ift6266: code_tutoriel/deep.py annotate

annotate code_tutoriel/deep.py @ 489:ee9836baade3

merge

author	dumitru@dumitru.mtv.corp.google.com
date	Mon, 31 May 2010 19:07:59 -0700
parents	4bc5eeec6394
children

rev	line source
165 4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	1 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	2 Draft of DBN, DAA, SDAA, RBM tutorial code
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	3
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	4 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	5 import sys
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	6 import numpy
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	7 import theano
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	8 import time
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	9 import theano.tensor as T
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	10 from theano.tensor.shared_randomstreams import RandomStreams
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	11 from theano import shared, function
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	12
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	13 import gzip
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	14 import cPickle
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	15 import pylearn.io.image_tiling
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	16 import PIL
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	17
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	18 # NNET STUFF
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	19
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	20 class LogisticRegression(object):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	21 """Multi-class Logistic Regression Class
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	22
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	23 The logistic regression is fully described by a weight matrix :math:`W`
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	24 and bias vector :math:`b`. Classification is done by projecting data
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	25 points onto a set of hyperplanes, the distance to which is used to
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	26 determine a class membership probability.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	27 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	28
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	29 def __init__(self, input, n_in, n_out):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	30 """ Initialize the parameters of the logistic regression
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	31 :param input: symbolic variable that describes the input of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	32 architecture (one minibatch)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	33 :type n_in: int
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	34 :param n_in: number of input units, the dimension of the space in
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	35 which the datapoints lie
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	36 :type n_out: int
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	37 :param n_out: number of output units, the dimension of the space in
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	38 which the labels lie
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	39 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	40
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	41 # initialize with 0 the weights W as a matrix of shape (n_in, n_out)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	42 self.W = theano.shared( value=numpy.zeros((n_in,n_out),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	43 dtype = theano.config.floatX) )
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	44 # initialize the baises b as a vector of n_out 0s
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	45 self.b = theano.shared( value=numpy.zeros((n_out,),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	46 dtype = theano.config.floatX) )
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	47 # compute vector of class-membership probabilities in symbolic form
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	48 self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W)+self.b)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	49
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	50 # compute prediction as class whose probability is maximal in
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	51 # symbolic form
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	52 self.y_pred=T.argmax(self.p_y_given_x, axis=1)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	53
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	54 # list of parameters for this layer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	55 self.params = [self.W, self.b]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	56
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	57 def negative_log_likelihood(self, y):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	58 """Return the mean of the negative log-likelihood of the prediction
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	59 of this model under a given target distribution.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	60 :param y: corresponds to a vector that gives for each example the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	61 correct label
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	62 Note: we use the mean instead of the sum so that
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	63 the learning rate is less dependent on the batch size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	64 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	65 return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y])
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	66
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	67 def errors(self, y):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	68 """Return a float representing the number of errors in the minibatch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	69 over the total number of examples of the minibatch ; zero one
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	70 loss over the size of the minibatch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	71 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	72 # check if y has same dimension of y_pred
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	73 if y.ndim != self.y_pred.ndim:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	74 raise TypeError('y should have the same shape as self.y_pred',
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	75 ('y', target.type, 'y_pred', self.y_pred.type))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	76
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	77 # check if y is of the correct datatype
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	78 if y.dtype.startswith('int'):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	79 # the T.neq operator returns a vector of 0s and 1s, where 1
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	80 # represents a mistake in prediction
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	81 return T.mean(T.neq(self.y_pred, y))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	82 else:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	83 raise NotImplementedError()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	84
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	85 class SigmoidalLayer(object):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	86 def __init__(self, rng, input, n_in, n_out):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	87 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	88 Typical hidden layer of a MLP: units are fully-connected and have
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	89 sigmoidal activation function. Weight matrix W is of shape (n_in,n_out)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	90 and the bias vector b is of shape (n_out,).
