def __init__(self,
style_number,
style_dim,
hidden_status_dim,
tparams=None,
prefix='style'):
"""
Init the style parameter: init_params.
"""
self.W_BBeta = theano.shared(value=util.uniform_random_weight(
(style_dim, style_dim), 0.01),
name=self._p(prefix, 'W_BBeta'),
borrow=True)
self.W_hBeta = theano.shared(value=util.uniform_random_weight(
(hidden_status_dim, style_dim), 0.01),
name=self._p(prefix, 'W_hBeta'),
borrow=True)
self.B = theano.shared(value=util.uniform_random_weight(
(style_number, style_dim), 0.01),
name=self._p(prefix, 'B'),
borrow=True)
self.vb = theano.shared(util.uniform_random_weight((style_dim, ),
0.01),
name=self._p(prefix, 'vb'),
borrow=True)
# parameters of the model
self.params = [self.W_BBeta, self.W_hBeta, self.B, self.vb]
if not tparams is None:
tparams[self._p(prefix, 'W_BBeta')] = self.W_BBeta
tparams[self._p(prefix, 'W_hBeta')] = self.W_hBeta
tparams[self._p(prefix, 'B')] = self.B
tparams[self._p(prefix, 'vb')] = self.vb
else:
print " tparams is None"
def __init__(self, style_number, style_dim, hidden_status_dim, tparams=None, prefix='style'):
"""
Init the style parameter: init_params.
"""
self.W_BBeta = theano.shared(
value=util.uniform_random_weight((style_dim, style_dim), 0.1),
name=self._p(prefix, 'W_BBeta'),
borrow=True
)
self.W_hBeta = theano.shared(
value=util.uniform_random_weight((hidden_status_dim, style_dim), 0.1),
name=self._p(prefix, 'W_hBeta'),
borrow=True
)
self.B = theano.shared(
value=util.uniform_random_weight((style_number, style_dim), 0.1),
name=self._p(prefix, 'B'),
borrow=True
)
self.vb = theano.shared(
util.uniform_random_weight((style_dim,), 0.1),
name=self._p(prefix, 'vb'),
borrow=True
)
# parameters of the model
self.params = [self.W_BBeta, self.W_hBeta, self.B, self.vb]
if not tparams is None:
tparams[self._p(prefix, 'W_BBeta')] = self.W_BBeta
tparams[self._p(prefix, 'W_hBeta')] = self.W_hBeta
tparams[self._p(prefix, 'B')] = self.B
tparams[self._p(prefix, 'vb')] = self.vb
else:
print " tparams is None"
def __init__(self, base_dim, refer_dim, tparams, prefix="maxout"):
self.W_t = theano.shared(
value=util.uniform_random_weight((refer_dim, 2 * refer_dim), 0.1),
name=self._p(prefix, 'W_t'),
borrow=True
)
self.W_o = theano.shared(
value=util.uniform_random_weight((base_dim, refer_dim), 0.1),
name=self._p(prefix, 'W_o'),
borrow=True
)
# initialize the biases b as a vector of n_out 0s
self.b = theano.shared(
value=numpy.zeros(
(2 * refer_dim,),
dtype=config.globalFloatType()
),
name=self._p(prefix, 'b'),
borrow=True
)
# parameters of the model
self.params = [self.W_t, self.W_o, self.b]
if not tparams is None:
tparams[self._p(prefix, 'W_t')] = self.W_t
tparams[self._p(prefix, 'W_o')] = self.W_o
tparams[self._p(prefix, 'b')] = self.b
else:
print " tparams is None"
def __init__(self, word_embedding_dim, hidden_status_dim, tparams=None, prefix='GRU'):
"""
Init the GRU parameter: init_params.
Updation in GRU :
step1. r(t) = f(W_r dot x(t) + U_r dot h(t-1) + C_r dot h_last).
step2. z(t) = f(W_z dot x(t) + U_z dot h(t-1) + C_z dot h_last).
step3. h_wave(t) = f(W dot x(t) + U dot (r(t) * h(t-1)) + C dot h_last).
step4. h(t) = (1-z(t)) * h(t-1) + z(t) * h_wave(t).
