|
1
|
|
|
import inspect |
|
2
|
|
|
from blocks.extensions import SimpleExtension |
|
3
|
|
|
|
|
4
|
|
|
|
|
5
|
|
|
class SharedVariableModifier(SimpleExtension): |
|
6
|
|
|
"""Adjusts shared variable parameter using some function. |
|
7
|
|
|
|
|
8
|
|
|
Applies a function to compute the new value of a shared parameter each |
|
9
|
|
|
iteration. |
|
10
|
|
|
|
|
11
|
|
|
This class can be used to adapt over the training process parameters |
|
12
|
|
|
like learning rate, momentum, etc. |
|
13
|
|
|
|
|
14
|
|
|
Parameters |
|
15
|
|
|
---------- |
|
16
|
|
|
parameter : :class:`~tensor.TensorSharedVariable` |
|
17
|
|
|
Shared variable to be adjusted |
|
18
|
|
|
function : callable |
|
19
|
|
|
A function which outputs a numeric value to which the |
|
20
|
|
|
given shared variable will be set and may take one or two |
|
21
|
|
|
arguments. |
|
22
|
|
|
|
|
23
|
|
|
In the first case, function that takes the total number of |
|
24
|
|
|
iterations done (``int``) as an input. |
|
25
|
|
|
|
|
26
|
|
|
In the second case, it is a function which takes number of |
|
27
|
|
|
iterations done (``int``) and old value of the shared variable |
|
28
|
|
|
(with the same dtype as `parameter`). |
|
29
|
|
|
num_args : int, optional |
|
30
|
|
|
The number of arguments to pass to the function. If unspecified, |
|
31
|
|
|
it will be inferred. This is useful if you are using function-like |
|
32
|
|
|
objects for which the arity of the function cannot be inferred. |
|
33
|
|
|
|
|
34
|
|
|
Notes |
|
35
|
|
|
----- |
|
36
|
|
|
This class includes a method ``function`` that calls the function |
|
37
|
|
|
passed in the constructor and a ``num_args`` property which computes |
|
38
|
|
|
the number of arguments to use by inspecting the function object. |
|
39
|
|
|
Subclasses may override a method called ``function`` and/or |
|
40
|
|
|
the ``num_args`` property and instead pass ``None`` to the superclass |
|
41
|
|
|
constructor. This can be used to bypass certain serialization issues |
|
42
|
|
|
on Legacy Python regarding the unpicklability of instance |
|
43
|
|
|
method objects. |
|
44
|
|
|
|
|
45
|
|
|
""" |
|
46
|
|
|
def __init__(self, parameter, function, num_args=None, **kwargs): |
|
47
|
|
|
kwargs.setdefault("after_batch", True) |
|
48
|
|
|
super(SharedVariableModifier, self).__init__(**kwargs) |
|
49
|
|
|
self.parameter = parameter |
|
50
|
|
|
self._function = function |
|
51
|
|
|
self._num_args = num_args |
|
52
|
|
|
|
|
53
|
|
|
@property |
|
54
|
|
|
def num_args(self): |
|
55
|
|
|
if self._num_args is None: |
|
56
|
|
|
self._num_args = len(inspect.getargspec(self._function).args) |
|
57
|
|
|
return self._num_args |
|
58
|
|
|
|
|
59
|
|
|
def function(self, *args): |
|
60
|
|
|
return self._function(*args) |
|
61
|
|
|
|
|
62
|
|
|
def do(self, which_callback, *args): |
|
63
|
|
|
iterations_done = self.main_loop.log.status['iterations_done'] |
|
64
|
|
|
if self.num_args == 1: |
|
65
|
|
|
new_value = self.function(iterations_done) |
|
66
|
|
|
else: |
|
67
|
|
|
old_value = self.parameter.get_value() |
|
68
|
|
|
new_value = self.function(iterations_done, old_value) |
|
69
|
|
|
self.parameter.set_value(new_value) |
|
70
|
|
|
|
|
71
|
|
|
|
|
72
|
|
|
class TrackTheBest(SimpleExtension): |
|
73
|
|
|
"""Check if a log quantity has the minimum/maximum value so far. |
|
74
|
|
|
|
|
75
|
|
|
Parameters |
|
76
|
|
|
---------- |
|
77
|
|
|
record_name : str |
|
78
|
|
|
The name of the record to track. |
|
79
|
|
|
notification_name : str, optional |
|
80
|
|
|
The name for the record to be made in the log when the current |
|
81
|
|
|
value of the tracked quantity is the best so far. It not given, |
|
82
|
|
|
'record_name' plus "best_so_far" suffix is used. |
|
83
|
|
|
choose_best : callable, optional |
|
84
|
|
|
A function that takes the current value and the best so far |
|
85
|
|
|
and return the best of two. By default :func:`min`, which |
|
86
|
|
|
corresponds to tracking the minimum value. |
|
87
|
|
|
|
|
88
|
|
|
Attributes |
|
89
|
|
|
---------- |
|
90
|
|
|
best_name : str |
|
91
|
|
|
The name of the status record to keep the best value so far. |
|
92
|
|
|
notification_name : str |
|
93
|
|
|
The name of the record written to the log when the current |
|
94
|
|
|
value of the tracked quantity is the best so far. |
|
95
|
|
|
|
|
96
|
|
|
Notes |
|
97
|
|
|
----- |
|
98
|
|
|
In the likely case that you are relying on another extension to |
|
99
|
|
|
add the tracked quantity to the log, make sure to place this |
|
100
|
|
|
extension *after* the extension that writes the quantity to the log |
|
101
|
|
|
in the `extensions` argument to :class:`blocks.main_loop.MainLoop`. |
|
102
|
|
|
|
|
103
|
|
|
""" |
|
104
|
|
|
def __init__(self, record_name, notification_name=None, |
|
105
|
|
|
choose_best=min, **kwargs): |
|
106
|
|
|
self.record_name = record_name |
|
107
|
|
|
if not notification_name: |
|
108
|
|
|
notification_name = record_name + "_best_so_far" |
|
109
|
|
|
self.notification_name = notification_name |
|
110
|
|
|
self.best_name = "best_" + record_name |
|
111
|
|
|
self.choose_best = choose_best |
|
112
|
|
|
kwargs.setdefault("after_epoch", True) |
|
113
|
|
|
super(TrackTheBest, self).__init__(**kwargs) |
|
114
|
|
|
|
|
115
|
|
|
def do(self, which_callback, *args): |
|
116
|
|
|
current_value = self.main_loop.log.current_row.get(self.record_name) |
|
117
|
|
|
if current_value is None: |
|
118
|
|
|
return |
|
119
|
|
|
best_value = self.main_loop.status.get(self.best_name, None) |
|
120
|
|
|
if (best_value is None or |
|
121
|
|
|
(current_value != best_value and |
|
122
|
|
|
self.choose_best(current_value, best_value) == |
|
123
|
|
|
current_value)): |
|
124
|
|
|
self.main_loop.status[self.best_name] = current_value |
|
125
|
|
|
self.main_loop.log.current_row[self.notification_name] = True |
|
126
|
|
|
|
mila-udem /
blocks
