|
1
|
|
|
import logging |
|
2
|
|
|
|
|
3
|
|
|
logger = logging.getLogger(__name__) |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
class ConstIterations: |
|
7
|
|
|
"""Stopping Criterion: After certain iterations |
|
8
|
|
|
|
|
9
|
|
|
Args: |
|
10
|
|
|
num_iters (:obj:`int`): Number of iterations |
|
11
|
|
|
|
|
12
|
|
|
Attributes: |
|
13
|
|
|
num_iters (:obj:`int`): Number of iterations |
|
14
|
|
|
cur_iter (:obj:`int`): Current number of iterations |
|
15
|
|
|
""" |
|
16
|
|
|
def __init__(self, num_iters): |
|
17
|
|
|
self.num_iters = num_iters |
|
18
|
|
|
self.cur_iter = 0 |
|
19
|
|
|
|
|
20
|
|
|
def reset(self): |
|
21
|
|
|
"""Reset internal iteration counter |
|
22
|
|
|
""" |
|
23
|
|
|
self.cur_iter = 0 |
|
24
|
|
|
|
|
25
|
|
|
def continue_learning(self): |
|
26
|
|
|
"""Determine whether learning should continue |
|
27
|
|
|
If so, return True, otherwise, return False. |
|
28
|
|
|
""" |
|
29
|
|
|
if self.cur_iter < self.num_iters: |
|
30
|
|
|
self.cur_iter += 1 |
|
31
|
|
|
return True |
|
32
|
|
|
else: |
|
33
|
|
|
return False |
|
34
|
|
|
|
|
35
|
|
|
|
|
36
|
|
|
class MonitorBased: |
|
37
|
|
|
"""Stop training based on the return of a monitoring function. |
|
38
|
|
|
|
|
39
|
|
|
If the monitoring result keep improving within past n_steps, keep learning. |
|
40
|
|
|
Otherwise, stop. |
|
41
|
|
|
If the monitoring result is the best at the moment, call the parameter save |
|
42
|
|
|
function. |
|
43
|
|
|
Once it is done, the parameters saved last is the training results. |
|
44
|
|
|
|
|
45
|
|
|
Args: |
|
46
|
|
|
n_steps (:obj:`int`): The amount of steps to look for improvement |
|
47
|
|
|
monitor_fn: Parameter monitor function. |
|
48
|
|
|
monitor_fn_args (:obj:`tuple`): Argument tuple (arg1, arg2, ...) for monitor function. |
|
49
|
|
|
save_fn: Parameter save function. |
|
50
|
|
|
save_fn_args (:obj:`tuple`): Argument tuple (arg1, arg2, ...) for save function. |
|
51
|
|
|
|
|
52
|
|
|
Attributes: |
|
53
|
|
|
n_steps (:obj:`int`): The amount of steps to look for improvement |
|
54
|
|
|
monitor_fn: Parameter monitor function. |
|
55
|
|
|
monitor_fn_args (:obj:`tuple`): Argument tuple (arg1, arg2, ...) for monitor function. |
|
56
|
|
|
save_fn: Parameter save function. |
|
57
|
|
|
save_fn_args (:obj:`tuple`): Argument tuple (arg1, arg2, ...) for save function. |
|
58
|
|
|
step_count (:obj:`int`): Number of steps that the parameter monitored is worse than the best value. |
|
59
|
|
|
best_value: Best value seen so far. |
|
60
|
|
|
""" |
|
61
|
|
|
def __init__(self, n_steps, monitor_fn, monitor_fn_args, save_fn, save_fn_args): |
|
62
|
|
|
self.n_steps = n_steps |
|
63
|
|
|
self.monitor_fn = monitor_fn |
|
64
|
|
|
self.monitor_fn_args = monitor_fn_args |
|
65
|
|
|
self.save_fn = save_fn |
|
66
|
|
|
self.save_fn_args = save_fn_args |
|
67
|
|
|
self.step_count = 0 |
|
68
|
|
|
self.best_value = None |
|
69
|
|
|
|
|
70
|
|
|
def reset(self): |
|
71
|
|
|
"""Reset internal step count |
|
72
|
|
|
""" |
|
73
|
|
|
self.step_count = 0 |
|
74
|
|
|
self.best_value = None |
|
75
|
|
|
|
|
76
|
|
|
def continue_learning(self): |
|
77
|
|
|
"""Determine whether learning should continue |
|
78
|
|
|
If so, return True, otherwise, return False. |
|
79
|
|
|
""" |
|
80
|
|
|
param = self.monitor_fn(*self.monitor_fn_args) |
|
81
|
|
|
if self.best_value is None: |
|
82
|
|
|
self.best_value = param |
|
83
|
|
|
self.save_fn(*self.save_fn_args) |
|
84
|
|
|
if param > self.best_value: |
|
85
|
|
|
self.step_count = 0 |
|
86
|
|
|
self.best_value = param |
|
87
|
|
|
self.save_fn(*self.save_fn_args) |
|
88
|
|
|
logger.info('New Best: %g' % self.best_value) |
|
89
|
|
|
else: |
|
90
|
|
|
self.step_count += 1 |
|
91
|
|
|
if self.step_count > self.n_steps: |
|
92
|
|
|
return False |
|
93
|
|
|
return True |
|
94
|
|
|
|
TinghuiWang /
pyActLearn
