def fprop(self,
state_below,
use_noise=True,
no_noise_bias=False,
first_only=False):
"""
Constructs the computational graph of this layer.
If the input is ints, we assume is an index, otherwise we assume is
a set of floats.
"""
if self.weight_noise and use_noise and self.noise_params:
W_ems = [(x + y) for x, y in zip(self.W_ems, self.nW_ems)]
if not no_noise_bias:
b_ems = [(x + y) for x, y in zip(self.b_ems, self.nb_ems)]
else:
b_ems = self.b_ems
else:
W_ems = self.W_ems
b_ems = self.b_ems
if self.rank_n_approx:
if first_only:
emb_val = self.rank_n_activ(utils.dot(state_below, W_ems[0]))
self.out = emb_val
return emb_val
emb_val = TT.dot(
self.rank_n_activ(utils.dot(state_below, W_ems[0])), W_ems[1])
if b_ems:
emb_val += b_ems[0]
st_pos = 1
else:
emb_val = utils.dot(state_below, W_ems[0])
if b_ems:
emb_val += b_ems[0]
st_pos = 0
emb_val = self.activation[0](emb_val)
if self.dropout < 1.:
if use_noise:
emb_val = emb_val * self.trng.binomial(
emb_val.shape, n=1, p=self.dropout, dtype=emb_val.dtype)
else:
emb_val = emb_val * self.dropout
for dx in xrange(1, self.n_layers):
emb_val = utils.dot(emb_val, W_ems[st_pos + dx])
if b_ems:
emb_val = self.activation[dx](emb_val + b_ems[dx])
else:
emb_val = self.activation[dx](emb_val)
if self.dropout < 1.:
if use_noise:
emb_val = emb_val * self.trng.binomial(emb_val.shape,
n=1,
p=self.dropout,
dtype=emb_val.dtype)
else:
emb_val = emb_val * self.dropout
self.out = emb_val
return emb_val
def fprop(self, state_below, use_noise=True, no_noise_bias=False,
first_only = False):
"""
Constructs the computational graph of this layer.
If the input is ints, we assume is an index, otherwise we assume is
a set of floats.
"""
print 'multilayer use noise:', use_noise
if self.weight_noise and use_noise and self.noise_params:
W_ems = [(x+y) for x, y in zip(self.W_ems, self.nW_ems)]
if not no_noise_bias:
b_ems = [(x+y) for x, y in zip(self.b_ems, self.nb_ems)]
else:
b_ems = self.b_ems
else:
W_ems = self.W_ems
b_ems = self.b_ems
if self.rank_n_approx:
if first_only:
emb_val = self.rank_n_activ(utils.dot(state_below, W_ems[0]))
self.out = emb_val
return emb_val
emb_val = TT.dot(
self.rank_n_activ(utils.dot(state_below, W_ems[0])),
W_ems[1])
if b_ems:
emb_val += b_ems[0]
st_pos = 1
else:
emb_val = utils.dot(state_below, W_ems[0])
if b_ems:
emb_val += b_ems[0]
st_pos = 0
emb_val = self.activation[0](emb_val)
if self.dropout < 1.:
if use_noise:
print 'training use noise'
emb_val = emb_val * self.trng.binomial(emb_val.shape, n=1, p=self.dropout, dtype=emb_val.dtype)
else:
print 'decoding not use noise'
emb_val = emb_val * self.dropout
for dx in xrange(1, self.n_layers):
emb_val = utils.dot(emb_val, W_ems[st_pos+dx])
if b_ems:
emb_val = self.activation[dx](emb_val+ b_ems[dx])
else:
emb_val = self.activation[dx](emb_val)
if self.dropout < 1.:
if use_noise:
emb_val = emb_val * self.trng.binomial(emb_val.shape, n=1, p=self.dropout, dtype=emb_val.dtype)
else:
emb_val = emb_val * self.dropout
self.out = emb_val
return emb_val
def fprop(self, state_below, use_noise=True, no_noise_bias=False,
first_only = False):
"""
Constructs the computational graph of this layer.
If the input is ints, we assume is an index, otherwise we assume is
a set of floats.
