
1. 什么是Hook
经常会听到钩子函数(hook function)这个概念,最近在看目标检测开源框架mmdetection,里面也出现大量Hook的编程方式,那到底什么是hook?hook的作用是什么?
-
what is hook ?钩子hook,顾名思义,可以理解是一个挂钩,作用是有需要的时候挂一个东西上去。具体的解释是:钩子函数是把我们自己实现的hook函数在某一时刻挂接到目标挂载点上。
-
hook函数的作用 举个例子,hook的概念在windows桌面软件开发很常见,特别是各种事件触发的机制; 比如C++的MFC程序中,要监听鼠标左键按下的时间,MFC提供了一个onLeftKeyDown的钩子函数。很显然,MFC框架并没有为我们实现onLeftKeyDown具体的操作,只是为我们提供一个钩子,当我们需要处理的时候,只要去重写这个函数,把我们需要操作挂载在这个钩子里,如果我们不挂载,MFC事件触发机制中执行的就是空的操作。
从上面可知
-
hook函数是程序中预定义好的函数,这个函数处于原有程序流程当中(暴露一个钩子出来)
-
我们需要再在有流程中钩子定义的函数块中实现某个具体的细节,需要把我们的实现,挂接或者注册(register)到钩子里,使得hook函数对目标可用
-
hook 是一种编程机制,和具体的语言没有直接的关系
-
如果从设计模式上看,hook模式是模板方法的扩展
-
钩子只有注册的时候,才会使用,所以原有程序的流程中,没有注册或挂载时,执行的是空(即没有执行任何操作)
本文用python来解释hook的实现方式,并展示在开源项目中hook的应用案例。hook函数和我们常听到另外一个名称:回调函数(callback function)功能是类似的,可以按照同种模式来理解。

2. hook实现例子
据我所知,hook函数最常使用在某种流程处理当中。这个流程往往有很多步骤。hook函数常常挂载在这些步骤中,为增加额外的一些操作,提供灵活性。
下面举一个简单的例子,这个例子的目的是实现一个通用往队列中插入内容的功能。流程步骤有2个
-
需要再插入队列前,对数据进行筛选
input_filter_fn -
插入队列
insert_queue
1class ContentStash(object): 2 """ 3 content stash for online operation 4 pipeline is 5 1. input_filter: filter some contents, no use to user 6 2. insert_queue(redis or other broker): insert useful content to queue 7 """ 8 9 def __init__(self): 10 self.input_filter_fn = None 11 self.broker = [] 12 13 def register_input_filter_hook(self, input_filter_fn): 14 """ 15 register input filter function, parameter is content dict 16 Args: 17 input_filter_fn: input filter function 18 19 Returns: 20 21 """ 22 self.input_filter_fn = input_filter_fn 23 24 def insert_queue(self, content): 25 """ 26 insert content to queue 27 Args: 28 content: dict 29 30 Returns: 31 32 """ 33 self.broker.append(content) 34 35 def input_pipeline(self, content, use=False): 36 """ 37 pipeline of input for content stash 38 Args: 39 use: is use, defaul False 40 content: dict 41 42 Returns: 43 44 """ 45 if not use: 46 return 47 48 # input filter 49 if self.input_filter_fn: 50 _filter = self.input_filter_fn(content) 51 52 # insert to queue 53 if not _filter: 54 self.insert_queue(content) 55 56 57 58# test 59## 实现一个你所需要的钩子实现:比如如果content 包含time就过滤掉,否则插入队列 60def input_filter_hook(content): 61 """ 62 test input filter hook 63 Args: 64 content: dict 65 66 Returns: None or content 67 68 """ 69 if content.get('time') is None: 70 return 71 else: 72 return content 73 74 75# 原有程序 76content = {'filename': 'test.jpg', 'b64_file': "#test", 'data': {"result": "cat", "probility": 0.9}} 77content_stash = ContentStash('audit', work_dir='') 78 79# 挂上钩子函数, 可以有各种不同钩子函数的实现,但是要主要函数输入输出必须保持原有程序中一致,比如这里是content 80content_stash.register_input_filter_hook(input_filter_hook) 81 82# 执行流程 83content_stash.input_pipeline(content) 84 85
3. hook在开源框架中的应用
3.1 keras
在深度学习训练流程中,hook函数体现的淋漓尽致。
一个训练过程(不包括数据准备),会轮询多次训练集,每次称为一个epoch,每个epoch又分为多个batch来训练。流程先后拆解成:
-
开始训练
-
训练一个epoch前
-
训练一个batch前
-
训练一个batch后
-
训练一个epoch后
-
评估验证集
-
结束训练
这些步骤是穿插在训练一个batch数据的过程中,这些可以理解成是钩子函数,我们可能需要在这些钩子函数中实现一些定制化的东西,比如在训练一个epoch后我们要保存下训练的模型,在结束训练时用最好的模型执行下测试集的效果等等。
keras中是通过各种回调函数来实现钩子hook功能的。这里放一个callback的父类,定制时只要继承这个父类,实现你过关注的钩子就可以了。
1@keras_export('keras.callbacks.Callback') 2class Callback(object): 3 """Abstract base class used to build new callbacks. 4 5 Attributes: 6 params: Dict. Training parameters 7 (eg. verbosity, batch size, number of epochs...). 8 model: Instance of `keras.models.Model`. 9 Reference of the model being trained. 10 11 The `logs` dictionary that callback methods 12 take as argument will contain keys for quantities relevant to 13 the current batch or epoch (see method-specific docstrings). 14 """ 15 16 def __init__(self): 17 self.validation_data = None # pylint: disable=g-missing-from-attributes 18 self.model = None 19 # Whether this Callback should only run on the chief worker in a 20 # Multi-Worker setting. 