Class implementing a metric-based early-stopping decision rule.
Source code in declearn/main/utils/_early_stop.py
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 | class EarlyStopping:
"""Class implementing a metric-based early-stopping decision rule."""
def __init__(
self,
tolerance: float = 0.0,
patience: int = 1,
decrease: bool = True,
relative: bool = False,
) -> None:
"""Instantiate the early stopping criterion.
Parameters
----------
tolerance: float, default=0.
Improvement value wrt the best previous value below
which the metric is deemed to be non-improving.
If negative, define a tolerance to punctual regression
of the metric. If positive, define an "intolerance" to
low improvements (in absolute or relative value).
patience: int, default=1
Number of consecutive non-improving epochs that trigger
early stopping.
decrease: bool, default=True
Whether the monitored metric is supposed to decrease
rather than increase with training.
relative: bool, default=False
Whether the `tolerance` threshold should be compared
to the `(last - best) / best` improvement ratio rather
than to the absolute improvement value `last - best`.
"""
self.tolerance = tolerance
self.patience = patience
self.decrease = decrease
self.relative = relative
self._best_metric = None # type: Optional[float]
self._n_iter_stuck = 0
def reset(
self,
) -> None:
"""Reset the early-stopping criterion to its initial state."""
self._best_metric = None
self._n_iter_stuck = 0
@property
def keep_training(self) -> bool:
"""Whether training should continue as per this criterion."""
return self._n_iter_stuck < self.patience
def update(
self,
metric: float,
) -> bool:
"""Update the early-stopping decision based on a new value.
Parameters
----------
metric: float
Value of the monitored metric at the current epoch.
Returns
-------
keep_training: bool
Whether training should continue.
"""
# Case when the input metric is the first to be received.
if self._best_metric is None:
self._best_metric = metric
return True
# Otherwise, compute the metric's improvement and act consequently.
diff = (metric - self._best_metric) * (-1 if self.decrease else 1)
if diff > 0:
self._best_metric = metric
if self.relative:
diff /= self._best_metric
if diff <= self.tolerance:
self._n_iter_stuck += 1
else:
self._n_iter_stuck = 0
return self.keep_training
|
keep_training: bool
property
Whether training should continue as per this criterion.
__init__(tolerance=0.0, patience=1, decrease=True, relative=False)
Instantiate the early stopping criterion.
Parameters:
| Name |
Type |
Description |
Default |
tolerance |
float
|
Improvement value wrt the best previous value below
which the metric is deemed to be non-improving.
If negative, define a tolerance to punctual regression
of the metric. If positive, define an "intolerance" to
low improvements (in absolute or relative value). |
0.0
|
patience |
int
|
Number of consecutive non-improving epochs that trigger
early stopping. |
1
|
decrease |
bool
|
Whether the monitored metric is supposed to decrease
rather than increase with training. |
True
|
relative |
bool
|
Whether the tolerance threshold should be compared
to the (last - best) / best improvement ratio rather
than to the absolute improvement value last - best. |
False
|
Source code in declearn/main/utils/_early_stop.py
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 | def __init__(
self,
tolerance: float = 0.0,
patience: int = 1,
decrease: bool = True,
relative: bool = False,
) -> None:
"""Instantiate the early stopping criterion.
Parameters
----------
tolerance: float, default=0.
Improvement value wrt the best previous value below
which the metric is deemed to be non-improving.
If negative, define a tolerance to punctual regression
of the metric. If positive, define an "intolerance" to
low improvements (in absolute or relative value).
patience: int, default=1
Number of consecutive non-improving epochs that trigger
early stopping.
decrease: bool, default=True
Whether the monitored metric is supposed to decrease
rather than increase with training.
relative: bool, default=False
Whether the `tolerance` threshold should be compared
to the `(last - best) / best` improvement ratio rather
than to the absolute improvement value `last - best`.
"""
self.tolerance = tolerance
self.patience = patience
self.decrease = decrease
self.relative = relative
self._best_metric = None # type: Optional[float]
self._n_iter_stuck = 0
|
reset()
Reset the early-stopping criterion to its initial state.
Source code in declearn/main/utils/_early_stop.py
| def reset(
self,
) -> None:
"""Reset the early-stopping criterion to its initial state."""
self._best_metric = None
self._n_iter_stuck = 0
|
update(metric)
Update the early-stopping decision based on a new value.
Parameters:
| Name |
Type |
Description |
Default |
metric |
float
|
Value of the monitored metric at the current epoch. |
required
|
Returns:
| Name | Type |
Description |
keep_training |
bool
|
Whether training should continue. |
Source code in declearn/main/utils/_early_stop.py
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 | def update(
self,
metric: float,
) -> bool:
"""Update the early-stopping decision based on a new value.
Parameters
----------
metric: float
Value of the monitored metric at the current epoch.
Returns
-------
keep_training: bool
Whether training should continue.
"""
# Case when the input metric is the first to be received.
if self._best_metric is None:
self._best_metric = metric
return True
# Otherwise, compute the metric's improvement and act consequently.
diff = (metric - self._best_metric) * (-1 if self.decrease else 1)
if diff > 0:
self._best_metric = metric
if self.relative:
diff /= self._best_metric
if diff <= self.tolerance:
self._n_iter_stuck += 1
else:
self._n_iter_stuck = 0
return self.keep_training
|