pylearn: pylearn/algorithms/mcRBM.py annotate

annotate pylearn/algorithms/mcRBM.py @ 978:ab4bc97ca060

mcRBM - particles initialized w randn instead of rand()

author	James Bergstra <bergstrj@iro.umontreal.ca>
date	Mon, 23 Aug 2010 16:04:31 -0400
parents	9cac1ecaeef7
children	2a53384d9742

rev	line source
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	1 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	2 This file implements the Mean & Covariance RBM discussed in
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	3
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	4 Ranzato, M. and Hinton, G. E. (2010)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	5 Modeling pixel means and covariances using factored third-order Boltzmann machines.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	6 IEEE Conference on Computer Vision and Pattern Recognition.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	7
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	8 and performs one of the experiments on CIFAR-10 discussed in that paper.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	9
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	10
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	11 Math
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	12 ====
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	13
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	14 Energy of "covariance RBM"
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	15
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	16 E = -0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i C_{if} v_i )^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	17 = -0.5 \sum_f (\sum_k P_{fk} h_k) ( \sum_i C_{if} v_i )^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	18 "vector element f" "vector element f"
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	19
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	20 In some parts of the paper, the P matrix is chosen to be a diagonal matrix with non-positive
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	21 diagonal entries, so it is helpful to see this as a simpler equation:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	22
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	23 E = \sum_f h_f ( \sum_i C_{if} v_i )^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	24
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	25
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	26
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	27 Full Energy of mean and Covariance RBM, with
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	28 :math:`h_k = h_k^{(c)}`,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	29 :math:`g_j = h_j^{(m)}`,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	30 :math:`b_k = b_k^{(c)}`,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	31 :math:`c_j = b_j^{(m)}`,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	32 :math:`U_{if} = C_{if}`,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	33
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	34 :
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	35
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	36 E (v, h, g) =
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	37 - 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / \|U_{*f}\|^2 \|v\|^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	38 - \sum_k b_k h_k
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	39 + 0.5 \sum_i v_i^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	40 - \sum_j \sum_i W_{ij} g_j v_i
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	41 - \sum_j c_j g_j
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	42
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	43 For the energy function to correspond to a probability distribution, P must be non-positive.
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	44
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	45
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	46 Conventions in this file
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	47 ========================
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	48
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	49 This file contains some global functions, as well as a class (MeanCovRBM) that makes using them a little
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	50 more convenient.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	51
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	52
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	53 Global functions like `free_energy` work on an mcRBM as parametrized in a particular way.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	54 Suppose we have
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	55 I input dimensions,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	56 F squared filters,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	57 J mean variables, and
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	58 K covariance variables.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	59 The mcRBM is parametrized by 5 variables:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	60
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	61 - `P`, a matrix (probably sparse) of pooling (F x K)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	62 - `U`, a matrix whose rows are visible covariance directions (I x F)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	63 - `W`, a matrix whose rows are visible mean directions (I x J)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	64 - `b`, a vector of hidden covariance biases (K)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	65 - `c`, a vector of hidden mean biases (J)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	66
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	67 Matrices are generally layed out according to a C-order convention.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	68
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	69 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	70
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	71 # Free energy is the marginal energy of visible units
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	72 # Recall:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	73 # Q(x) = exp(-E(x))/Z ==> -log(Q(x)) - log(Z) = E(x)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	74 #
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	75 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	76 # E (v, h, g) =
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	77 # - 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / \|U_{*f}\|^2 \|v\|^2
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	78 # - \sum_k b_k h_k
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	79 # + 0.5 \sum_i v_i^2
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	80 # - \sum_j \sum_i W_{ij} g_j v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	81 # - \sum_j c_j g_j
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	82 # - \sum_i a_i v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	83 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	84 #
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	85 # Derivation, in which partition functions are ignored.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	86 #
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	87 # E(v) = -\log(Q(v))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	88 # = -\log( \sum_{h,g} Q(v,h,g))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	89 # = -\log( \sum_{h,g} exp(-E(v,h,g)))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	90 # = -\log( \sum_{h,g} exp(-
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	91 # - 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / (\|U_{f}\| \|v\|)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	92 # - \sum_k b_k h_k
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	93 # + 0.5 \sum_i v_i^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	94 # - \sum_j \sum_i W_{ij} g_j v_i
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	95 # - \sum_j c_j g_j
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	96 # - \sum_i a_i v_i ))
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	97 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	98 # Get rid of double negs in exp
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	99 # = -\log( \sum_{h} exp(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	100 # + 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / (\|U_{f}\| \|v\|)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	101 # + \sum_k b_k h_k
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	102 # - 0.5 \sum_i v_i^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	103 # ) * \sum_{g} exp(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	104 # + \sum_j \sum_i W_{ij} g_j v_i
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	105 # + \sum_j c_j g_j))
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	106 # - \sum_i a_i v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	107 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	108 # Break up log
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	109 # = -\log( \sum_{h} exp(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	110 # + 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / (\|U_{f}\|\|v\|)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	111 # + \sum_k b_k h_k
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	112 # ))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	113 # -\log( \sum_{g} exp(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	114 # + \sum_j \sum_i W_{ij} g_j v_i
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	115 # + \sum_j c_j g_j )))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	116 # + 0.5 \sum_i v_i^2
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	117 # - \sum_i a_i v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	118 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	119 # Use domain h is binary to turn log(sum(exp(sum...))) into sum(log(..