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	91
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	92 Hidden unit activation is given by: sigmoid(dot(input,W) + b)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	93
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	94 :type rng: numpy.random.RandomState
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	95 :param rng: a random number generator used to initialize weights
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	96 :type input: theano.tensor.matrix
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	97 :param input: a symbolic tensor of shape (n_examples, n_in)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	98 :type n_in: int
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	99 :param n_in: dimensionality of input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	100 :type n_out: int
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	101 :param n_out: number of hidden units
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	102 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	103 self.input = input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	104
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	105 W_values = numpy.asarray( rng.uniform( \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	106 low = -numpy.sqrt(6./(n_in+n_out)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	107 high = numpy.sqrt(6./(n_in+n_out)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	108 size = (n_in, n_out)), dtype = theano.config.floatX)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	109 self.W = theano.shared(value = W_values)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	110
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	111 b_values = numpy.zeros((n_out,), dtype= theano.config.floatX)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	112 self.b = theano.shared(value= b_values)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	113
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	114 self.output = T.nnet.sigmoid(T.dot(input, self.W) + self.b)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	115 self.params = [self.W, self.b]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	116
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	117 # PRETRAINING LAYERS
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	118
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	119 class RBM(object):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	120 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	121 *** WRITE THE ENERGY FUNCTION USE SAME LETTERS AS VARIABLE NAMES IN CODE
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	122 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	123
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	124 def __init__(self, input=None, n_visible=None, n_hidden=None,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	125 W=None, hbias=None, vbias=None,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	126 numpy_rng=None, theano_rng=None):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	127 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	128 RBM constructor. Defines the parameters of the model along with
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	129 basic operations for inferring hidden from visible (and vice-versa),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	130 as well as for performing CD updates.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	131
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	132 :param input: None for standalone RBMs or symbolic variable if RBM is
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	133 part of a larger graph.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	134
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	135 :param n_visible: number of visible units (necessary when W or vbias is None)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	136
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	137 :param n_hidden: number of hidden units (necessary when W or hbias is None)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	138
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	139 :param W: weights to use for the RBM. None means that a shared variable will be
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	140 created with a randomly chosen matrix of size (n_visible, n_hidden).
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	141
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	142 :param hbias: ***
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	143
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	144 :param vbias: ***
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	145
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	146 :param numpy_rng: random number generator (necessary when W is None)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	147
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	148 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	149
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	150 params = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	151 if W is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	152 # choose initial values for weight matrix of RBM
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	153 initial_W = numpy.asarray(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	154 numpy_rng.uniform( \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	155 low=-numpy.sqrt(6./(n_hidden+n_visible)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	156 high=numpy.sqrt(6./(n_hidden+n_visible)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	157 size=(n_visible, n_hidden)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	158 dtype=theano.config.floatX)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	159 W = theano.shared(value=initial_W, name='W')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	160 params.append(W)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	161
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	162 if hbias is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	163 # theano shared variables for hidden biases
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	164 hbias = theano.shared(value=numpy.zeros(n_hidden,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	165 dtype=theano.config.floatX), name='hbias')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	166 params.append(hbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	167
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	168 if vbias is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	169 # theano shared variables for visible biases
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	170 vbias = theano.shared(value=numpy.zeros(n_visible,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	171 dtype=theano.config.floatX), name='vbias')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	172 params.append(vbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	173
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	174 if input is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	175 # initialize input layer for standalone RBM or layer0 of DBN
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	176 input = T.matrix('input')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	177
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	178 # setup theano random number generator
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	179 if theano_rng is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	180 theano_rng = RandomStreams(numpy_rng.randint(2**30))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	181
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	182 self.visible = self.input = input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	183 self.W = W
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	184 self.hbias = hbias
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	185 self.vbias = vbias
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	186 self.theano_rng = theano_rng
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	187 self.params = params
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	188 self.hidden_mean = T.nnet.sigmoid(T.dot(input, W)+hbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	189 self.hidden_sample = theano_rng.binomial(self.hidden_mean.shape, 1, self.hidden_mean)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	190
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	191 def gibbs_k(self, v_sample, k):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	192 ''' This function implements k steps of Gibbs sampling '''
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	193
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	194 # We compute the visible after k steps of Gibbs by iterating
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	195 # over ``gibs_1`` for k times; this can be done in Theano using
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	196 # the `scan op`. For a more comprehensive description of scan see
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	197 # http://deeplearning.net/software/theano/library/scan.html .
