def forward(self, context, question, context_char=None, question_char=None, context_f=None, question_f=None):
"""
context_char and question_char not used
"""
# get embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(context_encode, context_mask,
question_encode, question_mask)
vis_param = {'match': match_para}
# pointer net: (answer_len, batch, context_len)
ans_range_prop = self.pointer_net.forward(qt_aware_ct, context_mask)
ans_range_prop = ans_range_prop.transpose(0, 1)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self, context, question, context_char=None, question_char=None, context_f=None, question_f=None):
assert context_char is not None and question_char is not None and context_f is not None \
and question_f is not None
vis_param = {}
# (seq_len, batch, additional_feature_size)
context_f = context_f.transpose(0, 1)
question_f = question_f.transpose(0, 1)
# word-level embedding: (seq_len, batch, word_embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# char-level embedding: (seq_len, batch, char_embedding_size)
context_emb_char, context_char_mask = self.char_embedding.forward(context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(question_char)
context_vec_char = self.char_encoder.forward(context_emb_char, context_char_mask, context_mask)
question_vec_char = self.char_encoder.forward(question_emb_char, question_char_mask, question_mask)
# mix embedding: (seq_len, batch, embedding_size)
context_vec = torch.cat((context_vec, context_vec_char, context_f), dim=-1)
question_vec = torch.cat((question_vec, question_vec_char, question_f), dim=-1)
# encode: (seq_len, batch, hidden_size*2)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, zs = self.encoder.forward(question_vec, question_mask)
align_ct = context_encode
for i in range(self.num_align_hops):
# align: (seq_len, batch, hidden_size*2)
qt_align_ct, alpha = self.aligner[i](align_ct, question_encode, question_mask)
bar_ct = self.aligner_sfu[i](align_ct, torch.cat([qt_align_ct,
align_ct * qt_align_ct,
align_ct - qt_align_ct], dim=-1))
vis_param['match'] = alpha
# self-align: (seq_len, batch, hidden_size*2)
ct_align_ct, self_alpha = self.self_aligner[i](bar_ct, context_mask)
hat_ct = self.self_aligner_sfu[i](bar_ct, torch.cat([ct_align_ct,
bar_ct * ct_align_ct,
bar_ct - ct_align_ct], dim=-1))
vis_param['self-match'] = self_alpha
# aggregation: (seq_len, batch, hidden_size*2)
align_ct, _ = self.aggregation[i](hat_ct, context_mask)
# pointer net: (answer_len, batch, context_len)
for i in range(self.num_ptr_hops):
ans_range_prop, zs = self.ptr_net[i](align_ct, context_mask, zs)
# answer range
ans_range_prop = ans_range_prop.transpose(0, 1)
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self,
context,
question,
context_char=None,
question_char=None,
context_f=None,
question_f=None):
"""
context_char and question_char not used
"""
# get embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(
context_encode, context_mask, question_encode, question_mask)
vis_param = {'match': match_para}
# pointer net: (answer_len, batch, context_len)
ans_range_prop = self.pointer_net.forward(qt_aware_ct, context_mask)
ans_range_prop = ans_range_prop.transpose(0, 1)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self,
context,
question,
context_char=None,
question_char=None):
assert context_char is not None and question_char is not None
# word-level embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# char-level embedding: (seq_len, batch, char_embedding_size)
context_emb_char, context_char_mask = self.char_embedding.forward(
context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(
question_char)
# word-level encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# char-level encode: (seq_len, batch, hidden_size)
context_vec_char = self.char_encoder.forward(context_emb_char,
context_char_mask,
context_mask)
question_vec_char = self.char_encoder.forward(question_emb_char,
question_char_mask,
question_mask)
context_encode = torch.cat((context_encode, context_vec_char), dim=-1)
question_encode = torch.cat((question_encode, question_vec_char),
dim=-1)
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(
context_encode, context_mask, question_encode, question_mask)
vis_param = {'match': match_para}
# birnn after self match: (seq_len, batch, hidden_size)
qt_aware_ct_ag, _ = self.birnn_after_self.forward(
qt_aware_ct, context_mask)
# pointer net init hidden: (batch, hidden_size)
ptr_net_hidden = F.tanh(