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	120 # = -\log(\sum_{h} exp(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	121 # + 0.5 \sum_f \sum_k P_{fk} h_k ( \sum_i U_{if} v_i )^2 / (\|U_{f}\| \|v\|)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	122 # + \sum_k b_k h_k
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	123 # ))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	124 # - \sum_{j} \log(1 + exp(\sum_i W_{ij} v_i + c_j ))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	125 # + 0.5 \sum_i v_i^2
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	126 # - \sum_i a_i v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	127 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	128 # = - \sum_{k} \log(1 + exp(b_k + 0.5 \sum_f P_{fk}( \sum_i U_{if} v_i )^2 / (\|U_{f}\|\|v\|)))
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	129 # - \sum_{j} \log(1 + exp(\sum_i W_{ij} v_i + c_j ))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	130 # + 0.5 \sum_i v_i^2
972 0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	131 # - \sum_i a_i v_i
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	132 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	133 # For negative-one-diagonal P this gives:
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	134 #
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	135 # = - \sum_{k} \log(1 + exp(b_k - 0.5 \sum_i (U_{ik} v_i )^2 / (\|U_{k}\|\|v\|)))
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	136 # - \sum_{j} \log(1 + exp(\sum_i W_{ij} v_i + c_j ))
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	137 # + 0.5 \sum_i v_i^2
0b392d1401c5 mcRBM - adding math and comments James Bergstra <bergstrj@iro.umontreal.ca> parents: 967 diff changeset	138 # - \sum_i a_i v_i
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	139
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	140 import sys
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	141 import logging
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	142 import numpy as np
973 aa201f357d7b mcRBM - added numpy import James Bergstra <bergstrj@iro.umontreal.ca> parents: 972 diff changeset	143 import numpy
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	144 from theano import function, shared, dot
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	145 from theano import tensor as TT
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	146 import theano.sparse #installs the sparse shared var handler
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	147 floatX = theano.config.floatX
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	148
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	149 from pylearn.sampling.hmc import HMC_sampler
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	150 from pylearn.io import image_tiling
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	151
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	152 from sparse_coding import numpy_project_onto_ball
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	153
973 aa201f357d7b mcRBM - added numpy import James Bergstra <bergstrj@iro.umontreal.ca> parents: 972 diff changeset	154 print >> sys.stderr, "mcRBM IS NOT READY YET"
aa201f357d7b mcRBM - added numpy import James Bergstra <bergstrj@iro.umontreal.ca> parents: 972 diff changeset	155
aa201f357d7b mcRBM - added numpy import James Bergstra <bergstrj@iro.umontreal.ca> parents: 972 diff changeset	156
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	157 #TODO: This should be in the nnet part of the library
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	158 def sgd_updates(params, grads, lr):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	159 try:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	160 float(lr)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	161 lr = [lr for p in params]
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	162 except TypeError:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	163 pass
974 f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	164 updates = [(p, p - plr * gp) for (plr, p, gp) in zip(lr, params, grads)]
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	165 return updates
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	166
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	167 def as_shared(x, name=None, dtype=floatX):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	168 if hasattr(x, 'type'):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	169 return x
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	170 else:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	171 if 'float' in str(x.dtype):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	172 return shared(x.astype(floatX), name=name)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	173 else:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	174 return shared(x, name=name)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	175
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	176 def hidden_cov_units_preactivation_given_v(rbm, v, small=1e-8):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	177 (U,W,a,b,c) = rbm
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	178 unit_v = v / (TT.sqrt(TT.sum(v**2, axis=1))+small).dimshuffle(0,'x') # unit rows
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	179 unit_U = U # assuming unit cols!
