def forward(self, x, x_single_mask, x_char, x_char_mask, x_features, x_pos,
x_ent, x_bert, x_bert_mask, x_bert_offsets, q, q_mask, q_char,
q_char_mask, q_bert, q_bert_mask, q_bert_offsets, context_len):
batch_size = q.shape[0]
x_mask = x_single_mask.expand(batch_size, -1)
x_word_embed = self.vocab_embed(x).expand(
batch_size, -1, -1) # batch x x_len x vocab_dim
ques_word_embed = self.vocab_embed(q) # batch x q_len x vocab_dim
x_input_list = [
dropout(x_word_embed,
p=self.opt['dropout_emb'],
training=self.drop_emb)
] # batch x x_len x vocab_dim
ques_input_list = [
dropout(ques_word_embed,
p=self.opt['dropout_emb'],
training=self.drop_emb)
] # batch x q_len x vocab_dim
# contextualized embedding
x_cemb = ques_cemb = None
if 'BERT' in self.opt:
x_cemb = ques_cemb = None
if 'BERT_LINEAR_COMBINE' in self.opt:
x_bert_output = self.Bert(x_bert, x_bert_mask, x_bert_offsets,
x_single_mask)
x_cemb_mid = self.linear_sum(x_bert_output, self.alphaBERT,
self.gammaBERT)
ques_bert_output = self.Bert(q_bert, q_bert_mask,
q_bert_offsets, q_mask)
ques_cemb_mid = self.linear_sum(ques_bert_output,
self.alphaBERT, self.gammaBERT)
x_cemb_mid = x_cemb_mid.expand(batch_size, -1, -1)
else:
x_cemb_mid = self.Bert(x_bert, x_bert_mask, x_bert_offsets,
x_single_mask)
x_cemb_mid = x_cemb_mid.expand(batch_size, -1, -1)
ques_cemb_mid = self.Bert(q_bert, q_bert_mask, q_bert_offsets,
q_mask)
x_input_list.append(x_cemb_mid)
ques_input_list.append(ques_cemb_mid)
if 'CHAR_CNN' in self.opt:
x_char_final = self.character_cnn(x_char, x_char_mask)
x_char_final = x_char_final.expand(batch_size, -1, -1)
ques_char_final = self.character_cnn(q_char, q_char_mask)
x_input_list.append(x_char_final)
ques_input_list.append(ques_char_final)
x_prealign = self.pre_align(x_word_embed, ques_word_embed, q_mask)
x_input_list.append(
x_prealign) # batch x x_len x (vocab_dim + cdim + vocab_dim)
x_pos_emb = self.pos_embedding(x_pos).expand(
batch_size, -1, -1) # batch x x_len x pos_dim
x_ent_emb = self.ent_embedding(x_ent).expand(
batch_size, -1, -1) # batch x x_len x ent_dim
x_input_list.append(x_pos_emb)
x_input_list.append(x_ent_emb)
x_input_list.append(
x_features
) # batch x x_len x (vocab_dim + cdim + vocab_dim + pos_dim + ent_dim + feature_dim)
x_input = torch.cat(
x_input_list, 2
) # batch x x_len x (vocab_dim + cdim + vocab_dim + pos_dim + ent_dim + feature_dim)
ques_input = torch.cat(ques_input_list,
2) # batch x q_len x (vocab_dim + cdim)
# Multi-layer RNN
_, x_rnn_layers = self.context_rnn(
x_input, x_mask, return_list=True, x_additional=x_cemb
) # layer x batch x x_len x context_rnn_output_size
_, ques_rnn_layers = self.ques_rnn(
ques_input, q_mask, return_list=True, x_additional=ques_cemb
) # layer x batch x q_len x ques_rnn_output_size
# rnn with question only
ques_highlvl = self.high_lvl_ques_rnn(
torch.cat(ques_rnn_layers, 2),
q_mask) # batch x q_len x high_lvl_ques_rnn_output_size
ques_rnn_layers.append(ques_highlvl) # (layer + 1) layers
# deep multilevel inter-attention
if x_cemb is None:
x_long = x_word_embed
ques_long = ques_word_embed
else:
x_long = torch.cat([x_word_embed, x_cemb],
2) # batch x x_len x (vocab_dim + cdim)
ques_long = torch.cat([ques_word_embed, ques_cemb],
2) # batch x q_len x (vocab_dim + cdim)
x_rnn_after_inter_attn, x_inter_attn = self.deep_attn(
