5 分钟读懂Python 中的 Hook 钩子函数

时间:2022-11-25 11:34:54

5 分钟读懂Python 中的 Hook 钩子函数

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。

5 分钟读懂Python 中的 Hook 钩子函数

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

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
class ContentStash(object):
  """
  content stash for online operation
  pipeline is
  1. input_filter: filter some contents, no use to user
  2. insert_queue(redis or other broker): insert useful content to queue
  """
 
  def __init__(self):
    self.input_filter_fn = None
    self.broker = []
 
  def register_input_filter_hook(self, input_filter_fn):
    """
    register input filter function, parameter is content dict
    Args:
      input_filter_fn: input filter function
    Returns:
    """
    self.input_filter_fn = input_filter_fn
 
  def insert_queue(self, content):
    """
    insert content to queue
    Args:
      content: dict
    Returns:
    """
    self.broker.append(content)
 
  def input_pipeline(self, content, use=False):
    """
    pipeline of input for content stash
    Args:
      use: is use, defaul False
      content: dict
    Returns:
    """
    if not use:
      return
 
    # input filter
    if self.input_filter_fn:
      _filter = self.input_filter_fn(content)
      
    # insert to queue
    if not _filter:
      self.insert_queue(content)
 
# test
## 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
def input_filter_hook(content):
  """
  test input filter hook
  Args:
    content: dict
  Returns: None or content
  """
  if content.get('time') is None:
    return
  else:
    return content
 
# 原有程序
content = {'filename': 'test.jpg', 'b64_file': "#test", 'data': {"result": "cat", "probility": 0.9}}
content_stash = ContentStash('audit', work_dir='')
 
# 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
content_stash.register_input_filter_hook(input_filter_hook)
 
# 执行流程
content_stash.input_pipeline(content)

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
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
@keras_export('keras.callbacks.Callback')
class Callback(object):
 """Abstract base class used to build new callbacks.
 Attributes:
   params: Dict. Training parameters
     (eg. verbosity, batch size, number of epochs...).
   model: Instance of `keras.models.Model`.
     Reference of the model being trained.
 The `logs` dictionary that callback methods
 take as argument will contain keys for quantities relevant to
 the current batch or epoch (see method-specific docstrings).
 """
 
 def __init__(self):
  self.validation_data = None # pylint: disable=g-missing-from-attributes
  self.model = None
  # Whether this Callback should only run on the chief worker in a
  # Multi-Worker setting.
  # TODO(omalleyt): Make this attr public once solution is stable.
  self._chief_worker_only = None
  self._supports_tf_logs = False
 
 def set_params(self, params):
  self.params = params
 
 def set_model(self, model):
  self.model = model
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_batch_begin(self, batch, logs=None):
  """A backwards compatibility alias for `on_train_batch_begin`."""
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_batch_end(self, batch, logs=None):
  """A backwards compatibility alias for `on_train_batch_end`."""
 
 @doc_controls.for_subclass_implementers
 def on_epoch_begin(self, epoch, logs=None):
  """Called at the start of an epoch.
  Subclasses should override for any actions to run. This function should only
  be called during TRAIN mode.
  Arguments:
    epoch: Integer, index of epoch.
    logs: Dict. Currently no data is passed to this argument for this method
     but that may change in the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_epoch_end(self, epoch, logs=None):
  """Called at the end of an epoch.
  Subclasses should override for any actions to run. This function should only
  be called during TRAIN mode.
  Arguments:
    epoch: Integer, index of epoch.
    logs: Dict, metric results for this training epoch, and for the
     validation epoch if validation is performed. Validation result keys
     are prefixed with `val_`.
  """
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_train_batch_begin(self, batch, logs=None):
  """Called at the beginning of a training batch in `fit` methods.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict, contains the return value of `model.train_step`. Typically,
     the values of the `Model`'s metrics are returned. Example:
     `{'loss': 0.2, 'accuracy': 0.7}`.
  """
  # For backwards compatibility.
  self.on_batch_begin(batch, logs=logs)
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_train_batch_end(self, batch, logs=None):
  """Called at the end of a training batch in `fit` methods.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict. Aggregated metric results up until this batch.
  """
  # For backwards compatibility.
  self.on_batch_end(batch, logs=logs)
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_test_batch_begin(self, batch, logs=None):
  """Called at the beginning of a batch in `evaluate` methods.
  Also called at the beginning of a validation batch in the `fit`
  methods, if validation data is provided.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict, contains the return value of `model.test_step`. Typically,
     the values of the `Model`'s metrics are returned. Example:
     `{'loss': 0.2, 'accuracy': 0.7}`.
  """
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_test_batch_end(self, batch, logs=None):
  """Called at the end of a batch in `evaluate` methods.
  Also called at the end of a validation batch in the `fit`
  methods, if validation data is provided.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict. Aggregated metric results up until this batch.
  """
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_predict_batch_begin(self, batch, logs=None):
  """Called at the beginning of a batch in `predict` methods.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict, contains the return value of `model.predict_step`,
     it typically returns a dict with a key 'outputs' containing
     the model's outputs.
  """
 
 @doc_controls.for_subclass_implementers
 @generic_utils.default
 def on_predict_batch_end(self, batch, logs=None):
  """Called at the end of a batch in `predict` methods.
  Subclasses should override for any actions to run.
  Arguments:
    batch: Integer, index of batch within the current epoch.
    logs: Dict. Aggregated metric results up until this batch.
  """
 