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	198
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	199 def gibbs_1(v0_sample, W, hbias, vbias):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	200 ''' This function implements one Gibbs step '''
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	201
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	202 # compute the activation of the hidden units given a sample of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	203 # vissibles
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	204 h0_mean = T.nnet.sigmoid(T.dot(v0_sample, W) + hbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	205 # get a sample of the hiddens given their activation
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	206 h0_sample = self.theano_rng.binomial(h0_mean.shape, 1, h0_mean)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	207 # compute the activation of the visible given the hidden sample
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	208 v1_mean = T.nnet.sigmoid(T.dot(h0_sample, W.T) + vbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	209 # get a sample of the visible given their activation
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	210 v1_act = self.theano_rng.binomial(v1_mean.shape, 1, v1_mean)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	211 return [v1_mean, v1_act]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	212
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	213
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	214 # DEBUGGING TO DO ALL WITHOUT SCAN
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	215 if k == 1:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	216 return gibbs_1(v_sample, self.W, self.hbias, self.vbias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	217
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	218
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	219 # Because we require as output two values, namely the mean field
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	220 # approximation of the visible and the sample obtained after k steps,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	221 # scan needs to know the shape of those two outputs. Scan takes
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	222 # this information from the variables containing the initial state
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	223 # of the outputs. Since we do not need a initial state of ``v_mean``
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	224 # we provide a dummy one used only to get the correct shape
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	225 v_mean = T.zeros_like(v_sample)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	226
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	227 # ``outputs_taps`` is an argument of scan which describes at each
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	228 # time step what past values of the outputs the function applied
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	229 # recursively needs. This is given in the form of a dictionary,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	230 # where the keys are outputs indexes, and values are a list of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	231 # of the offsets used by the corresponding outputs
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	232 # In our case the function ``gibbs_1`` applied recursively, requires
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	233 # at time k the past value k-1 for the first output (index 0) and
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	234 # no past value of the second output
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	235 outputs_taps = { 0 : [-1], 1 : [] }
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	236
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	237 v_means, v_samples = theano.scan( fn = gibbs_1,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	238 sequences = [],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	239 initial_states = [v_sample, v_mean],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	240 non_sequences = [self.W, self.hbias, self.vbias],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	241 outputs_taps = outputs_taps,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	242 n_steps = k)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	243 return v_means[-1], v_samples[-1]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	244
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	245 def free_energy(self, v_sample):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	246 wx_b = T.dot(v_sample, self.W) + self.hbias
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	247 vbias_term = T.sum(T.dot(v_sample, self.vbias))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	248 hidden_term = T.sum(T.log(1+T.exp(wx_b)))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	249 return -hidden_term - vbias_term
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	250
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	251 def cd(self, visible = None, persistent = None, steps = 1):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	252 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	253 Return a 5-tuple of values related to contrastive divergence: (cost,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	254 end-state of negative-phase chain, gradient on weights, gradient on
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	255 hidden bias, gradient on visible bias)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	256
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	257 If visible is None, it defaults to self.input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	258 If persistent is None, it defaults to self.input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	259
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	260 CD aka CD1 - cd()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	261 CD-10 - cd(steps=10)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	262 PCD - cd(persistent=shared(numpy.asarray(initializer)))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	263 PCD-k - cd(persistent=shared(numpy.asarray(initializer)),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	264 steps=10)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	265 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	266 if visible is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	267 visible = self.input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	268
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	269 if visible is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	270 raise TypeError('visible argument is required when self.input is None')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	271
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	272 if steps is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	273 steps = self.gibbs_1
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	274
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	275 if persistent is None:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	276 chain_start = visible
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	277 else:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	278 chain_start = persistent
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	279
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	280 chain_end_mean, chain_end_sample = self.gibbs_k(chain_start, steps)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	281
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	282 #print >> sys.stderr, "WARNING: DEBUGGING with wrong FREE ENERGY"
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	283 #free_energy_delta = - self.free_energy(chain_end_sample)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	284 free_energy_delta = self.free_energy(visible) - self.free_energy(chain_end_sample)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	285
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	286 # we will return all of these regardless of what is in self.params
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	287 all_params = [self.W, self.hbias, self.vbias]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	288
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	289 gparams = T.grad(free_energy_delta, all_params,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	290 consider_constant = [chain_end_sample])
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	291
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	292 cross_entropy = T.mean(T.sum(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	293 visibleT.log(chain_end_mean) + (1 - visible)T.log(1-chain_end_mean),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	294 axis = 1))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	295
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	296 return (cross_entropy, chain_end_sample,) + tuple(gparams)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	297
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	298 def cd_updates(self, lr, visible = None, persistent = None, steps = 1):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	299 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	300 Return the learning updates for the RBM parameters that are shared variables.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	301
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	302 Also returns an update for the persistent if it is a shared variable.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	303
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	304 These updates are returned as a dictionary.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	305
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	306 :param lr: [scalar] learning rate for contrastive divergence learning
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	307 :param visible: see `cd_grad`
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	308 :param persistent: see `cd_grad`
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	309 :param steps: see `cd_grad`
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	310
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	311 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	312
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	313 cross_entropy, chain_end, gW, ghbias, gvbias = self.cd(visible,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	314 persistent, steps)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	315
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	316 updates = {}
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	317 if hasattr(self.W, 'value'):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	318 updates[self.W] = self.W - lr * gW
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	319 if hasattr(self.hbias, 'value'):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	320 updates[self.hbias] = self.hbias - lr * ghbias
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	321 if hasattr(self.vbias, 'value'):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	322 updates[self.vbias] = self.vbias - lr * gvbias
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	323 if persistent:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	324 #if persistent is a shared var, then it means we should use
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	325 updates[persistent] = chain_end
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	326
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	327 return updates
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	328
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	329 # DEEP MODELS
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	330
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	331 class DBN(object):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	332 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	333 *** WHAT IS A DBN?
