def __init__(self, word_vectors, hidden_size, char_vectors, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
char_vectors = char_vectors) # added last line
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
### start our code:
self.selfattention = layers.SelfAttention(input_size = 8 * hidden_size,
hidden_size=hidden_size,
dropout = 0.2)
### end our code
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.1):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
# self.enc = layers.RNNEncoder(input_size=hidden_size,
# hidden_size=hidden_size,
# num_layers=1,
# drop_prob=drop_prob)
self.emb_encoder = layers.EmbeddingEncoder(d_model=hidden_size, drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=4 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
# self.model_encoder = layers.ModelEncoder(d_model=hidden_size, drop_prob=drop_prob)
#
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size, use_pos, use_ner, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.use_pos = use_pos
self.use_ner = use_ner
rnn_input_size = hidden_size
if use_pos:
rnn_input_size += 1
if use_ner:
rnn_input_size += 1
self.enc = layers.RNNEncoder(input_size=rnn_input_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self,
word_vectors,
char_vectors,
hidden_size,
num_heads=8,
drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
hidden_size *= 2 # update hidden size for other layers due to char embeddings
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self,
word_vectors,
char_vectors,
pos_vectors,
ner_vectors,
iob_vectors,
hidden_size,
drop_prob=0.):
super(BiDAF_CharTag, self).__init__()
self.hidden_size = hidden_size
self.emb = layers.EmbeddingCharTag(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
iob_vectors=iob_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=self.hidden_size,
hidden_size=self.hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * self.hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * self.hidden_size,
hidden_size=self.hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=self.hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, args):
super(BiDAFExtra, self).__init__()
self.c_emb = layers.EmbeddingExtra(word_vectors=word_vectors,
args=args,
aux_feat=True)
self.q_emb = layers.EmbeddingExtra(word_vectors=word_vectors,
args=args,
aux_feat=False)
self.c_enc = layers.RNNEncoder(input_size=args.hidden_size + args.num_features,
hidden_size=args.hidden_size,
num_layers=1,
drop_prob=args.drop_prob if hasattr(args, 'drop_prob') else 0.)
self.q_enc = layers.RNNEncoder(input_size=args.hidden_size,
hidden_size=args.hidden_size,
num_layers=1,
drop_prob=args.drop_prob if hasattr(args, 'drop_prob') else 0.)
self.att = layers.BiDAFAttention(hidden_size=2 * args.hidden_size,
drop_prob=args.drop_prob if hasattr(args, 'drop_prob') else 0.)
self.mod = layers.RNNEncoder(input_size=8 * args.hidden_size,
hidden_size=args.hidden_size,
num_layers=2,
drop_prob=args.drop_prob if hasattr(args, 'drop_prob') else 0.)
self.out = layers.BiDAFOutput(hidden_size=args.hidden_size,
drop_prob=args.drop_prob if hasattr(args, 'drop_prob') else 0.)
self.args = args
def __init__(self, word_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
self.proj_bert_down = nn.Linear(in_features=768,
out_features=hidden_size,
bias=True)
nn.init.xavier_uniform_(self.proj_bert_down.weight, gain=1)
self.proj_glove_down = nn.Linear(in_features=300,
out_features=hidden_size,
bias=True)
nn.init.xavier_uniform_(self.proj_glove_down.weight, gain=1)
def __init__(self,
word_vectors,
char_vec,
word_len,
hidden_size,
emb_size=500,
drop_prob=0.):
super(BiDAFChar2, self).__init__()
self.emb = layers.EmbeddingWithChar(word_vectors=word_vectors,
hidden_size=emb_size,
char_vec=char_vec,
word_len=word_len,
drop_prob=drop_prob,
char_prop=0.4,
hwy_drop=drop_prob,
char_dim=200,
act='gelu')
self.resize = nn.Linear(emb_size, hidden_size)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size, char_vectors, drop_prob=0.):
super(SelfAttention_and_global, self).__init__()
self.hidden_size = hidden_size
self.emb = layers.Char_Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(
hidden_size=2 * hidden_size,
drop_prob=drop_prob) # replace this with yours
self.self_att = layers.SelfAttention(hidden_size=8 * hidden_size,
drop_prob=drop_prob)
self.second_mod = layers.RNNEncoder(input_size=16 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
att_size=16 * hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF_charCNN, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.char_emb = layers.CharEmbedding(char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.hwy = layers.HighwayEncoder(2, 2*hidden_size)