[x_long],
x_rnn_layers, [ques_long],
ques_rnn_layers,
x_mask,
q_mask,
return_bef_rnn=True)
# x_rnn_after_inter_attn: batch x x_len x deep_attn_output_size
# x_inter_attn: batch x x_len x deep_attn_input_size
# deep self attention
if x_cemb is None:
x_self_attn_input = torch.cat(
[x_rnn_after_inter_attn, x_inter_attn, x_word_embed], 2)
else:
x_self_attn_input = torch.cat(
[x_rnn_after_inter_attn, x_inter_attn, x_cemb, x_word_embed],
2)
# batch x x_len x (deep_attn_output_size + deep_attn_input_size + cdim + vocab_dim)
x_self_attn_output = self.highlvl_self_att(x_self_attn_input,
x_self_attn_input,
x_mask,
x3=x_rnn_after_inter_attn,
drop_diagonal=True)
# batch x x_len x deep_attn_output_size
x_highlvl_output = self.high_lvl_context_rnn(
torch.cat([x_rnn_after_inter_attn, x_self_attn_output], 2), x_mask)
# bach x x_len x high_lvl_context_rnn.output_size
x_final = x_highlvl_output
# question self attention
ques_final = self.ques_self_attn(
ques_highlvl, ques_highlvl, q_mask, x3=None, drop_diagonal=True
) # batch x q_len x high_lvl_ques_rnn_output_size
# merge questions
q_merge_weights = self.ques_merger(ques_final, q_mask)
ques_merged = weighted_avg(ques_final,
q_merge_weights) # batch x ques_final_size
# predict scores
score_s, score_e, score_no, score_yes, score_noanswer = self.get_answer(
x_final, ques_merged, x_mask)
return score_s, score_e, score_no, score_yes, score_noanswer
def forward(self, q_list, ocr_list, od_list, return_score=False):
if return_score:
att_score = {}
else:
att_score = None
batch_size = len(ocr_list['num_cnt'])
od_max_num = od_list['position'].size(1)
ocr_max_num = ocr_list['position'].size(1)
q_input = self.get_embedding_from_list(q_list, self.q_embedding, self.opt['q_emb_initial'])
ocr_input = self.get_embedding_from_list(ocr_list, self.ocr_embedding, self.opt['ocr_emb_initial'])
od_input = self.get_embedding_from_list(od_list, self.ocr_embedding, self.opt['ocr_emb_initial'])
if 'PRE_ALIGN_befor_rnn' in self.opt:
ocr_prealign, od_prealign = self.get_prealign_emb(q_list, ocr_list, od_list, batch_size)
ocr_input = torch.cat([ocr_input, ocr_prealign], dim=-1)
od_input = torch.cat([od_input, od_prealign], dim=-1)
if 'fasttext' in self.opt['ocr_embedding']:
multi2one_ocr_input = self.multi2one(ocr_input, ocr_list['fasttext_mask'])
multi2one_od_input = self.multi2one(od_input, od_list['fasttext_mask'])
elif 'glove' in self.opt['ocr_embedding']:
multi2one_ocr_input = self.multi2one(ocr_input, ocr_list['glove_mask'])
multi2one_od_input = self.multi2one(od_input, od_list['glove_mask'])
if 'img_feature' in self.opt:
img_fea = q_list['img_features']
img_spa = q_list['img_spatials']
if self.opt['img_fea_way'] == 'replace_od':
od_input = self.img_fea2od(img_fea)
od_mask = torch.ByteTensor(batch_size, self.img_fea_num).fill_(1).cuda()
elif self.opt['img_fea_way'] == 'final_att':
# img_fea = self.img_fea_linear(img_fea)
od_input = torch.FloatTensor(batch_size, od_max_num, self.multi2one_output_size).fill_(0).cuda()
od_mask = torch.ByteTensor(batch_size, od_max_num).fill_(0).cuda()
# img_fea_mask = torch.ByteTensor(batch_size, self.img_fea_num).fill_(1).cuda()
else:
od_input = torch.FloatTensor(batch_size, od_max_num, self.multi2one_output_size).fill_(0).cuda()