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	334 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	335
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	336 def __init__(self, input_len, hidden_layers_sizes, n_classes, rng):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	337 """ This class is made to support a variable number of layers.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	338
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	339 :param train_set_x: symbolic variable pointing to the training dataset
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	340
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	341 :param train_set_y: symbolic variable pointing to the labels of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	342 training dataset
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	343
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	344 :param input_len: dimension of the input to the sdA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	345
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	346 :param n_layers_sizes: intermidiate layers size, must contain
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	347 at least one value
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	348
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	349 :param n_classes: dimension of the output of the network
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	350
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	351 :param corruption_levels: amount of corruption to use for each
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	352 layer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	353
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	354 :param rng: numpy random number generator used to draw initial weights
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	355
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	356 :param pretrain_lr: learning rate used during pre-trainnig stage
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	357
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	358 :param finetune_lr: learning rate used during finetune stage
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	359 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	360
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	361 self.sigmoid_layers = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	362 self.rbm_layers = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	363 self.pretrain_functions = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	364 self.params = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	365
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	366 theano_rng = RandomStreams(rng.randint(2**30))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	367
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	368 # allocate symbolic variables for the data
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	369 index = T.lscalar() # index to a [mini]batch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	370 self.x = T.matrix('x') # the data is presented as rasterized images
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	371 self.y = T.ivector('y') # the labels are presented as 1D vector of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	372 # [int] labels
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	373 input = self.x
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	374
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	375 # The SdA is an MLP, for which all weights of intermidiate layers
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	376 # are shared with a different denoising autoencoders
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	377 # We will first construct the SdA as a deep multilayer perceptron,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	378 # and when constructing each sigmoidal layer we also construct a
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	379 # denoising autoencoder that shares weights with that layer, and
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	380 # compile a training function for that denoising autoencoder
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	381
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	382 for n_hid in hidden_layers_sizes:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	383 # construct the sigmoidal layer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	384
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	385 sigmoid_layer = SigmoidalLayer(rng, input, input_len, n_hid)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	386 self.sigmoid_layers.append(sigmoid_layer)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	387
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	388 self.rbm_layers.append(RBM(input=input,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	389 W=sigmoid_layer.W,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	390 hbias=sigmoid_layer.b,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	391 n_visible = input_len,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	392 n_hidden = n_hid,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	393 numpy_rng=rng,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	394 theano_rng=theano_rng))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	395
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	396 # its arguably a philosophical question...
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	397 # but we are going to only declare that the parameters of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	398 # sigmoid_layers are parameters of the StackedDAA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	399 # the hidden-layer biases in the daa_layers are parameters of those
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	400 # daa_layers, but not the StackedDAA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	401 self.params.extend(self.sigmoid_layers[-1].params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	402
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	403 # get ready for the next loop iteration
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	404 input_len = n_hid
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	405 input = self.sigmoid_layers[-1].output
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	406
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	407 # We now need to add a logistic layer on top of the MLP
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	408 self.logistic_regressor = LogisticRegression(input = input,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	409 n_in = input_len, n_out = n_classes)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	410
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	411 self.params.extend(self.logistic_regressor.params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	412
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	413 def pretraining_functions(self, train_set_x, batch_size, learning_rate, k=1):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	414 if k!=1:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	415 raise NotImplementedError()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	416 index = T.lscalar() # index to a [mini]batch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	417 n_train_batches = train_set_x.value.shape[0] / batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	418 batch_begin = (index % n_train_batches) * batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	419 batch_end = batch_begin+batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	420
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	421 print 'TRAIN_SET X', train_set_x.value.shape
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	422 rval = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	423 for rbm in self.rbm_layers:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	424 # N.B. these cd() samples are independent from the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	425 # samples used for learning
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	426 outputs = list(rbm.cd())[0:2]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	427 rval.append(function([index], outputs,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	428 updates = rbm.cd_updates(lr=learning_rate),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	429 givens = {self.x: train_set_x[batch_begin:batch_end]}))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	430 if rbm is self.rbm_layers[0]:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	431 f = rval[-1]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	432 AA=len(outputs)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	433 for i, implicit_out in enumerate(f.maker.env.outputs): #[len(outputs):]:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	434 print 'OUTPUT ', i
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	435 theano.printing.debugprint(implicit_out, file=sys.stdout)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	436
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	437 return rval
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	438
