class RelationRanking(nn.Module):
def __init__(self, word_vocab, rel_vocab, config):
super(RelationRanking, self).__init__()
self.config = config
rel1_vocab, rel2_vocab = rel_vocab
self.word_embed = Embeddings(word_vec_size=config.d_word_embed, dicts=word_vocab)
self.rel1_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel1_vocab)
self.rel2_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel2_vocab)
if self.config.rnn_type.lower() == 'gru':
self.rnn = nn.GRU(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
else:
self.rnn = nn.LSTM(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden
if self.config.birnn:
seq_in_size *= 2
self.question_attention = MLPWordSeqAttention(input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size, config.d_rel_embed, 1, bias=False)
self.seq_out = nn.Sequential(
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed)
)
self.conv = nn.Sequential(
nn.Conv2d(1, config.channel_size, (config.conv_kernel_1, config.conv_kernel_2), stride=1,
padding=(config.conv_kernel_1//2, config.conv_kernel_2//2)), #channel_in=1, channel_out=8, kernel_size=3*3
nn.ReLU(True))
self.seq_maxlen = config.seq_maxlen + (config.conv_kernel_1 + 1) % 2
self.rel_maxlen = config.rel_maxlen + (config.conv_kernel_2 + 1) % 2
self.pooling = nn.MaxPool2d((config.seq_maxlen, 1),
stride=(config.seq_maxlen, 1), padding=0)
self.pooling2 = nn.MaxPool2d((1, config.rel_maxlen),
stride=(1, config.rel_maxlen), padding=0)
self.fc = nn.Sequential(
nn.Linear(config.rel_maxlen * config.channel_size, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20, 1))
self.fc1 = nn.Sequential(
nn.Linear(config.seq_maxlen * config.channel_size, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20,1))
self.fc2 = nn.Sequential(
nn.Linear(4, 1))
def question_encoder(self, inputs):
'''
:param inputs: (batch, dim1)
'''
batch_size = inputs.size(0)
state_shape = self.config.n_cells, batch_size, self.config.d_hidden
if self.config.rnn_type.lower() == 'gru':
h0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.rnn(inputs, h0)
else:
h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.rnn(inputs, (h0, c0))
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
outputs.contiguous()
return outputs
def cal_score(self, outputs, seqs_len, rel_embed, pos=None):
'''
:param rel_embed: (batch, dim2) or (neg_size, batch, dim2)
return: (batch, 1)
'''
batch_size = outputs.size(0)
if pos:
neg_size = pos
else:
neg_size, batch_size, embed_size = rel_embed.size()
seq_len, seq_emb_size = outputs.size()[1:]
outputs = outputs.unsqueeze(0).expand(neg_size, batch_size, seq_len,
seq_emb_size).contiguous().view(neg_size*batch_size, seq_len, -1)
rel_embed = rel_embed.view(neg_size * batch_size, -1)
seqs_len = seqs_len.unsqueeze(0).expand(neg_size, batch_size).contiguous().view(neg_size*batch_size)
# `weight` - (batch, length)
seq_att, weight = self.question_attention.forward(rel_embed, outputs)
# `seq_encode` - (batch, hidden size X num directions)
seq_encode = self.seq_out(seq_att)
# `score` - (batch, 1) or (neg_size * batch, 1)
score = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
if pos:
score = score.unsqueeze(0).expand(neg_size, batch_size, 1)
else:
score = score.view(neg_size, batch_size, 1)
return score
def matchPyramid(self, seq, rel, seq_len, rel_len):
'''
param:
seq: (batch, _seq_len, embed_size)
rel: (batch, _rel_len, embed_size)
seq_len: (batch,)
rel_len: (batch,)
return:
score: (batch, 1)
'''
batch_size = seq.size(0)
rel_trans = torch.transpose(rel, 1, 2)
# (batch, 1, seq_len, rel_len)
seq_norm = torch.sqrt(torch.sum(seq*seq, dim=2, keepdim=True))
rel_norm = torch.sqrt(torch.sum(rel_trans*rel_trans, dim=1, keepdim=True))
cross = torch.bmm(seq/seq_norm, rel_trans/rel_norm).unsqueeze(1)
# (batch, channel_size, seq_len, rel_len)
conv1 = self.conv(cross)
channel_size = conv1.size(1)
# (batch, seq_maxlen)
# (batch, rel_maxlen)
dpool_index1, dpool_index2 = self.dynamic_pooling_index(seq_len, rel_len, self.seq_maxlen,
self.rel_maxlen)
dpool_index1 = dpool_index1.unsqueeze(1).unsqueeze(-1).expand(batch_size, channel_size,
self.seq_maxlen, self.rel_maxlen)
dpool_index2 = dpool_index2.unsqueeze(1).unsqueeze(2).expand_as(dpool_index1)
conv1_expand = torch.gather(conv1, 2, dpool_index1)
conv1_expand = torch.gather(conv1_expand, 3, dpool_index2)
# (batch, channel_size, p_size1, p_size2)
pool1 = self.pooling(conv1_expand).view(batch_size, -1)
# (batch, 1)
out = self.fc(pool1)
pool2 = self.pooling2(conv1_expand).view(batch_size, -1)
out2 = self.fc1(pool2)
return out, out2
def dynamic_pooling_index(self, len1, len2, max_len1, max_len2):
def dpool_index_(batch_idx, len1_one, len2_one, max_len1, max_len2):
stride1 = 1.0 * max_len1 / len1_one
stride2 = 1.0 * max_len2 / len2_one
idx1_one = [int(i/stride1) for i in range(max_len1)]
idx2_one = [int(i/stride2) for i in range(max_len2)]
return idx1_one, idx2_one
batch_size = len(len1)
index1, index2 = [], []
for i in range(batch_size):
idx1_one, idx2_one = dpool_index_(i, len1[i], len2[i], max_len1, max_len2)
