def make_chains(self, bias):
"""
.. todo::
WRITEME properly
Make the shared variable representing a layer of
the network for all negative chains
for now units are initialized randomly based on their
biases only
"""
assert not self.use_cd
b = bias.get_value(borrow=True)
nhid ,= b.shape
shape = (self.negative_chains, nhid)
driver = self.rng.uniform(0.0, 1.0, shape)
thresh = sigmoid_numpy(b)
value = driver < thresh
return sharedX(value)
def set_biases(self, biases, recenter = False):
self.b.set_value(biases)
if recenter:
assert self.center
if self.pool_size != 1:
raise NotImplementedError()
self.offset.set_value(sigmoid_numpy(self.b.get_value()))
def make_state(self, num_examples, numpy_rng):
driver = numpy_rng.uniform(0.,1., (num_examples, self.nvis))
on_prob = sigmoid_numpy(2. * self.ising_bias_numpy())
sample = 2. * (driver < on_prob) - 1.
rval = sharedX(sample, name = 'v_sample_shared')
return rval
def make_state(self, num_examples, numpy_rng):
driver = numpy_rng.uniform(0., 1., (num_examples, self.nvis))
mean = sigmoid_numpy(self.bias.get_value())
sample = driver < mean
rval = sharedX(sample, name='v_sample_shared')
return rval
def make_state(self, num_examples, numpy_rng):
driver = numpy_rng.uniform(0.,1., (num_examples, self.nvis))
mean = sigmoid_numpy(self.bias.get_value())
sample = driver < mean
rval = sharedX(sample, name = 'v_sample_shared')
return rval
def make_state(self, num_examples, numpy_rng):
""" Returns a shared variable containing an actual state
(not a mean field state) for this variable.
"""
driver = numpy_rng.uniform(0.,1., (num_examples, self.dim))
on_prob = sigmoid_numpy(2. * self.b.get_value())
sample = 2. * (driver < on_prob) - 1.
rval = sharedX(sample, name = 'v_sample_shared')
return rval
def set_biases(self, biases, recenter=False):
self.bias.set_value(biases)
if recenter:
assert self.center
self.offset.set_value(sigmoid_numpy(self.bias.get_value()))
