def __init__(self, param):
"""
:param param: embedding, hidden_size, dropout_p, encoder_dropout_p, encoder_direction_num, encoder_layer_num
"""
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim - 9
self.encoder_a = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=False)
input_size = self.embedding.embedding_dim
self.encoder_p_q = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# match rnn
input_size = self.hidden_size * 2
self.match_rnn = match_rnn.MatchRNN(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
gated_attention=True,
is_bn=self.is_bn)
# addition_rnn
input_size = self.hidden_size * 2
self.addition_rnn = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
bidirectional=True,
dropout_p=self.dropout_p,
layer_num=1,
is_bn=self.is_bn)
# mean passage based on attn
self.mean_p = pointer.AttentionPooling(input_size=self.hidden_size * 2,
output_size=self.hidden_size *
2)
# outputs
self.choose = choose.Choose(self.hidden_size * 2, self.hidden_size * 2)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
# embedding
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True
)
# align
input_size = self.hidden_size * 2
self.align_1 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, False)
self.align_2 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, False)
self.align_3 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, True)
# pointer
self.pointer = Pointer(input_size, self.dropout_p)
def __init__(self, param):
super(Model, self).__init__()
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.w2i = param['embedding'].shape[1]
self.hidden_size = param['hidden_size']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
self.flag = True
if self.flag:
self.highway_c = Highway(self.w2i+6)
self.highway_q = Highway(self.w2i+4)
self.content_conv = DepthwiseSeparableConv(self.w2i+6, self.hidden_size, 5)
self.question_conv = DepthwiseSeparableConv(self.w2i+4, self.hidden_size, 5)
self.c_enc = EncoderBlock(conv_num=2, d=self.hidden_size, k=5, length=Max_Content_len, dropout_p=self.dropout_p)
self.q_enc = EncoderBlock(conv_num=2, d=self.hidden_size, k=5, length=Max_Question_len, dropout_p=self.dropout_p)
self.cq_att = CQAttention(self.hidden_size, self.dropout_p)
self.cq_resizer = DepthwiseSeparableConv(self.hidden_size*4, self.hidden_size, 5)
self.model_enc_blks = nn.ModuleList([EncoderBlock(conv_num=2, d=self.hidden_size, k=5, length=Max_Content_len,
dropout_p=self.dropout_p) for _ in range(3)])
self.pointer = Pointer(self.hidden_size)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.k = param['k']
self.perspective_num = 10
self.lamda = 10
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder: p, q, a
input_size = self.embedding.embedding_dim - 9
self.encoder = encoder.Rnn(
mode='LSTM',
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=False,
)
# attention: q2p
self.q2p = MultiPerspective(self.hidden_size, self.perspective_num, self.dropout_p)
# attention: a2p
self.a2p = MultiPerspective(self.hidden_size, self.perspective_num, self.dropout_p)
# attention: Mq2Ma
self.mq2ma = MultiPerspective(self.perspective_num*4, self.perspective_num, self.dropout_p)
# attention: Ma2Mq
self.ma2mq = MultiPerspective(self.perspective_num*4, self.perspective_num, self.dropout_p)
# aggregation
self.aggregation = Aggregation(self.perspective_num, self.hidden_size, self.dropout_p)
# memory
self.q2p = MultiPerspective(self.hidden_size, self.perspective_num, self.dropout_p)
self.m_rnn = encoder.Rnn(
mode='LSTM',
input_size=self.perspective_num*8,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
layer_num=1,
is_bn=self.is_bn,
)
# answer score
self.answer_score = AnswerScore(self.hidden_size, self.k, self.lamda, self.dropout_p)
self.dropout = nn.Dropout(self.dropout_p)
def __init__(self, param):
"""
:param param: embedding, hidden_size, dropout_p, encoder_dropout_p, encoder_direction_num, encoder_layer_num