index1.append(idx1_one)
index2.append(idx2_one)
index1 = torch.LongTensor(index1)
index2 = torch.LongTensor(index2)
if self.config.cuda:
index1 = index1.cuda()
index2 = index2.cuda()
return Variable(index1), Variable(index2)
def forward(self, batch):
# shape of seqs (batch size, sequence length)
seqs, seq_len, pos_rel1, pos_rel2, neg_rel1, neg_rel2, pos_rel, pos_rel_len, neg_rel, neg_rel_len = batch
# shape (batch_size, sequence length, dimension of embedding)
inputs = self.word_embed.forward(seqs)
outputs = self.question_encoder(inputs)
# shape (batch_size, dimension of rel embedding)
pos_rel1_embed = self.rel1_embed.word_lookup_table(pos_rel1)
pos_rel2_embed = self.rel2_embed.word_lookup_table(pos_rel2)
pos_rel1_embed = self.dropout(pos_rel1_embed)
pos_rel2_embed = self.dropout(pos_rel2_embed)
# shape (neg_size, batch_size, dimension of rel embedding)
neg_rel1_embed = self.rel1_embed.word_lookup_table(neg_rel1)
neg_rel2_embed = self.rel2_embed.word_lookup_table(neg_rel2)
neg_rel1_embed = self.dropout(neg_rel1_embed)
neg_rel2_embed = self.dropout(neg_rel2_embed)
neg_size, batch, neg_len = neg_rel.size()
# shape of `score` - (neg_size, batch_size, 1)
pos_score1 = self.cal_score(outputs, seq_len, pos_rel1_embed, neg_size)
pos_score2 = self.cal_score(outputs, seq_len, pos_rel2_embed, neg_size)
neg_score1 = self.cal_score(outputs, seq_len, neg_rel1_embed)
neg_score2 = self.cal_score(outputs, seq_len, neg_rel2_embed)
# (batch, len, emb_size)
pos_embed = self.word_embed.forward(pos_rel)
# (batch, 20)
pos_score3, pos_score4 = self.matchPyramid(inputs, pos_embed, seq_len, pos_rel_len)
# (neg_size, batch, 20)
pos_score3 = pos_score3.unsqueeze(0).expand(neg_size, batch, pos_score3.size(1))
pos_score4 = pos_score4.unsqueeze(0).expand(neg_size, batch, pos_score4.size(1))
# (neg_size*batch, len, emb_size)
neg_embed = self.word_embed.forward(neg_rel.view(-1, neg_len))
seqs_embed = inputs.unsqueeze(0).expand(neg_size, batch, inputs.size(1),
inputs.size(2)).contiguous().view(-1, inputs.size(1), inputs.size(2))
# (neg_size*batch,)
neg_rel_len = neg_rel_len.view(-1)
seq_len = seq_len.unsqueeze(0).expand(neg_size, batch).contiguous().view(-1)
# (neg_size*batch, 20)
neg_score3, neg_score4 = self.matchPyramid(seqs_embed, neg_embed, seq_len, neg_rel_len)
# (neg_size, batch, 20)
neg_score3 = neg_score3.view(neg_size, batch, neg_score3.size(1))
neg_score4 = neg_score4.view(neg_size, batch, neg_score4.size(1))
pos_concat = torch.cat((pos_score1, pos_score2, pos_score3, pos_score4), 2)
neg_concat = torch.cat((neg_score1, neg_score2, neg_score3, neg_score4), 2)
pos_score = self.fc2(pos_concat).squeeze(-1)
neg_score = self.fc2(neg_concat).squeeze(-1)
return pos_score, neg_score
class RelationRanking(nn.Module):
def __init__(self, word_vocab, rel_vocab, config):
super(RelationRanking, self).__init__()
self.config = config
rel1_vocab, rel2_vocab = rel_vocab
self.word_embed = Embeddings(word_vec_size=config.d_word_embed, dicts=word_vocab)
self.rel1_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel1_vocab)
self.rel2_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel2_vocab)
# print(self.rel_embed.word_lookup_table.weight.data)
#rel_embed的初始化待改 rel_embed.lookup_table.weight.data.normal_(0, 0.1)
if self.config.rnn_type.lower() == 'gru':
self.rnn = nn.GRU(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
else:
self.rnn = nn.LSTM(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden
if self.config.birnn:
seq_in_size *= 2
self.question_attention = MLPWordSeqAttention(input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size, config.d_rel_embed, 1, bias=False)
self.seq_out = nn.Sequential(
# nn.BatchNorm1d(seq_in_size),
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed)
)
self.conv = nn.Sequential(
nn.Conv2d(1, config.channel_size, (config.conv_kernel_1, config.conv_kernel_2), stride=1,
padding=(config.conv_kernel_1//2, config.conv_kernel_2//2)), #channel_in=1, channel_out=8, kernel_size=3*3
nn.ReLU(True))
self.seq_maxlen = config.seq_maxlen + (config.conv_kernel_1 + 1) % 2
self.rel_maxlen = config.rel_maxlen + (config.conv_kernel_2 + 1) % 2
p_size1 = self.seq_maxlen // config.pool_kernel_1
p_size2 = self.rel_maxlen // config.pool_kernel_2
self.pooling = nn.MaxPool2d((config.pool_kernel_1, config.pool_kernel_2),
stride=(config.pool_kernel_1, config.pool_kernel_2), padding=0)
self.fc = nn.Sequential(
nn.Linear(p_size1*p_size2*config.channel_size, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20, 1),
nn.Sigmoid())
def question_encoder(self, inputs):
'''
:param inputs: (batch, dim1)
'''
batch_size = inputs.size(0)
state_shape = self.config.n_cells, batch_size, self.config.d_hidden
if self.config.rnn_type.lower() == 'gru':
h0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.rnn(inputs, h0)
else:
h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.rnn(inputs, (h0, c0))
outputs.contiguous()
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