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	180 #unit_U = U / (TT.sqrt(TT.sum(U**2, axis=0))+small) #unit cols
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	181 return b - 0.5 * dot(unit_v, unit_U)**2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	182
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	183 def free_energy_given_v(rbm, v):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	184 """Returns theano expression for free energy of visible vector `v` in an mcRBM
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	185
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	186 An mcRBM is parametrized
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	187 by `U`, `W`, `b`, `c`.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	188 See module - level documentation for explanations of the `U`, `W`, `b` and `c` parameters.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	189
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	190
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	191 The free energy of v is what we need for learning and hybrid Monte-carlo negative-phase
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	192 sampling.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	193
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	194 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	195 U, W, a, b, c = rbm
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	196
975 38e66e0da66a mcRBM - put softplus in directly for num. stability James Bergstra <bergstrj@iro.umontreal.ca> parents: 974 diff changeset	197 t0 = -TT.sum(TT.nnet.softplus(hidden_cov_units_preactivation_given_v(rbm, v)),axis=1)
38e66e0da66a mcRBM - put softplus in directly for num. stability James Bergstra <bergstrj@iro.umontreal.ca> parents: 974 diff changeset	198 t1 = -TT.sum(TT.nnet.softplus(c + dot(v,W)), axis=1)
38e66e0da66a mcRBM - put softplus in directly for num. stability James Bergstra <bergstrj@iro.umontreal.ca> parents: 974 diff changeset	199 t2 = 0.5 * TT.sum(v**2, axis=1)
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	200 t3 = -TT.dot(v, a)
976 4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	201 return t0 + t1 + t2 + t3, (t0, t1, t2, t3)
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	202
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	203 def expected_h_g_given_v(P, U, W, b, c, v):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	204 """Returns theano expression conditional expectations (`h`, `g`) in an mcRBM.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	205
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	206 An mcRBM is parametrized
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	207 by `U`, `W`, `b`, `c`.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	208 See module - level documentation for explanations of the `U`, `W`, `b` and `c` parameters.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	209
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	210
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	211 The conditional E[h, g \| v] is what we need to classify images.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	212 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	213 raise NotImplementedError()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	214
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	215 #TODO: check to see if these args should be negated?
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	216
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	217 if P is None:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	218 h = nnet.sigmoid(b + 0.5 * cosines(v,U))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	219 else:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	220 h = nnet.sigmoid(b + 0.5 * dot(cosines(v,U), P))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	221 g = nnet.sigmoid(c + dot(v,W))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	222 return (h, g)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	223
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	224 class MeanCovRBM(object):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	225 """Container for mcRBM parameters that gives more convenient access to mcRBM methods.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	226 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	227
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	228 params = property(lambda s: [s.U, s.W, s.a, s.b, s.c])
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	229
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	230 n_visible = property(lambda s: s.W.value.shape[0])
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	231
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	232 def __init__(self, U, W, a, b, c):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	233 self.U = as_shared(U, 'U')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	234 self.W = as_shared(W, 'W')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	235 self.a = as_shared(a, 'a')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	236 self.b = as_shared(b, 'b')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	237 self.c = as_shared(c, 'c')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	238
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	239 assert self.b.type.dtype == 'float32'
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	240
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	241 @classmethod
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	242 def new_from_dims(cls,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	243 n_I, # input dimensionality
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	244 n_K, # number of covariance hidden units
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	245 n_F, # number of covariance filters (squared)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	246 n_J, # number of mean filters (linear)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	247 seed = 8923402190,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	248 ):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	249 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	250 Return a MeanCovRBM instance with randomly-initialized parameters.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	251 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	252
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	253
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	254 if 0:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	255 if P_init == 'diag':
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	256 if n_K != n_F:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	257 raise ValueError('cannot use diagonal initialization of non-square P matrix')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	258 import scipy.sparse
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	259 P = -scipy.sparse.identity(n_K).tocsr()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	260 else:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	261 raise NotImplementedError()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	262
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	263 rng = np.random.RandomState(seed)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	264
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	265 # initialization taken from Marc'Aurelio
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	266
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	267 return cls(
977 9cac1ecaeef7 mcRBM - changed init of U to match M'A.R's code James Bergstra <bergstrj@iro.umontreal.ca> parents: 976 diff changeset	268 #U = numpy_project_onto_ball(rng.randn(n_I, n_F).T).T,
9cac1ecaeef7 mcRBM - changed init of U to match M'A.R's code James Bergstra <bergstrj@iro.umontreal.ca> parents: 976 diff changeset	269 U = 0.2 * rng.randn(n_I, n_F),
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	270 W = rng.randn(n_I, n_J)/np.sqrt((n_I+n_J)/2),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	271 a = np.ones(n_I)*(-2),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	272 b = np.ones(n_K)*2,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	273 c = np.zeros(n_J),)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	274
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	275 def __getstate__(self):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	276 # unpack shared containers, which may have references to Theano stuff
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	277 # and are not a long-term stable data type.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	278 return dict(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	279 U = self.U.value,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	280 W = self.W.value,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	281 b = self.b.value,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	282 c = self.c.value)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	283
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	284 def __setstate__(self, dct):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	285 self.__init__(**dct) # calls as_shared on pickled arrays
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	286
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	287 def hmc_sampler(self, n_particles=100, seed=7823748):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	288 return HMC_sampler(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	289 positions = [as_shared(
978 ab4bc97ca060 mcRBM - particles initialized w randn instead of rand() James Bergstra <bergstrj@iro.umontreal.ca> parents: 977 diff changeset	290 np.random.RandomState(seed^20893).randn(
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	291 n_particles,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	292 self.n_visible ))],
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	293 energy_fn = lambda p : self.free_energy_given_v(p[0]),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	294 seed=seed)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	295