od_mask = torch.ByteTensor(batch_size, od_max_num).fill_(0).cuda()
ocr_input = torch.FloatTensor(batch_size, ocr_max_num, self.multi2one_output_size).fill_(0).cuda()
if 'ES_ocr' in self.opt and self.opt['ES_using_way'] == 'post_process':
es_ocr_len = self.opt['ES_ocr_len']
ocr_mask = torch.ByteTensor(batch_size, ocr_max_num-self.opt['ES_ocr_len']).fill_(0).cuda()
else:
es_ocr_len = None
ocr_mask = torch.ByteTensor(batch_size, ocr_max_num).fill_(0).cuda()
od_idx = ocr_idx = 0
mask_copy = torch.ByteTensor(batch_size).fill_(0).cuda()
for i in range(batch_size):
if 'img_feature_replace_od' not in self.opt:
od_cnt = 0
for j in od_list['len_cnt'][i]:
od_input[i][od_cnt] = multi2one_od_input[od_idx][j-1]
od_cnt += 1
od_idx += 1
od_mask[i][0:od_cnt] = 1
ocr_cnt = 0
for j in ocr_list['len_cnt'][i]:
ocr_input[i][ocr_cnt] = multi2one_ocr_input[ocr_idx][j-1]
ocr_cnt += 1
ocr_idx += 1
if es_ocr_len != None and ocr_cnt >= es_ocr_len and self.opt['ES_using_way'] == 'post_process':
ocr_mask[i][0:ocr_cnt-es_ocr_len] = 1
# o_mask_pre[i][0:ocr_cnt-101] = 0
else:
ocr_mask[i][0:ocr_cnt] = 1
mask_copy[i] = ocr_cnt
if es_ocr_len != None and self.opt['ES_using_way'] == 'post_process':
es_emb = ocr_input[:, :es_ocr_len]
ocr_input = ocr_input[:, es_ocr_len:]
ocr_list['position'] = ocr_list['position'][:, es_ocr_len:]
es_mask = torch.ByteTensor(batch_size, es_ocr_len).fill_(1).cuda()
if 'fasttext' in self.opt['q_embedding']:
q_mask = q_list['fasttext_mask']
else:
q_mask = q_list['glove_mask']
if 'PRE_ALIGN_after_rnn' in self.opt:
if 'fasttext' in self.opt['q_embedding']:
ocr_prealign, ocr_word_leve_attention_score = self.pre_align(ocr_input, q_list['fasttext_emb'], q_mask)
od_prealign, od_word_leve_attention_score = self.pre_align(od_input, q_list['fasttext_emb'], q_mask)
else:
ocr_prealign, ocr_word_leve_attention_score = self.pre_align(ocr_input, q_list['glove_emb'], q_mask)
od_prealign, od_word_leve_attention_score = self.pre_align(od_input, q_list['glove_emb'], q_mask)
_, ocr_rnn_layers = self.context_rnn(ocr_input, ocr_mask, return_list=True, x_additional=None, LN=True) # layer x batch x x_len x context_rnn_output_size
_, q_rnn_layers = self.ques_rnn(q_input, q_mask, return_list=True, x_additional=None, LN=True) # layer x batch x q_len x ques_rnn_output_size
_, od_rnn_layers = self.context_rnn(od_input, od_mask, return_list=True, x_additional=None, LN=True)
# if 'LN' in self.opt:
# for i in range(len(ocr_rnn_layers)):
# ocr_rnn_layers[i] = self.ocr_rnn1_ln(ocr_rnn_layers[i])
# for i in range(len(od_rnn_layers)):
# od_rnn_layers[i] = self.od_rnn1_ln(od_rnn_layers[i])
# for i in range(len(q_rnn_layers)):
# q_rnn_layers[i] = self.q_rnn1_ln(q_rnn_layers[i])
# rnn with question only
q_highlvl = self.high_lvl_ques_rnn(torch.cat(q_rnn_layers, 2), q_mask, LN=True) # batch x q_len x high_lvl_ques_rnn_output_size
# if 'LN' in self.opt:
# q_highlvl = self.q_
q_rnn_layers.append(q_highlvl) # (layer + 1) layers
# deep multilevel inter-attention
if 'GLOVE' not in self.opt and 'FastText' not in self.opt:
ocr_long = []
q_long = []
od_long = []
elif 'PRE_ALIGN_after_rnn' in self.opt:
ocr_long = [ocr_prealign]
if 'fasttext' in self.opt['q_embedding']:
q_long = [q_list['fasttext_emb']]
else:
q_long = [q_list['glove_emb']]
od_long = [od_prealign]
else:
ocr_long = [ocr_input]
if 'fasttext' in self.opt['q_embedding']:
q_long = [q_list['fasttext_emb']]
else:
q_long = [q_list['glove_emb']]
od_long = [od_input]
ocr_rnn_after_inter_attn, ocr_inter_attn = self.deep_attn(ocr_long, ocr_rnn_layers, q_long, q_rnn_layers, ocr_mask, q_mask, return_bef_rnn=True)
od_rnn_after_inter_attn, od_inter_attn = self.deep_attn(od_long, od_rnn_layers, q_long, q_rnn_layers, od_mask, q_mask, return_bef_rnn=True)
# deep self attention
ocr_self_attn_input = torch.cat([ocr_rnn_after_inter_attn, ocr_inter_attn, ocr_input], 2)
od_self_attn_input = torch.cat([od_rnn_after_inter_attn, od_inter_attn, od_input], 2)
if 'no_Context_Self_Attention' in self.opt:
ocr_highlvl_output = self.high_lvl_context_rnn(ocr_rnn_after_inter_attn, ocr_mask, LN=True)
od_highlvl_output = self.high_lvl_context_rnn(od_rnn_after_inter_attn, od_mask, LN=True)
else:
ocr_self_attn_output = self.highlvl_self_att(ocr_self_attn_input, ocr_self_attn_input, ocr_mask, x3=ocr_rnn_after_inter_attn, drop_diagonal=False)
od_self_attn_output = self.highlvl_self_att(od_self_attn_input, od_self_attn_input, od_mask, x3=od_rnn_after_inter_attn, drop_diagonal=False)
ocr_highlvl_output = self.high_lvl_context_rnn(torch.cat([ocr_rnn_after_inter_attn, ocr_self_attn_output], 2), ocr_mask, LN=True)
od_highlvl_output = self.high_lvl_context_rnn(torch.cat([od_rnn_after_inter_attn, od_self_attn_output], 2), od_mask, LN=True)
if 'position_dim' in self.opt:
ocr_position = ocr_list['position']
od_position = od_list['position']
if 'img_feature' in self.opt and self.opt['img_fea_way'] == 'replace_od':
od_position = img_spa
if self.opt['position_mod'] == 'qk+':
x_od_ocr = self.od_ocr_attn(ocr_highlvl_output, od_highlvl_output, od_mask)
pos_att = self.position_attn(ocr_position, od_position, od_mask, x3 = od_highlvl_output)
x_od_ocr += pos_att
elif self.opt['position_mod'] == 'cat':
x_od_ocr = self.od_ocr_attn(torch.cat([ocr_highlvl_output, ocr_position],dim=2), torch.cat([od_highlvl_output, od_position],dim=2), od_mask)
if self.opt['pos_att_merge_mod'] == 'cat':
ocr_final = torch.cat([ocr_highlvl_output, x_od_ocr], 2)
elif self.opt['pos_att_merge_mod'] == 'atted':
ocr_final = x_od_ocr
elif self.opt['pos_att_merge_mod'] == 'original':
ocr_final = ocr_highlvl_output
# question self attention
q_final = self.ques_self_attn(q_highlvl, q_highlvl, q_mask, drop_diagonal=False) # batch x q_len x high_lvl_ques_rnn_output_size
# merge questions
q_merge_weights = self.ques_merger(q_final, q_mask)
q_merged = weighted_avg(q_final, q_merge_weights) # batch x ques_final_size
# predict scores
if es_ocr_len != None and self.opt['ES_using_way'] == 'post_process':
es_mid = self.ES_linear(es_emb)
es_final = self.ES_ocr_att(es_mid, ocr_final, ocr_mask)
ocr_final = torch.cat([es_final, ocr_final], dim=-2)
ocr_mask = torch.cat([es_mask, ocr_mask], dim=-1)
if 'img_feature' in self.opt and self.opt['img_fea_way'] == 'final_att':
img_fea = self.image_feature_model(q_merged, img_fea)
# q_merged = torch.cat([q_merged, img_fea], dim=-1)
#ocr_fea = self.ocr_final_model(q_merged, ocr_final, mask=ocr_mask)
#q_merged = torch.cat([q_merged, ocr_fea, img_fea], dim=-1)
if 'useES' in self.opt:
score_s = self.get_answer(ocr_final, q_merged, ocr_mask, self.opt['ES_ocr_len'], mask_flag='mask_score' in self.opt)
