def __init__(self, n_in, n_hidden):
self.n_in = int(n_in)
self.n_hidden = int(n_hidden)
self.input = T.tensor3()
self.W_hh = glorot_uniform((n_hidden, n_hidden))
self.W_in = glorot_uniform((n_in, n_hidden))
self.bh = zero((n_hidden, ))
self.params = [self.W_hh, self.W_in, self.bh]
self.L1 = T.sum(abs(self.W_hh)) + T.sum(abs(self.W_in))
self.L2_sqr = T.sum(self.W_hh**2) + T.sum(self.W_in**2)
def __init__(self,
n_in,
n_hidden,
n_decoder,
n_out,
lr=0.001,
n_epochs=400,
n_batch=16,
maxlen=20,
L1_reg=0,
L2_reg=0):
self.n_in = int(n_in)
self.n_hidden = int(n_hidden)
self.n_decoder = int(n_decoder)
self.n_out = int(n_out)
self.lr = float(lr)
self.n_batch = int(n_batch)
self.maxlen = int(maxlen)
self.x = T.tensor3(name='x', dtype=theano.config.floatX)
self.y = T.tensor3(name='y', dtype=theano.config.floatX)
self.W_hy = glorot_uniform((self.n_decoder, self.n_out))
self.b_hy = zero((n_out, ))
self.W_hi = glorot_uniform((self.n_hidden, self.n_decoder))
self.b_hi = zero((n_decoder, ))
self.layers = []
self.decoder = []
self.params = []
self.errors = []
self.n_epochs = n_epochs
self.initial_momentum = 0.5
self.final_momentum = 0.9
self.momentum_switchover = 5
self.learning_rate_decay = 0.999
self.updates = {}
self.n_layers = 0
self.L1_reg = L1_reg
self.L2_reg = L2_reg
self.L1 = 0
self.L2_sqr = 0
def __init__(self,n_in,n_hidden):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.input= T.tensor3()
self.W_hh=glorot_uniform((n_hidden,n_hidden))
self.W_in=glorot_uniform((n_in,n_hidden))
self.bh=zero((n_hidden,))
self.params=[self.W_hh,self.W_in,self.bh]
self.L1 = T.sum(abs(self.W_hh))+T.sum(abs(self.W_in))
self.L2_sqr = T.sum(self.W_hh**2) + T.sum(self.W_in**2)
def __init__(self,n_in,n_hidden,n_out,lr=0.001,n_epochs=400,L1_reg=0,L2_reg=0):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.n_out=int(n_out)
self.lr=float(lr)
self.x = T.matrix(name = 'x', dtype = theano.config.floatX)
self.layers = []
self.params=[]
self.errors=[]
self.n_epochs=n_epochs
self.W_hy = glorot_uniform((self.n_hidden,self.n_out))
self.b_hy = zero((self.n_out,))
self.initial_momentum=0.5
self.final_momentum=0.9
self.momentum_switchover=5
self.learning_rate_decay=0.999
self.updates = {}
self.n_layers=0
self.L1_reg=L1_reg
self.L2_reg=L2_reg
self.L1= 0
self.L2_sqr= 0
def __init__(self,n_in,n_hidden):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.input= T.tensor3()
self.W_z = glorot_uniform((n_in,n_hidden))
self.U_z = glorot_uniform((n_hidden,n_hidden))
self.b_z = zero((n_hidden,))
self.W_r = glorot_uniform((n_in,n_hidden))
self.U_r = glorot_uniform((n_hidden,n_hidden))
self.b_r = zero((n_hidden,))
self.W_h = glorot_uniform((n_in,n_hidden))
self.U_h = glorot_uniform((n_hidden,n_hidden))
self.b_h = zero((n_hidden,))
self.params = [
self.W_z, self.U_z, self.b_z,
self.W_r, self.U_r, self.b_r,
self.W_h, self.U_h, self.b_h,
]
self.L1 = T.sum(abs(self.W_z))+T.sum(abs(self.U_z))+\
T.sum(abs(self.W_r))+T.sum(abs(self.U_r))+\