"""
super(Model, self).__init__()
self.w2v_size = param['embedding'].shape[1]
self.vocab_size = param['embedding'].shape[0]
self.embedding_type = param['embedding_type']
self.embedding_is_training = param['embedding_is_training']
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_bidirectional = param['encoder_bidirectional']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
if self.embedding_type == 'standard':
self.embedding = embedding.Embedding(param['embedding'])
is_bn = False
else:
self.embedding = embedding.ExtendEmbedding(param['embedding'])
is_bn = True
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=self.encoder_bidirectional,
layer_num=self.encoder_layer_num,
is_bn=is_bn
)
# match rnn
input_size = self.hidden_size * 2 if self.encoder_bidirectional else self.hidden_size
self.match_rnn = match_rnn.MatchRNN(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
gated_attention=False,
is_bn=self.is_bn
)
# pointer
self.pointer_net = pointer.BoundaryPointer(
mode=self.mode,
input_size=self.hidden_size*2,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
is_bn=self.is_bn
)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
# embedding
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True
)
# attention flow layer
self.att_c = nn.Linear(self.hidden_size * 2, 1)
self.att_q = nn.Linear(self.hidden_size * 2, 1)
self.att_cq = nn.Linear(self.hidden_size * 2, 1)
# modeling layer
self.modeling_rnn = encoder.Rnn(
mode=self.mode,
input_size=self.hidden_size * 8,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
layer_num=2,
is_bn=self.is_bn
)
# outputs
self.p1 = nn.Linear(self.hidden_size * 10, 1)
self.p2 = nn.Linear(self.hidden_size * 10, 1)
self.rnn = encoder.Rnn(
mode=self.mode,
input_size=self.hidden_size * 2,
hidden_size=self.hidden_size,
bidirectional=True,
dropout_p=self.dropout_p,
layer_num=1,
is_bn=self.is_bn
)
self.dropout = nn.Dropout(self.dropout_p)
def __init__(self, param):
"""
:param param: embedding, hidden_size, dropout_p, encoder_dropout_p, encoder_direction_num, encoder_layer_num
"""
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# match rnn
input_size = self.hidden_size * 2
self.match_rnn = match_rnn.MatchRNN(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
gated_attention=True,
is_bn=self.is_bn)
# addition_rnn
input_size = self.hidden_size * 2
self.addition_rnn = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
bidirectional=True,
dropout_p=self.dropout_p,
layer_num=1,
is_bn=self.is_bn)
# init state of pointer
self.init_state = pointer.AttentionPooling(
input_size=self.hidden_size * 2, output_size=self.hidden_size)
# pointer
self.pointer_net = pointer.BoundaryPointer(
mode=self.mode,
input_size=self.hidden_size * 2,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
is_bn=self.is_bn)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder_p: 双向rnn
input_size = self.embedding.embedding_dim
self.encoder_p = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# encoder_q: 单向rnn
self.encoder_q = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=False,
layer_num=self.encoder_layer_num,
is_bn=True)
# encoder_a: 单向rnn
input_size = self.embedding.sd_embedding.embedding_dim
self.encoder_a = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=False,
layer_num=self.encoder_layer_num,
is_bn=False)
# similar W
self.sim_w = nn.Linear(self.hidden_size, self.hidden_size * 2)
# outputs
self.choose = choose.Choose(self.hidden_size * 2, self.hidden_size)
self.dropout = nn.Dropout(self.dropout_p)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
# embedding
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder: p
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=self.is_bn
)
self.mean_q = pointer.AttentionPooling(self.hidden_size*2, self.hidden_size)
self.mean_a = pointer.AttentionPooling(self.hidden_size*2, self.hidden_size)