else:
score_s = self.get_answer(ocr_final, q_merged, ocr_mask, None, mask_flag='mask_score' in self.opt)
if 'fixed_answers' in self.opt:
fixed_ans_logits = self.fixed_ans_classifier(q_merged)
fixed_ans_logits = self.fixed_ocr_alpha * fixed_ans_logits
score_s = (1 - self.fixed_ocr_alpha) * score_s
score_s = torch.cat([fixed_ans_logits, score_s], dim=-1)
return score_s, att_score
def forward(self, x, x_single_mask, x_char, x_char_mask, x_features, x_pos,
x_ent, x_bert, x_bert_mask, x_bert_offsets, q, q_mask, q_char,
q_char_mask, q_bert, q_bert_mask, q_bert_offsets, context_len):
"""
forward()前向计算函数以BatchGen()产生的批次数据作为输入，经过编码层、交互层和输出层计算得到最终的打分结果
:param x: [1, x_len] (word_ids)
:param x_single_mask: [1, x_len]
:param x_char: [1, x_len, char_len] (char_ids)
:param x_char_mask: [1, x_len, char_len]
:param x_features: [batch_size, x_len, feature_len] (5 if answer_span_in_context_feature 4 otherwise)
:param x_pos: [1, x_len] (POS id)
:param x_ent: [1, x_len] (ENT id)
:param x_bert: [1, x_bert_token_len]
:param x_bert_mask: [1, x_bert_token_len]
:param x_bert_offsets: [1, x_len, 2]
:param q: [batch, q_len] (word_ids)
:param q_mask: [batch, q_len]
:param q_char: [batch, q_len, char_len] (char ids)
:param q_char_mask: [batch, q_len, char_len]
:param q_bert: [1, q_bert_token_len]
:param q_bert_mask: [1, q_bert_token_len]
:param q_bert_offsets: [1, q_len, 2]
:param context_len: number of words in context (only one per batch)
:return:
score_s: [batch, context_len]
score_e: [batch, context_len]
score_no: [batch, 1]
score_yes: [batch, 1]
score_noanswer: [batch, 1]
"""
batch_size = q.shape[0]
# 由于同一个batch中的问答共享一篇文章，x_single_mask只有一行，这里将x_single_mask重复batch_size行，与问题数据对齐
x_mask = x_single_mask.expand(batch_size, -1)
# 获得文章单词编码，同样重复batch_size行
x_word_embed = self.vocab_embed(x).expand(
batch_size, -1, -1) # [batch, x_len, vocab_dim]
# 获得问题单词编码
ques_word_embed = self.vocab_embed(q) # [batch, q_len, vocab_dim]
# 文章单词历史
x_input_list = [
dropout(x=x_word_embed,
p=self.opt['dropout_emb'],
training=self.drop_emb)
] # [batch, x_len, vocab_dim]
# 问题单词历史
ques_input_list = [
dropout(x=x_word_embed,
p=self.opt['dropout_emb'],
training=self.drop_emb)
] # [batch, q_len, vocab_dim]
# 上下文编码层
x_cemb = ques_cemb = None
if 'BERT' in self.opt:
x_cemb = ques_cemb = None
if 'BERT_LINEAR_COMBINE' in self.opt:
# 得到BERT每一层输出的文章单词编码
x_bert_output = self.Bert(x_bert, x_bert_mask, x_bert_offsets,
x_single_mask)
# 计算加权和
x_cemb_mid = self.linear_sum(x_bert_output, self.alphaBERT,
self.gammaBERT)
# 得到BERT每一层输出的问题单词编码
ques_bert_output = self.Bert(q_bert, q_bert_mask,
q_bert_offsets, q_mask)
# 计算加权和
ques_cemb_mid = self.linear_sum(ques_bert_output,
self.alphaBERT, self.gammaBERT)
x_cemb_mid = x_cemb_mid.expand(batch_size, -1, -1)
else:
# 不计算加权和的情况
x_cemb_mid = self.Bert(x_bert, x_bert_mask, x_bert_offsets,
x_single_mask)
x_cemb_mid = x_cemb_mid.expand(batch_size, -1, -1)
ques_cemb_mid = self.Bert(q_bert, q_bert_mask, q_bert_offsets,
q_mask)
# 上下文编码加入单词历史
x_input_list.append(x_cemb_mid)
ques_input_list.append(ques_cemb_mid)
if 'CHAR_CNN' in self.opt:
x_char_final = self.character_cnn(x_char, x_char_mask)
x_char_final = x_char_final.expand(batch_size, -1, -1)
ques_char_final = self.character_cnn(q_char, q_char_mask)
x_input_list.append(x_char_final)