self.init_ptr_hidden.forward(qt_aware_last_hidden))
# pointer net: (answer_len, batch, context_len)
ans_range_prop = self.pointer_net.forward(qt_aware_ct_ag, context_mask,
ptr_net_hidden)
ans_range_prop = ans_range_prop.transpose(0, 1)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self, context, question, context_char=None, question_char=None, context_f=None, question_f=None):
assert context_char is not None and question_char is not None
# (seq_len, batch, additional_feature_size)
context_f = context_f.transpose(0, 1)
question_f = question_f.transpose(0, 1)
# word-level embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# char-level embedding: (seq_len, batch, char_embedding_size)
context_emb_char, context_char_mask = self.char_embedding.forward(context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(question_char)
# word-level encode: (seq_len, batch, hidden_size)
context_vec = torch.cat([context_vec, context_f], dim=-1)
question_vec = torch.cat([question_vec, question_f], dim=-1)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# char-level encode: (seq_len, batch, hidden_size)
context_vec_char = self.char_encoder.forward(context_emb_char, context_char_mask, context_mask)
question_vec_char = self.char_encoder.forward(question_emb_char, question_char_mask, question_mask)
context_encode = torch.cat((context_encode, context_vec_char), dim=-1)
question_encode = torch.cat((question_encode, question_vec_char), dim=-1)
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(context_encode, context_mask,
question_encode, question_mask)
vis_param = {'match': match_para}
# birnn after self match: (seq_len, batch, hidden_size)
qt_aware_ct_ag, _ = self.birnn_after_self.forward(qt_aware_ct, context_mask)
# pointer net init hidden: (batch, hidden_size)
ptr_net_hidden = F.tanh(self.init_ptr_hidden.forward(qt_aware_last_hidden))
# pointer net: (answer_len, batch, context_len)
ans_range_prop = self.pointer_net.forward(qt_aware_ct_ag, context_mask, ptr_net_hidden)
ans_range_prop = ans_range_prop.transpose(0, 1)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self,
context,
question,
context_char=None,
question_char=None,
context_f=None,
question_f=None):
assert context_char is not None and question_char is not None and context_f is not None \
and question_f is not None
vis_param = {}
# (seq_len, batch, additional_feature_size)
context_f = context_f.transpose(0, 1)
question_f = question_f.transpose(0, 1)
# word-level embedding: (seq_len, batch, word_embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# char-level embedding: (seq_len, batch, char_embedding_size)
context_emb_char, context_char_mask = self.char_embedding.forward(
context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(
question_char)
context_vec_char = self.char_encoder.forward(context_emb_char,
context_char_mask,
context_mask)
question_vec_char = self.char_encoder.forward(question_emb_char,
question_char_mask,
question_mask)
# mix embedding: (seq_len, batch, embedding_size)
context_vec = torch.cat((context_vec, context_vec_char, context_f),
dim=-1)
question_vec = torch.cat((question_vec, question_vec_char, question_f),
dim=-1)
# encode: (seq_len, batch, hidden_size*2)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, zs = self.encoder.forward(question_vec, question_mask)
align_ct = context_encode
for i in range(self.num_align_hops):
# align: (seq_len, batch, hidden_size*2)
qt_align_ct, alpha = self.aligner[i](align_ct, question_encode,
question_mask)
bar_ct = self.aligner_sfu[i](
align_ct,
torch.cat([
qt_align_ct, align_ct * qt_align_ct, align_ct - qt_align_ct
],
dim=-1))
vis_param['match'] = alpha
# self-align: (seq_len, batch, hidden_size*2)
ct_align_ct, self_alpha = self.self_aligner[i](bar_ct,
context_mask)
hat_ct = self.self_aligner_sfu[i](
bar_ct,
torch.cat(
[ct_align_ct, bar_ct * ct_align_ct, bar_ct - ct_align_ct],
dim=-1))
vis_param['self-match'] = self_alpha
# aggregation: (seq_len, batch, hidden_size*2)
align_ct, _ = self.aggregation[i](hat_ct, context_mask)
# pointer net: (answer_len, batch, context_len)
for i in range(self.num_ptr_hops):
ans_range_prop, zs = self.ptr_net[i](align_ct, context_mask, zs)
# answer range
ans_range_prop = ans_range_prop.transpose(0, 1)
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self,
context,
question,
context_char=None,
question_char=None,
context_f=None,
question_f=None):
if self.enable_char:
assert context_char is not None and question_char is not None
if self.enable_features:
assert context_f is not None and question_f is not None
# get embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
if self.enable_features:
assert context_f is not None and question_f is not None
# (seq_len, batch, additional_feature_size)
context_f = context_f.transpose(0, 1)
question_f = question_f.transpose(0, 1)
context_vec = torch.cat([context_vec, context_f], dim=-1)
question_vec = torch.cat([question_vec, question_f], dim=-1)
# char-level embedding: (seq_len, batch, char_embedding_size)
if self.enable_char:
context_emb_char, context_char_mask = self.char_embedding.forward(
context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(
question_char)
context_vec_char = self.char_encoder.forward(
context_emb_char, context_char_mask, context_mask)
question_vec_char = self.char_encoder.forward(
question_emb_char, question_char_mask, question_mask)
if self.mix_encode:
context_vec = torch.cat((context_vec, context_vec_char),
dim=-1)
question_vec = torch.cat((question_vec, question_vec_char),
dim=-1)
# encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# char-level encode: (seq_len, batch, hidden_size)
if self.enable_char and not self.mix_encode:
context_encode = torch.cat((context_encode, context_vec_char),
dim=-1)
question_encode = torch.cat((question_encode, question_vec_char),
dim=-1)
# question match-lstm
match_rnn_in_question = question_encode
if self.enable_question_match:
ct_aware_qt, _, _ = self.question_match_rnn.forward(
question_encode, question_mask, context_encode, context_mask)
match_rnn_in_question = ct_aware_qt
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(
context_encode, context_mask, match_rnn_in_question, question_mask)
vis_param = {'match': match_para}
# self match lstm: (seq_len, batch, hidden_size)
if self.enable_self_match:
qt_aware_ct, qt_aware_last_hidden, self_para = self.self_match_rnn.forward(
qt_aware_ct, context_mask, qt_aware_ct, context_mask)
vis_param['self'] = self_para
# birnn after self match: (seq_len, batch, hidden_size)
if self.enable_birnn_after_self:
qt_aware_ct, _ = self.birnn_after_self.forward(
qt_aware_ct, context_mask)
# self gated
if self.enable_self_gated:
qt_aware_ct = self.self_gated(qt_aware_ct)
# pointer net init hidden: (batch, hidden_size)
ptr_net_hidden = None
if self.init_ptr_hidden_mode == 'pooling':
ptr_net_hidden = self.init_ptr_hidden.forward(
question_encode, question_mask)
elif self.init_ptr_hidden_mode == 'linear':
ptr_net_hidden = self.init_ptr_hidden.forward(qt_aware_last_hidden)
ptr_net_hidden = F.tanh(ptr_net_hidden)
# pointer net: (answer_len, batch, context_len)
# ans_range_prop = self.pointer_net.forward(qt_aware_ct, context_mask, ptr_net_hidden)
# ans_range_prop = ans_range_prop.transpose(0, 1)
ans_range_prop = multi_scale_ptr(self.pointer_net, ptr_net_hidden,
qt_aware_ct, context_mask,
self.scales)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self, context, question, context_char=None, question_char=None, context_f=None, question_f=None):
if self.enable_char:
assert context_char is not None and question_char is not None
if self.enable_features:
assert context_f is not None and question_f is not None
# get embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
if self.enable_features:
assert context_f is not None and question_f is not None
# (seq_len, batch, additional_feature_size)
context_f = context_f.transpose(0, 1)
question_f = question_f.transpose(0, 1)
context_vec = torch.cat([context_vec, context_f], dim=-1)
question_vec = torch.cat([question_vec, question_f], dim=-1)
# char-level embedding: (seq_len, batch, char_embedding_size)
if self.enable_char:
context_emb_char, context_char_mask = self.char_embedding.forward(context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(question_char)
context_vec_char = self.char_encoder.forward(context_emb_char, context_char_mask, context_mask)
question_vec_char = self.char_encoder.forward(question_emb_char, question_char_mask, question_mask)
if self.mix_encode:
context_vec = torch.cat((context_vec, context_vec_char), dim=-1)
question_vec = torch.cat((question_vec, question_vec_char), dim=-1)
# encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# char-level encode: (seq_len, batch, hidden_size)
if self.enable_char and not self.mix_encode:
context_encode = torch.cat((context_encode, context_vec_char), dim=-1)
question_encode = torch.cat((question_encode, question_vec_char), dim=-1)