We can combine W and C into one tensor W
"""
self.hidden_status_dim = hidden_status_dim
self.params = tparams
self.prefix = prefix
self.word_embedding_dim = word_embedding_dim
W_bound = 0.1
# combine step1~3 W dot t, so W's dimension is (word_embedding_dim, hidden_status_dim, 3)
W = uniform_random_weight(size=(word_embedding_dim, 3 * hidden_status_dim), bound=W_bound)
# combine step1~2 U dot h, so U's dimension is (hidden_status_dim, 2)
# connot combine step1~3, so split U_rh
U = uniform_random_weight(size=(hidden_status_dim, 2 * hidden_status_dim), bound=W_bound)
U_rh = uniform_random_weight(size=(hidden_status_dim, hidden_status_dim), bound=W_bound)
U_union = uniform_random_weight(size=(3 * hidden_status_dim, hidden_status_dim), bound=W_bound)
if tparams is not None:
tparams[self._p(prefix, 'W')] = theano.shared(W, name=self._p(prefix, 'W'))
tparams[self._p(prefix, 'U')] = theano.shared(U, name=self._p(prefix, 'U'))
tparams[self._p(prefix, 'U_rh')] = theano.shared(U_rh, name=self._p(prefix, 'U_rh'))
tparams[self._p(prefix, 'U_union')] = theano.shared(U_union, name=self._p(prefix, 'U_union'))
else:
print ' tparams is None'
def __init__(self,
word_embedding_dim,
hidden_status_dim,
tparams=None,
prefix='RNN'):
"""
Init the RNN parameter: init_params.
Updation in GRU :
step1. h(t) = f(W_r dot x(t) + U_r dot h(t-1)).
We can combine W and C into one tensor W
"""
self.hidden_status_dim = hidden_status_dim
self.params = tparams
self.prefix = prefix
W_bound = 0.01
# combine step1~3 W dot t, so W's dimension is (word_embedding_dim, hidden_status_dim)
W = uniform_random_weight(size=(word_embedding_dim, hidden_status_dim),
bound=W_bound)
# combine step1~2 U dot h, so U's dimension is (hidden_status_dim, 2)
# connot combine step1~3, so split U_rh
U = numpy.concatenate([ortho_weight(hidden_status_dim)], axis=1)
if tparams is not None:
tparams[self._p(prefix,
'W')] = theano.shared(W, name=self._p(prefix, 'W'))
tparams[self._p(prefix,
'U')] = theano.shared(U, name=self._p(prefix, 'U'))
else:
print 'tparams is None'
def __init__(self, word_embedding_dim, hidden_status_dim, encoder_hidden_dim,
tparams=None, prefix='Attention'):
"""
Init the GRU parameter: init_params.
Updation in GRU :
step1. r(t) = f(W_r dot x(t) + U_r dot h(t-1) + C_r dot h_last).
step2. z(t) = f(W_z dot x(t) + U_z dot h(t-1) + C_z dot h_last).
step3. h_wave(t) = f(W dot x(t) + U dot (r(t) * h(t-1)) + C dot h_last).
step4. h(t) = (1-z(t)) * h(t-1) + z(t) * h_wave(t).
We can combine W and C into one tensor W
"""
self.hidden_status_dim = hidden_status_dim
self.params = tparams
self.prefix = prefix
W_bound = numpy.sqrt(6. / (hidden_status_dim))
# combine step1~3 W dot t, so W's dimension is (word_embedding_dim, hidden_status_dim, 3)
W = uniform_random_weight(size=(hidden_status_dim, hidden_status_dim), bound=W_bound)
# combine step1~2 U dot h, so U's dimension is (hidden_status_dim, 2)
# connot combine step1~3, so split U_rh
U = numpy.concatenate([ortho_weight(hidden_status_dim)]*int(encoder_hidden_dim/hidden_status_dim),
axis=0)
# U = uniform_random_weight(height=2*hidden_status_dim, width=hidden_status_dim, bound=W_bound)
va = numpy.zeros((hidden_status_dim,), dtype=globalFloatType())
if tparams is not None:
tparams[self._p(prefix, 'W')] = theano.shared(W, name=self._p(prefix, 'W'))
tparams[self._p(prefix, 'U')] = theano.shared(U, name=self._p(prefix, 'U'))
tparams[self._p(prefix, 'va')] = theano.shared(va, name=self._p(prefix, 'va'))
else:
print ' tparams is None'
def __init__(self, word_embedding_dim, hidden_status_dim, encoder_hidden_dim,
tparams=None, prefix='Attention'):
"""
Init the GRU parameter: init_params.
Updation in GRU :
step1. r(t) = f(W_r dot x(t) + U_r dot h(t-1) + C_r dot h_last).
step2. z(t) = f(W_z dot x(t) + U_z dot h(t-1) + C_z dot h_last).
step3. h_wave(t) = f(W dot x(t) + U dot (r(t) * h(t-1)) + C dot h_last).
step4. h(t) = (1-z(t)) * h(t-1) + z(t) * h_wave(t).
