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, 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,
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,
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, 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, 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_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, 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_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, 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,
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, 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, 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, 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, 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, kernel_size, filters, drop_prob=0.):
super(QANet, self).__init__()
# Input embedding layer
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
char_vectors=char_vectors) # added character vectors
# resize input embedding layer output size to fit embedding encoder layer input size
self.resize_emb_pe = nn.Linear(in_features = hidden_size, out_features = filters, bias=False)
# Embedding encoder layer
self.emb_enc = qa.EncoderBlock(input_size=filters, kernel_size=kernel_size,
filters=filters, num_conv_layers=4, drop_prob=drop_prob)
# Context-Query attention layer
self.att = layers.BiDAFAttention(hidden_size = filters,
drop_prob = drop_prob)
# Model encoder layer:
mod_enc = qa.EncoderBlock(input_size=4*filters, kernel_size=5,
filters=filters, num_conv_layers=2, drop_prob=drop_prob)
self.mod_enc = nn.ModuleList([mod_enc]*7) # 7 number of blocks
# QANet Output layer
self.output = qa.QANetOutput(input_size=4*filters, drop_prob=drop_prob)
def __init__(self,
word_vectors,
char_vectors,
hidden_size,
num_heads,
char_embed_drop_prob,
drop_prob=0.1):
super(SketchyReader, self).__init__()
'''class QANet(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, device, drop_prob=0.):
super(QANet, self).__init__()
self.device = device'''
self.emb = layers.Embedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
char_embed_drop_prob=char_embed_drop_prob,
word_embed_drop_prob=drop_prob)
hidden_size *= 2 # update hidden size for other layers due to char embeddings
self.c_resizer = layers.Initialized_Conv1d(hidden_size, 128)
self.q_resizer = layers.Initialized_Conv1d(hidden_size, 128)
self.model_resizer = layers.Initialized_Conv1d(512, 128)
self.enc = layers.StackedEncoder(
num_conv_blocks=4,
kernel_size=7,
num_heads=num_heads,
dropout=drop_prob) # embedding encoder layer
self.att = layers.BiDAFAttention(
hidden_size=128,
drop_prob=drop_prob) # context-query attention layer
# self.mod1 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
# self.mod2 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
# self.mod3 = layers.StackedEncoder(num_conv_blocks=2,
# kernel_size=7,
# dropout=drop_prob) # model layer
self.model_encoder_layers = nn.ModuleList([
layers.StackedEncoder(num_conv_blocks=2,
kernel_size=7,
dropout=drop_prob) for _ in range(7)
])
self.out = layers.SketchyOutput(hidden_size=128) # output layer
def test2():
batch = 4
c_len = 10
q_len = 6
hidden_size = 8
c = torch.randn(batch, c_len, hidden_size * 2)
q = torch.randn(batch, q_len, hidden_size * 2)
BiDAFA = layers.BiDAFAttention(hidden_size * 2)
ret = BiDAFA.forward(c, q, None, None)
return ret
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_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)
def __init__(self,
word_vectors,
char_vectors,
hidden_size,
drop_prob=0.,
args=None):
super(BiDAF, self).__init__()
self.emb = layers.QAEmbedding(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.context_enc = layers.EncoderBlock(
conv_layers=args.num_conv_layers,
k=args.encoder_k,
hidden_dim=hidden_size,
dropout_prob=drop_prob,
attention_heads=args.attention_heads)
self.question_enc = layers.EncoderBlock(
conv_layers=args.num_conv_layers,
k=args.encoder_k,
hidden_dim=hidden_size,
dropout_prob=drop_prob,
attention_heads=args.attention_heads)
self.context_conv = DepthwiseSeparableConv2(500, hidden_size, 5)
self.question_conv = DepthwiseSeparableConv2(500, hidden_size, 5)
self.att = layers.BiDAFAttention(hidden_size=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.mod_enc = layers.EncoderBlock(
conv_layers=2,
k=5,
hidden_dim=hidden_size,
dropout_prob=drop_prob,
attention_heads=args.attention_heads)
self.cq_resizer = DepthwiseSeparableConv2(hidden_size * 4, hidden_size,
5)
self.out = layers.BiDAFOutput2(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, hidden_size, drop_prob=0.):
super(BiDAFBERTEmbeddings, self).__init__()
self.emb = None # TODO: BertEmbeddings.
self.enc = layers.RNNEncoder(input_size=None,
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 = layer.BiDAFOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def __init__(self, word_vectors, char_vectors, args):
super(QANetRevised, self).__init__()
self.hidden_size = args.hidden_size
self.emb = layers.QAEmbedding(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=self.hidden_size,
drop_prob=args.drop_prob)
self.att = layers.BiDAFAttention(hidden_size=self.hidden_size,
drop_prob=args.drop_prob)
self.context_encoder = layers.EncoderBlock(
conv_layers=args.num_conv_layers,
k=args.encoder_k,
hidden_dim=self.hidden_size,
dropout_prob=args.drop_prob,
attention_heads=args.attention_heads)
self.query_encoder = layers.EncoderBlock(
conv_layers=args.num_conv_layers,
k=args.encoder_k,
hidden_dim=self.hidden_size,
dropout_prob=args.drop_prob,
attention_heads=args.attention_heads)
self.encoder_block_1 = layers.EncoderBlock(
conv_layers=2,
k=5,
hidden_dim=self.hidden_size,
dropout_prob=args.drop_prob,
attention_heads=args.attention_heads)
self.encoder_block_2 = layers.EncoderBlock(
conv_layers=2,
k=5,
hidden_dim=self.hidden_size,
dropout_prob=args.drop_prob,
attention_heads=args.attention_heads)
self.encoder_block_3 = layers.EncoderBlock(
conv_layers=2,
k=5,
hidden_dim=self.hidden_size,
dropout_prob=args.drop_prob,
attention_heads=args.attention_heads)
self.stacked_encoder_blocks = nn.ModuleList(
[self.encoder_block_1, self.encoder_block_2, self.encoder_block_3])
self.output = layers.QANetOutput(hidden_size=self.hidden_size,
dropout_prob=args.drop_prob)
self.resize_attn = layers.DepthwiseSeparableConv(
self.hidden_size * 4, self.hidden_size, 5)
self.resize_context = DepthwiseSeparableConv2(500, self.hidden_size, 7)
self.resize_query = DepthwiseSeparableConv2(500, self.hidden_size, 7)
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, bert_base_model, hidden_size, drop_prob=0.):
super(BertBidaf, self).__init__()
BERT_BASE_HIDDEN_SIZE = 768
self.bert = bert_base_model
self.enc = layers.RNNEncoder(input_size=BERT_BASE_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=2 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = BertBidafOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
