Mercurial > pylearn

comparison doc/v2_planning/plugin_RP_GD.py @ 1256:bf41991692ea

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
new plugin approach
author gdesjardins
date Fri, 24 Sep 2010 12:53:53 -0400
parents
children c88db30f4e08
comparison
equal deleted inserted replaced
1253:826d78f0135f 1256:bf41991692ea
1 ####
2 # H1: everything works in term of iterator
3 # everything has a step() and end() method
4 ####
5
6 # Construct counter plugin that keeps track of number of epochs
7 class Counter(Plugin):
8
9 def __init__(self, sch, name, threshold):
10 super(self, Counter).__init__(sch, name)
11 self.n = 0
12 self.threshold = threshold
13
14 def execute(self, msg):
15 self.n += 1
16 if self.n > self.threshold:
17 self.fire(Event('terminate', value = self.n))
18
19 def fixed_epoch_trainer(model, save, n_epochs):
20 sched = Scheduler()
21
22 # define plugins
23 [model, validate, save] = map(pluggin_wrapper, [model, validate, save])
24
25 counter = Counter(sched, 'epoch', n_epochs)
26
27 # register actions
28 model.act(sched, on=[sched.begin(), model.step(), counter.step()])
29 counter.act(sched, on=model.end())
30 save_model.act(sched, on=counter.end())
31
32 sched.terminate(on=counter.end())
33
34 return sched
35
36 def early_stop_trainer(model, validate, save, **kw):
37 sched = Scheduler()
38
39 # define plugins
40 [model, validate, save] = map(pluggin_wrapper, [model, validate, save])
41
42 early_stop = Stopper(**kw)
43
44 # register actions
45 model.act(sched, on=[sched.begin(), model.step(), validate.step()])
46 validate.act(sched, on=model.end())
47 early_stop.act(sched, on=validate.step())
48 save_model.act(sched, on=[early_stop.step(), early_stop.end()])
49
50 sched.terminate(on=early_stop.end())
51
52 return sched
53
54 def dbn_trainer(rbm1, rbm2):
55 sched = Scheduler()
56
57 pretrain_layer1 = fixed_epoch_trainer(rbm1, save)
58 pretrain_layer1.act(sched, on=sched.begin())
59
60 pretrain_layer2 = fixed_epoch_trainer(rbm2, save)
61 pretrain_layer2.act(sched, on=pretrain_layer1.end())
62
63 ## TBD: by the layer committee
64 mlp = function(rbm1, rbm2)
65
66 fine_tuning = early_stop_trainer(mlp, validate_mlp, save_mlp)
67 fine_tuning.act(sched, on=pretrain_layer2.end())
68
69 return sched
70
71 def single_crossval_run(trainer, kfold_plugin, kfold_measure)
72
73 sched = Scheduler()
74
75 # k-fold plugin will call rbm.change_dataset using various splits of the data
76 kfold_plugin.act(sched, on=[sched.begin(), trainer.end()])
77 trainer.act(sched, on=[kfold_plugin.step()])
78
79 # trainer terminates on early_stop.end(). This means that trainer.end() will forward
80 # the early-stopping message which contains the best validation error.
81 kfold_measure.act(sched, on=[trainer.end(), kill=kfold_plugin.end()]
82
83 # this best validation error is then forwarded by single_crossval_run
84 sched.terminate(on=kfold_measure.end())
85
86 return sched
87
88
89 #### MAIN LOOP ####
90 rbm1 = ...
91 rbm2 = ...
92 dataset = ....
93 dbn_trainer = dbn_trainer(rbm1, rbm2)
94 kfold_plugin = KFold([rbm1, rbm2], dataset)
95 kfold_measure = ...
96
97 # manually add "hook" to monitor early stopping statistics
98 # NB: advantage of plugins is that this code can go anywhere ...
99 print_stat.act(pretrain_layer1, on=pretrain_layer1.plugins['early_stop'].step())
100
101 #### THIS SHOULD CORRESPOND TO THE OUTER LOOP ####
102 sched = Scheduler()
103
104 hyperparam_change = DBN_HyperParam([rbm1, rbm2])
105 hyperparam_test = single_crossval_run(dbn_trainer, kfold_plugin, kfold_measure)
106
107 hyperparam_change.act(sched, on=[sched.begin(), hyperparam_test.end()])
108 hyperparam_test.act(sched, on=hyperparam_change.step())
109
110 sched.terminate(hyperparam_change.end())
111
112
113 ##### RUN THE WHOLE DAMN THING #####
114 sched.run()