Mercurial > pylearn

comparison pylearn/algorithms/rbm.py @ 544:de6de7c2c54b

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
merged and changed state to dictionary
author James Bergstra <bergstrj@iro.umontreal.ca>
date Mon, 17 Nov 2008 20:05:31 -0500
parents ee5324c21e60 5b4ccbf022c8
children 40cae12a9bb8
comparison
equal deleted inserted replaced
543:34aba0efa3e9 544:de6de7c2c54b
9 import numpy as N 9 import numpy as N
10 10
11 from ..datasets import make_dataset 11 from ..datasets import make_dataset
12 from .minimizer import make_minimizer 12 from .minimizer import make_minimizer
13 from .stopper import make_stopper 13 from .stopper import make_stopper
14
15 from ..dbdict.experiment import subdict 14 from ..dbdict.experiment import subdict
16 15
17 class RBM(module.FancyModule): 16 class RBM(T.RModule):
18 17
19 # is it really necessary to pass ALL of these ? - GD 18 # is it really necessary to pass ALL of these ? - GD
20 def __init__(self, 19 def __init__(self,
21 nvis=None, nhid=None, 20 nvis=None, nhid=None,
22 input=None, 21 input=None,
23 w=None, hidb=None, visb=None): 22 w=None, hidb=None, visb=None,
23 seed=0, lr=0.1):
24
24 super(RBM, self).__init__() 25 super(RBM, self).__init__()
26 self.nhid, self.nvis = nhid, nvis
27 self.lr = lr
25 28
26 # symbolic theano stuff 29 # symbolic theano stuff
27 # what about multidimensional inputs/outputs ? do they have to be 30 # what about multidimensional inputs/outputs ? do they have to be
28 # flattened or should we used tensors instead ? 31 # flattened or should we used tensors instead ?
29 self.w = w if w is not None else module.Member(T.dmatrix()) 32 self.w = w if w is not None else module.Member(T.dmatrix())
30 self.visb = visb if visb is not None else module.Member(T.dvector()) 33 self.visb = visb if visb is not None else module.Member(T.dvector())
31 self.hidb = hidb if hidb is not None else module.Member(T.dvector()) 34 self.hidb = hidb if hidb is not None else module.Member(T.dvector())
35 self.seed = seed;
36
37 # 1-step Markov chain
38 vis = T.dmatrix()
39 hid = sigmoid(T.dot(vis, self.w) + self.hidb)
40 hid_sample = self.random.binomial(T.shape(hid), 1, hid)
41 neg_vis = sigmoid(T.dot(hid_sample, self.w.T) + self.visb)
42 neg_vis_sample = self.random.binomial(T.shape(neg_vis), 1, neg_vis)
43 neg_hid = sigmoid(T.dot(neg_vis_sample, self.w) + self.hidb)
32 44
33 # 1-step Markov chain 45 # function which execute 1-step Markov chain (with and without cd updates)
34 self.hid = sigmoid(T.dot(self.w,self.input) + self.hidb) 46 self.updownup = module.Method([vis], [hid, neg_vis_sample, neg_hid])
35 self.hid_sample = self.hid #TODO: sample!
36 self.vis = sigmoid(T.dot(self.w.T, self.hid) + self.visb)
37 self.vis_sample = self.vis #TODO: sample!
38 self.neg_hid = sigmoid(T.dot(self.w, self.vis) + self.hidb)
39 47
40 # cd1 updates: 48 # function to perform manual cd update given 2 visible and 2 hidden values
41 self.params = [self.w, self.visb, self.hidb] 49 vistemp = T.dmatrix()
42 self.gradients = [ 50 hidtemp = T.dmatrix()
43 T.dot(self.hid, self.input) - T.dot(self.neg_hid, self.vis), 51 nvistemp = T.dmatrix()
44 self.input - self.vis, 52 nhidtemp = T.dmatrix()
45 self.hid - self.neg_hid ] 53 self.cd_update = module.Method([vistemp, hidtemp, nvistemp, nhidtemp],
54 [],
55 updates = {self.w: self.w + self.lr *
56 (T.dot(vistemp.T, hidtemp) -
57 T.dot(nvistemp.T, nhidtemp)),
58 self.visb: self.visb + self.lr *
59 (T.sum(vistemp - nvistemp,axis=0)),
60 self.hidb: self.hidb + self.lr *
61 (T.sum(hidtemp - nhidtemp,axis=0))});
46 62
47 def __instance_initialize(self, obj): 63 # TODO: add parameter for weigth initialization
48 obj.w = N.random.standard_normal((self.nhid,self.nvis)) 64 def _instance_initialize(self, obj):
49 obj.genb = N.zeros(self.nvis) 65 obj.w = N.random.standard_normal((self.nvis,self.nhid))
66 obj.visb = N.zeros(self.nvis)
50 obj.hidb = N.zeros(self.nhid) 67 obj.hidb = N.zeros(self.nhid)
68 obj.seed(self.seed);
51 69
52 def RBM_cd(): 70 def _instance_cd1(self, obj, input, k=1):
53 pass; 71 poshid, negvissample, neghid = obj.updownup(input)
72 for i in xrange(k-1):
73 ahid, negvissample, neghid = obj.updownup(negvissample)
74 # CD-k update
75 obj.cd_update(input, poshid, negvissample, neghid)
76
54 77
55 def train_rbm(state, channel=lambda *args, **kwargs:None): 78 def train_rbm(state, channel=lambda *args, **kwargs:None):
56 dataset = make_dataset(**subdict_copy(state, prefix='dataset_')) 79 dataset = make_dataset(**subdict_copy(state, prefix='dataset_'))
57 train = dataset.train 80 train = dataset.train
58 81
59 rbm_module = RBM( 82 rbm_module = RBM(
60 nvis=train.x.shape[1], 83 nvis=train.x.shape[1],
61 nhid=state['size_hidden']) 84 nhid=state['nhid'])
85 rbm = rbm_module.make()
62 86
63 batchsize = state['batchsize'] 87 batchsize = state.get('batchsize', 1)
64 verbose = state['verbose'] 88 verbose = state.get('verbose', 1)
65 iter = [0] 89 iter = [0]
66 90
67 while iter[0] != state['max_iters']: 91 while iter[0] != state['max_iters']:
68 for j in xrange(0,len(train.x)-batchsize+1,batchsize): 92 for j in xrange(0,len(train.x)-batchsize+1,batchsize):
69 rbm.cd1(train.x[j:j+batchsize]) 93 rbm.cd1(train.x[j:j+batchsize])