def __init__(self, n_visible, n_hidden, n_context):
super(BinaryBinaryCRBM, self).__init__(n_visible, n_hidden)
# data shape
self.n_context = n_context
# units
self.x = units.Units(self, name='x') # context
# parameters
self.A = parameters.ProdParameters(
self, [self.x, self.v],
theano.shared(value=self._initial_A(), name='A'),
name='A') # context-to-visible weights
self.B = parameters.ProdParameters(
self, [self.x, self.h],
theano.shared(value=self._initial_B(), name='B'),
name='B') # context-to-hidden weights
from morb import base, units, parameters, stats, param_updaters, trainers, monitors
# This example shows how the FIOTRBM model from "Facial Expression Transfer with
# Input-Output Temporal Restricted Boltzmann Machines" by Zeiler et al. (NIPS
# 2011) can be recreated in Morb.
rbm = base.RBM()
rbm.v = units.GaussianUnits(rbm) # output (visibles)
rbm.h = units.BinaryUnits(rbm) # latent (hiddens)
rbm.s = units.Units(rbm) # input (context)
rbm.vp = units.Units(rbm) # output history (context)
initial_A = ...
initial_B = ...
initial_bv = ...
initial_bh = ...
initial_Wv = ...
initial_Wh = ...
initial_Ws = ...
parameters.FixedBiasParameters(
rbm, rbm.v.precision_units) # add precision term to the energy function
rbm.A = parameters.ProdParameters(
rbm, [rbm.vp, rbm.v],
initial_A) # weights from past output to current output
rbm.B = parameters.ProdParameters(
rbm, [rbm.vp, rbm.h], initial_B) # weights from past output to hiddens
rbm.bv = parameters.BiasParameters(rbm, rbm.v, initial_bv) # visible bias
rbm.bh = parameters.BiasParameters(rbm, rbm.h, initial_bh) # hidden bias
rbm.W = parameters.ThirdOrderFactoredParameters(
rbm, [rbm.v, rbm.h, rbm.s],
n_hidden = 20
print ">> Constructing RBM..."
numpy_rng = np.random.RandomState(123)
initial_W = np.asarray(np.random.uniform(
low=-4 * np.sqrt(6. / (n_hidden + n_visible + n_context)),
high=4 * np.sqrt(6. / (n_hidden + n_visible + n_context)),
size=(n_visible, n_hidden, n_context)),
dtype=theano.config.floatX)
initial_bv = np.zeros(n_visible, dtype=theano.config.floatX)
initial_bh = np.zeros(n_hidden, dtype=theano.config.floatX)
rbm = morb.base.RBM()
rbm.v = units.BinaryUnits(rbm, name='v') # visibles
rbm.h = units.BinaryUnits(rbm, name='h') # hiddens
rbm.x = units.Units(rbm, name='x') # context
rbm.W = parameters.ThirdOrderParameters(rbm, [rbm.v, rbm.h, rbm.x],
theano.shared(value=initial_W,
name='W'),
name='W') # weights
rbm.bv = parameters.BiasParameters(rbm,
rbm.v,
theano.shared(value=initial_bv, name='bv'),
name='bv') # visible bias
rbm.bh = parameters.BiasParameters(rbm,
rbm.h,
theano.shared(value=initial_bh, name='bh'),
name='bh') # hidden bias
initial_vmap = {rbm.v: T.matrix('v'), rbm.x: T.matrix('x')}