# question match-lstm
match_rnn_in_question = question_encode
if self.enable_question_match:
ct_aware_qt, _, _ = self.question_match_rnn.forward(question_encode, question_mask,
context_encode, context_mask)
match_rnn_in_question = ct_aware_qt
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_para = self.match_rnn.forward(context_encode, context_mask,
match_rnn_in_question, question_mask)
vis_param = {'match': match_para}
# self match lstm: (seq_len, batch, hidden_size)
if self.enable_self_match:
qt_aware_ct, qt_aware_last_hidden, self_para = self.self_match_rnn.forward(qt_aware_ct, context_mask,
qt_aware_ct, context_mask)
vis_param['self'] = self_para
# birnn after self match: (seq_len, batch, hidden_size)
if self.enable_birnn_after_self:
qt_aware_ct, _ = self.birnn_after_self.forward(qt_aware_ct, context_mask)
# self gated
if self.enable_self_gated:
qt_aware_ct = self.self_gated(qt_aware_ct)
# pointer net init hidden: (batch, hidden_size)
ptr_net_hidden = None
if self.init_ptr_hidden_mode == 'pooling':
ptr_net_hidden = self.init_ptr_hidden.forward(question_encode, question_mask)
elif self.init_ptr_hidden_mode == 'linear':
ptr_net_hidden = self.init_ptr_hidden.forward(qt_aware_last_hidden)
ptr_net_hidden = F.tanh(ptr_net_hidden)
# pointer net: (answer_len, batch, context_len)
# ans_range_prop = self.pointer_net.forward(qt_aware_ct, context_mask, ptr_net_hidden)
# ans_range_prop = ans_range_prop.transpose(0, 1)
ans_range_prop = multi_scale_ptr(self.pointer_net, ptr_net_hidden, qt_aware_ct, context_mask, self.scales)
# answer range
if not self.training and self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
_, ans_range = torch.max(ans_range_prop, dim=2)
return ans_range_prop, ans_range, vis_param
def forward(self,
context,
question,
context_char=None,
question_char=None):
if self.enable_char:
assert context_char is not None and question_char is not None
# get embedding: (seq_len, batch, embedding_size)
context_vec, context_mask = self.embedding.forward(context)
question_vec, question_mask = self.embedding.forward(question)
# char-level embedding: (seq_len, batch, char_embedding_size)
if self.enable_char:
context_emb_char, context_char_mask = self.char_embedding.forward(
context_char)
question_emb_char, question_char_mask = self.char_embedding.forward(
question_char)
context_vec_char = self.char_encoder.forward(
context_emb_char, context_char_mask, context_mask)
question_vec_char = self.char_encoder.forward(
question_emb_char, question_char_mask, question_mask)
if self.mix_encode:
context_vec = torch.cat((context_vec, context_vec_char),
dim=-1)
question_vec = torch.cat((question_vec, question_vec_char),
dim=-1)
# encode: (seq_len, batch, hidden_size)
context_encode, _ = self.encoder.forward(context_vec, context_mask)
question_encode, _ = self.encoder.forward(question_vec, question_mask)
# char-level encode: (seq_len, batch, hidden_size)
if self.enable_char and not self.mix_encode:
context_encode = torch.cat((context_encode, context_vec_char),
dim=-1)
question_encode = torch.cat((question_encode, question_vec_char),
dim=-1)
# match lstm: (seq_len, batch, hidden_size)
qt_aware_ct, qt_aware_last_hidden, match_alpha = self.match_rnn.forward(
context_encode, context_mask, question_encode, question_mask)
vis_param = {'match': match_alpha}
# self match lstm: (seq_len, batch, hidden_size)
if self.enable_self_match:
qt_aware_ct, qt_aware_last_hidden, self_alpha = self.self_match_rnn.forward(
qt_aware_ct, context_mask, qt_aware_ct, context_mask)
vis_param['self'] = self_alpha
# birnn after self match: (seq_len, batch, hidden_size)
if self.enable_birnn_after_self:
qt_aware_ct, _ = self.birnn_after_self.forward(
qt_aware_ct, context_mask)
# pointer net init hidden: (batch, hidden_size)
ptr_net_hidden = None
if self.init_ptr_hidden_mode == 'pooling':
ptr_net_hidden = self.init_ptr_hidden.forward(
question_encode, question_mask)
elif self.init_ptr_hidden_mode == 'linear':
ptr_net_hidden = self.init_ptr_hidden.forward(qt_aware_last_hidden)
ptr_net_hidden = F.tanh(ptr_net_hidden)
# pointer net: (answer_len, batch, context_len)
ans_range_prop = self.pointer_net.forward(qt_aware_ct, context_mask,
ptr_net_hidden)
ans_range_prop = ans_range_prop.transpose(0, 1)
# answer range
if self.enable_search:
ans_range = answer_search(ans_range_prop, context_mask)
else:
ans_range = torch.max(ans_range_prop, 2)[1]
return ans_range_prop, ans_range, vis_param