T.sum(abs(self.W_h))+T.sum(abs(self.U_h))
self.L2_sqr = T.sum(self.W_z**2) + T.sum(self.U_z**2)+\
T.sum(self.W_r**2) + T.sum(self.U_r**2)+\
T.sum(self.W_h**2) + T.sum(self.U_h**2)
def __init__(self, n_in, n_hidden):
self.n_in = int(n_in)
self.n_hidden = int(n_hidden)
self.input = T.tensor3()
self.W_i = glorot_uniform((n_in, n_hidden))
self.U_i = glorot_uniform((n_hidden, n_hidden))
self.b_i = zero((n_hidden, ))
self.W_f = glorot_uniform((n_in, n_hidden))
self.U_f = glorot_uniform((n_hidden, n_hidden))
self.b_f = zero((n_hidden, ))
self.W_c = glorot_uniform((n_in, n_hidden))
self.U_c = glorot_uniform((n_hidden, n_hidden))
self.b_c = zero((n_hidden, ))
self.W_o = glorot_uniform((n_in, n_hidden))
self.U_o = glorot_uniform((n_hidden, n_hidden))
self.b_o = zero((n_hidden, ))
self.params = [
self.W_i,
self.U_i,
self.b_i,
self.W_c,
self.U_c,
self.b_c,
self.W_f,
self.U_f,
self.b_f,
self.W_o,
self.U_o,
self.b_o,
]
self.L1 = T.sum(abs(self.W_i))+T.sum(abs(self.U_i))+\
T.sum(abs(self.W_f))+T.sum(abs(self.U_f))+\
T.sum(abs(self.W_c))+T.sum(abs(self.U_c))+\
T.sum(abs(self.W_o))+T.sum(abs(self.U_o))
self.L2_sqr = T.sum(self.W_i**2) + T.sum(self.U_i**2)+\
T.sum(self.W_f**2) + T.sum(self.U_f**2)+\
T.sum(self.W_c**2) + T.sum(self.U_c**2)+\
T.sum(self.W_o**2) + T.sum(self.U_o**2)
def __init__(self,n_in,n_hidden):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.input= T.tensor3()
self.W_i = glorot_uniform((n_in,n_hidden))
self.U_i = glorot_uniform((n_hidden,n_hidden))
self.b_i = zero((n_hidden,))
self.W_f = glorot_uniform((n_in,n_hidden))
self.U_f = glorot_uniform((n_hidden,n_hidden))
self.b_f = zero((n_hidden,))
self.W_c = glorot_uniform((n_in,n_hidden))
self.U_c = glorot_uniform((n_hidden,n_hidden))
self.b_c = zero((n_hidden,))
self.W_o = glorot_uniform((n_in,n_hidden))
self.U_o = glorot_uniform((n_hidden,n_hidden))
self.b_o = zero((n_hidden,))
self.params = [
self.W_i, self.U_i, self.b_i,
self.W_c, self.U_c, self.b_c,
self.W_f, self.U_f, self.b_f,
self.W_o, self.U_o, self.b_o,
]
self.L1 = T.sum(abs(self.W_i))+T.sum(abs(self.U_i))+\
T.sum(abs(self.W_f))+T.sum(abs(self.U_f))+\
T.sum(abs(self.W_c))+T.sum(abs(self.U_c))+\
T.sum(abs(self.W_o))+T.sum(abs(self.U_o))
self.L2_sqr = T.sum(self.W_i**2) + T.sum(self.U_i**2)+\
T.sum(self.W_f**2) + T.sum(self.U_f**2)+\
T.sum(self.W_c**2) + T.sum(self.U_c**2)+\
T.sum(self.W_o**2) + T.sum(self.U_o**2)
def __init__(self,n_in,n_hidden,n_decoder,n_out,
n_epochs=400,n_chapter=100,n_batch=16,maxlen=20,n_words_x=10000,n_words_y=10000,dim_word=100,
momentum_switchover=5,lr=0.001,learning_rate_decay=0.999,snapshot=100,sample_Freq=100,val_Freq=100,L1_reg=0,L2_reg=0):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.n_decoder=int(n_decoder)
self.n_out=int(n_out)
self.n_batch=int(n_batch)
if n_chapter is not None:self.n_chapter=int(n_chapter)
else:self.n_chapter=None
self.n_epochs=n_epochs
self.maxlen= int(maxlen)
self.dim_word=dim_word
self.n_words_x=n_words_x
self.n_words_y=n_words_y