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	439 def finetune(self, datasets, lr, batch_size):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	440
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	441 # unpack the various datasets
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	442 (train_set_x, train_set_y) = datasets[0]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	443 (valid_set_x, valid_set_y) = datasets[1]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	444 (test_set_x, test_set_y) = datasets[2]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	445
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	446 # compute number of minibatches for training, validation and testing
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	447 assert train_set_x.value.shape[0] % batch_size == 0
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	448 assert valid_set_x.value.shape[0] % batch_size == 0
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	449 assert test_set_x.value.shape[0] % batch_size == 0
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	450 n_train_batches = train_set_x.value.shape[0] / batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	451 n_valid_batches = valid_set_x.value.shape[0] / batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	452 n_test_batches = test_set_x.value.shape[0] / batch_size
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	453
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	454 index = T.lscalar() # index to a [mini]batch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	455 target = self.y
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	456
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	457 train_index = index % n_train_batches
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	458
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	459 classifier = self.logistic_regressor
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	460 cost = classifier.negative_log_likelihood(target)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	461 # compute the gradients with respect to the model parameters
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	462 gparams = T.grad(cost, self.params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	463
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	464 # compute list of fine-tuning updates
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	465 updates = [(param, param - gparam*finetune_lr)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	466 for param,gparam in zip(self.params, gparams)]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	467
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	468 train_fn = theano.function([index], cost,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	469 updates = updates,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	470 givens = {
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	471 self.x : train_set_x[train_indexbatch_size:(train_index+1)batch_size],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	472 target : train_set_y[train_indexbatch_size:(train_index+1)batch_size]})
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	473
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	474 test_score_i = theano.function([index], classifier.errors(target),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	475 givens = {
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	476 self.x: test_set_x[indexbatch_size:(index+1)batch_size],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	477 target: test_set_y[indexbatch_size:(index+1)batch_size]})
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	478
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	479 valid_score_i = theano.function([index], classifier.errors(target),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	480 givens = {
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	481 self.x: valid_set_x[indexbatch_size:(index+1)batch_size],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	482 target: valid_set_y[indexbatch_size:(index+1)batch_size]})
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	483
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	484 def test_scores():
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	485 return [test_score_i(i) for i in xrange(n_test_batches)]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	486
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	487 def valid_scores():
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	488 return [valid_score_i(i) for i in xrange(n_valid_batches)]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	489
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	490 return train_fn, valid_scores, test_scores
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	491
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	492 def load_mnist(filename):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	493 f = gzip.open(filename,'rb')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	494 train_set, valid_set, test_set = cPickle.load(f)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	495 f.close()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	496
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	497 def shared_dataset(data_xy):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	498 data_x, data_y = data_xy
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	499 shared_x = theano.shared(numpy.asarray(data_x, dtype=theano.config.floatX))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	500 shared_y = theano.shared(numpy.asarray(data_y, dtype=theano.config.floatX))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	501 return shared_x, T.cast(shared_y, 'int32')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	502
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	503 n_train_examples = train_set[0].shape[0]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	504 datasets = shared_dataset(train_set), shared_dataset(valid_set), shared_dataset(test_set)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	505
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	506 return n_train_examples, datasets
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	507
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	508 def dbn_main(finetune_lr = 0.01,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	509 pretraining_epochs = 10,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	510 pretrain_lr = 0.1,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	511 training_epochs = 1000,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	512 batch_size = 20,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	513 mnist_file='mnist.pkl.gz'):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	514 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	515 Demonstrate stochastic gradient descent optimization for a multilayer perceptron
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	516
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	517 This is demonstrated on MNIST.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	518
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	519 :param learning_rate: learning rate used in the finetune stage
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	520 (factor for the stochastic gradient)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	521
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	522 :param pretraining_epochs: number of epoch to do pretraining
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	523
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	524 :param pretrain_lr: learning rate to be used during pre-training
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	525
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	526 :param n_iter: maximal number of iterations ot run the optimizer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	527
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	528 :param mnist_file: path the the pickled mnist_file
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	529
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	530 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	531
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	532 n_train_examples, train_valid_test = load_mnist(mnist_file)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	533
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	534 print "Creating a Deep Belief Network"
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	535 deep_model = DBN(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	536 input_len=28*28,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	537 hidden_layers_sizes = [500, 150, 100],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	538 n_classes=10,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	539 rng = numpy.random.RandomState())
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	540
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	541 ####
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	542 #### Phase 1: Pre-training
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	543 ####
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	544 print "Pretraining (unsupervised learning) ..."
