1
|
|
|
"""The event-based main loop of Blocks.""" |
2
|
|
|
from abc import ABCMeta |
3
|
|
|
from collections import defaultdict |
4
|
|
|
from numbers import Integral |
5
|
|
|
from uuid import uuid4 |
6
|
|
|
|
7
|
|
|
import six |
8
|
|
|
|
9
|
|
|
|
10
|
|
|
@six.add_metaclass(ABCMeta) |
11
|
|
|
class TrainingLogBase(object): |
12
|
|
|
"""Base class for training log. |
13
|
|
|
|
14
|
|
|
A training log stores the training timeline, statistics and other |
15
|
|
|
auxiliary information. Training logs can use different backends e.g. |
16
|
|
|
in-memory Python objects or an SQLite database. |
17
|
|
|
|
18
|
|
|
Information is stored similar to a nested dictionary, so use |
19
|
|
|
``log[time][key]`` to read data. An entry without stored data will |
20
|
|
|
return an empty dictionary-like object that can be written to, |
21
|
|
|
``log[time][key] = value``. |
22
|
|
|
|
23
|
|
|
Depending on the backend, ``log[time] = {'key': 'value'}`` could fail. |
24
|
|
|
Use ``log[time].update({'key': 'value'})`` for compatibility across |
25
|
|
|
backends. |
26
|
|
|
|
27
|
|
|
In addition to the set of records displaying training dynamics, a |
28
|
|
|
training log has a :attr:`status` attribute, which is a dictionary with |
29
|
|
|
data that is not bound to a particular time. |
30
|
|
|
|
31
|
|
|
.. warning:: |
32
|
|
|
|
33
|
|
|
Changes to mutable objects might not be reflected in the log, |
34
|
|
|
depending on the backend. So don't use |
35
|
|
|
``log.status['key'].append(...)``, use ``log.status['key'] = ...`` |
36
|
|
|
instead. |
37
|
|
|
|
38
|
|
|
Parameters |
39
|
|
|
---------- |
40
|
|
|
uuid : :class:`uuid.UUID`, optional |
41
|
|
|
The UUID of this log. For persistent log backends, passing the UUID |
42
|
|
|
will result in an old log being loaded. Otherwise a new, random |
43
|
|
|
UUID will be created. |
44
|
|
|
|
45
|
|
|
Attributes |
46
|
|
|
---------- |
47
|
|
|
status : dict |
48
|
|
|
A dictionary with data representing the current state of training. |
49
|
|
|
By default it contains ``iterations_done``, ``epochs_done`` and |
50
|
|
|
``_epoch_ends`` (a list of time stamps when epochs ended). |
51
|
|
|
|
52
|
|
|
""" |
53
|
|
|
def __init__(self, uuid=None): |
54
|
|
|
if uuid is None: |
55
|
|
|
self.uuid = uuid4() |
56
|
|
|
else: |
57
|
|
|
self.uuid = uuid |
58
|
|
|
if uuid is None: |
59
|
|
|
self.status.update({ |
|
|
|
|
60
|
|
|
'iterations_done': 0, |
61
|
|
|
'epochs_done': 0, |
62
|
|
|
'_epoch_ends': [], |
63
|
|
|
'resumed_from': None |
64
|
|
|
}) |
65
|
|
|
|
66
|
|
|
@property |
67
|
|
|
def h_uuid(self): |
68
|
|
|
"""Return a hexadecimal version of the UUID bytes. |
69
|
|
|
|
70
|
|
|
This is necessary to store ids in an SQLite database. |
71
|
|
|
|
72
|
|
|
""" |
73
|
|
|
return self.uuid.hex |
74
|
|
|
|
75
|
|
|
def resume(self): |
76
|
|
|
"""Resume a log by setting a new random UUID. |
77
|
|
|
|
78
|
|
|
Keeps a record of the old log that this is a continuation of. It |
79
|
|
|
copies the status of the old log into the new log. |
80
|
|
|
|
81
|
|
|
""" |
82
|
|
|
old_uuid = self.h_uuid |
83
|
|
|
old_status = dict(self.status) |
|
|
|
|
84
|
|
|
self.uuid = uuid4() |
85
|
|
|
self.status.update(old_status) |
|
|
|
|
86
|
|
|
self.status['resumed_from'] = old_uuid |
|
|
|
|
87
|
|
|
|
88
|
|
|
def _check_time(self, time): |
89
|
|
|
if not isinstance(time, Integral) or time < 0: |
90
|
|
|
raise ValueError("time must be a non-negative integer") |
91
|
|
|
|
92
|
|
|
@property |
93
|
|
|
def current_row(self): |
94
|
|
|
return self[self.status['iterations_done']] |
|
|
|
|
95
|
|
|
|
96
|
|
|
@property |
97
|
|
|
def previous_row(self): |
98
|
|
|
return self[self.status['iterations_done'] - 1] |
|
|
|
|
99
|
|
|
|
100
|
|
|
@property |
101
|
|
|
def last_epoch_row(self): |
102
|
|
|
return self[self.status['_epoch_ends'][-1]] |
|
|
|
|
103
|
|
|
|
104
|
|
|
|
105
|
|
|
class TrainingLog(defaultdict, TrainingLogBase): |
106
|
|
|
"""Training log using a `defaultdict` as backend. |
107
|
|
|
|
108
|
|
|
Notes |
109
|
|
|
----- |
110
|
|
|
For analysis of the logs, it can be useful to convert the log to a |
111
|
|
|
Pandas_ data frame: |
112
|
|
|
|
113
|
|
|
.. code:: python |
114
|
|
|
|
115
|
|
|
df = DataFrame.from_dict(log, orient='index') |
116
|
|
|
|
117
|
|
|
.. _Pandas: http://pandas.pydata.org |
118
|
|
|
|
119
|
|
|
""" |
120
|
|
|
def __init__(self): |
121
|
|
|
defaultdict.__init__(self, dict) |
122
|
|
|
self.status = {} |
123
|
|
|
TrainingLogBase.__init__(self) |
124
|
|
|
|
125
|
|
|
def __reduce__(self): |
126
|
|
|
constructor, args, _, _, items = super(TrainingLog, self).__reduce__() |
|
|
|
|
127
|
|
|
return constructor, (), self.__dict__, _, items |
128
|
|
|
|
129
|
|
|
def __getitem__(self, time): |
130
|
|
|
self._check_time(time) |
131
|
|
|
return super(TrainingLog, self).__getitem__(time) |
132
|
|
|
|
133
|
|
|
def __setitem__(self, time, value): |
134
|
|
|
self._check_time(time) |
135
|
|
|
return super(TrainingLog, self).__setitem__(time, value) |
136
|
|
|
|
This check looks for calls to members that are non-existent. These calls will fail.
The member could have been renamed or removed.