976 4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	296 def free_energy_given_v(self, v, extra=False):
4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	297 rval = free_energy_given_v(self.params, v)
4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	298 if extra:
4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	299 return rval
4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	300 else:
4cbd65cf902d mcRBM - added extra free_energy param James Bergstra <bergstrj@iro.umontreal.ca> parents: 975 diff changeset	301 return rval[0]
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	302
974 f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	303 def contrastive_gradient(self, pos_v, neg_v, U_l1_penalty=0, W_l1_penalty=0):
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	304 """Return a list of gradient expressions for self.params
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	305
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	306 :param pos_v: positive-phase sample of visible units
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	307 :param neg_v: negative-phase sample of visible units
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	308 """
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	309 pos_FE = self.free_energy_given_v(pos_v)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	310 neg_FE = self.free_energy_given_v(neg_v)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	311
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	312 gpos_FE = theano.tensor.grad(pos_FE.sum(), self.params)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	313 gneg_FE = theano.tensor.grad(neg_FE.sum(), self.params)
974 f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	314 rval = [ gp - gn for (gp,gn) in zip(gpos_FE, gneg_FE)]
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	315 rval[0] = rval[0] - TT.sign(self.U)*U_l1_penalty
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	316 rval[1] = rval[1] - TT.sign(self.W)*W_l1_penalty
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	317 return rval
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	318
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	319 from pylearn.dataset_ops.protocol import TensorFnDataset
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	320 from pylearn.dataset_ops.memo import memo
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	321 import scipy.io
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	322 @memo
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	323 def load_mcRBM_demo_patches():
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	324 d = scipy.io.loadmat('/u/bergstrj/cvs/articles/2010/spike_slab_RBM/src/marcaurelio/training_colorpatches_16x16_demo.mat')
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	325 totnumcases = d["whitendata"].shape[0]
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	326 #d = d["whitendata"][0:np.floor(totnumcases/batch_size)*batch_size,:].copy()
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	327 d = d["whitendata"].copy()
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	328 return d
f2cdcc71ece1 mcRBM - added L1 penalties and normal sign convention to contrastive grad James Bergstra <bergstrj@iro.umontreal.ca> parents: 973 diff changeset	329
967 90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	330
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	331 if __name__ == '__main__':
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	332
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	333 print >> sys.stderr, "TODO: use P matrix (aka FH matrix)"
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	334
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	335 R,C= 8,8 # the size of image patches
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	336 l1_penalty=1e-3
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	337 no_l1_epochs = 10
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	338
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	339 epoch_size=50000
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	340 batchsize = 128
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	341 lr = 0.075 / batchsize
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	342 s_lr = TT.scalar()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	343 n_K=256
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	344 n_F=256
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	345 n_J=100
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	346
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	347 rbm = MeanCovRBM.new_from_dims(n_I=R*C,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	348 n_K=n_K,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	349 n_J=n_J,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	350 n_F=n_F,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	351 )
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	352
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	353 sampler = rbm.hmc_sampler(n_particles=100)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	354
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	355 from pylearn.dataset_ops import image_patches
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	356
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	357 batch_idx = TT.iscalar()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	358 train_batch = image_patches.image_patches(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	359 s_idx = (batch_idx * batchsize + np.arange(batchsize)),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	360 dims = (1000,R,C),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	361 dtype=floatX,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	362 rasterized=True)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	363
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	364 grads = rbm.contrastive_gradient(pos_v=train_batch, neg_v=sampler.positions[0])
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	365
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	366 learn_fn = function([batch_idx, s_lr],
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	367 outputs=[
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	368 grads[0].norm(2),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	369 rbm.U.norm(2)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	370 ],
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	371 updates = sgd_updates(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	372 rbm.params,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	373 grads,
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	374 lr=[2s_lr, .2s_lr, .02s_lr, .1s_lr, .02*s_lr ]))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	375
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	376 for jj in xrange(10000):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	377 sampler.simulate()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	378 l2_of_Ugrad = learn_fn(jj, lr/max(1, jj/(20*epoch_size/batchsize)))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	379
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	380 if jj > no_l1_epochs * epoch_size/batchsize:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	381 rbm.U.value -= l1_penalty * np.sign(rbm.U.value)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	382 rbm.W.value -= l1_penalty * np.sign(rbm.W.value)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	383
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	384 if jj % 5 == 0:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	385 rbm.U.value = numpy_project_onto_ball(rbm.U.value.T).T
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	386
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	387 if ((jj < 10)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	388 or (jj < 100 and 0==jj%10)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	389 or (jj < 1000 and 0==jj%100)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	390 or (jj < 10000 and 0==jj%1000)):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	391 print 'saving samples', jj, 'epoch', jj/(epoch_size/batchsize), l2_of_Ugrad
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	392 print 'neg particles', sampler.positions[0].value.min(), sampler.positions[0].value.max()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	393 image_tiling.save_tiled_raster_images(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	394 image_tiling.tile_raster_images(sampler.positions[0].value, (R,C)),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	395 "sample_%06i.png"%jj)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	396 image_tiling.save_tiled_raster_images(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	397 image_tiling.tile_raster_images(rbm.U.value.T, (R,C)),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	398 "U_%06i.png"%jj)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	399 image_tiling.save_tiled_raster_images(
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	400 image_tiling.tile_raster_images(rbm.W.value.T, (R,C)),
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	401 "W_%06i.png"%jj)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	402
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	403
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	404
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	405 #
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	406 #
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	407 # Marc'Aurelio Ranzato's code
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	408 #
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	409 ######################################################################
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	410 # compute the value of the free energy at a given input
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	411 # F = - sum log(1+exp(- .5 FH (VF data/norm(data))^2 + bias_cov)) +...
