def decode_step(self, trg_words, train):
e = F.pick(self.trg_lookup_, trg_words, 1)
e = F.dropout(e, self.dropout_rate_, train)
h = self.trg_lstm_.forward(F.concat([e, self.feed_], 0))
h = F.dropout(h, self.dropout_rate_, train)
atten_probs = F.softmax(self.t_concat_fb_ @ h, 0)
c = self.concat_fb_ @ atten_probs
self.feed_ = F.tanh(self.whj_ @ F.concat([h, c], 0) + self.bj_)
return self.wjy_ @ self.feed_ + self.by_
def decode_step(self, trg_words, train):
"""One step decoding."""
e = F.pick(self.trg_lookup, trg_words, 1)
e = F.dropout(e, self.dropout_rate, train)
h = self.trg_lstm.forward(F.concat([e, self.feed], 0))
h = F.dropout(h, self.dropout_rate, train)
atten_probs = F.softmax(self.t_concat_fb @ h, 0)
c = self.concat_fb @ atten_probs
self.feed = F.tanh(self.whj @ F.concat([h, c], 0) + self.bj)
return self.wjy @ self.feed + self.by
def forward(self, inputs, train):
batch_size = len(inputs[0])
lookup = F.parameter(self.plookup)
self.rnn1.restart()
self.rnn2.restart()
self.hy.reset()
outputs = []
for i in range(len(inputs) - 1):
x = F.pick(lookup, inputs[i], 1)
x = F.dropout(x, DROPOUT_RATE, train)
h1 = self.rnn1.forward(x)
h1 = F.dropout(h1, DROPOUT_RATE, train)
h2 = self.rnn2.forward(h1)
h2 = F.dropout(h2, DROPOUT_RATE, train)
outputs.append(self.hy.forward(h2))
return outputs
def encode(self, src_batch, train):
"""Encodes source sentences and prepares internal states."""
# Embedding lookup.
src_lookup = F.parameter(self.psrc_lookup)
e_list = []
for x in src_batch:
e = F.pick(src_lookup, x, 1)
e = F.dropout(e, self.dropout_rate, train)
e_list.append(e)
# Forward encoding
self.src_fw_lstm.restart()
f_list = []
for e in e_list:
f = self.src_fw_lstm.forward(e)
f = F.dropout(f, self.dropout_rate, train)
f_list.append(f)
# Backward encoding
self.src_bw_lstm.restart()
b_list = []
for e in reversed(e_list):
b = self.src_bw_lstm.forward(e)
b = F.dropout(b, self.dropout_rate, train)
b_list.append(b)
b_list.reverse()
# Concatenates RNN states.
fb_list = [f_list[i] + b_list[i] for i in range(len(src_batch))]
self.concat_fb = F.concat(fb_list, 1)
self.t_concat_fb = F.transpose(self.concat_fb)
# Initializes decode states.
embed_size = self.psrc_lookup.shape()[0]
self.trg_lookup = F.parameter(self.ptrg_lookup)
self.whj = F.parameter(self.pwhj)
self.bj = F.parameter(self.pbj)
self.wjy = F.parameter(self.pwjy)
self.by = F.parameter(self.pby)
self.feed = F.zeros([embed_size])
self.trg_lstm.restart(
self.src_fw_lstm.get_c() + self.src_bw_lstm.get_c(),
self.src_fw_lstm.get_h() + self.src_bw_lstm.get_h())
def encode(self, src_batch, train):
# Embedding lookup.
src_lookup = F.parameter(self.psrc_lookup_)
e_list = []
for x in src_batch:
e = F.pick(src_lookup, x, 1)
e = F.dropout(e, self.dropout_rate_, train)
e_list.append(e)
# Forward encoding
self.src_fw_lstm_.init()
f_list = []
for e in e_list:
f = self.src_fw_lstm_.forward(e)
f = F.dropout(f, self.dropout_rate_, train)
f_list.append(f)
# Backward encoding
self.src_bw_lstm_.init()
b_list = []
for e in reversed(e_list):
b = self.src_bw_lstm_.forward(e)
b = F.dropout(b, self.dropout_rate_, train)
b_list.append(b)
b_list.reverse()
# Concatenates RNN states.
fb_list = [f_list[i] + b_list[i] for i in range(len(src_batch))]
self.concat_fb_ = F.concat(fb_list, 1)
self.t_concat_fb_ = F.transpose(self.concat_fb_)
# Initializes decode states.
self.trg_lookup_ = F.parameter(self.ptrg_lookup_)
self.whj_ = F.parameter(self.pwhj_)
self.bj_ = F.parameter(self.pbj_)
self.wjy_ = F.parameter(self.pwjy_)
self.by_ = F.parameter(self.pby_)
self.feed_ = F.zeros([self.embed_size_])
self.trg_lstm_.init(
self.src_fw_lstm_.get_c() + self.src_bw_lstm_.get_c(),
self.src_fw_lstm_.get_h() + self.src_bw_lstm_.get_h())
def forward(self, inputs, train):
batch_size = len(inputs[0])
lookup = F.parameter(self.plookup_)
self.rnn1_.init()
self.rnn2_.init()
self.hy_.init()
xs = [
F.dropout(F.pick(lookup, inputs[i], 1), DROPOUT_RATE, train)
for i in range(len(inputs) - 1)
]
hs1 = self.rnn1_.forward(xs)
for i in range(len(inputs) - 1):
hs1[i] = F.dropout(hs1[i], DROPOUT_RATE, train)
hs2 = self.rnn2_.forward(hs1)
outputs = [
self.hy_.forward(F.dropout(hs2[i], DROPOUT_RATE, train))
for i in range(len(inputs) - 1)
]
return outputs
def make_graph(inputs, train):
x = F.input(inputs)
w1 = F.parameter(pw1)
b1 = F.parameter(pb1)
h = F.relu(w1 @ x + b1)
h = F.dropout(h, .5, train)
w2 = F.parameter(pw2)
b2 = F.parameter(pb2)
return w2 @ h + b2
def encode(self, src_batch, train):
# Reversed encoding.
src_lookup = F.parameter(self.psrc_lookup_)
self.src_lstm_.init()
for it in src_batch:
x = F.pick(src_lookup, it, 1)
x = F.dropout(x, self.dropout_rate_, train)
self.src_lstm_.forward(x)
# Initializes decoder states.
self.trg_lookup_ = F.parameter(self.ptrg_lookup_)
self.why_ = F.parameter(self.pwhy_)
self.by_ = F.parameter(self.pby_)
self.trg_lstm_.init(self.src_lstm_.get_c(), self.src_lstm_.get_h())
def make_graph(inputs, train):
# Stores input values.
x = F.input(data=inputs)
# Calculates the hidden layer.
w1 = F.input(param=pw1)
b1 = F.input(param=pb1)
h = F.relu(F.matmul(w1, x) + b1)
# Batch normalization
#Node beta = F::input(pbeta);
#Node gamma = F::input(pgamma);
#h = F::batch::normalize(h) * gamma + beta;
# Dropout
h = F.dropout(h, .5, train)
# Calculates the output layer.
w2 = F.input(param=pw2)
b2 = F.input(param=pb2)
return F.matmul(w2, h) + b2
def decode_step(self, trg_words, train):
x = F.pick(self.trg_lookup_, trg_words, 1)
x = F.dropout(x, self.dropout_rate_, train)
h = self.trg_lstm_.forward(x)
h = F.dropout(h, self.dropout_rate_, train)
return self.why_ @ h + self.by_