# shape of `encoder` - (batch size, hidden size X num directions)
# encoder = ht[-1] if not self.config.birnn else ht[-2:].transpose(0,1).contiguous().view(batch_size, -1)
# seq_encode = self.seq_out(encoder)
return outputs
def cal_score(self, outputs, seqs_len, rel_embed, pos=None):
'''
:param rel_embed: (batch, dim2) or (neg_size, batch, dim2)
return: (batch, 1)
'''
batch_size = outputs.size(0)
if pos:
neg_size = pos
else: # neg的要扩展
neg_size, batch_size, embed_size = rel_embed.size()
seq_len, seq_emb_size = outputs.size()[1:]
outputs = outputs.unsqueeze(0).expand(neg_size, batch_size, seq_len,
seq_emb_size).contiguous().view(neg_size*batch_size, seq_len, -1)
rel_embed = rel_embed.view(neg_size * batch_size, -1)
seqs_len = seqs_len.unsqueeze(0).expand(neg_size, batch_size).contiguous().view(neg_size*batch_size)
# `seq_encode` - (batch, hidden size X num directions)
# `weight` - (batch, length)
# seq_att, weight = self.question_attention.forward(rel_embed, outputs, seqs_len)
seq_att, weight = self.question_attention.forward(rel_embed, outputs)
# if pos:
# print('weight:', weight)
# seq_encode = self.dropout(seq_att)
seq_encode = self.seq_out(seq_att)
# `score` - (batch, 1) or (neg_size * batch, 1)
# score = self.bilinear(seq_encode, rel_embed)
score = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
'''
dot = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
dis = seq_encode - rel_embed
euclidean = torch.sqrt(torch.sum(dis * dis, 1, keepdim=True))
score = (1/(1+euclidean)) * (1/1+torch.exp(-(dot+1)))
'''
if pos: # pos要把结果扩展
score = score.squeeze(1).unsqueeze(0).expand(neg_size, batch_size)
else:
score = score.view(neg_size, batch_size)
return score
def matchPyramid(self, seq, rel, seq_len, rel_len):
'''
param:
seq: (batch, _seq_len, embed_size)
rel: (batch, _rel_len, embed_size)
seq_len: (batch,)
rel_len: (batch,)
return:
score: (batch, 1)
'''
batch_size = seq.size(0)
rel_trans = torch.transpose(rel, 1, 2)
# (batch, 1, seq_len, rel_len)
# cross = torch.bmm(seq, rel_trans).unsqueeze(1)
# 将内积改为cos 相似度
seq_norm = torch.sqrt(torch.sum(seq*seq, dim=2, keepdim=True))
rel_norm = torch.sqrt(torch.sum(rel_trans*rel_trans, dim=1, keepdim=True))
cross = torch.bmm(seq/seq_norm, rel_trans/rel_norm).unsqueeze(1)
# print('cross: ', cross.size())
# print(cross.squeeze(1).squeeze(0))
# (batch, channel_size, seq_len, rel_len)
conv1 = self.conv(cross)
channel_size = conv1.size(1)
# print('conv: ', conv1.size())
# (batch, seq_maxlen)
# (batch, rel_maxlen)
dpool_index1, dpool_index2 = self.dynamic_pooling_index(seq_len, rel_len, self.seq_maxlen,
self.rel_maxlen)
dpool_index1 = dpool_index1.unsqueeze(1).unsqueeze(-1).expand(batch_size, channel_size,
self.seq_maxlen, self.rel_maxlen)
dpool_index2 = dpool_index2.unsqueeze(1).unsqueeze(2).expand_as(dpool_index1)
# print('d1: ', dpool_index1.size())
# print('d2: ', dpool_index2.size())
conv1_expand = torch.gather(conv1, 2, dpool_index1)
conv1_expand = torch.gather(conv1_expand, 3, dpool_index2)
# print(conv1_expand.size())
# (batch, channel_size, p_size1, p_size2)
pool1 = self.pooling(conv1_expand).view(batch_size, -1)
# print('pool: ', pool1.size())
# (batch, 1)
out = self.fc(pool1)
return out
def dynamic_pooling_index(self, len1, len2, max_len1, max_len2):
def dpool_index_(batch_idx, len1_one, len2_one, max_len1, max_len2):
stride1 = 1.0 * max_len1 / len1_one
stride2 = 1.0 * max_len2 / len2_one
idx1_one = [int(i/stride1) for i in range(max_len1)]
idx2_one = [int(i/stride2) for i in range(max_len2)]
# mesh1, mesh2 = np.meshgrid(idx1_one, idx2_one)
# index_one = np.transpose(np.stack([np.ones(mesh1.shape) * batch_idx, mesh1, mesh2]), (2,1,0))
return idx1_one, idx2_one
batch_size = len(len1)
index1, index2 = [], []
for i in range(batch_size):
idx1_one, idx2_one = dpool_index_(i, len1[i], len2[i], max_len1, max_len2)
index1.append(idx1_one)
index2.append(idx2_one)
# print(index1)
# print(index2)
index1 = torch.LongTensor(index1)
index2 = torch.LongTensor(index2)
if self.config.cuda:
index1 = index1.cuda()
index2 = index2.cuda()
return Variable(index1), Variable(index2)
def forward(self, batch):
# shape of seqs (batch size, sequence length)
seqs, seq_len, pos_rel1, pos_rel2, neg_rel1, neg_rel2, pos_rel, pos_rel_len, neg_rel, neg_rel_len = batch
# print('seqs:', seqs)
# shape (batch_size, sequence length, dimension of embedding)
inputs = self.word_embed.forward(seqs)
outputs = self.question_encoder(inputs)
# shape (batch_size, dimension of rel embedding)
pos_rel1_embed = self.rel1_embed.word_lookup_table(pos_rel1)
pos_rel2_embed = self.rel2_embed.word_lookup_table(pos_rel2)
pos_rel1_embed = self.dropout(pos_rel1_embed)
pos_rel2_embed = self.dropout(pos_rel2_embed)
# shape (neg_size, batch_size, dimension of rel embedding)
neg_rel1_embed = self.rel1_embed.word_lookup_table(neg_rel1)
neg_rel2_embed = self.rel2_embed.word_lookup_table(neg_rel2)
neg_rel1_embed = self.dropout(neg_rel1_embed)
neg_rel2_embed = self.dropout(neg_rel2_embed)
neg_size, batch, neg_len = neg_rel.size()