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	412 # - sum log(1+exp(w_mean data + bias_mean)) + ...
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	413 # - bias_vis data + 0.5 data^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	414 # NOTE: FH is constrained to be positive
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	415 # (in the paper the sign is negative but the sign in front of it is also flipped)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	416 def compute_energy_mcRBM(data,normdata,vel,energy,VF,FH,bias_cov,bias_vis,w_mean,bias_mean,t1,t2,t6,feat,featsq,feat_mean,length,lengthsq,normcoeff,small,num_vis):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	417 # normalize input data vectors
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	418 data.mult(data, target = t6) # DxP (nr input dims x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	419 t6.sum(axis = 0, target = lengthsq) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	420 lengthsq.mult(0.5, target = energy) # energy of quadratic regularization term
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	421 lengthsq.mult(1./num_vis) # normalize by number of components (like std)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	422
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	423 lengthsq.add(small) # small prevents division by 0
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	424 # energy_j = \sum_i 0.5 data_ij ^2
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	425 # lengthsq_j = 1/ (\sum_i data_ij ^2 + small)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	426 cmt.sqrt(lengthsq, target = length)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	427 # length_j = sqrt(lengthsq_j)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	428 length.reciprocal(target = normcoeff) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	429 # normcoef_j = 1/sqrt(lengthsq_j)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	430 data.mult_by_row(normcoeff, target = normdata) # normalized data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	431 # normdata is like data, but cols have unit L2 norm
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	432
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	433 ## potential
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	434 # covariance contribution
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	435 cmt.dot(VF.T, normdata, target = feat) # HxP (nr factors x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	436 feat.mult(feat, target = featsq) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	437
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	438 # featsq is the squared cosines (VF with data)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	439 cmt.dot(FH.T,featsq, target = t1) # OxP (nr cov hiddens x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	440 t1.mult(-0.5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	441 t1.add_col_vec(bias_cov) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	442 cmt.exp(t1) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	443 t1.add(1, target = t2) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	444 cmt.log(t2)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	445 t2.mult(-1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	446 energy.add_sums(t2, axis=0)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	447 # mean contribution
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	448 cmt.dot(w_mean.T, data, target = feat_mean) # HxP (nr mean hiddens x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	449 feat_mean.add_col_vec(bias_mean) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	450 cmt.exp(feat_mean)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	451 feat_mean.add(1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	452 cmt.log(feat_mean)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	453 feat_mean.mult(-1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	454 energy.add_sums(feat_mean, axis=0)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	455 # visible bias term
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	456 data.mult_by_col(bias_vis, target = t6)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	457 t6.mult(-1) # DxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	458 energy.add_sums(t6, axis=0) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	459 # kinetic
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	460 vel.mult(vel, target = t6)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	461 energy.add_sums(t6, axis = 0, mult = .5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	462
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	463 ######################################################
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	464 # mcRBM trainer: sweeps over the training set.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	465 # For each batch of samples compute derivatives to update the parameters
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	466 # at the training samples and at the negative samples drawn calling HMC sampler.
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	467 def train_mcRBM():
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	468
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	469 config = ConfigParser()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	470 config.read('input_configuration')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	471
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	472 verbose = config.getint('VERBOSITY','verbose')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	473
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	474 num_epochs = config.getint('MAIN_PARAMETER_SETTING','num_epochs')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	475 batch_size = config.getint('MAIN_PARAMETER_SETTING','batch_size')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	476 startFH = config.getint('MAIN_PARAMETER_SETTING','startFH')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	477 startwd = config.getint('MAIN_PARAMETER_SETTING','startwd')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	478 doPCD = config.getint('MAIN_PARAMETER_SETTING','doPCD')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	479
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	480 # model parameters
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	481 num_fac = config.getint('MODEL_PARAMETER_SETTING','num_fac')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	482 num_hid_cov = config.getint('MODEL_PARAMETER_SETTING','num_hid_cov')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	483 num_hid_mean = config.getint('MODEL_PARAMETER_SETTING','num_hid_mean')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	484 apply_mask = config.getint('MODEL_PARAMETER_SETTING','apply_mask')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	485
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	486 # load data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	487 data_file_name = config.get('DATA','data_file_name')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	488 d = loadmat(data_file_name) # input in the format PxD (P vectorized samples with D dimensions)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	489 totnumcases = d["whitendata"].shape[0]
