Mercurial > pylearn

comparison algorithms/stacker.py @ 476:8fcd0f3d9a17

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added a few algorithms
author Olivier Breuleux <breuleuo@iro.umontreal.ca>
date Mon, 27 Oct 2008 17:26:00 -0400
parents
children bb6bdd3b7ff3
comparison
equal deleted inserted replaced
475:11e0357f06f4 476:8fcd0f3d9a17
1
2 import theano
3 from theano import tensor as T
4 import sys
5 import numpy as N
6
7 class Stacker(T.RModule):
8
9 def __init__(self, submodules, input = None, regularize = False):
10 super(Stacker, self).__init__()
11
12 current = input
13 layers = []
14 for i, (submodule, outname) in enumerate(submodules):
15 layer = submodule(current, regularize = regularize)
16 layers.append(layer)
17 current = layer[outname]
18 self.layers = layers
19
20 self.input = self.layers[0].input
21 self.output = current
22
23 local_update = []
24 global_update = []
25 to_update = []
26 all_kits = []
27 for layer in layers:
28 u = layer.update
29 u.resolve_all()
30 to_update += u.updates.keys()
31 all_kits += u.kits
32 # the input is the whole deep model's input instead of the layer's own
33 # input (which is previous_layer[outname])
34 inputs = [self.input] + u.inputs[1:]
35 method = theano.Method(inputs, u.outputs, u.updates, u.kits)
36 local_update.append(method)
37 global_update.append(
38 theano.Method(inputs,
39 u.outputs,
40 # we update the params of the previous layers too but wrt
41 # this layer's cost
42 dict((param, param - layer.lr * T.grad(layer.cost, param))
43 for param in to_update),
44 list(all_kits)))
45
46 self.local_update = local_update
47 self.global_update = global_update
48 self.update = self.global_update[-1]
49 self.compute = theano.Method(self.input, self.output)
50 ll = self.layers[-1]
51 for name, method in ll.components_map():
52 if isinstance(method, theano.Method) and not hasattr(self, name):
53 m = method.dup()
54 m.resolve_all()
55 m.inputs = [self.input if x is ll.input else x for x in m.inputs]
56 setattr(self, name, m)
57
58 def _instance_initialize(self, obj, nunits = None, lr = 0.01, seed = None, **kwargs):
59 super(Stacker, self)._instance_initialize(obj, **kwargs)
60 if seed is not None:
61 R = N.random.RandomState(seed)
62 else:
63 R = N.random
64 for layer in obj.layers:
65 if layer.lr is None:
66 layer.lr = lr
67 if nunits:
68 if len(nunits) != len(obj.layers) + 1:
69 raise ValueError('You should give exactly one more unit numbers as there are layers.')
70 for ni, no, layer in zip(nunits[:-1], nunits[1:], obj.layers):
71 if seed is not None:
72 layer.initialize(ni, no, seed = R.random_integers(sys.maxint - 1))
73 else:
74 layer.initialize(ni, no)
75 if seed is not None:
76 obj.seed(seed)
77
78 def _instance_flops_approx(self, obj):
79 rval = 0
80 for layer in obj.layers:
81 rval += layer.flops_approx()
82 return rval
83