def __init__(self, n_vocab, n_char, n_tag, args):
feature_dim = args[
'word_embedding_dim'] + 2 * args['char_embedding_dim']
super(ModelBase,
self).__init__(char_embed=L.EmbedID(n_char,
args['char_embedding_dim'],
ignore_label=-1),
bi_char=L.NStepBiLSTM(1,
args['char_embedding_dim'],
args['char_embedding_dim'],
0),
word_embed=L.EmbedID(n_vocab,
args['word_embedding_dim'],
ignore_label=-1),
bi_word=L.NStepBiLSTM(1, feature_dim,
int(feature_dim / 2), 0),
l=L.Linear(feature_dim, n_tag),
crf=L.CRF1d(n_tag))
# Initialize value for hyper parameters
self.char_embedding_dim = args['char_embedding_dim']
self.tag_embedding_dim = args['tag_embedding_dim']
self.dropout_ratio = args['dropout_ratio']
self.lr_param = args['lr_param']
self.threshold = args['threshold']
self.decay_rate = args['decay_rate']
self.batch_size = args['batch_size']
if args['mode'] == 'train':
for w in self.bi_char:
w.b1.data[:] = 1.0
w.b5.data[:] = 1.0
for w in self.bi_word:
w.b1.data[:] = 1.0
w.b5.data[:] = 1.0
def __init__(
self,
n_speakers=4,
dropout=0.25,
in_size=513,
hidden_size=256,
n_layers=1,
embedding_layers=1,
embedding_size=20,
dc_loss_ratio=0.5,
):
""" BLSTM-based diarization model.
Args:
n_speakers (int): Number of speakers in recording
dropout (float): dropout ratio
in_size (int): Dimension of input feature vector
hidden_size (int): Number of hidden units in LSTM
n_layers (int): Number of LSTM layers after embedding
embedding_layers (int): Number of LSTM layers for embedding
embedding_size (int): Dimension of embedding vector
dc_loss_ratio (float): mixing parameter for DPCL loss
"""
super(BLSTMDiarization, self).__init__()
with self.init_scope():
self.bi_lstm1 = L.NStepBiLSTM(n_layers, hidden_size * 2,
hidden_size, dropout)
self.bi_lstm_emb = L.NStepBiLSTM(embedding_layers, in_size,
hidden_size, dropout)
self.linear1 = L.Linear(hidden_size * 2, n_speakers)
self.linear2 = L.Linear(hidden_size * 2, embedding_size)
self.dc_loss_ratio = dc_loss_ratio
self.n_speakers = n_speakers
def __init__(self, n_cell, size_hidden, rate_dropout):
super(ONT_BiLSTM, self).__init__()
self.rate_dropout = rate_dropout
with self.init_scope():
self.rnn_a = L.NStepBiLSTM(n_cell, 300, size_hidden, rate_dropout)
self.rnn_b = L.NStepBiLSTM(n_cell, 300, size_hidden, rate_dropout)
self.l1 = L.Linear(size_hidden * 4, size_hidden * 4)
self.l2 = L.Linear(size_hidden * 4, 4)
def __init__(self, hidden_dim):
super(InferenceCompositionLayer, self).__init__()
with self.init_scope():
self.hypo_encoder = L.NStepBiLSTM(1,
hidden_dim,
int(hidden_dim / 8),
dropout=0.3)
self.premise_encoder = L.NStepBiLSTM(1,
hidden_dim,
int(hidden_dim / 8),
dropout=0.3)
def __init__(self, n_vocab, embed_dim, hidden_dim):
super(InputEncodingLayer, self).__init__()
with self.init_scope():
self.embed_mat = L.EmbedID(n_vocab, embed_dim)
self.hypo_encoder = L.NStepBiLSTM(1,
hidden_dim,
int(hidden_dim / 2),
dropout=0.3)
self.premise_encoder = L.NStepBiLSTM(1,
hidden_dim,
int(hidden_dim / 2),
dropout=0.3)
def __init__(self,
n_layers=1,
n_inputs=3,
n_outputs=3,