"""
if self.weight_noise and use_noise and self.noise_params:
W_ems = [(x+y) for x, y in zip(self.W_ems, self.nW_ems)]
if not no_noise_bias:
b_ems = [(x+y) for x, y in zip(self.b_ems, self.nb_ems)]
else:
b_ems = self.b_ems
else:
W_ems = self.W_ems
b_ems = self.b_ems
#FIXME one bias for the whole layer? or we need different biases for each component
emb_val1 = utils.dot(state_below, W_ems[0])
emb_val2 = utils.dot(state_below, W_ems[1])
if b_ems:
emb_val1 += b_ems[0]
emb_val2 += b_ems[1]
emb_val1 = self.activation[0](emb_val1)
emb_val2 = self.activation[0](emb_val2)
emb_val = emb_val1 * emb_val2
#FIXME make sure how the dropout for tensor networks works
if self.dropout < 1.:
if use_noise:
emb_val = emb_val * self.trng.binomial(emb_val.shape, n=1, p=self.dropout, dtype=emb_val.dtype)
else:
emb_val = emb_val * self.dropout
for dx in xrange(1, self.n_layers):
emb_val1 = utils.dot(emb_val, W_ems[2*dx])
emb_val2 = utils.dot(emb_val, W_ems[2*dx+1])
if b_ems:
emb_val1 = emb_val1+ b_ems[2*dx]
emb_val2 = emb_val2+ b_ems[2*dx+1]
emb_val1 = self.activation[dx](emb_val1)
emb_val2 = self.activation[dx](emb_val2)
emb_val = emb_val1 * emb_val2
#FIXME make sure how the dropout for tensor networks works
if self.dropout < 1.:
if use_noise:
emb_val = emb_val * self.trng.binomial(emb_val.shape, n=1, p=self.dropout, dtype=emb_val.dtype)
else:
emb_val = emb_val * self.dropout
self.out = emb_val
return emb_val
def fprop(self,
state_below,
temp=numpy.float32(1),
use_noise=True,
additional_inputs=None,
no_noise_bias=False):
"""
Forward pass through the cost layer.
:type state_below: tensor or layer
:param state_below: The theano expression (or groundhog layer)
representing the input of the cost layer
:type temp: float or tensor scalar
:param temp: scalar representing the temperature that should be used
when sampling from the output distribution
:type use_noise: bool
:param use_noise: flag. If true, noise is used when computing the
output of the model
:type no_noise_bias: bool
:param no_noise_bias: flag, stating if weight noise should be added
to the bias as well, or only to the weights
"""
if self.rank_n_approx:
if use_noise and self.noise_params:
emb_val = self.rank_n_activ(
utils.dot(state_below, self.W_em1 + self.nW_em1))
emb_val = TT.dot(self.W_em2 + self.nW_em2, emb_val)
else:
emb_val = self.rank_n_activ(utils.dot(state_below, self.W_em1))
emb_val = TT.dot(self.W_em2, emb_val)
else:
if use_noise and self.noise_params:
emb_val = utils.dot(state_below, self.W_em + self.nW_em)
else:
emb_val = utils.dot(state_below, self.W_em)
if additional_inputs:
if use_noise and self.noise_params:
for inp, weight, noise_weight in zip(
additional_inputs, self.additional_weights,
self.noise_additional_weights):
emb_val += utils.dot(inp, (noise_weight + weight))
else:
for inp, weight in zip(additional_inputs,
self.additional_weights):
emb_val += utils.dot(inp, weight)
self.preactiv = emb_val
if use_noise and self.noise_params and not no_noise_bias:
emb_val = TT.nnet.sigmoid(temp *
(emb_val + self.b_em + self.nb_em))
else:
emb_val = TT.nnet.sigmoid(temp * (emb_val + self.b_em))
self.out = emb_val
self.state_below = state_below
self.model_output = emb_val
return emb_val
def fprop(self,
state_below,
temp=numpy.float32(1),
use_noise=True,
additional_inputs=None,
no_noise_bias=False):
"""
Forward pass through the cost layer.