self.x = T.matrix(name = 'x', dtype = 'int32')
self.y = T.matrix(name = 'y', dtype = 'int32')
self.x_mask = T.matrix(name = 'x_mask', dtype = 'float32')
self.y_mask = T.matrix(name = 'y_mask', dtype = 'float32')
self.x_emb = T.tensor3(name = 'x', dtype = 'float32')
self.y_emb = T.tensor3(name = 'y', dtype = 'float32')
self.W_hy = glorot_uniform((self.n_out,self.n_words_y))
self.b_hy = zero((self.n_words_y,))
self.W_hi = glorot_uniform((self.n_hidden,self.n_decoder))
self.b_hi = zero((n_decoder,))
self.Wemb=glorot_normal((self.n_words_x,self.dim_word))
#self.Wemb_dec=glorot_normal((self.n_words_y,self.dim_word))
self.layers = []
self.decoder=[]
self.params=[]
self.errors=[]
#self.updates = {}
self.initial_momentum=0.5
self.final_momentum=0.9
self.lr=float(lr)
self.momentum_switchover=int(momentum_switchover)
self.learning_rate_decay=learning_rate_decay
self.snapshot=int(snapshot)
self.sample_Freq=int(sample_Freq)
self.val_Freq=int(val_Freq)
self.L1_reg=L1_reg
self.L2_reg=L2_reg
self.L1= 0
self.L2_sqr= 0
## word embedding
self.x_emb=self.Wemb[T.cast(self.x.flatten(),'int32')].reshape((self.x.shape[0], self.x.shape[1], self.dim_word))
self.y_emb=self.Wemb[T.cast(self.y.flatten(),'int32')].reshape((self.y.shape[0], self.y.shape[1], self.dim_word))
def __init__(self,n_in,n_hidden,output_mode='concat'):
self.n_in=int(n_in)
self.n_hidden=int(n_hidden)
self.output_mode = output_mode
self.input= T.tensor3()
# forward weights
self.W_i = glorot_uniform((n_in,n_hidden))
self.U_i = glorot_uniform((n_hidden,n_hidden))
self.b_i = zero((n_hidden,))
self.W_f = glorot_uniform((n_in,n_hidden))
self.U_f = glorot_uniform((n_hidden,n_hidden))
self.b_f = zero((n_hidden,))
self.W_c = glorot_uniform((n_in,n_hidden))
self.U_c = glorot_uniform((n_hidden,n_hidden))
self.b_c = zero((n_hidden,))
self.W_o = glorot_uniform((n_in,n_hidden))
self.U_o = glorot_uniform((n_hidden,n_hidden))
self.b_o = zero((n_hidden,))
# backward weights
self.Wb_i = glorot_uniform((n_in,n_hidden))
self.Ub_i = glorot_uniform((n_hidden,n_hidden))
self.bb_i = zero((n_hidden,))
self.Wb_f = glorot_uniform((n_in,n_hidden))
self.Ub_f = glorot_uniform((n_hidden,n_hidden))
self.bb_f = zero((n_hidden,))
self.Wb_c = glorot_uniform((n_in,n_hidden))
self.Ub_c = glorot_uniform((n_hidden,n_hidden))
self.bb_c = zero((n_hidden,))
self.Wb_o = glorot_uniform((n_in,n_hidden))
self.Ub_o = glorot_uniform((n_hidden,n_hidden))
self.bb_o = zero((n_hidden,))
self.params = [
self.W_i, self.U_i, self.b_i,
self.W_c, self.U_c, self.b_c,
self.W_f, self.U_f, self.b_f,
self.W_o, self.U_o, self.b_o,
self.Wb_i, self.Ub_i, self.bb_i,
self.Wb_c, self.Ub_c, self.bb_c,
self.Wb_f, self.Ub_f, self.bb_f,
self.Wb_o, self.Ub_o, self.bb_o,
]
self.L1 = T.sum(abs(self.W_i))+T.sum(abs(self.U_i))+\
T.sum(abs(self.W_f))+T.sum(abs(self.U_f))+\
T.sum(abs(self.W_c))+T.sum(abs(self.U_c))+\
T.sum(abs(self.W_o))+T.sum(abs(self.U_o))+\
T.sum(abs(self.Wb_i))+T.sum(abs(self.Ub_i))+\