# shape of `score` - (neg_size, batch_size)
pos_score1 = self.cal_score(outputs, seq_len, pos_rel1_embed, neg_size)
pos_score2 = self.cal_score(outputs, seq_len, pos_rel2_embed, neg_size)
neg_score1 = self.cal_score(outputs, seq_len, neg_rel1_embed)
neg_score2 = self.cal_score(outputs, seq_len, neg_rel2_embed)
# (batch, len, emb_size)
pos_embed = self.word_embed.forward(pos_rel)
# (batch, 1)
pos_score3 = self.matchPyramid(inputs, pos_embed, seq_len, pos_rel_len)
# (neg_size, batch)
pos_score3 = pos_score3.squeeze(-1).unsqueeze(0).expand(neg_size, batch)
# (neg_size*batch, len, emb_size)
neg_embed = self.word_embed.forward(neg_rel.view(-1, neg_len))
seqs_embed = inputs.unsqueeze(0).expand(neg_size, batch, inputs.size(1),
inputs.size(2)).contiguous().view(-1, inputs.size(1), inputs.size(2))
# (neg_size*batch,)
neg_rel_len = neg_rel_len.view(-1)
seq_len = seq_len.unsqueeze(0).expand(neg_size, batch).contiguous().view(-1)
# (neg_size*batch, 1)
neg_score3 = self.matchPyramid(seqs_embed, neg_embed, seq_len, neg_rel_len)
# (neg_size, batch)
neg_score3 = neg_score3.squeeze(-1).view(neg_size, batch)
return pos_score1+pos_score2+pos_score3, neg_score1+neg_score2+neg_score3
class ZYK(nn.Module):
def __init__(self, word_vocab, rel_vocab, config):
super(ZYK, self).__init__()
self.config = config
self.word_embed = Embeddings(word_vec_size=config.d_word_embed, dicts=word_vocab)
self.rel_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel_vocab)
if self.config.rnn_type.lower() == 'gru':
self.question_encoder_rnn = nn.GRU(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn, batch_first=True)
self.relation_encoder_rnn = nn.GRU(input_size=config.d_word_embed, hidden_size=config.d_rel_embed,
num_layers=1)
else:
self.question_encoder_rnn = nn.LSTM(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn, batch_first=True)
self.relation_encoder_rnn = nn.LSTM(input_size=config.d_word_embed, hidden_size=config.d_rel_embed,
num_layers=1)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden*2 if self.config.birnn else config.d_hidden
self.question_attention = MLPWordSeqAttention(input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size, config.d_rel_embed, 1, bias=False)
self.seq_out = nn.Sequential(
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed)
)
self.conv = nn.Sequential(
nn.Conv2d(1, config.channel_size, (config.conv_kernel_1, config.conv_kernel_2), stride=1,
padding=(config.conv_kernel_1//2, config.conv_kernel_2//2)), #channel_in=1, channel_out=8, kernel_size=3*3
nn.ReLU(True))
self.seq_maxlen = config.seq_maxlen + (config.conv_kernel_1 + 1) % 2
self.rel_maxlen = config.rel_maxlen + (config.conv_kernel_2 + 1) % 2
self.pooling = nn.MaxPool2d((config.seq_maxlen, 1),
stride=(config.seq_maxlen, 1), padding=0)
self.pooling2 = nn.MaxPool2d((1, config.rel_maxlen),
stride=(1, config.rel_maxlen), padding=0)
self.fc = nn.Sequential(
nn.Linear(config.rel_maxlen * config.channel_size, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20, 1))
self.fc1 = nn.Sequential(
nn.Linear(config.seq_maxlen * config.channel_size, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20,1))
self.fc2 = nn.Sequential(
nn.Linear(3, 1))
# we use rnn to encode the question
def question_encoder(self, inputs, _):
'''
:param inputs: (batch, max_seq_len, word_dim)
'''
batch_size = inputs.size(0)
state_shape = self.config.n_cells, batch_size, self.config.d_hidden
if self.config.rnn_type.lower() == 'gru':
h0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.question_encoder_rnn(inputs, h0)
else:
h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.question_encoder_rnn(inputs, (h0, c0))
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
outputs.contiguous()
return outputs, _
# def question_encoder(self, inputs, inputs_length):
# '''
# :param inputs: (batch, max_len, dim)
# '''
# batch_size = inputs.size(0)
# state_shape = self.config.n_cells, batch_size, self.config.d_hidden
# packed = torch.nn.utils.rnn.pack_padded_sequence(inputs, inputs_length)
# if self.config.rnn_type.lower() == 'gru':
# h0 = Variable(inputs.data.new(*state_shape).zero_())
# outputs, ht = self.question_encoder_rnn(packed, h0)
# else:
# h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
# outputs, (ht, ct) = self.question_encoder_rnn(packed, (h0, c0))
# outputs, output_lengths = torch.nn.utils.rnn.pad_packed_sequence(outputs)
# # shape of `outputs` - (batch size, sequence length, hidden size X num directions)
# return outputs, ht
def relation_encoder(self, inputs, inputs_length, h0):
'''
:param inputs: (batch, max_len, dim)
hidden: (batch, hidden_dim)
'''
device, batch_size = inputs.device, inputs.size(0)
inputs = inputs.transpose(0, 1).contiguous()
input_length_sorted = sorted(inputs_length, reverse=True)
sort_index = np.argsort(-np.array(inputs_length)).tolist()
input_sorted = Variable(torch.zeros(inputs.size())).to(device)
for b in range(batch_size):
input_sorted[:, b, :] = inputs[:, sort_index[b], :]
state_shape = 1, batch_size, self.config.d_rel_embed
packed = torch.nn.utils.rnn.pack_padded_sequence(inputs, input_length_sorted)