:type state_below: tensor or layer
:param state_below: The theano expression (or groundhog layer)
representing the input of the cost layer
:type temp: float or tensor scalar
:param temp: scalar representing the temperature that should be used
when sampling from the output distribution
:type use_noise: bool
:param use_noise: flag. If true, noise is used when computing the
output of the model
:type no_noise_bias: bool
:param no_noise_bias: flag, stating if weight noise should be added
to the bias as well, or only to the weights
"""
if self.rank_n_approx:
if use_noise and self.noise_params:
emb_val = self.rank_n_activ(utils.dot(state_below,
self.W_em1 + self.nW_em1))
emb_val = TT.dot(self.W_em2 + self.nW_em2, emb_val)
else:
emb_val = self.rank_n_activ(utils.dot(state_below, self.W_em1))
emb_val = TT.dot(self.W_em2, emb_val)
else:
if use_noise and self.noise_params:
emb_val = utils.dot(state_below, self.W_em + self.nW_em)
else:
emb_val = utils.dot(state_below, self.W_em)
if additional_inputs:
if use_noise and self.noise_params:
for inp, weight, noise_weight in zip(
additional_inputs, self.additional_weights,
self.noise_additional_weights):
emb_val += utils.dot(inp, (noise_weight + weight))
else:
for inp, weight in zip(additional_inputs, self.additional_weights):
emb_val += utils.dot(inp, weight)
self.preactiv = emb_val
if use_noise and self.noise_params and not no_noise_bias:
emb_val = TT.nnet.sigmoid(temp *
(emb_val + self.b_em + self.nb_em))
else:
emb_val = TT.nnet.sigmoid(temp * (emb_val + self.b_em))
self.out = emb_val
self.state_below = state_below
self.model_output = emb_val
return emb_val
def fprop(self,
state_below,
temp=numpy.float32(1),
use_noise=True,
additional_inputs=None,
no_noise_bias=False,
target=None,
full_softmax=True):
"""
Forward pass through the cost layer.
:type state_below: tensor or layer
:param state_below: The theano expression (or groundhog layer)
representing the input of the cost layer
:type temp: float or tensor scalar
:param temp: scalar representing the temperature that should be used
when sampling from the output distribution
:type use_noise: bool
:param use_noise: flag. If true, noise is used when computing the
output of the model
:type no_noise_bias: bool
:param no_noise_bias: flag, stating if weight noise should be added
to the bias as well, or only to the weights
"""
if not full_softmax:
assert target is not None, 'target must be given'
if self.rank_n_approx:
if self.weight_noise and use_noise and self.noise_params:
emb_val = self.rank_n_activ(utils.dot(state_below,
self.W_em1 + self.nW_em1))
nW_em = self.nW_em2
else:
emb_val = self.rank_n_activ(utils.dot(state_below, self.W_em1))
W_em = self.W_em2
else:
W_em = self.W_em
if self.weight_noise:
nW_em = self.nW_em
emb_val = state_below
if full_softmax:
if self.weight_noise and use_noise and self.noise_params:
emb_val = TT.dot(emb_val, W_em + nW_em)
else:
emb_val = TT.dot(emb_val, W_em)
if additional_inputs:
if use_noise and self.noise_params:
for inp, weight, noise_weight in zip(
additional_inputs, self.additional_weights,
self.noise_additional_weights):
emb_val += utils.dot(inp, (noise_weight + weight))
else:
for inp, weight in zip(additional_inputs, self.additional_weights):
emb_val += utils.dot(inp, weight)
if self.weight_noise and use_noise and self.noise_params and \
not no_noise_bias:
emb_val = temp * (emb_val + self.b_em + self.nb_em)
else:
emb_val = temp * (emb_val + self.b_em)
else:
W_em = W_em[:, target]
if self.weight_noise:
nW_em = nW_em[:, target]
W_em += nW_em
if emb_val.ndim == 3:
emb_val = emb_val.reshape([emb_val.shape[0] * emb_val.shape[1], emb_val.shape[2]])
emb_val = (W_em.T * emb_val).sum(1) + self.b_em[target]
if self.weight_noise and use_noise:
emb_val += self.nb_em[target]
emb_val = temp * emb_val
self.preactiv = emb_val
if full_softmax:
emb_val = utils.softmax(emb_val)
else:
emb_val = TT.nnet.sigmoid(emb_val)
self.out = emb_val
self.state_below = state_below
self.model_output = emb_val
return emb_val
def fprop(self,
state_below,
temp=numpy.float32(1),
use_noise=True,
additional_inputs=None,
no_noise_bias=False,
target=None,
full_softmax=True):
"""
Forward pass through the cost layer.