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	490 d = d["whitendata"][0:floor(totnumcases/batch_size)*batch_size,:].copy()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	491 totnumcases = d.shape[0]
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	492 num_vis = d.shape[1]
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	493 num_batches = int(totnumcases/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	494 dev_dat = cmt.CUDAMatrix(d.T) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	495
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	496 # training parameters
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	497 epsilon = config.getfloat('OPTIMIZER_PARAMETERS','epsilon')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	498 epsilonVF = 2*epsilon
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	499 epsilonFH = 0.02*epsilon
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	500 epsilonb = 0.02*epsilon
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	501 epsilonw_mean = 0.2*epsilon
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	502 epsilonb_mean = 0.1*epsilon
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	503 weightcost_final = config.getfloat('OPTIMIZER_PARAMETERS','weightcost_final')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	504
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	505 # HMC setting
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	506 hmc_step_nr = config.getint('HMC_PARAMETERS','hmc_step_nr')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	507 hmc_step = 0.01
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	508 hmc_target_ave_rej = config.getfloat('HMC_PARAMETERS','hmc_target_ave_rej')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	509 hmc_ave_rej = hmc_target_ave_rej
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	510
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	511 # initialize weights
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	512 VF = cmt.CUDAMatrix(np.array(0.02 * np.random.randn(num_vis, num_fac), dtype=np.float32, order='F')) # VxH
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	513 if apply_mask == 0:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	514 FH = cmt.CUDAMatrix( np.array( np.eye(num_fac,num_hid_cov), dtype=np.float32, order='F') ) # HxO
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	515 else:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	516 dd = loadmat('your_FHinit_mask_file.mat') # see CVPR2010paper_material/topo2D_3x3_stride2_576filt.mat for an example
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	517 FH = cmt.CUDAMatrix( np.array( dd["FH"], dtype=np.float32, order='F') )
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	518 bias_cov = cmt.CUDAMatrix( np.array(2.0*np.ones((num_hid_cov, 1)), dtype=np.float32, order='F') )
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	519 bias_vis = cmt.CUDAMatrix( np.array(np.zeros((num_vis, 1)), dtype=np.float32, order='F') )
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	520 w_mean = cmt.CUDAMatrix( np.array( 0.05 * np.random.randn(num_vis, num_hid_mean), dtype=np.float32, order='F') ) # VxH
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	521 bias_mean = cmt.CUDAMatrix( np.array( -2.0*np.ones((num_hid_mean,1)), dtype=np.float32, order='F') )
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	522
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	523 # initialize variables to store derivatives
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	524 VFinc = cmt.CUDAMatrix( np.array(np.zeros((num_vis, num_fac)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	525 FHinc = cmt.CUDAMatrix( np.array(np.zeros((num_fac, num_hid_cov)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	526 bias_covinc = cmt.CUDAMatrix( np.array(np.zeros((num_hid_cov, 1)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	527 bias_visinc = cmt.CUDAMatrix( np.array(np.zeros((num_vis, 1)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	528 w_meaninc = cmt.CUDAMatrix( np.array(np.zeros((num_vis, num_hid_mean)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	529 bias_meaninc = cmt.CUDAMatrix( np.array(np.zeros((num_hid_mean, 1)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	530
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	531 # initialize temporary storage
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	532 data = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F')) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	533 normdata = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F')) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	534 negdataini = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F')) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	535 feat = cmt.CUDAMatrix( np.array(np.empty((num_fac, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	536 featsq = cmt.CUDAMatrix( np.array(np.empty((num_fac, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	537 negdata = cmt.CUDAMatrix( np.array(np.random.randn(num_vis, batch_size), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	538 old_energy = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	539 new_energy = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	540 gradient = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F')) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	541 normgradient = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F')) # VxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	542 thresh = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	543 feat_mean = cmt.CUDAMatrix( np.array(np.empty((num_hid_mean, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	544 vel = cmt.CUDAMatrix( np.array(np.random.randn(num_vis, batch_size), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	545 length = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F')) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	546 lengthsq = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F')) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	547 normcoeff = cmt.CUDAMatrix( np.array(np.zeros((1, batch_size)), dtype=np.float32, order='F')) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	548 if apply_mask==1: # this used to constrain very large FH matrices only allowing to change values in a neighborhood
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	549 dd = loadmat('your_FHinit_mask_file.mat')
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	550 mask = cmt.CUDAMatrix( np.array(dd["mask"], dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	551 normVF = 1
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	552 small = 0.5
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	553
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	554 # other temporary vars
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	555 t1 = cmt.CUDAMatrix( np.array(np.empty((num_hid_cov, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	556 t2 = cmt.CUDAMatrix( np.array(np.empty((num_hid_cov, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	557 t3 = cmt.CUDAMatrix( np.array(np.empty((num_fac, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	558 t4 = cmt.CUDAMatrix( np.array(np.empty((1,batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	559 t5 = cmt.CUDAMatrix( np.array(np.empty((1,1)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	560 t6 = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	561 t7 = cmt.CUDAMatrix( np.array(np.empty((num_vis, batch_size)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	562 t8 = cmt.CUDAMatrix( np.array(np.empty((num_vis, num_fac)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	563 t9 = cmt.CUDAMatrix( np.array(np.zeros((num_fac, num_hid_cov)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	564 t10 = cmt.CUDAMatrix( np.array(np.empty((1,num_fac)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	565 t11 = cmt.CUDAMatrix( np.array(np.empty((1,num_hid_cov)), dtype=np.float32, order='F'))