21 # TODO(omalleyt): Make this attr public once solution is stable. 22 self._chief_worker_only = None 23 self._supports_tf_logs = False 24 25 def set_params(self, params): 26 self.params = params 27 28 def set_model(self, model): 29 self.model = model 30 31 @doc_controls.for_subclass_implementers 32 @generic_utils.default 33 def on_batch_begin(self, batch, logs=None): 34 """A backwards compatibility alias for `on_train_batch_begin`.""" 35 36 @doc_controls.for_subclass_implementers 37 @generic_utils.default 38 def on_batch_end(self, batch, logs=None): 39 """A backwards compatibility alias for `on_train_batch_end`.""" 40 41 @doc_controls.for_subclass_implementers 42 def on_epoch_begin(self, epoch, logs=None): 43 """Called at the start of an epoch. 44 45 Subclasses should override for any actions to run. This function should only 46 be called during TRAIN mode. 47 48 Arguments: 49 epoch: Integer, index of epoch. 50 logs: Dict. Currently no data is passed to this argument for this method 51 but that may change in the future. 52 """ 53 54 @doc_controls.for_subclass_implementers 55 def on_epoch_end(self, epoch, logs=None): 56 """Called at the end of an epoch. 57 58 Subclasses should override for any actions to run. This function should only 59 be called during TRAIN mode. 60 61 Arguments: 62 epoch: Integer, index of epoch. 63 logs: Dict, metric results for this training epoch, and for the 64 validation epoch if validation is performed. Validation result keys 65 are prefixed with `val_`. 66 """ 67 68 @doc_controls.for_subclass_implementers 69 @generic_utils.default 70 def on_train_batch_begin(self, batch, logs=None): 71 """Called at the beginning of a training batch in `fit` methods. 72 73 Subclasses should override for any actions to run. 74 75 Arguments: 76 batch: Integer, index of batch within the current epoch. 77 logs: Dict, contains the return value of `model.train_step`. Typically, 78 the values of the `Model`'s metrics are returned. Example: 79 `{'loss': 0.2, 'accuracy': 0.7}`. 80 """ 81 # For backwards compatibility. 82 self.on_batch_begin(batch, logs=logs) 83 84 @doc_controls.for_subclass_implementers 85 @generic_utils.default 86 def on_train_batch_end(self, batch, logs=None): 87 """Called at the end of a training batch in `fit` methods. 88 89 Subclasses should override for any actions to run. 90 91 Arguments: 92 batch: Integer, index of batch within the current epoch. 93 logs: Dict. Aggregated metric results up until this batch. 94 """ 95 # For backwards compatibility. 96 self.on_batch_end(batch, logs=logs) 97 98 @doc_controls.for_subclass_implementers 99 @generic_utils.default 100 def on_test_batch_begin(self, batch, logs=None): 101 """Called at the beginning of a batch in `evaluate` methods. 102 103 Also called at the beginning of a validation batch in the `fit` 104 methods, if validation data is provided. 105 106 Subclasses should override for any actions to run. 107 108 Arguments: 109 batch: Integer, index of batch within the current epoch. 110 logs: Dict, contains the return value of `model.test_step`. Typically, 111 the values of the `Model`'s metrics are returned. Example: 112 `{'loss': 0.2, 'accuracy': 0.7}`. 113 """ 114 115 @doc_controls.for_subclass_implementers 116 @generic_utils.default 117 def on_test_batch_end(self, batch, logs=None): 118 """Called at the end of a batch in `evaluate` methods. 119 120 Also called at the end of a validation batch in the `fit` 121 methods, if validation data is provided. 122 123 Subclasses should override for any actions to run. 124 125 Arguments: 126 batch: Integer, index of batch within the current epoch. 