if self.config.rnn_type.lower() == 'gru':
if h0 is not None:
h0 = h0.unsqueeze(0)
else:
h0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.relation_encoder_rnn(packed, h0)
else:
if h0 is not None:
h0 = c0 = h0.unsqueeze(0)
else:
h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.relation_encoder_rnn(packed, (h0, c0))
outputs, output_lengths = torch.nn.utils.rnn.pad_packed_sequence(outputs)
outputs_resorted = Variable(torch.zeros(outputs.size())).to(device)
hidden_resorted = Variable(torch.zeros(ht.size())).to(device)
for b in range(batch_size):
outputs_resorted[:, sort_index[b], :] = outputs[:, b, :]
hidden_resorted[:, sort_index[b], :] = ht[:, b, :]
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
outputs_resorted = outputs_resorted.transpose(0, 1).contiguous()
return outputs_resorted, hidden_resorted.squeeze(0)
def cal_score(self, outputs, seqs_len, rel_embed, pos=None):
'''
:param rel_embed: (batch, dim2) or (neg_size, batch, dim2)
return: (batch, 1)
'''
batch_size = outputs.size(0)
if pos:
neg_size = pos
else:
neg_size, batch_size, embed_size = rel_embed.size()
seq_len, seq_emb_size = outputs.size()[1:]
outputs = outputs.unsqueeze(0).expand(neg_size, batch_size, seq_len,
seq_emb_size).contiguous().view(neg_size*batch_size, seq_len, -1)
rel_embed = rel_embed.view(neg_size * batch_size, -1)
seqs_len = seqs_len.unsqueeze(0).expand(neg_size, batch_size).contiguous().view(neg_size*batch_size)
# `weight` - (batch, length)
seq_att, weight = self.question_attention.forward(rel_embed, outputs)
# `seq_encode` - (batch, hidden size X num directions)
seq_encode = self.seq_out(seq_att)
# `score` - (batch, 1) or (neg_size * batch, 1)
score = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
if pos:
score = score.unsqueeze(0).expand(neg_size, batch_size, 1)
else:
score = score.view(neg_size, batch_size, 1)
return score
def matchPyramid(self, seq, rel, seq_len, rel_len):
'''
param:
seq: (batch, _seq_len, embed_size)
rel: (batch, _rel_len, embed_size)
seq_len: (batch,)
rel_len: (batch,)
return:
score: (batch, 1)
'''
batch_size = seq.size(0)
rel_trans = torch.transpose(rel, 1, 2)
# (batch, 1, seq_len, rel_len)
seq_norm = torch.sqrt(torch.sum(seq*seq, dim=2, keepdim=True))
rel_norm = torch.sqrt(torch.sum(rel_trans*rel_trans, dim=1, keepdim=True))
cross = torch.bmm(seq/seq_norm, rel_trans/rel_norm).unsqueeze(1)
# (batch, channel_size, seq_len, rel_len)
conv1 = self.conv(cross)
channel_size = conv1.size(1)
# (batch, seq_maxlen)
# (batch, rel_maxlen)
device = seq.device
dpool_index1, dpool_index2 = self.dynamic_pooling_index(seq_len, rel_len, self.seq_maxlen,
self.rel_maxlen)
dpool_index1 = dpool_index1.unsqueeze(1).unsqueeze(-1).expand(batch_size, channel_size,
self.seq_maxlen, self.rel_maxlen).to(device)
dpool_index2 = dpool_index2.unsqueeze(1).unsqueeze(2).expand_as(dpool_index1).to(device)
conv1_expand = torch.gather(conv1, 2, dpool_index1)
conv1_expand = torch.gather(conv1_expand, 3, dpool_index2)
# (batch, channel_size, p_size1, p_size2)
pool1 = self.pooling(conv1_expand).view(batch_size, -1)
# (batch, 1)
out = self.fc(pool1)
pool2 = self.pooling2(conv1_expand).view(batch_size, -1)
out2 = self.fc1(pool2)
return out, out2
def dynamic_pooling_index(self, len1, len2, max_len1, max_len2):
def dpool_index_(batch_idx, len1_one, len2_one, max_len1, max_len2):
# stride1 = 1.0 * max_len1 / len1_one
# stride2 = 1.0 * max_len2 / len2_one
stride1 = 1.0 * max_len1 / len1_one.item()
stride2 = 1.0 * max_len2 / len2_one.item()
idx1_one = [int(i/stride1) for i in range(max_len1)]
idx2_one = [int(i/stride2) for i in range(max_len2)]
return idx1_one, idx2_one
batch_size = len(len1)
index1, index2 = [], []
for i in range(batch_size):
idx1_one, idx2_one = dpool_index_(i, len1[i], len2[i], max_len1, max_len2)
index1.append(idx1_one)
index2.append(idx2_one)
index1 = torch.LongTensor(index1)
index2 = torch.LongTensor(index2)
return Variable(index1), Variable(index2)
def forward(self, batch):
"""
:param seqs: (batch_size,max_seq_len ), seq_len:(batch_size)
pos_rel:(batch_size)
pos_rel_word:(batch_size, max_rel_len), pos_rel_word_len:(batch_size)
neg_rel:(neg_size, batch_size)
neg_rel_word:(neg_size, batch_size, max_rel_len), neg_rel_word_len:(neg_size, batch_size)
:return:
"""
seqs, seq_len, pos_rel, pos_rel_word, pos_rel_word_len, neg_rel, neg_rel_word, neg_rel_word_len = batch
seqs = self.word_embed.forward(seqs) # seqs: (batch_size, max_seq_len, d_word_embed)
neg_size, batch, max_rel_len = neg_rel_word.size()
pos_rel_word = self.word_embed.forward(pos_rel_word) # pos_rel_word: (batch_size, max_rel_len, d_rel_embedding)
_, _, d_rel_embed = pos_rel_word.size()
neg_rel_word = self.word_embed.forward(neg_rel_word.view(-1, max_rel_len)) # neg_rel_word: (neg_size, batch_size, max_rel_len, d_rel_embedding)
neg_rel_word_len = neg_rel_word_len.view(-1) # neg_rel_word_len: (neg_size*batch_size)
pos_cnn_score1, pos_cnn_score2 = self.matchPyramid(seqs, pos_rel_word, seq_len, pos_rel_word_len)