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	566
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	567 # start training
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	568 for epoch in range(num_epochs):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	569
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	570 print "Epoch " + str(epoch + 1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	571
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	572 # anneal learning rates
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	573 epsilonVFc = epsilonVF/max(1,epoch/20)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	574 epsilonFHc = epsilonFH/max(1,epoch/20)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	575 epsilonbc = epsilonb/max(1,epoch/20)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	576 epsilonw_meanc = epsilonw_mean/max(1,epoch/20)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	577 epsilonb_meanc = epsilonb_mean/max(1,epoch/20)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	578 weightcost = weightcost_final
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	579
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	580 if epoch <= startFH:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	581 epsilonFHc = 0
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	582 if epoch <= startwd:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	583 weightcost = 0
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	584
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	585 for batch in range(num_batches):
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	586
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	587 # get current minibatch
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	588 data = dev_dat.slice(batchbatch_size,(batch + 1)batch_size) # DxP (nr dims x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	589
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	590 # normalize input data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	591 data.mult(data, target = t6) # DxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	592 t6.sum(axis = 0, target = lengthsq) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	593 lengthsq.mult(1./num_vis) # normalize by number of components (like std)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	594 lengthsq.add(small) # small avoids division by 0
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	595 cmt.sqrt(lengthsq, target = length)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	596 length.reciprocal(target = normcoeff) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	597 data.mult_by_row(normcoeff, target = normdata) # normalized data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	598 ## compute positive sample derivatives
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	599 # covariance part
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	600 cmt.dot(VF.T, normdata, target = feat) # HxP (nr facs x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	601 feat.mult(feat, target = featsq) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	602 cmt.dot(FH.T,featsq, target = t1) # OxP (nr cov hiddens x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	603 t1.mult(-0.5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	604 t1.add_col_vec(bias_cov) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	605 t1.apply_sigmoid(target = t2) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	606 cmt.dot(featsq, t2.T, target = FHinc) # HxO
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	607 cmt.dot(FH,t2, target = t3) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	608 t3.mult(feat)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	609 cmt.dot(normdata, t3.T, target = VFinc) # VxH
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	610 t2.sum(axis = 1, target = bias_covinc)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	611 bias_covinc.mult(-1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	612 # visible bias
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	613 data.sum(axis = 1, target = bias_visinc)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	614 bias_visinc.mult(-1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	615 # mean part
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	616 cmt.dot(w_mean.T, data, target = feat_mean) # HxP (nr mean hiddens x nr samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	617 feat_mean.add_col_vec(bias_mean) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	618 feat_mean.apply_sigmoid() # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	619 feat_mean.mult(-1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	620 cmt.dot(data, feat_mean.T, target = w_meaninc)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	621 feat_mean.sum(axis = 1, target = bias_meaninc)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	622
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	623 # HMC sampling: draw an approximate sample from the model
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	624 if doPCD == 0: # CD-1 (set negative data to current training samples)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	625 hmc_step, hmc_ave_rej = draw_HMC_samples(data,negdata,normdata,vel,gradient,normgradient,new_energy,old_energy,VF,FH,bias_cov,bias_vis,w_mean,bias_mean,hmc_step,hmc_step_nr,hmc_ave_rej,hmc_target_ave_rej,t1,t2,t3,t4,t5,t6,t7,thresh,feat,featsq,batch_size,feat_mean,length,lengthsq,normcoeff,small,num_vis)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	626 else: # PCD-1 (use previous negative data as starting point for chain)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	627 negdataini.assign(negdata)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	628 hmc_step, hmc_ave_rej = draw_HMC_samples(negdataini,negdata,normdata,vel,gradient,normgradient,new_energy,old_energy,VF,FH,bias_cov,bias_vis,w_mean,bias_mean,hmc_step,hmc_step_nr,hmc_ave_rej,hmc_target_ave_rej,t1,t2,t3,t4,t5,t6,t7,thresh,feat,featsq,batch_size,feat_mean,length,lengthsq,normcoeff,small,num_vis)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	629
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	630 # compute derivatives at the negative samples
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	631 # normalize input data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	632 negdata.mult(negdata, target = t6) # DxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	633 t6.sum(axis = 0, target = lengthsq) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	634 lengthsq.mult(1./num_vis) # normalize by number of components (like std)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	635 lengthsq.add(small)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	636 cmt.sqrt(lengthsq, target = length)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	637 length.reciprocal(target = normcoeff) # 1xP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	638 negdata.mult_by_row(normcoeff, target = normdata) # normalized data