T.sum(abs(self.Wb_f))+T.sum(abs(self.Ub_f))+\
T.sum(abs(self.Wb_c))+T.sum(abs(self.Ub_c))+\
T.sum(abs(self.Wb_o))+T.sum(abs(self.Ub_o))
self.L2_sqr = T.sum(self.W_i**2) + T.sum(self.U_i**2)+\
T.sum(self.W_f**2) + T.sum(self.U_f**2)+\
T.sum(self.W_c**2) + T.sum(self.U_c**2)+\
T.sum(self.W_o**2) + T.sum(self.U_o**2)+\
T.sum(self.Wb_i**2) + T.sum(self.Ub_i**2)+\
T.sum(self.Wb_f**2) + T.sum(self.Ub_f**2)+\
T.sum(self.Wb_c**2) + T.sum(self.Ub_c**2)+\
T.sum(self.Wb_o**2) + T.sum(self.Ub_o**2)
def __init__(self,n_in,n_hidden,output_mode='concat'):
self.n_in=int(n_in)
n_hidden=int(n_hidden/2)
self.n_hidden=int(n_hidden)
self.output_mode = output_mode
self.input= T.tensor3()
self.x_mask=T.matrix()
# forward weights
self.W_z = glorot_uniform((n_in,n_hidden))
self.U_z = glorot_uniform((n_hidden,n_hidden))
self.b_z = zero((n_hidden,))
self.W_r = glorot_uniform((n_in,n_hidden))
self.U_r = glorot_uniform((n_hidden,n_hidden))
self.b_r = zero((n_hidden,))
self.W_h = glorot_uniform((n_in,n_hidden))
self.U_h = glorot_uniform((n_hidden,n_hidden))
self.b_h = zero((n_hidden,))
# backward weights
self.Wb_z = glorot_uniform((n_in,n_hidden))
self.Ub_z = glorot_uniform((n_hidden,n_hidden))
self.bb_z = zero((n_hidden,))
self.Wb_r = glorot_uniform((n_in,n_hidden))
self.Ub_r = glorot_uniform((n_hidden,n_hidden))
self.bb_r = zero((n_hidden,))
self.Wb_h = glorot_uniform((n_in,n_hidden))
self.Ub_h = glorot_uniform((n_hidden,n_hidden))
self.bb_h = zero((n_hidden,))
self.params = [
self.W_z, self.U_z, self.b_z,
self.W_r, self.U_r, self.b_r,
self.W_h, self.U_h, self.b_h,
self.Wb_z, self.Ub_z, self.bb_z,
self.Wb_r, self.Ub_r, self.bb_r,
self.Wb_h, self.Ub_h, self.bb_h
]
self.L1 = T.sum(abs(self.W_z))+T.sum(abs(self.U_z))+\
T.sum(abs(self.W_r))+T.sum(abs(self.U_r))+\
T.sum(abs(self.W_h))+T.sum(abs(self.U_h))+\
T.sum(abs(self.Wb_z))+T.sum(abs(self.Ub_z))+\
T.sum(abs(self.Wb_r))+T.sum(abs(self.Ub_r))+\
T.sum(abs(self.Wb_h))+T.sum(abs(self.Ub_h))
self.L2_sqr = T.sum(self.W_z**2) + T.sum(self.U_z**2)+\
T.sum(self.W_r**2) + T.sum(self.U_r**2)+\
T.sum(self.W_h**2) + T.sum(self.U_h**2)+\
T.sum(self.Wb_z**2) + T.sum(self.Ub_z**2)+\
T.sum(self.Wb_r**2) + T.sum(self.Ub_r**2)+\
T.sum(self.Wb_h**2) + T.sum(self.Ub_h**2)
def __init__(self, n_in, n_hidden, output_mode='concat'):
self.n_in = int(n_in)
n_hidden = int(n_hidden / 2)
self.n_hidden = int(n_hidden)
self.output_mode = output_mode
self.input = T.tensor3()
self.x_mask = T.matrix()
# forward weights
self.W_z = glorot_uniform((n_in, n_hidden))
self.U_z = glorot_uniform((n_hidden, n_hidden))
self.b_z = zero((n_hidden, ))
self.W_r = glorot_uniform((n_in, n_hidden))
self.U_r = glorot_uniform((n_hidden, n_hidden))
self.b_r = zero((n_hidden, ))
self.W_h = glorot_uniform((n_in, n_hidden))
self.U_h = glorot_uniform((n_hidden, n_hidden))
self.b_h = zero((n_hidden, ))
# backward weights
self.Wb_z = glorot_uniform((n_in, n_hidden))
self.Ub_z = glorot_uniform((n_hidden, n_hidden))
