpu,#1
            workers_per_gpu=cfg.data.workers_per_gpu,#4
            dist=distributed,
            shuffle=False)
        #注册hook
        '''
        eval_cfg =
        evaluation = dict(
        save_best='auto',
        interval=10,
        dynamic_intervals=[(300 -15, 1)],
        metric='bbox') 
        '''
        eval_cfg = cfg.get('evaluation', {})
        #eval_cfg['by_epoch']=True
        eval_cfg['by_epoch'] = cfg.runner['type'] != 'IterBasedRunner'
        #eval_hook=EvalHook
        eval_hook = DistEvalHook if distributed else EvalHook
        # In this PR (https://github.com/open-mmlab/mmcv/pull/1193), the
        # priority of IterTimerHook has been modified from 'NORMAL' to 'LOW'.
        # 这样看这里面只有1个hook，将它插入到hook队列
        runner.register_hook(
            eval_hook(val_dataloader, **eval_cfg), priority='LOW')

这是里面的EvalHook在mmmdet.core.evaluation.eval_hooks

#没有参数全都是默认的
class EvalHook(BaseEvalHook):

    def __init__(self, *args, dynamic_intervals=None, **kwargs):
        super(EvalHook, self).__init__(*args, **kwargs)

        self.use_dynamic_intervals = dynamic_intervals is not None
        if self.use_dynamic_intervals:
            self.dynamic_milestones, self.dynamic_intervals = \
                _calc_dynamic_intervals(self.interval, dynamic_intervals)

    def _decide_interval(self, runner):
        if self.use_dynamic_intervals:
            progress = runner.epoch if self.by_epoch else runner.iter
            step = bisect.bisect(self.dynamic_milestones, (progress + 1))
            # Dynamically modify the evaluation interval
            self.interval = self.dynamic_intervals[step - 1]

    def before_train_epoch(self, runner):
        """Evaluate the model only at the start of training by epoch."""
        self._decide_interval(runner)
        super().before_train_epoch(runner)

    def before_train_iter(self, runner):
        self._decide_interval(runner)
        super().before_train_iter(runner)

    def _do_evaluate(self, runner):
        """perform evaluation and save ckpt."""
        if not self._should_evaluate(runner):
            return

        from mmdet.apis import single_gpu_test
        results = single_gpu_test(runner.model, self.dataloader, show=False)
        runner.log_buffer.output['eval_iter_num'] = len(self.dataloader)
        key_score = self.evaluate(runner, results)
        if self.save_best:
            self._save_ckpt(runner, key_score)

继承的evalhook在/mmcv/runner/hooks/evaluation

5）runner.run()

在run的过程中，当执行after_train_epoch 方法时候，调用evalhook中的after_train_epoch （1.执行_should_evaluate判断是否进行评估，2.执行_do_evaluate: 1的结果是false：return；1的结果是true：执行mmdet.api.test.py,single_gpu_test进行评估）

三.模块

1.class YOLOX

在mmdel/models/detectors/yolox.py

2.class CSPDarknet

在 mmdel/models/backbones/csp_darknet.py

3.class YOLOXPAFPN

在 mmdel/models/necks/yolox_pafpn.py

4.class YOLOXHEAD

在 mmdel/models/denseheads/yolox_head.py