def decode(self, hiddens):
"""
Map inputs through the encoder function.
Parameters
----------
hiddens : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the input
minibatch(es) to be encoded. Assumed to be 2-tensors, with the
first dimension indexing training examples and the second
indexing data dimensions.
Returns
-------
decoded : tensor_like or list of tensor_like
Theano symbolic (or list thereof) representing the corresponding
minibatch(es) after decoding.
"""
if self.act_dec is None:
act_dec = lambda x: x
else:
act_dec = self.act_dec
if isinstance(hiddens, tensor.Variable):
X = act_dec(self.visbias + tensor.dot(hiddens, self.w_prime))
else:
X = [self.decode(v) for v in hiddens]
g = T.dvector()
for i in range(self.N):
start = i * 21
end = start + 20
indexis = numpy.arange(start, end + 1)
if i == 0:
g = rescaled_softmax(theano.tensor.subtensor.take(
X, indexis, axis=1, mode='raise'),
min_val=1e-5)
else:
g = T.concatenate([
g,
rescaled_softmax(theano.tensor.subtensor.take(
X, indexis, axis=1, mode='raise'),
min_val=1e-5)
],
axis=1)
return g
def _corrupt(self, x):
noise = self.s_rng.normal(size=x.shape,
avg=0.,
std=self.corruption_level,
dtype=theano.config.floatX)
return rescaled_softmax(x + noise)
def _corrupt(self, x):
noise = self.s_rng.normal(
size=x.shape,
avg=0.,
std=self.corruption_level,
dtype=theano.config.floatX
)
return rescaled_softmax(x + noise)
def decode(self, hiddens):
"""
Map inputs through the encoder function.
Parameters
----------
hiddens : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the input
minibatch(es) to be encoded. Assumed to be 2-tensors, with the
first dimension indexing training examples and the second
indexing data dimensions.
Returns
-------
decoded : tensor_like or list of tensor_like
Theano symbolic (or list thereof) representing the corresponding
minibatch(es) after decoding.
"""
if self.act_dec is None:
act_dec = lambda x: x
else:
act_dec = self.act_dec
if isinstance(hiddens, tensor.Variable):
X = act_dec(self.visbias + tensor.dot(hiddens, self.w_prime))
else:
X= [self.decode(v) for v in hiddens]
g = T.dvector()
for i in range(self.N):
start = i * 21
end = start + 20
indexis = numpy.arange(start,end+1)
if i==0:
g = rescaled_softmax(theano.tensor.subtensor.take(X,indexis,axis=1,mode='raise'),min_val=1e-5)
else:
g =T.concatenate([g,rescaled_softmax(theano.tensor.subtensor.take(X,indexis,axis=1,mode='raise'),min_val=1e-5)],axis=1)
return g
def _corrupt(self, x):
"""
Corrupts a single tensor_like object.
Parameters
----------
x : tensor_like
Theano symbolic representing a (mini)batch of inputs to be
corrupted, with the first dimension indexing training
examples and the second indexing data dimensions.
Returns
-------
corrupted : tensor_like
Theano symbolic representing the corresponding corrupted input.
"""
noise = self.s_rng.normal(size=x.shape, avg=0.0, std=self.corruption_level, dtype=theano.config.floatX)
return rescaled_softmax(x + noise)
def _corrupt(self, x):
"""
Corrupts a single tensor_like object.
Parameters
----------
x : tensor_like
Theano symbolic representing a (mini)batch of inputs to be
corrupted, with the first dimension indexing training
examples and the second indexing data dimensions.
Returns
-------
corrupted : tensor_like
Theano symbolic representing the corresponding corrupted input.
"""
noise = self.s_rng.normal(size=x.shape,
avg=0.,
std=self.corruption_level,
dtype=theano.config.floatX)
return rescaled_softmax(x + noise)