ques_input_list.append(ques_char_final)
# 单词注意力层
x_prealign = self.pre_align(x_word_embed, ques_word_embed, q_mask)
x_input_list.append(
x_prealign) # [batch, x_len, vocab_dim + cdim + vocab_dim]
# 词性编码
x_pos_emb = self.pos_embedding(x_pos).expand(
batch_size, -1, -1) # [batch, x_len, pos_dim]
# 命名实体编码
x_ent_emb = self.ent_embedding(x_ent).expand(
batch_size, -1, -1) # [batch, x_len, ent_dim]
x_input_list.append(x_pos_emb)
x_input_list.append(x_ent_emb)
# 加入文章单词的词频和精确匹配特征
x_input_list.append(
x_features
) # [batch_size, x_len, vocab_dim + cdim + vocab_dim + pos_dim, ent_dim, feature_dim]
# 将文章答案的单词历史向量拼接起来
x_input = torch.cat(
x_input_list, 2
) # [batch_size, x_len, vocab_dim + cdim + vocab_dim + pos_dim + ent_dim + feature_dim]
# 将问题答案的单词历史向量拼接起来
ques_input = torch.cat(ques_input_list,
2) # [batch_size, q_len, vocab_dim + cdim]
# Multi-layer RNN, 获得文章和问题RNN层的输出
_, x_rnn_layers = self.context_rnn(
x_input, x_mask, return_list=True, x_additional=x_cemb
) # [layer, batch, x_len, context_rnn_output_size]
_, ques_rnn_layers = self.ques_rnn(
ques_input, q_mask, return_list=True, x_additional=ques_cemb
) # [layer, batch, q_len, ques_rnn_output_size]
# 问题理解层
ques_highlvl = self.high_lvl_ques_rnn(
torch.cat(ques_rnn_layers, 2),
q_mask) # [batch, q_len, high_lvl_ques_rnn_output_size]
ques_rnn_layers.append(ques_highlvl) # (layer + 1) layers
# deep multilevel inter-attention, 全关注互注意力层的输入
if x_cemb is None:
x_long = x_word_embed
ques_long = ques_word_embed
else:
x_long = torch.cat([x_word_embed, x_cemb],
2) # [batch, x_len, vocab_dim + cdim]
ques_long = torch.cat([ques_word_embed, ques_cemb],
2) # [batch, q_len, vocab_dim + cdim]
# 文章单词经过全关注互注意力层, x_rnn_after_inter_attn: [batch, x_len, deep_attn_output_size], x_inter_attn: [batch, x_len, deep_attn_input_size]
x_rnn_after_inter_attn, x_inter_attn = self.deep_attn(
[x_long],
x_rnn_layers, [ques_long],
ques_rnn_layers,
x_mask,
q_mask,
return_bef_rnn=True)
# deep self attention, 全关注自注意力层的输入, x_self_attn_input: [batch, x_len, deep_attn_output_size + deep_attn_input_size + cdim + vocab_dim]
if x_cemb is None:
x_self_attn_input = torch.cat(
[x_rnn_after_inter_attn, x_inter_attn, x_word_embed], 2)
else:
x_self_attn_input = torch.cat(
[x_rnn_after_inter_attn, x_inter_attn, x_cemb, x_word_embed],
2)
# 文章经过全关注自注意力层
x_self_attn_output = self.highlvl_self_attn(
x_self_attn_input,
x_self_attn_input,
x_mask,
x3=x_rnn_after_inter_attn,
drop_diagonal=True) # [batch, x_len, deep_attn_output_size]
# 文章单词经过高级RNN层
x_highlvl_output = self.high_lvl_context_rnn(
torch.cat([x_rnn_after_inter_attn, x_self_attn_output], 2), x_mask)
# 文章单词的最终编码x_final
x_final = x_highlvl_output # [batch, x_len, high_lvl_context_rnn_output_size]
# 问题单词的自注意力层
ques_final = self.ques_self_attn(
ques_highlvl, ques_highlvl, q_mask, x3=None, drop_diagonal=True
) # [batch, q_len, high_lvl_ques_rnn_output_size]
# merge questions, 获得问题的向量表示
q_merge_weights = self.ques_merger(ques_final, q_mask)
ques_merged = weighted_avg(
ques_final, q_merge_weights
) # [batch, ques_final_size], 按照q_merge_weights计算ques_final的加权和
# 获得答案在文章每个位置开始和结束的概率以及三种特殊答案“是/否/没有答案”的概率
score_s, score_e, score_no, score_yes, score_noanswer = self.get_answer(
x_final, ques_merged, x_mask)
return score_s, score_e, score_no, score_yes, score_noanswer