def get_hidden_values(self, input):
""" Computes the values of the hidden layer """
return af(T.dot(input, self.W) + self.b)
def get_hidden_values(self, input):
""" Computes the values of the hidden layer """
return af(T.dot(input, self.W) + self.b)
def get_reconstructed_input(self, hidden):
"""Computes the reconstructed input given the values of the
hidden layer
"""
return af(T.dot(hidden, self.W_prime) + self.b_prime)
def get_reconstructed_input(self, hidden):
"""Computes the reconstructed input given the values of the
hidden layer
"""
return af(T.dot(hidden, self.W_prime) + self.b_prime)
def __init__(self,
numpy_rng,
theano_rng=None,
input1=None,
input2=None,
cor_reg=None,
n_visible1=784 / 2,
n_visible2=784 / 2,
n_hidden=500,
W1=None,
bhid1=None,
bvis1=None,
W2=None,
bhid2=None,
bvis2=None):
self.n_visible1 = n_visible1
self.n_visible2 = n_visible2
self.n_hidden = n_hidden
# create a Theano random generator that gives symbolic random values
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2**30))
# note : W' was written as `W_prime` and b' as `b_prime`
if not W1:
# W is initialized with `initial_W` which is uniformely sampled
# from -4*sqrt(6./(n_visible+n_hidden)) and
# 4*sqrt(6./(n_hidden+n_visible))the output of uniform if
# converted using asarray to dtype
# theano.config.floatX so that the code is runable on GPU
initial_W1 = numpy.asarray(numpy_rng.uniform(
low=-4 * numpy.sqrt(6. / (n_hidden + n_visible1)),
high=4 * numpy.sqrt(6. / (n_hidden + n_visible1)),
size=(n_visible1, n_hidden)),
dtype=theano.config.floatX)
W1 = theano.shared(value=initial_W1, name='W1', borrow=True)
if not W2:
# W is initialized with `initial_W` which is uniformely sampled
# from -4*sqrt(6./(n_visible+n_hidden)) and
# 4*sqrt(6./(n_hidden+n_visible))the output of uniform if
# converted using asarray to dtype
# theano.config.floatX so that the code is runable on GPU
initial_W2 = numpy.asarray(numpy_rng.uniform(
low=-4 * numpy.sqrt(6. / (n_hidden + n_visible2)),
high=4 * numpy.sqrt(6. / (n_hidden + n_visible2)),
size=(n_visible2, n_hidden)),
dtype=theano.config.floatX)
W2 = theano.shared(value=initial_W2, name='W2', borrow=True)
if not bvis1:
bvis1 = theano.shared(value=numpy.zeros(
n_visible1, dtype=theano.config.floatX),
name='b1p',
borrow=True)
if not bvis2:
bvis2 = theano.shared(value=numpy.zeros(
n_visible2, dtype=theano.config.floatX),
name='b2p',
borrow=True)
if not bhid1:
bhid1 = theano.shared(value=numpy.zeros(
n_hidden, dtype=theano.config.floatX),
name='b1',
borrow=True)
if not bhid2:
bhid2 = theano.shared(value=numpy.zeros(
n_hidden, dtype=theano.config.floatX),
name='b2',
borrow=True)
self.W1 = W1
self.W2 = W2
# b corresponds to the bias of the hidden
self.b1 = bhid1
self.b2 = bhid2
# b_prime corresponds to the bias of the visible
self.b1_prime = bvis1
self.b2_prime = bvis2
# tied weights, therefore W_prime is W transpose
self.W1_prime = self.W1.T
self.W2_prime = self.W2.T
self.theano_rng = theano_rng
self.L1 = (
abs(self.W1).sum() + abs(self.W2).sum(
) #+abs(self.b1).sum()+abs(self.b2).sum()+abs(self.b1_prime).sum()+abs(self.b2_prime).sum()
)
self.L2_sqr = ((self.W1**2).sum(
) #+(self.W2**2).sum()#+abs(self.b1**2).sum()+abs(self.b2**2).sum()+abs(self.b1_prime**2).sum()+abs(self.b2_prime**2).sum()
)
# if no input is given, generate a variable representing the input
if input1 is None:
# we use a matrix because we expect a minibatch of several
# examples, each example being a row
self.x1 = T.dmatrix(name='input1')
self.x2 = T.dmatrix(name='input2')
else:
self.x1 = input1
self.x2 = input2
self.params = [
self.W1, self.b1, self.b1_prime, self.W2, self.b2, self.b2_prime
]
# end-snippet-1
self.output1 = af(T.dot(self.x1, self.W1) + self.b1)
self.output2 = af(T.dot(self.x2, self.W2) + self.b2)
self.rec1 = (T.dot(self.output1, self.W1_prime) + self.b1_prime)
self.rec2 = (T.dot(self.output2, self.W2_prime) + self.b2_prime)