:type state_below: tensor or layer
:param state_below: The theano expression (or groundhog layer)
representing the input of the cost layer
:type temp: float or tensor scalar
:param temp: scalar representing the temperature that should be used
when sampling from the output distribution
:type use_noise: bool
:param use_noise: flag. If true, noise is used when computing the
output of the model
:type no_noise_bias: bool
:param no_noise_bias: flag, stating if weight noise should be added
to the bias as well, or only to the weights
"""
if not full_softmax:
assert target != None, 'target must be given'
if self.rank_n_approx:
if self.weight_noise and use_noise and self.noise_params:
emb_val = self.rank_n_activ(
utils.dot(state_below, self.W_em1 + self.nW_em1))
nW_em = self.nW_em2
else:
emb_val = self.rank_n_activ(utils.dot(state_below, self.W_em1))
W_em = self.W_em2
else:
W_em = self.W_em
if self.weight_noise:
nW_em = self.nW_em
emb_val = state_below
if full_softmax:
if self.weight_noise and use_noise and self.noise_params:
emb_val = TT.dot(emb_val, W_em + nW_em)
else:
emb_val = TT.dot(emb_val, W_em)
if additional_inputs:
if use_noise and self.noise_params:
for inp, weight, noise_weight in zip(
additional_inputs, self.additional_weights,
self.noise_additional_weights):
emb_val += utils.dot(inp, (noise_weight + weight))
else:
for inp, weight in zip(additional_inputs,
self.additional_weights):
emb_val += utils.dot(inp, weight)
if self.weight_noise and use_noise and self.noise_params and \
not no_noise_bias:
emb_val = temp * (emb_val + self.b_em + self.nb_em)
else:
emb_val = temp * (emb_val + self.b_em)
else:
W_em = W_em[:, target]
if self.weight_noise:
nW_em = nW_em[:, target]
W_em += nW_em
if emb_val.ndim == 3:
emb_val = emb_val.reshape(
[emb_val.shape[0] * emb_val.shape[1], emb_val.shape[2]])
emb_val = (W_em.T * emb_val).sum(1) + self.b_em[target]
if self.weight_noise and use_noise:
emb_val += self.nb_em[target]
emb_val = temp * emb_val
self.preactiv = emb_val
if full_softmax:
emb_val = utils.softmax(emb_val)
else:
emb_val = TT.nnet.sigmoid(emb_val)
self.out = emb_val
self.state_below = state_below
self.model_output = emb_val
return emb_val
def fprop(self,
state_below,
mask=None,
init_state=None,
gater_below=None,
reseter_below=None,
c=None,
c_mask=None,
nsteps=None,
batch_size=None,
use_noise=True,
truncate_gradient=-1,
no_noise_bias=False,
return_alignment=False):
updater_below = gater_below
if theano.config.floatX=='float32':
floatX = numpy.float32
else:
floatX = numpy.float64
if nsteps is None:
nsteps = state_below.shape[0]
if batch_size and batch_size != 1:
nsteps = nsteps / batch_size
if batch_size is None and state_below.ndim == 3:
batch_size = state_below.shape[1]
if state_below.ndim == 2 and \
(not isinstance(batch_size,int) or batch_size > 1):
state_below = state_below.reshape((nsteps, batch_size, self.n_in))
if updater_below:
updater_below = updater_below.reshape((nsteps, batch_size, self.n_in))
if reseter_below:
reseter_below = reseter_below.reshape((nsteps, batch_size, self.n_in))
if not init_state:
if not isinstance(batch_size, int) or batch_size != 1:
init_state = TT.alloc(floatX(0), batch_size, self.n_hids)
else:
init_state = TT.alloc(floatX(0), self.n_hids)
p_from_c = utils.dot(c, self.A_cp).reshape(#attention weights for each hidden state
(c.shape[0], c.shape[1], self.n_hids))
if mask:
sequences = [state_below, mask, updater_below, reseter_below]
non_sequences = [c, c_mask, p_from_c]
# seqs | out | non_seqs
fn = lambda x, m, g, r, h, c1, cm, pc : self.step_fprop(x, h, mask=m,
gater_below=g, reseter_below=r,
c=c1, p_from_c=pc, c_mask=cm,
use_noise=use_noise, no_noise_bias=no_noise_bias,
return_alignment=return_alignment)
else:
sequences = [state_below, updater_below, reseter_below]
non_sequences = [c, p_from_c]
# seqs | out | non_seqs
fn = lambda x, g, r, h, c1, pc : self.step_fprop(x, h,
gater_below=g, reseter_below=r,
c=c1, p_from_c=pc,
use_noise=use_noise, no_noise_bias=no_noise_bias,
return_alignment=return_alignment)
outputs_info = [init_state, None]
if return_alignment:
outputs_info.append(None)
#use scan to repetitively update hidden states
rval, updates = theano.scan(fn,
sequences=sequences,
non_sequences=non_sequences,
outputs_info=outputs_info,
name='layer_%s'%self.name,
truncate_gradient=truncate_gradient,
n_steps=nsteps)
self.out = rval
self.rval = rval
self.updates = updates
return self.out
def step_fprop(self,
state_below,
state_before,
gater_below=None,
reseter_below=None,
mask=None,
c=None,
c_mask=None,
p_from_c=None,
use_noise=True,
no_noise_bias=False,
step_num=None,
return_alignment=False):
"""
Constructs the computational graph of this layer.