self.enc = layers.RNNEncoder(input_size=2*hidden_size,
hidden_size=2*hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * 2*hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * 2*hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self,
word_vectors,
char_vectors,
pos_vectors,
ner_vectors,
hidden_size,
drop_prob=0.,
freeze_tag=True):
super(BiDAF_tag_ext, self).__init__()
self.emb = layers.Embedding_Tag_Ext(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
freeze_tag=freeze_tag)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
# print("vectors: ", word_vectors)
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.self_att = layers.SelfAttention(hidden_size=8 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
self.batch_size = 64
self.hidden_size = hidden_size
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.hidden_size = 2 * hidden_size # As we concatinating word vectors and Char
# vectors
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = Encoder(dim=self.hidden_size,
depth=1,
heads=3,
ff_glu=True,
ff_dropout=self.drop_prob,
attn_dropout=self.drop_prob,
use_scalenorm=True,
position_infused_attn=True)
self.att = layers.TBiDAFAttention(hidden_size=self.hidden_size,
drop_prob=drop_prob)
self.mod = Encoder(dim=2 * self.hidden_size,
depth=3,
heads=6,
ff_glu=True,
ff_dropout=self.drop_prob,
attn_dropout=self.drop_prob,
use_scalenorm=True,
position_infused_attn=True)
self.out = layers.BiDAFOutput(hidden_size=self.hidden_size,
drop_prob=self.drop_prob)
def __init__(self,
word_vectors,
word_vectors_char,
hidden_size,
drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
word_vectors_char=word_vectors_char,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size, char_vectors, drop_prob=0.):
super(BiDAF, self).__init__()
self.hidden_size = hidden_size
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
char_vectors = char_vectors) # added character vectors
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
### start our code:
self.selfattention = layers.SelfAttention(input_size = 8 * hidden_size,
hidden_size=hidden_size,
dropout = 0.2)
### end our code
self.linear = nn.Linear(in_features = 8*self.hidden_size, out_features = 2*self.hidden_size, bias=True)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=4,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size,
char_dict_size, char_emb_size, \
conv_kernel_size, conv_depth1, \
conv_output_hidden_size, drop_prob=0.):
super(BiDAF_CBE, self).__init__()
word_vectors, hidden_size, drop_prob, \
char_dict_size, char_emb_size, \
conv_kernel_size, conv_depth1, \
conv_output_hidden_size
self.emb = layers.EmbeddingWithCharLevel(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
char_dict_size=char_dict_size,
char_emb_size=char_emb_size,
conv_kernel_size=conv_kernel_size,
conv_depth1=conv_depth1,
conv_output_hidden_size=conv_output_hidden_size)
self.enc = layers.RNNEncoder(input_size=2*hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
# self.enc = layers.RNNEncoder(input_size=hidden_size,
# hidden_size=hidden_size,
# num_layers=1,
# drop_prob=drop_prob)
# self.transformer = make_model(word_vectors, drop_prob, hidden_size)
self.emb_enc = EncoderBlock(conv_num=4, ch_num=64, k=7)
self.att = layers.BiDAFAttention(hidden_size=hidden_size,
drop_prob=drop_prob)
# TODO
self.mod = layers.RNNEncoder(input_size=4 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.,twist_embeddings=False):
super(BiDAF_charCNN_BERTEnc_BERTMod, self).__init__()
###
self.twist_embeddings = twist_embeddings
idx_list = []
for i in range(hidden_size):
idx_list.append(i)
idx_list.append(hidden_size+i)
self.register_buffer('idx_twist',torch.tensor(idx_list))
###
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.char_emb = layers.CharEmbedding(char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.hwy = layers.HighwayEncoder(2, 2*hidden_size)
self.enc = bert_layers.BertEncoder(n_layers=3, #n_layers=4,
d_feature=2*hidden_size,
n_heads=8,
out_size=2*hidden_size,
#d_ff=2048,
d_ff = 2*hidden_size,
dropout_prob=0.1,
#dropout_prob=drop_prob,
ff_activation=F.relu)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = bert_layers.BertEncoder(n_layers=3, #n_layers=3,
d_feature=8*hidden_size,
n_heads=8,
out_size=2*hidden_size,
#d_ff=2048,
d_ff = 2*hidden_size,
dropout_prob=0.1,
#dropout_prob=drop_prob,
ff_activation=F.relu)
# self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
# hidden_size=hidden_size,
# num_layers=2,
# drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, vectors, hidden_size, char_limit, use_transformer, use_GRU, drop_prob=.1, **kwargs):
super(BiDAF, self).__init__()