pos_cnn_score1 = pos_cnn_score1.unsqueeze(0).expand(neg_size, batch, pos_cnn_score1.size(1))
pos_cnn_score2 = pos_cnn_score2.unsqueeze(0).expand(neg_size, batch, pos_cnn_score2.size(1))
seqs_extend = seqs.unsqueeze(0).expand(neg_size, batch, seqs.size(1),
seqs.size(2)).contiguous().view(-1, seqs.size(1), seqs.size(2))
seq_len_extend = seq_len.unsqueeze(0).expand(neg_size, batch).contiguous().view(-1)
neg_cnn_score1, neg_cnn_score2 = self.matchPyramid(seqs_extend, neg_rel_word, seq_len_extend, neg_rel_word_len)
neg_cnn_score1 = neg_cnn_score1.view(neg_size, batch, neg_cnn_score1.size(1))
neg_cnn_score2 = neg_cnn_score2.view(neg_size, batch, neg_cnn_score2.size(1))
seq_encoded, _ = self.question_encoder(seqs, seq_len)
pos_rel = self.rel_embed.word_lookup_table(pos_rel)
neg_rel = self.rel_embed.word_lookup_table(neg_rel.view(-1))
if hasattr(self.config, "rel_direct") and self.config.rel_direct == True:
pos_rel_encoded = pos_rel
neg_rel_encoded = neg_rel
else:
_, pos_rel_encoded = self.relation_encoder(pos_rel_word, pos_rel_word_len, pos_rel)
_, neg_rel_encoded = self.relation_encoder(neg_rel_word, neg_rel_word_len, neg_rel)
pos_rnn_score = self.cal_score(seq_encoded, seq_len, pos_rel_encoded, neg_size)
neg_rnn_score = self.cal_score(seq_encoded, seq_len, neg_rel_encoded.view(neg_size, batch, d_rel_embed))
if hasattr(self.config, "cnn") and self.config.cnn == False:
pos_cnn_score1 = pos_cnn_score1.fill_(0)
neg_cnn_score1 = neg_cnn_score1.fill_(0)
pos_concat = torch.cat((pos_cnn_score1, pos_cnn_score2, pos_rnn_score), 2)
neg_concat = torch.cat((neg_cnn_score1, neg_cnn_score2, neg_rnn_score), 2)
pos_score = self.fc2(pos_concat).squeeze(-1)
neg_score = self.fc2(neg_concat).squeeze(-1)
return pos_score, neg_score
class ZYK2(nn.Module):
def __init__(self, word_vocab, rel_vocab, config):
super(ZYK2, self).__init__()
self.config = config
self.word_embed = Embeddings(word_vec_size=config.d_word_embed, dicts=word_vocab)
self.rel_embed = Embeddings(word_vec_size=config.d_rel_embed, dicts=rel_vocab)
if self.config.rnn_type.lower() == 'gru':
self.question_encoder_rnn = nn.GRU(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn, batch_first=True)
else:
self.question_encoder_rnn = nn.LSTM(input_size=config.d_word_embed, hidden_size=config.d_hidden,
num_layers=config.n_layers, dropout=config.dropout_prob,
bidirectional=config.birnn, batch_first=True)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden*2 if self.config.birnn else config.d_hidden
self.question_attention = MLPWordSeqAttention(input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size, config.d_rel_embed, 1, bias=False)
self.seq_out = nn.Sequential(
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed)
)
# self.conv = nn.Sequential(
# nn.Conv2d(1, config.channel_size, (config.conv_kernel_1, config.conv_kernel_2), stride=1,
# padding=(config.conv_kernel_1//2, config.conv_kernel_2//2)), #channel_in=1, channel_out=8, kernel_size=3*3
# nn.ReLU(True))
self.seq_maxlen = config.seq_maxlen + (config.conv_kernel_1 + 1) % 2
self.rel_maxlen = config.rel_maxlen + (config.conv_kernel_2 + 1) % 2
self.pooling = nn.MaxPool2d((config.seq_maxlen, 1),
stride=(config.seq_maxlen, 1), padding=0)
self.pooling2 = nn.MaxPool2d((1, config.rel_maxlen),
stride=(1, config.rel_maxlen), padding=0)
self.fc = nn.Sequential(
nn.Linear(config.rel_maxlen, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20, 1))
self.fc1 = nn.Sequential(
nn.Linear(config.seq_maxlen, 20),
nn.ReLU(),
nn.Dropout(p=config.dropout_prob),
nn.Linear(20, 1))
self.fc2 = nn.Sequential(
nn.Linear(2, 1))
# we use rnn to encode the question
def question_encoder(self, inputs, _):
'''
:param inputs: (batch, max_seq_len, word_dim)
'''
batch_size = inputs.size(0)
state_shape = self.config.n_cells, batch_size, self.config.d_hidden
if self.config.rnn_type.lower() == 'gru':
h0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.question_encoder_rnn(inputs, h0)
else:
h0 = c0 = Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.question_encoder_rnn(inputs, (h0, c0))
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
outputs.contiguous()
return outputs, _
def cal_score(self, outputs, seqs_len, rel_embed, pos=None):
'''
:param rel_embed: (batch, dim2) or (neg_size, batch, dim2)
return: (batch, 1)
'''
batch_size = outputs.size(0)
if pos:
neg_size = pos
else:
neg_size, batch_size, embed_size = rel_embed.size()
seq_len, seq_emb_size = outputs.size()[1:]
outputs = outputs.unsqueeze(0).expand(neg_size, batch_size, seq_len,
seq_emb_size).contiguous().view(neg_size*batch_size, seq_len, -1)
rel_embed = rel_embed.view(neg_size * batch_size, -1)
seqs_len = seqs_len.unsqueeze(0).expand(neg_size, batch_size).contiguous().view(neg_size*batch_size)
# `weight` - (batch, length)
seq_att, weight = self.question_attention.forward(rel_embed, outputs)
# `seq_encode` - (batch, hidden size X num directions)
seq_encode = self.seq_out(seq_att)
# `score` - (batch, 1) or (neg_size * batch, 1)
score = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