127 logs: Dict. Aggregated metric results up until this batch. 128 """ 129 130 @doc_controls.for_subclass_implementers 131 @generic_utils.default 132 def on_predict_batch_begin(self, batch, logs=None): 133 """Called at the beginning of a batch in `predict` methods. 134 135 Subclasses should override for any actions to run. 136 137 Arguments: 138 batch: Integer, index of batch within the current epoch. 139 logs: Dict, contains the return value of `model.predict_step`, 140 it typically returns a dict with a key 'outputs' containing 141 the model's outputs. 142 """ 143 144 @doc_controls.for_subclass_implementers 145 @generic_utils.default 146 def on_predict_batch_end(self, batch, logs=None): 147 """Called at the end of a batch in `predict` methods. 148 149 Subclasses should override for any actions to run. 150 151 Arguments: 152 batch: Integer, index of batch within the current epoch. 153 logs: Dict. Aggregated metric results up until this batch. 154 """ 155 156 @doc_controls.for_subclass_implementers 157 def on_train_begin(self, logs=None): 158 """Called at the beginning of training. 159 160 Subclasses should override for any actions to run. 161 162 Arguments: 163 logs: Dict. Currently no data is passed to this argument for this method 164 but that may change in the future. 165 """ 166 167 @doc_controls.for_subclass_implementers 168 def on_train_end(self, logs=None): 169 """Called at the end of training. 170 171 Subclasses should override for any actions to run. 172 173 Arguments: 174 logs: Dict. Currently the output of the last call to `on_epoch_end()` 175 is passed to this argument for this method but that may change in 176 the future. 177 """ 178 179 @doc_controls.for_subclass_implementers 180 def on_test_begin(self, logs=None): 181 """Called at the beginning of evaluation or validation. 182 183 Subclasses should override for any actions to run. 184 185 Arguments: 186 logs: Dict. Currently no data is passed to this argument for this method 187 but that may change in the future. 188 """ 189 190 @doc_controls.for_subclass_implementers 191 def on_test_end(self, logs=None): 192 """Called at the end of evaluation or validation. 193 194 Subclasses should override for any actions to run. 195 196 Arguments: 197 logs: Dict. Currently the output of the last call to 198 `on_test_batch_end()` is passed to this argument for this method 199 but that may change in the future. 200 """ 201 202 @doc_controls.for_subclass_implementers 203 def on_predict_begin(self, logs=None): 204 """Called at the beginning of prediction. 205 206 Subclasses should override for any actions to run. 207 208 Arguments: 209 logs: Dict. Currently no data is passed to this argument for this method 210 but that may change in the future. 211 """ 212 213 @doc_controls.for_subclass_implementers 214 def on_predict_end(self, logs=None): 215 """Called at the end of prediction. 216 217 Subclasses should override for any actions to run. 218 219 Arguments: 220 logs: Dict. Currently no data is passed to this argument for this method 221 but that may change in the future. 222 """ 223 224 def _implements_train_batch_hooks(self): 225 """Determines if this Callback should be called for each train batch.""" 226 return (not generic_utils.is_default(self.on_batch_begin) or 227 not generic_utils.is_default(self.on_batch_end) or 228 not generic_utils.is_default(self.on_train_batch_begin) or 229 not generic_utils.is_default(self.on_train_batch_end)) 230
这些钩子的原始程序是在模型训练流程中的
keras源码位置: tensorflow\python\keras\engine\training.py
部分摘录如下(## I am hook):