self.reg = (T.dot(self.output1, self.W2_prime) + self.b2_prime)
self.cor_reg = theano.shared(numpy.float32(1.0), name='reg')
def get_hidden_values(self, input1, input2):
""" Computes the values of the hidden layer """
return af(T.dot(input1, self.W1) +
self.b1), af(T.dot(input2, self.W2) + self.b2)
def __init__(
self,
numpy_rng,
theano_rng=None,
input1=None,
input2=None,
cor_reg=None,
n_visible1=784/2,
n_visible2=784/2,
n_hidden=500,
W1=None,
bhid1=None,
bvis1=None,
W2=None,
bhid2=None,
bvis2=None
):
self.n_visible1 = n_visible1
self.n_visible2 = n_visible2
self.n_hidden = n_hidden
# create a Theano random generator that gives symbolic random values
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# note : W' was written as `W_prime` and b' as `b_prime`
if not W1:
# W is initialized with `initial_W` which is uniformely sampled
# from -4*sqrt(6./(n_visible+n_hidden)) and
# 4*sqrt(6./(n_hidden+n_visible))the output of uniform if
# converted using asarray to dtype
# theano.config.floatX so that the code is runable on GPU
initial_W1 = numpy.asarray(
numpy_rng.uniform(
low=-4 * numpy.sqrt(6. / (n_hidden + n_visible1)),
high=4 * numpy.sqrt(6. / (n_hidden + n_visible1)),
size=(n_visible1, n_hidden)
),
dtype=theano.config.floatX
)
W1 = theano.shared(value=initial_W1, name='W1', borrow=True)
if not W2:
# W is initialized with `initial_W` which is uniformely sampled
# from -4*sqrt(6./(n_visible+n_hidden)) and
# 4*sqrt(6./(n_hidden+n_visible))the output of uniform if
# converted using asarray to dtype
# theano.config.floatX so that the code is runable on GPU
initial_W2 = numpy.asarray(
numpy_rng.uniform(
low=-4 * numpy.sqrt(6. / (n_hidden + n_visible2)),
high=4 * numpy.sqrt(6. / (n_hidden + n_visible2)),
size=(n_visible2, n_hidden)
),
dtype=theano.config.floatX
)
W2 = theano.shared(value=initial_W2, name='W2', borrow=True)
if not bvis1:
bvis1 = theano.shared(
value=numpy.zeros(
n_visible1,
dtype=theano.config.floatX
),
name='b1p',
borrow=True
)
if not bvis2:
bvis2 = theano.shared(
value=numpy.zeros(
n_visible2,
dtype=theano.config.floatX
),
name='b2p',
borrow=True
)
if not bhid1:
bhid1 = theano.shared(
value=numpy.zeros(
n_hidden,
dtype=theano.config.floatX
),
name='b1',
borrow=True
)
if not bhid2:
bhid2 = theano.shared(
value=numpy.zeros(
n_hidden,
dtype=theano.config.floatX
),
name='b2',
borrow=True
)
self.W1 = W1
self.W2 = W2
# b corresponds to the bias of the hidden
self.b1 = bhid1
self.b2 = bhid2
# b_prime corresponds to the bias of the visible
self.b1_prime = bvis1
self.b2_prime = bvis2
# tied weights, therefore W_prime is W transpose
self.W1_prime = self.W1.T
self.W2_prime = self.W2.T
self.theano_rng = theano_rng
self.L1 = (
abs(self.W1).sum()+abs(self.W2).sum()#+abs(self.b1).sum()+abs(self.b2).sum()+abs(self.b1_prime).sum()+abs(self.b2_prime).sum()
)
self.L2_sqr = (
(self.W1**2).sum()#+(self.W2**2).sum()#+abs(self.b1**2).sum()+abs(self.b2**2).sum()+abs(self.b1_prime**2).sum()+abs(self.b2_prime**2).sum()
)
# if no input is given, generate a variable representing the input
if input1 is None:
# we use a matrix because we expect a minibatch of several
# examples, each example being a row
self.x1 = T.dmatrix(name='input1')
self.x2 = T.dmatrix(name='input2')
else:
self.x1 = input1
self.x2 = input2
self.params = [self.W1, self.b1, self.b1_prime,
self.W2, self.b2, self.b2_prime
]
# end-snippet-1
self.output1 = af(T.dot(self.x1, self.W1) + self.b1)
self.output2 = af(T.dot(self.x2, self.W2) + self.b2)
self.rec1 = (T.dot(self.output1, self.W1_prime) + self.b1_prime)
self.rec2 = (T.dot(self.output2, self.W2_prime) + self.b2_prime)
self.reg = (T.dot(self.output1, self.W2_prime) + self.b2_prime)
self.cor_reg = theano.shared(numpy.float32(1.0),name='reg')
def get_hidden_values(self, input1, input2):
""" Computes the values of the hidden layer """
return af(T.dot(input1, self.W1) + self.b1), af(T.dot(input2, self.W2) + self.b2)