if pos:
score = score.unsqueeze(0).expand(neg_size, batch_size, 1)
else:
score = score.view(neg_size, batch_size, 1)
return score
def matchPyramid(self, seq, rel, seq_len, rel_len):
'''
param:
seq: (batch, _seq_len, embed_size)
rel: (batch, _rel_len, embed_size)
seq_len: (batch,)
rel_len: (batch,)
return:
score: (batch, 1)
'''
batch_size = seq.size(0)
rel_trans = torch.transpose(rel, 1, 2)
# (batch, 1, seq_len, rel_len)
seq_norm = torch.sqrt(torch.sum(seq*seq, dim=2, keepdim=True))
rel_norm = torch.sqrt(torch.sum(rel_trans*rel_trans, dim=1, keepdim=True))
cross = torch.bmm(seq/seq_norm, rel_trans/rel_norm).unsqueeze(1)
# (batch, channel_size, seq_len, rel_len)
conv1 = cross # self.conv(cross)
channel_size = conv1.size(1)
# (batch, seq_maxlen)
# (batch, rel_maxlen)
device = seq.device
dpool_index1, dpool_index2 = self.dynamic_pooling_index(seq_len, rel_len, self.seq_maxlen,
self.rel_maxlen)
dpool_index1 = dpool_index1.unsqueeze(1).unsqueeze(-1).expand(batch_size, channel_size,
self.seq_maxlen, self.rel_maxlen).to(device)
dpool_index2 = dpool_index2.unsqueeze(1).unsqueeze(2).expand_as(dpool_index1).to(device)
conv1_expand = torch.gather(conv1, 2, dpool_index1)
conv1_expand = torch.gather(conv1_expand, 3, dpool_index2)
# (batch, channel_size, p_size1, p_size2)
pool1 = self.pooling(conv1_expand).view(batch_size, -1)
# (batch, 1)
out = self.fc(pool1)
pool2 = self.pooling2(conv1_expand).view(batch_size, -1)
out2 = self.fc1(pool2)
return out, out2
def dynamic_pooling_index(self, len1, len2, max_len1, max_len2):
def dpool_index_(batch_idx, len1_one, len2_one, max_len1, max_len2):
# stride1 = 1.0 * max_len1 / len1_one
# stride2 = 1.0 * max_len2 / len2_one
stride1 = 1.0 * max_len1 / len1_one.item()
stride2 = 1.0 * max_len2 / len2_one.item()
idx1_one = [int(i/stride1) for i in range(max_len1)]
idx2_one = [int(i/stride2) for i in range(max_len2)]
return idx1_one, idx2_one
batch_size = len(len1)
index1, index2 = [], []
for i in range(batch_size):
idx1_one, idx2_one = dpool_index_(i, len1[i], len2[i], max_len1, max_len2)
index1.append(idx1_one)
index2.append(idx2_one)
index1 = torch.LongTensor(index1)
index2 = torch.LongTensor(index2)
return Variable(index1), Variable(index2)
def forward(self, batch):
"""
:param seqs: (batch_size,max_seq_len ), seq_len:(batch_size)
pos_rel:(batch_size)
pos_rel_word:(batch_size, max_rel_len), pos_rel_word_len:(batch_size)
neg_rel:(neg_size, batch_size)
neg_rel_word:(neg_size, batch_size, max_rel_len), neg_rel_word_len:(neg_size, batch_size)
:return:
"""
seqs, seq_len, pos_rel, pos_rel_word, pos_rel_word_len, neg_rel, neg_rel_word, neg_rel_word_len = batch
seqs = self.word_embed.forward(seqs) # seqs: (batch_size, max_seq_len, d_word_embed)
neg_size, batch, max_rel_len = neg_rel_word.size()
pos_rel_word = self.word_embed.forward(pos_rel_word) # pos_rel_word: (batch_size, max_rel_len, d_rel_embedding)
_, _, d_word_embed = pos_rel_word.size()
neg_rel_word = self.word_embed.forward(neg_rel_word.view(-1, max_rel_len)) # neg_rel_word: (neg_size, batch_size, max_rel_len, d_rel_embedding)
neg_rel_word_len = neg_rel_word_len.view(-1) # neg_rel_word_len: (neg_size*batch_size)
pos_cnn_score1, pos_cnn_score2 = self.matchPyramid(seqs, pos_rel_word, seq_len, pos_rel_word_len)
pos_cnn_score1 = pos_cnn_score1.unsqueeze(0).expand(neg_size, batch, pos_cnn_score1.size(1))
pos_cnn_score2 = pos_cnn_score2.unsqueeze(0).expand(neg_size, batch, pos_cnn_score2.size(1))
seqs_extend = seqs.unsqueeze(0).expand(neg_size, batch, seqs.size(1),
seqs.size(2)).contiguous().view(-1, seqs.size(1), seqs.size(2))
seq_len_extend = seq_len.unsqueeze(0).expand(neg_size, batch).contiguous().view(-1)
neg_cnn_score1, neg_cnn_score2 = self.matchPyramid(seqs_extend, neg_rel_word, seq_len_extend, neg_rel_word_len)
neg_cnn_score1 = neg_cnn_score1.view(neg_size, batch, neg_cnn_score1.size(1))
# neg_cnn_score2 = neg_cnn_score2.view(neg_size, batch, neg_cnn_score2.size(1))
seq_encoded, _ = self.question_encoder(seqs, seq_len)
pos_rel = self.rel_embed.word_lookup_table(pos_rel)
_, d_rel_embed = pos_rel.size()
neg_rel = self.rel_embed.word_lookup_table(neg_rel.view(-1))
pos_rnn_score = self.cal_score(seq_encoded, seq_len, pos_rel, neg_size)
neg_rnn_score = self.cal_score(seq_encoded, seq_len, neg_rel.view(neg_size, batch, d_rel_embed))
pos_concat = torch.cat((pos_cnn_score1, pos_rnn_score), 2)
neg_concat = torch.cat((neg_cnn_score1, neg_rnn_score), 2)
pos_score = self.fc2(pos_concat).squeeze(-1)
neg_score = self.fc2(neg_concat).squeeze(-1)
return pos_score, neg_score
# import torch.nn as nn
# import sys
# import torch
# from pytorch_pretrained_bert import BertTokenizer, BertForPreTraining
# sys.path.append("../tools/")
# word_vocab = torch.load("../../data/vocab/vocab.seq_rel.pt")
# model = BertForPreTraining.from_pretrained('bert-base-uncased')
# tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
# model.eval()
#
#
# device = "cuda:2"
# model.to(device)
# vector_size, words = 768, []
# pretrained = torch.zeros(len(word_vocab), vector_size)
# for index, word in word_vocab.index2word.items():
# if word in tokenizer.vocab:
# tokens_tensor = torch.tensor([tokenizer.convert_tokens_to_ids([word])]).to(device)
# encoded_layers, _ = model(tokens_tensor)
# pretrained[index] = encoded_layers[0][0].cpu()
# else:
# nn.init.uniform(pretrained[index], 0, 0)
class RelationRanking(nn.Module):