We can combine W and C into one tensor W
"""
self.hidden_status_dim = hidden_status_dim
self.params = tparams
self.prefix = prefix
W_bound = 0.01
# combine step1~3 W dot t, so W's dimension is (word_embedding_dim, hidden_status_dim, 3)
W = uniform_random_weight(size=(hidden_status_dim, hidden_status_dim), bound=W_bound)
# combine step1~2 U dot h, so U's dimension is (hidden_status_dim, 2)
# connot combine step1~3, so split U_rh
U = numpy.concatenate([ortho_weight(hidden_status_dim)]*int(encoder_hidden_dim/hidden_status_dim),
axis=0)
# U = uniform_random_weight(height=2*hidden_status_dim, width=hidden_status_dim, bound=W_bound)
va = numpy.zeros((hidden_status_dim,), dtype=globalFloatType())
if tparams is not None:
tparams[self._p(prefix, 'W')] = theano.shared(W, name=self._p(prefix, 'W'))
tparams[self._p(prefix, 'U')] = theano.shared(U, name=self._p(prefix, 'U'))
tparams[self._p(prefix, 'va')] = theano.shared(va, name=self._p(prefix, 'va'))
else:
print ' tparams is None'
def __init__(self, word_embedding_dim, hidden_status_dim, tparams=None, prefix='GRU'):
"""
Init the GRU parameter: init_params.
Updation in GRU :
step1. r(t) = f(W_r dot x(t) + U_r dot h(t-1) + C_r dot h_last).
step2. z(t) = f(W_z dot x(t) + U_z dot h(t-1) + C_z dot h_last).
step3. h_wave(t) = f(W dot x(t) + U dot (r(t) * h(t-1)) + C dot h_last).
step4. h(t) = (1-z(t)) * h(t-1) + z(t) * h_wave(t).
We can combine W and C into one tensor W
"""
self.hidden_status_dim = hidden_status_dim
self.params = tparams
self.prefix = prefix
W_bound = numpy.sqrt(6. / (hidden_status_dim + word_embedding_dim))
# combine step1~3 W dot t, so W's dimension is (word_embedding_dim, hidden_status_dim, 3)
W = uniform_random_weight(size=(word_embedding_dim, 3 * hidden_status_dim), bound=W_bound)
# combine step1~2 U dot h, so U's dimension is (hidden_status_dim, 2)
# connot combine step1~3, so split U_rh
U = numpy.concatenate([ortho_weight(hidden_status_dim),
ortho_weight(hidden_status_dim)], axis=1)
U_rh = ortho_weight(hidden_status_dim)
if tparams is not None:
tparams[self._p(prefix, 'W')] = theano.shared(W, name=self._p(prefix, 'W'))
tparams[self._p(prefix, 'U')] = theano.shared(U, name=self._p(prefix, 'U'))
tparams[self._p(prefix, 'U_rh')] = theano.shared(U_rh, name=self._p(prefix, 'U_rh'))
else:
print ' tparams is None'
def __init__(self, base_dim, refer_dim, tparams, prefix="maxout"):
self.W_t = theano.shared(value=util.uniform_random_weight(
(refer_dim, 2 * refer_dim), 0.1),
name=self._p(prefix, 'W_t'),
borrow=True)
self.W_o = theano.shared(value=util.uniform_random_weight(
(base_dim, refer_dim), 0.1),
name=self._p(prefix, 'W_o'),
borrow=True)
# parameters of the model
self.params = [self.W_t, self.W_o]
if not tparams is None:
tparams[self._p(prefix, 'W_t')] = self.W_t
tparams[self._p(prefix, 'W_o')] = self.W_o
else:
print " tparams is None"
def __init__(self,
word_embedding_dim,
hidden_status_dim,
encoder_hidden_dim,
tparams,
prefix='Decoder',
node_type=GRUNode):
"""
Init the Decoder parameter: init_params.