:type state_below: theano variable
:param state_below: the input to the layer
:type mask: None or theano variable
:param mask: mask describing the length of each sequence in a minibatch
:type state_before: theano variable
:param state_before: the previous value of the hidden state of the layer
:type updater_below: theano variable
:param updater_below: the input to the update gate
:type reseter_below: theano variable
:param reseter_below: the input to the reset gate
:type use_noise: bool
:param use_noise: flag saying if weight noise should be used in
computing the output of this layer
:type no_noise_bias: bool
:param no_noise_bias: flag saying if weight noise should be added to
the bias as well
"""
updater_below = gater_below
W_hh = self.W_hh
G_hh = self.G_hh
R_hh = self.R_hh
A_cp = self.A_cp
B_hp = self.B_hp
D_pe = self.D_pe
# The code works only with 3D tensors
cndim = c.ndim
if cndim == 2:
c = c[:, None, :]
# Warning: either source_num or target_num should be equal,
# or one of them sould be 1 (they have to broadcast)
# for the following code to make any sense.
source_len = c.shape[0]#sequence length
source_num = c.shape[1]#number of sequences in a batch
target_num = state_before.shape[0]
dim = self.n_hids
# Form projection to the tanh layer from the previous hidden state
# Shape: (source_len, target_num, dim)
p_from_h = ReplicateLayer(source_len)(utils.dot(state_before, B_hp)).out
# Form projection to the tanh layer from the source annotation.
if not p_from_c:
p_from_c = utils.dot(c, A_cp).reshape((source_len, source_num, dim))
# Sum projections - broadcasting happens at the dimension 1.
p = p_from_h + p_from_c
# Apply non-linearity and project to energy.
energy = TT.exp(utils.dot(TT.tanh(p), D_pe)).reshape((source_len, target_num))
if c_mask:
# This is used for batches only, that is target_num == source_num
energy *= c_mask
# Calculate energy sums.
normalizer = energy.sum(axis=0)
# Get probabilities. (softmax?)
probs = energy / normalizer
# Calculate weighted sums of source annotations.
# If target_num == 1, c shoulds broadcasted at the 1st dimension.
# Probabilities are broadcasted at the 2nd dimension.
ctx = (c * probs.dimshuffle(0, 1, 'x')).sum(axis=0)#averaged context (see the picture)
state_below += self.c_inputer(ctx).out
reseter_below += self.c_reseter(ctx).out
updater_below += self.c_updater(ctx).out
# Reset gate:
# optionally reset the hidden state.
reseter = self.reseter_activation(TT.dot(state_before, R_hh)+reseter_below)
reseted_state_before = reseter * state_before
# Feed the input to obtain potential new state.
preactiv = TT.dot(reseted_state_before, W_hh) + state_below
h = self.activation(preactiv)
# Update gate:
# optionally reject the potential new state and use the new one.
updater = self.updater_activation(TT.dot(state_before, G_hh) + updater_below)
h = updater * h + (1-updater) * state_before #h_t=z*h_{t-1}+(1-z)*h_t
if mask is not None:
if h.ndim ==2 and mask.ndim==1:
mask = mask.dimshuffle(0,'x')
h = mask * h + (1-mask) * state_before
results = [h, ctx]
if return_alignment:
results += [probs]
return results