# align
input_size = self.hidden_size * 2
self.align_1 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, False)
self.align_2 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, False)
self.align_3 = Aligner(input_size, self.dropout_p, self.mode, self.is_bn, True)
# p_rep, choosing
self.wp1 = nn.Linear(self.hidden_size*2, self.hidden_size)
self.wp2 = nn.Linear(self.hidden_size*2, self.hidden_size)
self.vp = nn.Linear(self.hidden_size, 1)
self.bi_linear = nn.Linear(self.hidden_size*2, self.hidden_size*2)
self.dropout = nn.Dropout(self.dropout_p)
self.reset_param()
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder: p
input_size = self.embedding.embedding_dim
self.encoder_p = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# encoder: q
self.encoder_q = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# encoder: a
input_size = self.embedding.embedding_dim - 9
self.encoder_a = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size // 2,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=False)
# self-att: a
self.a_att = nn.Linear(self.hidden_size, 1, bias=False)
# attention flow layer
self.att_c = nn.Linear(self.hidden_size * 2, 1)
self.att_q = nn.Linear(self.hidden_size * 2, 1)
self.att_cq = nn.Linear(self.hidden_size * 2, 1)
# modeling layer
self.modeling_rnn = encoder.Rnn(mode=self.mode,
input_size=self.hidden_size * 8,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
layer_num=2,
is_bn=self.is_bn)
# prediction
self.wq = nn.Linear(self.hidden_size * 2, self.hidden_size, bias=False)
self.vq = nn.Linear(self.hidden_size, 1, bias=False)
self.wp1 = nn.Linear(self.hidden_size * 10,
self.hidden_size,
bias=False)
self.wp2 = nn.Linear(self.hidden_size * 2,
self.hidden_size,
bias=False)
self.vp = nn.Linear(self.hidden_size, 1, bias=False)
self.predict = nn.Linear(self.hidden_size * 10,
self.hidden_size,
bias=False)
self.dropout = nn.Dropout(self.dropout_p)
self.reset_param()
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.k = 2
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder_a
input_size = self.embedding.embedding_dim - 9
self.encoder_a = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=False)
# encoder p, q
self.doc_list = nn.ModuleList()
self.query_list = nn.ModuleList()
input_size = self.embedding.embedding_dim
for i in range(self.k):
di_enc = encoder.Rnn(
mode=self.mode,
input_size=input_size if i == 0 else self.hidden_size * 2,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
layer_num=1,
is_bn=self.is_bn)
qi_enc = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.dropout_p,
bidirectional=True,
layer_num=1,
is_bn=self.is_bn)
self.doc_list.append(di_enc)
self.query_list.append(qi_enc)
# mean passage based on attn
self.mean_p = pointer.AttentionPooling(input_size=self.hidden_size * 2,
output_size=self.hidden_size *
2)
# mean answer based on attn
self.mean_a = pointer.AttentionPooling(input_size=self.hidden_size * 2,
output_size=self.hidden_size *
2)
# outputs
self.choose = choose.Choose(self.hidden_size * 2, self.hidden_size * 2)
self.dropout = nn.Dropout(self.dropout_p)
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.num_align_hops = param['num_align_hops'] # 2
# embedding
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=self.is_bn # 应该修改一下
)
self.mean_q = pointer.AttentionPooling(self.hidden_size * 2,
self.hidden_size)
self.mean_a = pointer.AttentionPooling(self.hidden_size * 2,
self.hidden_size)
# merge q into p
self.aligner = nn.ModuleList(
[SeqToSeqAtten() for _ in range(self.num_align_hops)])
self.aligner_sfu = nn.ModuleList([
SFU(self.hidden_size * 2,
self.hidden_size * 2 * 3,
dropout_p=self.dropout_p) for _ in range(self.num_align_hops)
])
# self align
self.self_aligner = nn.ModuleList(
[SelfSeqAtten() for _ in range(self.num_align_hops)])
self.self_aligner_sfu = nn.ModuleList([
SFU(self.hidden_size * 2,
self.hidden_size * 2 * 3,