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	545
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	546 pretrain_functions = deep_model.pretraining_functions(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	547 batch_size=batch_size,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	548 train_set_x=train_valid_test[0][0],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	549 learning_rate=pretrain_lr,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	550 )
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	551
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	552 start_time = time.clock()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	553 for layer_idx, pretrain_fn in enumerate(pretrain_functions):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	554 # go through pretraining epochs
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	555 print 'Pre-training layer %i'% layer_idx
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	556 for i in xrange(pretraining_epochs * n_train_examples / batch_size):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	557 outstuff = pretrain_fn(i)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	558 xe, negsample = outstuff[:2]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	559 print (layer_idx, i,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	560 n_train_examples / batch_size,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	561 float(xe),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	562 'Wmin', deep_model.rbm_layers[0].W.value.min(),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	563 'Wmax', deep_model.rbm_layers[0].W.value.max(),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	564 'vmin', deep_model.rbm_layers[0].vbias.value.min(),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	565 'vmax', deep_model.rbm_layers[0].vbias.value.max(),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	566 #'x>0.3', (input_i>0.3).sum(),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	567 )
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	568 sys.stdout.flush()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	569 if i % 1000 == 0:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	570 PIL.Image.fromarray(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	571 pylearn.io.image_tiling.tile_raster_images(negsample, (28,28), (10,10),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	572 tile_spacing=(1,1))).save('samples_%i_%i.png'%(layer_idx,i))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	573
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	574 PIL.Image.fromarray(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	575 pylearn.io.image_tiling.tile_raster_images(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	576 deep_model.rbm_layers[0].W.value.T,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	577 (28,28), (10,10),
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	578 tile_spacing=(1,1))).save('filters_%i_%i.png'%(layer_idx,i))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	579 end_time = time.clock()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	580 print 'Pretraining took %f minutes' %((end_time - start_time)/60.)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	581
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	582 return
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	583
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	584 print "Fine tuning (supervised learning) ..."
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	585 train_fn, valid_scores, test_scores =\
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	586 deep_model.finetune_functions(train_valid_test[0][0],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	587 learning_rate=finetune_lr, # the learning rate
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	588 batch_size = batch_size) # number of examples to use at once
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	589
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	590 ####
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	591 #### Phase 2: Fine Tuning
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	592 ####
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	593
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	594 patience = 10000 # look as this many examples regardless
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	595 patience_increase = 2. # wait this much longer when a new best is
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	596 # found
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	597 improvement_threshold = 0.995 # a relative improvement of this much is
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	598 # considered significant
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	599 validation_frequency = min(n_train_examples, patience/2)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	600 # go through this many
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	601 # minibatche before checking the network
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	602 # on the validation set; in this case we
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	603 # check every epoch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	604
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	605 patience_max = n_train_examples * training_epochs
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	606
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	607 best_epoch = None
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	608 best_epoch_test_score = None
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	609 best_epoch_valid_score = float('inf')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	610 start_time = time.clock()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	611
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	612 for i in xrange(patience_max):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	613 if i >= patience:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	614 break
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	615
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	616 cost_i = train_fn(i)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	617
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	618 if i % validation_frequency == 0:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	619 validation_i = numpy.mean([score for score in valid_scores()])
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	620
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	621 # if we got the best validation score until now
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	622 if validation_i < best_epoch_valid_score:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	623
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	624 # improve patience if loss improvement is good enough
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	625 threshold_i = best_epoch_valid_score * improvement_threshold
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	626 if validation_i < threshold_i:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	627 patience = max(patience, i * patience_increase)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	628
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	629 # save best validation score and iteration number
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	630 best_epoch_valid_score = validation_i
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	631 best_epoch = i/validation_i
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	632 best_epoch_test_score = numpy.mean(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	633 [score for score in test_scores()])
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	634
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	635 print('epoch %i, validation error %f %%, test error %f %%'%(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	636 i/validation_frequency, validation_i*100.,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	637 best_epoch_test_score*100.))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	638 else:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	639 print('epoch %i, validation error %f %%' % (
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	640 i/validation_frequency, validation_i*100.))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	641 end_time = time.clock()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	642
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	643 print(('Optimization complete with best validation score of %f %%,'
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	644 'with test performance %f %%') %
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	645 (finetune_status['best_validation_loss']*100.,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	646 finetune_status['test_score']*100.))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	647 print ('The code ran for %f minutes' % ((finetune_status['duration'])/60.))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	648
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	649 def rbm_main():
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	650 rbm = RBM(n_visible=20, n_hidden=30,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	651 numpy_rng = numpy.random.RandomState(34))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	652
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	653 cd_updates = rbm.cd_updates(lr=0.25)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	654
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	655 print cd_updates
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	656
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	657 f = function([rbm.input], [],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	658 updates={rbm.W:cd_updates[rbm.W]})
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	659
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	660 theano.printing.debugprint(f.maker.env.outputs[0],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	661 file=sys.stdout)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	662
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	663
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	664 if __name__ == '__main__':
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	665 dbn_main()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	666 #rbm_main()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	667
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	668