 @doc_controls.for_subclass_implementers
 def on_train_begin(self, logs=None):
  """Called at the beginning of training.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently no data is passed to this argument for this method
     but that may change in the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_train_end(self, logs=None):
  """Called at the end of training.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently the output of the last call to `on_epoch_end()`
     is passed to this argument for this method but that may change in
     the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_test_begin(self, logs=None):
  """Called at the beginning of evaluation or validation.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently no data is passed to this argument for this method
     but that may change in the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_test_end(self, logs=None):
  """Called at the end of evaluation or validation.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently the output of the last call to
     `on_test_batch_end()` is passed to this argument for this method
     but that may change in the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_predict_begin(self, logs=None):
  """Called at the beginning of prediction.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently no data is passed to this argument for this method
     but that may change in the future.
  """
 
 @doc_controls.for_subclass_implementers
 def on_predict_end(self, logs=None):
  """Called at the end of prediction.
  Subclasses should override for any actions to run.
  Arguments:
    logs: Dict. Currently no data is passed to this argument for this method
     but that may change in the future.
  """
 
 def _implements_train_batch_hooks(self):
  """Determines if this Callback should be called for each train batch."""
  return (not generic_utils.is_default(self.on_batch_begin) or
      not generic_utils.is_default(self.on_batch_end) or
      not generic_utils.is_default(self.on_train_batch_begin) or
      not generic_utils.is_default(self.on_train_batch_end))

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
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
# Container that configures and calls `tf.keras.Callback`s.
   if not isinstance(callbacks, callbacks_module.CallbackList):
    callbacks = callbacks_module.CallbackList(
      callbacks,
      add_history=True,
      add_progbar=verbose != 0,
      model=self,
      verbose=verbose,
      epochs=epochs,
      steps=data_handler.inferred_steps)
 
   ## I am hook
   callbacks.on_train_begin()
   training_logs = None
   # Handle fault-tolerance for multi-worker.
   # TODO(omalleyt): Fix the ordering issues that mean this has to
   # happen after `callbacks.on_train_begin`.
   data_handler._initial_epoch = ( # pylint: disable=protected-access
     self._maybe_load_initial_epoch_from_ckpt(initial_epoch))
   for epoch, iterator in data_handler.enumerate_epochs():
    self.reset_metrics()
    callbacks.on_epoch_begin(epoch)
    with data_handler.catch_stop_iteration():
     for step in data_handler.steps():
      with trace.Trace(
        'TraceContext',
        graph_type='train',
        epoch_num=epoch,
        step_num=step,
        batch_size=batch_size):
       ## I am hook
       callbacks.on_train_batch_begin(step)
       tmp_logs = train_function(iterator)
       if data_handler.should_sync:
        context.async_wait()
       logs = tmp_logs # No error, now safe to assign to logs.
       end_step = step + data_handler.step_increment
       callbacks.on_train_batch_end(end_step, logs)
    epoch_logs = copy.copy(logs)
 
    # Run validation.
 
    ## I am hook
    callbacks.on_epoch_end(epoch, epoch_logs)

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

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
def train_detector(model,
          dataset,
          cfg,
          distributed=False,
          validate=False,
          timestamp=None,
          meta=None):
  logger = get_root_logger(cfg.log_level)
 
  # prepare data loaders
 
  # put model on gpus
 
  # build runner
  optimizer = build_optimizer(model, cfg.optimizer)
  runner = EpochBasedRunner(
    model,
    optimizer=optimizer,
    work_dir=cfg.work_dir,
    logger=logger,
    meta=meta)
  # an ugly workaround to make .log and .log.json filenames the same
  runner.timestamp = timestamp
 
  # fp16 setting
  # register hooks
  runner.register_training_hooks(cfg.lr_config, optimizer_config,
                  cfg.checkpoint_config, cfg.log_config,
                  cfg.get('momentum_config', None))
  if distributed:
    runner.register_hook(DistSamplerSeedHook())
 
  # register eval hooks
  if validate:
    # Support batch_size > 1 in validation
    eval_cfg = cfg.get('evaluation', {})
    eval_hook = DistEvalHook if distributed else EvalHook
    runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
 
  # user-defined hooks
  if cfg.get('custom_hooks', None):
    custom_hooks = cfg.custom_hooks
    assert isinstance(custom_hooks, list), \
      f'custom_hooks expect list type, but got {type(custom_hooks)}'
    for hook_cfg in cfg.custom_hooks:
      assert isinstance(hook_cfg, dict), \
        'Each item in custom_hooks expects dict type, but got ' \
        f'{type(hook_cfg)}'
      hook_cfg = hook_cfg.copy()
      priority = hook_cfg.pop('priority', 'NORMAL')
      hook = build_from_cfg(hook_cfg, HOOKS)
      runner.register_hook(hook, priority=priority)

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

到此这篇关于5 分钟读懂Python 中的 Hook 钩子函数的文章就介绍到这了,更多相关Python Hook 钩子函数内容请搜索服务器之家以前的文章或继续浏览下面的相关文章希望大家以后多多支持服务器之家!

原文链接:https://blog.csdn.net/BF02jgtRS00XKtCx/article/details/110458293