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	639 # covariance part
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	640 cmt.dot(VF.T, normdata, target = feat) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	641 feat.mult(feat, target = featsq) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	642 cmt.dot(FH.T,featsq, target = t1) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	643 t1.mult(-0.5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	644 t1.add_col_vec(bias_cov) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	645 t1.apply_sigmoid(target = t2) # OxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	646 FHinc.subtract_dot(featsq, t2.T) # HxO
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	647 FHinc.mult(0.5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	648 cmt.dot(FH,t2, target = t3) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	649 t3.mult(feat)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	650 VFinc.subtract_dot(normdata, t3.T) # VxH
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	651 bias_covinc.add_sums(t2, axis = 1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	652 # visible bias
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	653 bias_visinc.add_sums(negdata, axis = 1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	654 # mean part
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	655 cmt.dot(w_mean.T, negdata, target = feat_mean) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	656 feat_mean.add_col_vec(bias_mean) # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	657 feat_mean.apply_sigmoid() # HxP
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	658 w_meaninc.add_dot(negdata, feat_mean.T)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	659 bias_meaninc.add_sums(feat_mean, axis = 1)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	660
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	661 # update parameters
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	662 VFinc.add_mult(VF.sign(), weightcost) # L1 regularization
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	663 VF.add_mult(VFinc, -epsilonVFc/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	664 # normalize columns of VF: normalize by running average of their norm
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	665 VF.mult(VF, target = t8)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	666 t8.sum(axis = 0, target = t10)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	667 cmt.sqrt(t10)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	668 t10.sum(axis=1,target = t5)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	669 t5.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	670 normVF = .95normVF + (.05/num_fac) t5.numpy_array[0,0] # estimate norm
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	671 t10.reciprocal()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	672 VF.mult_by_row(t10)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	673 VF.mult(normVF)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	674 bias_cov.add_mult(bias_covinc, -epsilonbc/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	675 bias_vis.add_mult(bias_visinc, -epsilonbc/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	676
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	677 if epoch > startFH:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	678 FHinc.add_mult(FH.sign(), weightcost) # L1 regularization
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	679 FH.add_mult(FHinc, -epsilonFHc/batch_size) # update
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	680 # set to 0 negative entries in FH
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	681 FH.greater_than(0, target = t9)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	682 FH.mult(t9)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	683 if apply_mask==1:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	684 FH.mult(mask)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	685 # normalize columns of FH: L1 norm set to 1 in each column
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	686 FH.sum(axis = 0, target = t11)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	687 t11.reciprocal()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	688 FH.mult_by_row(t11)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	689 w_meaninc.add_mult(w_mean.sign(),weightcost)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	690 w_mean.add_mult(w_meaninc, -epsilonw_meanc/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	691 bias_mean.add_mult(bias_meaninc, -epsilonb_meanc/batch_size)
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	692
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	693 if verbose == 1:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	694 print "VF: " + '%3.2e' % VF.euclid_norm() + ", DVF: " + '%3.2e' % (VFinc.euclid_norm()(epsilonVFc/batch_size)) + ", FH: " + '%3.2e' % FH.euclid_norm() + ", DFH: " + '%3.2e' % (FHinc.euclid_norm()(epsilonFHc/batch_size)) + ", bias_cov: " + '%3.2e' % bias_cov.euclid_norm() + ", Dbias_cov: " + '%3.2e' % (bias_covinc.euclid_norm()(epsilonbc/batch_size)) + ", bias_vis: " + '%3.2e' % bias_vis.euclid_norm() + ", Dbias_vis: " + '%3.2e' % (bias_visinc.euclid_norm()(epsilonbc/batch_size)) + ", wm: " + '%3.2e' % w_mean.euclid_norm() + ", Dwm: " + '%3.2e' % (w_meaninc.euclid_norm()(epsilonw_meanc/batch_size)) + ", bm: " + '%3.2e' % bias_mean.euclid_norm() + ", Dbm: " + '%3.2e' % (bias_meaninc.euclid_norm()(epsilonb_meanc/batch_size)) + ", step: " + '%3.2e' % hmc_step + ", rej: " + '%3.2e' % hmc_ave_rej
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	695 sys.stdout.flush()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	696 # back-up every once in a while
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	697 if np.mod(epoch,10) == 0:
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	698 VF.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	699 FH.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	700 bias_cov.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	701 w_mean.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	702 bias_mean.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	703 bias_vis.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	704 savemat("ws_temp", {'VF':VF.numpy_array,'FH':FH.numpy_array,'bias_cov': bias_cov.numpy_array, 'bias_vis': bias_vis.numpy_array,'w_mean': w_mean.numpy_array, 'bias_mean': bias_mean.numpy_array, 'epoch':epoch})
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	705 # final back-up
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	706 VF.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	707 FH.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	708 bias_cov.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	709 bias_vis.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	710 w_mean.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	711 bias_mean.copy_to_host()
90e11d5d0a41 adding algorithms/mcRBM, but it is not done yet James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	712 savemat("ws_fac" + str(num_fac) + "_cov" + str(num_hid_cov) + "_mean" + str(num_hid_mean), {'VF':VF.numpy_array,'FH':FH.numpy_array,'bias_cov': bias_cov.numpy_array, 'bias_vis': bias_vis.numpy_array, 'w_mean': w_mean.numpy_array, 'bias_mean': bias_mean.numpy_array, 'epoch':epoch})

Mercurial > pylearn

annotate pylearn/algorithms/mcRBM.py @ 978:ab4bc97ca060