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	669 if 0:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	670 class DAA(object):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	671 def __init__(self, n_visible= 784, n_hidden= 500, corruption_level = 0.1,\
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	672 input = None, shared_W = None, shared_b = None):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	673 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	674 Initialize the dA class by specifying the number of visible units (the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	675 dimension d of the input ), the number of hidden units ( the dimension
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	676 d' of the latent or hidden space ) and the corruption level. The
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	677 constructor also receives symbolic variables for the input, weights and
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	678 bias. Such a symbolic variables are useful when, for example the input is
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	679 the result of some computations, or when weights are shared between the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	680 dA and an MLP layer. When dealing with SdAs this always happens,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	681 the dA on layer 2 gets as input the output of the dA on layer 1,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	682 and the weights of the dA are used in the second stage of training
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	683 to construct an MLP.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	684
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	685 :param n_visible: number of visible units
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	686
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	687 :param n_hidden: number of hidden units
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	688
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	689 :param input: a symbolic description of the input or None
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	690
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	691 :param corruption_level: the corruption mechanism picks up randomly this
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	692 fraction of entries of the input and turns them to 0
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	693
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	694
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	695 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	696 self.n_visible = n_visible
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	697 self.n_hidden = n_hidden
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	698
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	699 # create a Theano random generator that gives symbolic random values
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	700 theano_rng = RandomStreams()
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	701
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	702 if shared_W != None and shared_b != None :
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	703 self.W = shared_W
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	704 self.b = shared_b
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	705 else:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	706 # initial values for weights and biases
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	707 # note : W' was written as `W_prime` and b' as `b_prime`
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	708
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	709 # W is initialized with `initial_W` which is uniformely sampled
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	710 # from -6./sqrt(n_visible+n_hidden) and 6./sqrt(n_hidden+n_visible)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	711 # the output of uniform if converted using asarray to dtype
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	712 # theano.config.floatX so that the code is runable on GPU
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	713 initial_W = numpy.asarray( numpy.random.uniform( \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	714 low = -numpy.sqrt(6./(n_hidden+n_visible)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	715 high = numpy.sqrt(6./(n_hidden+n_visible)), \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	716 size = (n_visible, n_hidden)), dtype = theano.config.floatX)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	717 initial_b = numpy.zeros(n_hidden, dtype = theano.config.floatX)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	718
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	719
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	720 # theano shared variables for weights and biases
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	721 self.W = theano.shared(value = initial_W, name = "W")
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	722 self.b = theano.shared(value = initial_b, name = "b")
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	723
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	724
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	725 initial_b_prime= numpy.zeros(n_visible)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	726 # tied weights, therefore W_prime is W transpose
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	727 self.W_prime = self.W.T
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	728 self.b_prime = theano.shared(value = initial_b_prime, name = "b'")
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	729
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	730 # if no input is given, generate a variable representing the input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	731 if input == None :
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	732 # we use a matrix because we expect a minibatch of several examples,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	733 # each example being a row
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	734 self.x = T.matrix(name = 'input')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	735 else:
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	736 self.x = input
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	737 # Equation (1)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	738 # keep 90% of the inputs the same and zero-out randomly selected subset of 10% of the inputs
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	739 # note : first argument of theano.rng.binomial is the shape(size) of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	740 # random numbers that it should produce
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	741 # second argument is the number of trials
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	742 # third argument is the probability of success of any trial
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	743 #
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	744 # this will produce an array of 0s and 1s where 1 has a
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	745 # probability of 1 - ``corruption_level`` and 0 with
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	746 # ``corruption_level``
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	747 self.tilde_x = theano_rng.binomial( self.x.shape, 1, 1 - corruption_level) * self.x
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	748 # Equation (2)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	749 # note : y is stored as an attribute of the class so that it can be
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	750 # used later when stacking dAs.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	751 self.y = T.nnet.sigmoid(T.dot(self.tilde_x, self.W ) + self.b)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	752 # Equation (3)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	753 self.z = T.nnet.sigmoid(T.dot(self.y, self.W_prime) + self.b_prime)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	754 # Equation (4)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	755 # note : we sum over the size of a datapoint; if we are using minibatches,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	756 # L will be a vector, with one entry per example in minibatch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	757 self.L = - T.sum( self.xT.log(self.z) + (1-self.x)T.log(1-self.z), axis=1 )
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	758 # note : L is now a vector, where each element is the cross-entropy cost
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	759 # of the reconstruction of the corresponding example of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	760 # minibatch. We need to compute the average of all these to get
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	761 # the cost of the minibatch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	762 self.cost = T.mean(self.L)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	763
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	764 self.params = [ self.W, self.b, self.b_prime ]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	765
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	766 class StackedDAA(DeepLayerwiseModel):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	767 """Stacked denoising auto-encoder class (SdA)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	768
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	769 A stacked denoising autoencoder model is obtained by stacking several
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	770 dAs. The hidden layer of the dA at layer `i` becomes the input of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	771 the dA at layer `i+1`. The first layer dA gets as input the input of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	772 the SdA, and the hidden layer of the last dA represents the output.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	773 Note that after pretraining, the SdA is dealt with as a normal MLP,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	774 the dAs are only used to initialize the weights.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	775 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	776
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	777 def __init__(self, n_ins, hidden_layers_sizes, n_outs,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	778 corruption_levels, rng, ):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	779 """ This class is made to support a variable number of layers.