1# Container that configures and calls `tf.keras.Callback`s. 2 if not isinstance(callbacks, callbacks_module.CallbackList): 3 callbacks = callbacks_module.CallbackList( 4 callbacks, 5 add_history=True, 6 add_progbar=verbose != 0, 7 model=self, 8 verbose=verbose, 9 epochs=epochs, 10 steps=data_handler.inferred_steps) 11 12 ## I am hook 13 callbacks.on_train_begin() 14 training_logs = None 15 # Handle fault-tolerance for multi-worker. 16 # TODO(omalleyt): Fix the ordering issues that mean this has to 17 # happen after `callbacks.on_train_begin`. 18 data_handler._initial_epoch = ( # pylint: disable=protected-access 19 self._maybe_load_initial_epoch_from_ckpt(initial_epoch)) 20 for epoch, iterator in data_handler.enumerate_epochs(): 21 self.reset_metrics() 22 callbacks.on_epoch_begin(epoch) 23 with data_handler.catch_stop_iteration(): 24 for step in data_handler.steps(): 25 with trace.Trace( 26 'TraceContext', 27 graph_type='train', 28 epoch_num=epoch, 29 step_num=step, 30 batch_size=batch_size): 31 ## I am hook 32 callbacks.on_train_batch_begin(step) 33 tmp_logs = train_function(iterator) 34 if data_handler.should_sync: 35 context.async_wait() 36 logs = tmp_logs # No error, now safe to assign to logs. 37 end_step = step + data_handler.step_increment 38 callbacks.on_train_batch_end(end_step, logs) 39 epoch_logs = copy.copy(logs) 40 41 # Run validation. 42 43 ## I am hook 44 callbacks.on_epoch_end(epoch, epoch_logs) 45
3.2 mmdetection
mmdetection是一个目标检测的开源框架,集成了许多不同的目标检测深度学习算法(pytorch版),如faster-rcnn, fpn, retianet等。里面也大量使用了hook,暴露给应用实现流程中具体部分。
详见https://github.com/open-mmlab/mmdetection
这里看一个训练的调用例子(摘录)(https://github.com/open-mmlab/mmdetection/blob/5d592154cca589c5113e8aadc8798bbc73630d98/mmdet/apis/train.py)
1def train_detector(model, 2 dataset, 3 cfg, 4 distributed=False, 5 validate=False, 6 timestamp=None, 7 meta=None): 8 logger = get_root_logger(cfg.log_level) 9 10 # prepare data loaders 11 12 # put model on gpus 13 14 # build runner 15 optimizer = build_optimizer(model, cfg.optimizer) 16 runner = EpochBasedRunner( 17 model, 18 optimizer=optimizer, 19 work_dir=cfg.work_dir, 20 logger=logger, 21 meta=meta) 22 # an ugly workaround to make .log and .log.json filenames the same 23 runner.timestamp = timestamp 24 25 # fp16 setting 26 # register hooks 27 runner.register_training_hooks(cfg.lr_config, optimizer_config, 28 cfg.checkpoint_config, cfg.log_config, 29 cfg.get('momentum_config', None)) 30 if distributed: 31 runner.register_hook(DistSamplerSeedHook()) 32 33 # register eval hooks 34 if validate: 35 # Support batch_size > 1 in validation 36 eval_cfg = cfg.get('evaluation', {}) 37 eval_hook = DistEvalHook if distributed else EvalHook 38 runner.register_hook(eval_hook(val_dataloader, **eval_cfg)) 39 40 # user-defined hooks 41 if cfg.get('custom_hooks', None): 42 custom_hooks = cfg.custom_hooks 43 assert isinstance(custom_hooks, list), \ 44 f'custom_hooks expect list type, but got {type(custom_hooks)}' 45 for hook_cfg in cfg.custom_hooks: 46 assert isinstance(hook_cfg, dict), \ 47 'Each item in custom_hooks expects dict type, but got ' \ 48 f'{type(hook_cfg)}' 49 hook_cfg = hook_cfg.copy() 50 priority = hook_cfg.pop('priority', 'NORMAL') 51 hook = build_from_cfg(hook_cfg, HOOKS) 52 runner.register_hook(hook, priority=priority) 53
4. 总结
本文介绍了hook的概念和应用,并给出了python的实现细则。希望对比有帮助。总结如下:
-
hook函数是流程中预定义好的一个步骤,没有实现
-
挂载或者注册时, 流程执行就会执行这个钩子函数
-
回调函数和hook函数功能上是一致的
-
hook设计方式带来灵活性,如果流程中有一个步骤,你想让调用方来实现,你可以用hook函数
作者简介:wedo实验君, 数据分析师;热爱生活,热爱写作
-------------------********************************** End **********-------------**-----********-**********************************
往期精彩文章推荐:

欢迎各位大佬点击链接加入群聊【helloworld开发者社区】:https://jq.qq.com/?_wv=1027&k=mBlk6nzX进群交流IT技术热点。
本文转自 https://mp.weixin.qq.com/s/1ELA1lu5x19OFOdjB_iBJQ,如有侵权,请联系删除。
