Mercurial > pylearn

comparison doc/v2_planning/plugin_RP_GD.py @ 1259:6f76ecef869e

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
a few changes
author Razvan Pascanu <r.pascanu@gmail.com>
date Fri, 24 Sep 2010 14:10:05 -0400
parents c88db30f4e08
children
comparison
equal deleted inserted replaced
1258:c88db30f4e08 1259:6f76ecef869e
7 2. Components are then "wrapped" into **plugins** 7 2. Components are then "wrapped" into **plugins**
8 3. Plugins are registered through a **scheduler** to act on certain **events** 8 3. Plugins are registered through a **scheduler** to act on certain **events**
9 4. We have defined two basic types of events: 9 4. We have defined two basic types of events:
10 4.1 scheduler events: these are emitted by the scheduler itself. For now we have defined: 10 4.1 scheduler events: these are emitted by the scheduler itself. For now we have defined:
11 scheduler.begin() : sent at the very beginning of scheduler execution 11 scheduler.begin() : sent at the very beginning of scheduler execution
12 scheduler.end() : sent when the scheduler ends 12 scheduler.end() : sent when the scheduler ends ( this makes sense since there is
13 at least from the user perspective a hierarchy of schedulars)
13 4.2 plugin generate events: plugins act as iterators and generate 2 basic events, 14 4.2 plugin generate events: plugins act as iterators and generate 2 basic events,
14 plugin.next() : sent of every "iteration". Each plugin is free to define what an 15 plugin.next() : sent of every "iteration". Each plugin is free to define what an
15 iteration actually means. 16 iteration actually means.
16 plugin.end() : sent when the plugin is done iterating. 17 plugin.end() : sent when the plugin is done iterating.
17 5. Using Olivier Breuleux's schedular, plugins can decide to act on every event, every n-th 18 5. Using Olivier Breuleux's schedular, plugins can decide to act on every event, every n-th
29 """ 30 """
30 This is an example of a plugin which takes care of counting "stuff". It is generic enough 31 This is an example of a plugin which takes care of counting "stuff". It is generic enough
31 that it would probably be included in the library somewhere. 32 that it would probably be included in the library somewhere.
32 """ 33 """
33 34
34 def __init__(self, name, next_count, end_count): 35 def __init__(self, name, end_count, next_count = 1):
35 """ 36 """
36 :param name: name of the event we are counting (could be useful for debugging) 37 :param name: name of the event we are counting (could be useful for debugging)
37 :param next_count: number of iterations before triggering a "next" event 38 :param next_count: number of iterations before triggering a "next" event
38 :param end_count: number of iterations before triggering an "end" event 39 :param end_count: number of iterations before triggering an "end" event
39 """ 40 """
61 they themselves contain other plugins and their own schedulers. A meta-plugin would replace 62 they themselves contain other plugins and their own schedulers. A meta-plugin would replace
62 code-blocks which are usually found inside for-loop or if-else statements. 63 code-blocks which are usually found inside for-loop or if-else statements.
63 64
64 The fixed_epoch_trainer meta-plugin takes care of training a given model, for a fixed 65 The fixed_epoch_trainer meta-plugin takes care of training a given model, for a fixed
65 number of epochs and then saving the resulting model. 66 number of epochs and then saving the resulting model.
67
68 Other arguments for having meta-plugin :
69 * they can define a semantically separable block of code
70 * they are what the library provides 99% of the code ( so you can define a certain
71 template of connecting plugins as a meta-plugin and ship it without worry of things)
72 * they can be breaked apart by the main schedular ( so you would not have "different"
73 schedulars running in the same time; is just one schedular, this way of constructing
74 things is just to improve understanding and intuitions of the syste
75 * they help pushing all the complexity over the backbone of the library ( i.e. the
76 schedular )
77 * all plugins registered inside a hyper-plugin are active only when the hyper-plugin is
78 active; In this sense they can help definining scopes ( as for variables ) - optional
66 """ 79 """
67 # we start by defining our own private scheduler 80 # we start by defining our own private scheduler
68 sched = Scheduler() 81 sched = Scheduler()
69 82
70 # convert "components" to plugins. Maybe this could be automated or done in another way... 83 # convert "components" to plugins. Maybe this could be automated or done in another way...