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	780
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	781 :param train_set_x: symbolic variable pointing to the training dataset
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	782
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	783 :param train_set_y: symbolic variable pointing to the labels of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	784 training dataset
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	785
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	786 :param n_ins: dimension of the input to the sdA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	787
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	788 :param n_layers_sizes: intermidiate layers size, must contain
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	789 at least one value
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	790
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	791 :param n_outs: dimension of the output of the network
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	792
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	793 :param corruption_levels: amount of corruption to use for each
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	794 layer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	795
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	796 :param rng: numpy random number generator used to draw initial weights
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	797
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	798 :param pretrain_lr: learning rate used during pre-trainnig stage
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	799
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	800 :param finetune_lr: learning rate used during finetune stage
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	801 """
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	802
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	803 self.sigmoid_layers = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	804 self.daa_layers = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	805 self.pretrain_functions = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	806 self.params = []
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	807 self.n_layers = len(hidden_layers_sizes)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	808
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	809 if len(hidden_layers_sizes) < 1 :
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	810 raiseException (' You must have at least one hidden layer ')
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	811
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	812 theano_rng = RandomStreams(rng.randint(2**30))
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	813
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	814 # allocate symbolic variables for the data
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	815 index = T.lscalar() # index to a [mini]batch
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	816 self.x = T.matrix('x') # the data is presented as rasterized images
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	817 self.y = T.ivector('y') # the labels are presented as 1D vector of
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	818 # [int] labels
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	819
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	820 # The SdA is an MLP, for which all weights of intermidiate layers
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	821 # are shared with a different denoising autoencoders
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	822 # We will first construct the SdA as a deep multilayer perceptron,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	823 # and when constructing each sigmoidal layer we also construct a
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	824 # denoising autoencoder that shares weights with that layer, and
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	825 # compile a training function for that denoising autoencoder
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	826
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	827 for i in xrange( self.n_layers ):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	828 # construct the sigmoidal layer
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	829
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	830 sigmoid_layer = SigmoidalLayer(rng,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	831 self.layers[-1].output if i else self.x,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	832 hidden_layers_sizes[i-1] if i else n_ins,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	833 hidden_layers_sizes[i])
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	834
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	835 daa_layer = DAA(corruption_level = corruption_levels[i],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	836 input = sigmoid_layer.input,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	837 W = sigmoid_layer.W,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	838 b = sigmoid_layer.b)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	839
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	840 # add the layer to the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	841 self.sigmoid_layers.append(sigmoid_layer)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	842 self.daa_layers.append(daa_layer)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	843
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	844 # its arguably a philosophical question...
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	845 # but we are going to only declare that the parameters of the
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	846 # sigmoid_layers are parameters of the StackedDAA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	847 # the hidden-layer biases in the daa_layers are parameters of those
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	848 # daa_layers, but not the StackedDAA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	849 self.params.extend(sigmoid_layer.params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	850
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	851 # We now need to add a logistic layer on top of the MLP
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	852 self.logistic_regressor = LogisticRegression(
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	853 input = self.sigmoid_layers[-1].output,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	854 n_in = hidden_layers_sizes[-1],
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	855 n_out = n_outs)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	856
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	857 self.params.extend(self.logLayer.params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	858
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	859 def pretraining_functions(self, train_set_x, batch_size):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	860
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	861 # compiles update functions for each layer, and
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	862 # returns them as a list
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	863 #
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	864 # Construct a function that trains this dA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	865 # compute gradients of layer parameters
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	866 gparams = T.grad(dA_layer.cost, dA_layer.params)
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	867 # compute the list of updates
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	868 updates = {}
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	869 for param, gparam in zip(dA_layer.params, gparams):
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	870 updates[param] = param - gparam * pretrain_lr
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	871
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	872 # create a function that trains the dA
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	873 update_fn = theano.function([index], dA_layer.cost, \
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	874 updates = updates,
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	875 givens = {
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	876 self.x : train_set_x[indexbatch_size:(index+1)batch_size]})
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	877 # collect this function into a list
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	878 self.pretrain_functions += [update_fn]
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	879
4bc5eeec6394 Updating the tutorial code to the latest revisions. Dumitru Erhan <dumitru.erhan@gmail.com> parents: diff changeset	880

Mercurial > ift6266

annotate code_tutoriel/deep.py @ 489:ee9836baade3