self.bb_z = zero((n_hidden, ))
self.Wb_r = glorot_uniform((n_in, n_hidden))
self.Ub_r = glorot_uniform((n_hidden, n_hidden))
self.bb_r = zero((n_hidden, ))
self.Wb_h = glorot_uniform((n_in, n_hidden))
self.Ub_h = glorot_uniform((n_hidden, n_hidden))
self.bb_h = zero((n_hidden, ))
self.params = [
self.W_z, self.U_z, self.b_z, self.W_r, self.U_r, self.b_r,
self.W_h, self.U_h, self.b_h, self.Wb_z, self.Ub_z, self.bb_z,
self.Wb_r, self.Ub_r, self.bb_r, self.Wb_h, self.Ub_h, self.bb_h
]
self.L1 = T.sum(abs(self.W_z))+T.sum(abs(self.U_z))+\
T.sum(abs(self.W_r))+T.sum(abs(self.U_r))+\
T.sum(abs(self.W_h))+T.sum(abs(self.U_h))+\
T.sum(abs(self.Wb_z))+T.sum(abs(self.Ub_z))+\
T.sum(abs(self.Wb_r))+T.sum(abs(self.Ub_r))+\
T.sum(abs(self.Wb_h))+T.sum(abs(self.Ub_h))
self.L2_sqr = T.sum(self.W_z**2) + T.sum(self.U_z**2)+\
T.sum(self.W_r**2) + T.sum(self.U_r**2)+\
T.sum(self.W_h**2) + T.sum(self.U_h**2)+\
T.sum(self.Wb_z**2) + T.sum(self.Ub_z**2)+\
T.sum(self.Wb_r**2) + T.sum(self.Ub_r**2)+\
T.sum(self.Wb_h**2) + T.sum(self.Ub_h**2)
def __init__(
self,
n_in,
n_hidden,
n_decoder,
n_out,
n_epochs=400,
n_chapter=100,
n_batch=16,
maxlen=20,
n_words_x=10000,
n_words_y=10000,
dim_word=100,
momentum_switchover=5,
lr=0.001,
learning_rate_decay=0.999,
snapshot=100,
sample_Freq=100,
val_Freq=100,
L1_reg=0,
L2_reg=0,
):
self.n_in = int(n_in)
self.n_hidden = int(n_hidden)
self.n_decoder = int(n_decoder)
self.n_out = int(n_out)
self.n_batch = int(n_batch)
if n_chapter is not None:
self.n_chapter = int(n_chapter)
else:
self.n_chapter = None
self.n_epochs = n_epochs
self.maxlen = int(maxlen)
self.dim_word = dim_word
self.n_words_x = n_words_x
self.n_words_y = n_words_y
self.x = T.matrix(name="x", dtype="int32")
self.y = T.matrix(name="y", dtype="int32")
self.x_mask = T.matrix(name="x_mask", dtype="float32")
self.y_mask = T.matrix(name="y_mask", dtype="float32")
self.x_emb = T.tensor3(name="x", dtype="float32")
self.y_emb = T.tensor3(name="y", dtype="float32")
self.W_hy = glorot_uniform((self.n_out, self.n_words_y))
self.b_hy = zero((self.n_words_y,))
self.W_hi = glorot_uniform((self.n_hidden, self.n_decoder))
self.b_hi = zero((n_decoder,))
self.Wemb = glorot_normal((self.n_words_x, self.dim_word))
# self.Wemb_dec=glorot_normal((self.n_words_y,self.dim_word))
self.layers = []
self.decoder = []
self.params = []
self.errors = []
# self.updates = {}
self.initial_momentum = 0.5
self.final_momentum = 0.9
self.lr = float(lr)
self.momentum_switchover = int(momentum_switchover)
self.learning_rate_decay = learning_rate_decay
self.snapshot = int(snapshot)
self.sample_Freq = int(sample_Freq)
self.val_Freq = int(val_Freq)
self.L1_reg = L1_reg
self.L2_reg = L2_reg
self.L1 = 0
self.L2_sqr = 0
## word embedding
self.x_emb = self.Wemb[T.cast(self.x.flatten(), "int32")].reshape(
(self.x.shape[0], self.x.shape[1], self.dim_word)
)
self.y_emb = self.Wemb[T.cast(self.y.flatten(), "int32")].reshape(
(self.y.shape[0], self.y.shape[1], self.dim_word)
)