dropout_p=self.dropout_p) for _ in range(self.num_align_hops)
])
# aggregation
self.choose_agg = nn.ModuleList([
encoder.Rnn(mode=self.mode,
input_size=self.hidden_size * 2,
hidden_size=self.hidden_size,
bidirectional=True,
dropout_p=self.dropout_p,
layer_num=1,
is_bn=self.is_bn) for _ in range(self.num_align_hops)
])
# p_rep, choosing
self.wp1 = nn.Linear(self.hidden_size * 2, self.hidden_size)
self.wp2 = nn.Linear(self.hidden_size * 2, self.hidden_size)
self.vp = nn.Linear(self.hidden_size, 1)
self.bi_linear = nn.Linear(self.hidden_size * 2, self.hidden_size * 2)
self.dropout = nn.Dropout(self.dropout_p)
self.reset_param()
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder: p
input_size = self.embedding.embedding_dim
self.encoder_p = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True
)
# encoder: q
self.encoder_q = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True
)
# encoder: a
input_size = self.embedding.embedding_dim - 9
self.encoder_a = encoder.Rnn(
mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size//2,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=False
)
# self-att: a
self.a_att = nn.Linear(self.hidden_size, 1, bias=False)
# Concat Attention
self.Wc1 = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.Wc2 = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.vc = nn.Linear(self.hidden_size, 1, bias=False)
# Bilinear Attention
self.Wb = nn.Linear(self.hidden_size*2, self.hidden_size*2, bias=False)
# Dot Attention :
self.Wd = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.vd = nn.Linear(self.hidden_size, 1, bias=False)
# Minus Attention :
self.Wm = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.vm = nn.Linear(self.hidden_size, 1, bias=False)
self.gru_agg = encoder.Rnn(
mode=self.mode,
input_size=self.hidden_size*10,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=self.is_bn
)
# prediction
self.wq = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.vq = nn.Linear(self.hidden_size, 1, bias=False)
self.wp1 = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.wp2 = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.vp = nn.Linear(self.hidden_size, 1, bias=False)
self.predict = nn.Linear(self.hidden_size*2, self.hidden_size, bias=False)
self.dropout = nn.Dropout(self.dropout_p)
self.reset_param()
def __init__(self, param):
super(Model, self).__init__()
self.mode = param['mode']
self.hidden_size = param['hidden_size']
self.dropout_p = param['dropout_p']
self.encoder_dropout_p = param['encoder_dropout_p']
self.encoder_layer_num = param['encoder_layer_num']
self.is_bn = param['is_bn']
# embedding
self.embedding = embedding.ExtendEmbedding(param['embedding'])
# encoder
input_size = self.embedding.embedding_dim
self.encoder = encoder.Rnn(mode=self.mode,
input_size=input_size,
hidden_size=self.hidden_size,
dropout_p=self.encoder_dropout_p,
bidirectional=True,
layer_num=self.encoder_layer_num,
is_bn=True)
# align
self.aligner = nn.ModuleList(
[SeqToSeqAtten() for _ in range(num_align_hops)])
self.aligner_sfu = nn.ModuleList([
SFU(self.hidden_size * 2,
self.hidden_size * 2 * 3,
dropout_p=self.dropout_p) for _ in range(num_align_hops)
])
# self align
self.self_aligner = nn.ModuleList(
[SelfSeqAtten() for _ in range(num_align_hops)])
self.self_aligner_sfu = nn.ModuleList([
SFU(self.hidden_size * 2,
self.hidden_size * 2 * 3,
dropout_p=self.dropout_p) for _ in range(num_align_hops)
])
# aggregation
self.aggregation = nn.ModuleList([
encoder.Rnn(mode=self.mode,
input_size=self.hidden_size * 2,
hidden_size=self.hidden_size,
bidirectional=True,
dropout_p=self.dropout_p,
layer_num=1,
is_bn=self.is_bn) for _ in range(num_align_hops)
])
# init zs
self.init_state = pointer.AttentionPooling(
input_size=self.hidden_size * 2, output_size=self.hidden_size * 2)
# pointer
self.ptr_net = nn.ModuleList([
Pointer(self.hidden_size * 2, self.hidden_size, self.dropout_p)
for _ in range(num_ptr_hops)
])