def __init__(self, word_vocab, rel_vocab, config):
super(RelationRanking, self).__init__()
self.config = config
rel1_vocab, rel2_vocab = rel_vocab
self.word_embed = Embeddings(word_vec_size=config.d_word_embed,
dicts=word_vocab)
self.rel1_embed = Embeddings(word_vec_size=config.d_rel_embed,
dicts=rel1_vocab)
self.rel2_embed = Embeddings(word_vec_size=config.d_rel_embed,
dicts=rel2_vocab)
# print(self.rel_embed.word_lookup_table.weight.data)
#rel_embed的初始化待改 rel_embed.lookup_table.weight.data.normal_(0, 0.1)
if self.config.rnn_type.lower() == 'gru':
self.rnn = nn.GRU(input_size=config.d_word_embed,
hidden_size=config.d_hidden,
num_layers=config.n_layers,
dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
else:
self.rnn = nn.LSTM(input_size=config.d_word_embed,
hidden_size=config.d_hidden,
num_layers=config.n_layers,
dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden
if self.config.birnn:
seq_in_size *= 2
self.question_attention = MLPWordSeqAttention(
input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size,
config.d_rel_embed,
1,
bias=False)
self.seq_out = nn.Sequential(
# nn.BatchNorm1d(seq_in_size),
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed))
def question_encoder(self, inputs):
'''
:param inputs: (batch, dim1)
'''
batch_size = inputs.size(0)
state_shape = self.config.n_cells, batch_size, self.config.d_hidden
if self.config.rnn_type.lower() == 'gru':
h0 = autograd.Variable(inputs.data.new(*state_shape).zero_())
outputs, ht = self.rnn(inputs, h0)
else:
h0 = c0 = autograd.Variable(inputs.data.new(*state_shape).zero_())
outputs, (ht, ct) = self.rnn(inputs, (h0, c0))
outputs.contiguous()
# shape of `outputs` - (batch size, sequence length, hidden size X num directions)
# shape of `encoder` - (batch size, hidden size X num directions)
# encoder = ht[-1] if not self.config.birnn else ht[-2:].transpose(0,1).contiguous().view(batch_size, -1)
# seq_encode = self.seq_out(encoder)
return outputs
def cal_score(self, outputs, seqs_len, rel_embed, pos=None):
'''
:param rel_embed: (batch, dim2) or (neg_size, batch, dim2)
return: (batch, 1)
'''
batch_size = outputs.size(0)
if pos:
neg_size = pos
else: # neg的要扩展
neg_size, batch_size, embed_size = rel_embed.size()
seq_len, seq_emb_size = outputs.size()[1:]
outputs = outputs.unsqueeze(0).expand(
neg_size, batch_size, seq_len,
seq_emb_size).contiguous().view(neg_size * batch_size, seq_len,
-1)
rel_embed = rel_embed.view(neg_size * batch_size, -1)
seqs_len = seqs_len.unsqueeze(0).expand(
neg_size, batch_size).contiguous().view(neg_size * batch_size)
# `seq_encode` - (batch, hidden size X num directions)
# `weight` - (batch, length)
# seq_att, weight = self.question_attention.forward(rel_embed, outputs, seqs_len)
seq_att, weight = self.question_attention.forward(rel_embed, outputs)
# if pos:
# print('weight:', weight)
# seq_encode = self.dropout(seq_att)
seq_encode = self.seq_out(seq_att)
# `score` - (batch, 1) or (neg_size * batch, 1)
# score = self.bilinear(seq_encode, rel_embed)
# score = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
dot = torch.sum(seq_encode * rel_embed, 1, keepdim=True)
dis = seq_encode - rel_embed
euclidean = torch.sqrt(torch.sum(dis * dis, 1, keepdim=True))
score = (1 / (1 + euclidean)) * (1 / 1 + torch.exp(-(dot + 1)))
if pos: # pos要把结果扩展
score = score.squeeze(1).unsqueeze(0).expand(neg_size, batch_size)
else:
score = score.view(neg_size, batch_size)
return score
def forward(self, batch):
# shape of seqs (batch size, sequence length)
seqs, seq_len, pos_rel1, pos_rel2, neg_rel1, neg_rel2 = batch
# print('seqs:', seqs)
# shape (batch_size, sequence length, dimension of embedding)
inputs = self.word_embed.forward(seqs)
outputs = self.question_encoder(inputs)
# shape (batch_size, dimension of rel embedding)
pos_rel1_embed = self.rel1_embed.word_lookup_table(pos_rel1)
pos_rel2_embed = self.rel2_embed.word_lookup_table(pos_rel2)
pos_rel1_embed = self.dropout(pos_rel1_embed)
pos_rel2_embed = self.dropout(pos_rel2_embed)
# shape (neg_size, batch_size, dimension of rel embedding)
neg_rel1_embed = self.rel1_embed.word_lookup_table(neg_rel1)
neg_rel2_embed = self.rel2_embed.word_lookup_table(neg_rel2)
neg_rel1_embed = self.dropout(neg_rel1_embed)
neg_rel2_embed = self.dropout(neg_rel2_embed)
neg_size = neg_rel1_embed.size(0)
# shape of `score` - (neg_size, batch_size)
pos_score1 = self.cal_score(outputs, seq_len, pos_rel1_embed, neg_size)
pos_score2 = self.cal_score(outputs, seq_len, pos_rel2_embed, neg_size)
neg_score1 = self.cal_score(outputs, seq_len, neg_rel1_embed)
neg_score2 = self.cal_score(outputs, seq_len, neg_rel2_embed)
return pos_score1, pos_score2, neg_score1, neg_score2