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 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, xs):
x = F.concat(xs, 1)
u = self.w_ @ x
j = F.slice(u, 0, 0, self.out_size_)
f = F.sigmoid(
F.slice(u, 0, self.out_size_, 2 * self.out_size_) +
F.broadcast(self.bf_, 1, len(xs)))
r = F.sigmoid(
F.slice(u, 0, 2 * self.out_size_, 3 * self.out_size_) +
F.broadcast(self.bf_, 1, len(xs)))
c = F.zeros([self.out_size_])
hs = []
for i in range(len(xs)):
ji = F.slice(j, 1, i, i + 1)
fi = F.slice(f, 1, i, i + 1)
ri = F.slice(r, 1, i, i + 1)
c = fi * c + (1 - fi) * ji
hs.append(ri * F.tanh(c) + (1 - ri) * xs[i])
return hs