"""
self.params = tparams
W_bound = numpy.sqrt(6. / (hidden_status_dim))
W = uniform_random_weight(size=(hidden_status_dim, hidden_status_dim),
bound=W_bound)
# if tparams is not None:
# tparams[self._p(prefix, 'Ws')] = theano.shared(W, name=self._p(prefix, 'Ws'))
self.hidden_status_dim = hidden_status_dim
self.prefix = prefix
self.node = node_type(word_embedding_dim=word_embedding_dim,
hidden_status_dim=hidden_status_dim,
tparams=tparams,
prefix=self._p(self.prefix, 'GRU'))
self.attention_node = AttentionNode(
word_embedding_dim=word_embedding_dim,
hidden_status_dim=hidden_status_dim,
encoder_hidden_dim=encoder_hidden_dim,
tparams=tparams,
prefix=self._p(self.prefix, 'Attention'))
def __init__(self, base_dim, refer_dim, tparams, prefix="maxout"):
self.W_t = theano.shared(
value=util.uniform_random_weight((refer_dim, 2 * refer_dim), 0.1),
name=self._p(prefix, 'W_t'),
borrow=True
)
self.W_o = theano.shared(
value=util.uniform_random_weight((base_dim, refer_dim), 0.1),
name=self._p(prefix, 'W_o'),
borrow=True
)
# parameters of the model
self.params = [self.W_t, self.W_o]
if not tparams is None:
tparams[self._p(prefix, 'W_t')] = self.W_t
tparams[self._p(prefix, 'W_o')] = self.W_o
else:
print " tparams is None"
def __init__(self, n_words, word_embedding_dim, represent, isfirst, tparams, prefix):
"""
Init the Memory parameter: init_params.
"""
self.word_embedding_dim = word_embedding_dim
self.params = tparams
self.prefix = prefix
EmbA = uniform_random_weight(size=(n_words, word_embedding_dim), bound=0.01)
EmbB = uniform_random_weight(size=(n_words, word_embedding_dim), bound=0.01)
EmbC = uniform_random_weight(size=(n_words, word_embedding_dim), bound=0.01)
TemA = uniform_random_weight(size=(500, word_embedding_dim), bound=0.01)
TemC = uniform_random_weight(size=(500, word_embedding_dim), bound=0.01)
if tparams is not None:
tparams[self._p(prefix, 'EmbA')] = theano.shared(EmbA, name=self._p(prefix, 'EmbA'))
if isfirst == True :
tparams[self._p(prefix, 'EmbB')] = theano.shared(EmbB, name=self._p(prefix, 'EmbB'))
tparams[self._p(prefix, 'EmbC')] = theano.shared(EmbC, name=self._p(prefix, 'EmbC'))
if represent == 'PE' :
tparams[self._p(prefix, 'TemA')] = theano.shared(TemA, name=self._p(prefix, 'TemA'))
tparams[self._p(prefix, 'TemC')] = theano.shared(TemC, name=self._p(prefix, 'TemC'))
def __init__(self, base_dim, refer_dim, tparams, prefix="maxout"):
self.W_t = theano.shared(value=util.uniform_random_weight(
(refer_dim, 2 * refer_dim), 0.01),
name=self._p(prefix, 'W_t'),
borrow=True)
self.W_o = theano.shared(value=util.uniform_random_weight(
(base_dim, refer_dim), 0.01),
name=self._p(prefix, 'W_o'),
borrow=True)
# initialize the biases b as a vector of n_out 0s
self.b = theano.shared(value=numpy.zeros(
(2 * refer_dim, ), dtype=config.globalFloatType()),
name=self._p(prefix, 'b'),
borrow=True)
# parameters of the model
self.params = [self.W_t, self.W_o, self.b]
if not tparams is None:
tparams[self._p(prefix, 'W_t')] = self.W_t
tparams[self._p(prefix, 'W_o')] = self.W_o
tparams[self._p(prefix, 'b')] = self.b
else:
print " tparams is None"
def __init__(self, word_embedding_dim, hidden_status_dim, encoder_hidden_dim,
tparams, prefix='Decoder', node_type=GRUNode):
"""
Init the Decoder parameter: init_params.
"""
self.params = tparams
W_bound = numpy.sqrt(6. / (hidden_status_dim))
W = uniform_random_weight(size=(hidden_status_dim, hidden_status_dim), bound=W_bound)
# if tparams is not None:
# tparams[self._p(prefix, 'Ws')] = theano.shared(W, name=self._p(prefix, 'Ws'))
self.hidden_status_dim = hidden_status_dim
self.prefix = prefix
self.node = node_type(word_embedding_dim=word_embedding_dim,
hidden_status_dim=hidden_status_dim,
tparams=tparams, prefix=self._p(self.prefix, 'GRU'))
self.attention_node = AttentionNode(word_embedding_dim=word_embedding_dim,
hidden_status_dim=hidden_status_dim,
encoder_hidden_dim = encoder_hidden_dim,
tparams=tparams, prefix=self._p(self.prefix, 'Attention'))