self.use_transformer = use_transformer
self.use_GRU = use_GRU
self.hidden_size = hidden_size
self.emb = layers.Embedding(vectors=vectors,
c2w_size=kwargs['c2w_size'],
hidden_size=hidden_size,
drop_prob=drop_prob,
char_limit=char_limit)
if not use_transformer:
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size, # output = 2*hidden_size
num_layers=1,
drop_prob=drop_prob,
use_GRU=use_GRU)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size, # output = 2*hidden_size
num_layers=2,
drop_prob=drop_prob,
use_GRU=use_GRU)
self.out = layers.BiDAFOutput(hidden_size=2 * hidden_size, drop_prob=drop_prob,
use_transformer=use_transformer)
else:
self.heads = kwargs['heads']
self.inter_size = kwargs['inter_size']
self.enc = layers.TransformerEncoderStack(
N=kwargs['enc_blocks'],
heads=self.heads,
input_size=hidden_size,
output_size=hidden_size,
inter_size=self.inter_size,
num_conv=kwargs['enc_convs'],
drop_prob=drop_prob,
p_sdd=kwargs['p_sdd']
)
self.squeeze = layers.InitializedLayer(4*hidden_size, hidden_size, bias=False)
self.mod = layers.TransformerEncoderStack(
N=kwargs['mod_blocks'],
heads=self.heads,
input_size=hidden_size,
output_size=hidden_size,
inter_size=self.inter_size,
num_conv=kwargs['mod_convs'],
drop_prob=drop_prob,
p_sdd=kwargs['p_sdd']
)
self.out = layers.QAOutput(2*hidden_size)
self.att = layers.BiDAFAttention(hidden_size=(1 if self.use_transformer else 2)*hidden_size,
drop_prob=drop_prob) # (batch_size, seq_len, 4*input_hidden_size)
def __init__(self, word_vectors, hidden_size, char_vectors, drop_prob=0.):
super(Final_Model, self).__init__()
self.hidden_size = hidden_size
self.emb = layers.Char_Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.pointnetGlobal = layers.PointNet(hidden_size=hidden_size,
kernel_size=1)
self.WordCNN = layers.WordCNN(hidden_size=hidden_size,
kernel_size=5,
padding=2)
self.enc_global = layers.RNNEncoder(input_size=2 * hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
# self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
# drop_prob=drop_prob) # replace this with yours
self.global_att = layers.GlobalBiDAFAttention(hidden_size=2 *
hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=10 * hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.self_att = layers.SelfAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.second_mod = layers.RNNEncoder(input_size=4 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
att_size=4 * hidden_size,
drop_prob=drop_prob)
def __init__(self, weights_matrix, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(weights_matrix=weights_matrix,
hidden_size=hidden_size)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size)
def __init__(self, word_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.TPRRNN(word_emb_size=(8 * hidden_size),
n_symbols=100,
d_symbols=10,
n_roles=20,
d_roles=10,
hidden_size=hidden_size)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0., enable_EM=True, enable_posner=True, enable_selfatt=True):
super(BiDAF, self).__init__()
self.embd_size = hidden_size
self.d = self.embd_size * 2 # word_embedding + char_embedding
self.enable_EM = enable_EM
if enable_EM:
self.d += 2 # word_feature
if enable_posner:
self.d += 10 # word_feature
self.emb = layers.Embedding(word_vectors=word_vectors, char_vectors=char_vectors,
hidden_size=self.embd_size,
drop_prob=drop_prob, enable_posner=enable_posner)
self.enc = layers.RNNEncoder(input_size=self.d,
hidden_size=self.d,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * self.d,
drop_prob=drop_prob)
self.enable_selfatt = enable_selfatt
if enable_selfatt:
# self.selfMatch = layers.SelfMatcher(in_size = 8 * self.d,
# drop_prob=drop_prob)
self.selfMatch = layers.StaticDotAttention(memory_size = 2 * self.d,
input_size = 2 * self.d, attention_size = 2 * self.d,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=4 * self.d,
hidden_size=self.d,
num_layers=2,
drop_prob=drop_prob)
else:
self.mod = layers.RNNEncoder(input_size=2 * self.d,
hidden_size=self.d,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=self.d,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.hidden_size = hidden_size
self.word_emb = layers.WordEmbedding(word_vectors, hidden_size)
self.char_emb = layers.CharEmbedding(char_vectors, hidden_size)
# assert hidden_size * 2 == (char_channel_size + word_dim)
# highway network
self.hwy = layers.HighwayEncoder(2, hidden_size * 2)
# highway network
# for i in range(2):
# setattr(self, f'highway_linear{i}', nn.Sequential(
# nn.Linear(hidden_size * 2, hidden_size * 2), nn.ReLU()))
# setattr(self, f'hightway_gate{i}', nn.Sequential(
# nn.Linear(hidden_size * 2, hidden_size * 2), nn.Sigmoid()))
# self.emb = layers.Embedding(word_vectors=word_vectors,
# hidden_size=hidden_size,
# drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size * 2,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self,
word_vectors,
char_vocab_size,
char_dim,
hidden_size,
drop_prob=0.,
kernel_size=5,
padding=1):
super(BiDAF_Char, self).__init__()
self.char_emb = layers.Char_Embedding(char_vocab_size=char_vocab_size,
char_dim=char_dim,
drop_prob=drop_prob,
hidden_size=hidden_size,
kernel_size=kernel_size,
padding=padding)
self.word_emb = layers.Word_Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.hwy = layers.HighwayEncoder(num_layers=2, hidden_size=hidden_size)
self.enc = layers.RNNEncoder(input_size=2 * hidden_size, # 08/09 注意这里改了input_size。因为经过highway后char_emb+word_emb的concatenation (bs, seq_len, 2*h)
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self,
word_mat,
w_embedding_size,
c_embeding_size,
c_vocab_size,
hidden_size,
num_head=1,
drop_prob=0.2):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_mat, w_embedding_size,
c_embeding_size, c_vocab_size, hidden_size,
drop_prob)
self.enc = layers.RNNEncoder(input_size=w_embedding_size + hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.var_dropout = layers.VariationalDropout(drop_prob,
batch_first=True)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.linear_trans = nn.Sequential(
nn.Linear(8 * hidden_size, 2 * hidden_size), nn.ReLU())
self.attn_mod = layers.RNNEncoder(hidden_size * 2,
hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.self_attn = layers.BiDAFSelfAttention(num_head, 2 * hidden_size)
self.linear_attn = nn.Sequential(
nn.Linear(2 * hidden_size, 2 * hidden_size), nn.ReLU())
self.mod = layers.RNNEncoder(input_size=2 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, ch_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
torch.cuda.empty_cache()
self.emb = layers.Embedding(word_vectors=word_vectors, ch_vectors=ch_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.chunk = layers.ChunkLayer(hidden_size=hidden_size, max_ans_len=10)
self.out = layers.BiDAFOutput(hidden_size=hidden_size, max_ans_len=10)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF_RNet, self).__init__()
self.emb = layers.WordCharEmbedding(word_vectors=word_vectors,
char_vectors=char_vectors,
cnn_size=16,
hidden_size=hidden_size,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.selfatt = layers.SelfMatchingAttention(input_size=8 * hidden_size,
hidden_size=4 * hidden_size,
num_layers=3,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0.):
super(BiDAF, self).__init__()
self.hidden_size = hidden_size * 2 # adding the char embedding, double the hidden_size.
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob)
#input_size=self.hidden_size+2 is due to we add two extra features (avg_attention) to both char embedding
#and word embedding to boost the performance. The avg_attention is use the attention mechanism to learn
#a weighted average among the vectors by the model itself.
self.enc = layers.RNNEncoder(
input_size=self.hidden_size + 2,
hidden_size=self.hidden_size,
# num_layers=2, # The number of layer can be changed, but less or no improvement.
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * self.hidden_size,
drop_prob=drop_prob)
#Add extra layer of self-attention based on the paper 'Simple and Effective Multi-Paragraph Reading Comprehension'
#URL: https://arxiv.org/pdf/1710.10723.pdf
self.self_att = layers.SelfAtt(hidden_size=2 * self.hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * self.hidden_size,
hidden_size=self.hidden_size,
num_layers=2,
drop_prob=drop_prob)
#Planned to use cosine similarity or TF-IDF to add extra feature to the embedding but need more thoughts to well
#implement
# self.sim = nn.CosineSimilarity(dim=1, eps=1e-6)
self.out = layers.BiDAFOutput(hidden_size=self.hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size, char_vocab_size, drop_prob=0., bidaf_layers = 2):
super(BiDAF, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
char_vocab_size=char_vocab_size,
char_embedding_size=100,
kernel_size=5,
drop_prob=drop_prob)
self.enc = layers.RNNEncoder(input_size=2 * hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.encs_att = nn.ModuleList([layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob) for _ in range(bidaf_layers)])
self.atts = nn.ModuleList([layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob) for _ in range(bidaf_layers)])
self.gates = nn.ModuleList([nn.Linear(8 * hidden_size, 8 * hidden_size) for _ in range(bidaf_layers)])
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
self.drop_out = nn.Dropout(drop_prob)