linear_num=786,
cell_num=128,
atten1=64,
atten2=1,
hidden1=64,
dropout_rate=0.1):
self.dropout_rate = dropout_rate
super(SER, self).__init__()
with self.init_scope():
self.c1 = L.Convolution2D(n_inputs,
128,
ksize=(5, 3),
stride=(1, 1),
pad=(2, 1))
self.c2 = L.Convolution2D(128,
512,
ksize=(5, 3),
stride=(1, 1),
pad=(2, 1))
self.b1 = L.BatchNormalization(128)
self.b2 = L.BatchNormalization(512)
self.cl = L.Linear(None, linear_num)
self.lstm = L.NStepBiLSTM(n_layers, linear_num, cell_num,
dropout_rate)
self.a1 = L.Linear(cell_num * 2, atten1)
self.a2 = L.Linear(atten1, atten2)
self.l3 = L.Linear(cell_num * 2, hidden1)
self.b3 = L.BatchNormalization(hidden1)
self.l4 = L.Linear(hidden1, n_outputs)
self.af = F.leaky_relu
def __init__(self,
n_layers,
insize,
outsize,
initialW=None,
use_bi_lstm=False):
super(EdgeRNN, self).__init__()
self.n_layer = n_layers
self.outsize = outsize
if use_bi_lstm:
assert outsize % 2 == 0, outsize
if not initialW:
initialW = initializers.HeNormal()
with self.init_scope():
self.fc1 = L.Linear(insize, 256, initialW=initialW)
self.fc2 = L.Linear(256, 256, initialW=initialW)
if use_bi_lstm:
self.lstm3 = L.NStepBiLSTM(self.n_layer,
256,
outsize // 2,
initialW=initialW,
dropout=0.1) #dropout = 0.0
else:
self.lstm3 = L.NStepLSTM(self.n_layer,
256,
outsize,
initialW=initialW,
dropout=0.1)
def __init__(self,
n_layers,
insize,
outsize,
initialW=None,
use_bi_lstm=False):
super(ConnLabelRNN, self).__init__()
if not initialW:
initialW = initializers.HeNormal()
self.n_layer = n_layers
with self.init_scope():
if use_bi_lstm:
self.lstm1 = L.NStepBiLSTM(self.n_layer,
insize,
512,
initialW=initialW,
dropout=0.1)
else:
self.lstm1 = L.NStepLSTM(self.n_layer,
insize,
1024,
initialW=initialW,
dropout=0.1)
self.fc2 = L.Linear(1024, 512)
self.fc3 = L.Linear(512, outsize)
def __init__(self,
n_layers,
insize,
outsize,
initialW=None,
use_bi_lstm=False):
super(NodeRNN, self).__init__()
if not initialW:
initialW = initializers.HeNormal()
self.n_layer = n_layers
with self.init_scope():
if use_bi_lstm:
self.lstm = L.NStepBiLSTM(self.n_layer,
1024,
outsize // 2,
initialW=initialW,
dropout=0.1) #dropout = 0.0
else:
self.lstm = L.NStepLSTM(self.n_layer,
1024,
outsize,
initialW=initialW,
dropout=0.1)
self.fc1 = L.Linear(insize, 1024, initialW=initialW)
self.fc2 = L.Linear(1024, 1024, initialW=initialW)
def __init__(self, n_word_types, n_trig_types, n_role_types,
n_entity_types, params):
super(SearchBasedModel, self).__init__()
with self.init_scope():
self.dim_embed = params['dim_embed']
self.dim_bilstm = params['dim_bilstm']
self.dim_role_type = params['dim_role_type']
self.dim_arg_type = params['dim_arg_type']
self.dropout = params['dropout']
self.threshold = params['threshold']
self.n_best = params['n_best']
self.margin = params['margin']
self.action_dim = params['dim_action']
self.max_pos = params['max_pos']
self.dim_pos = params['dim_pos']
self.arg_type_and_word_dim = self.dim_arg_type + (self.dim_bilstm *
2) + self.dim_pos
self.relation_dim = self.arg_type_and_word_dim + self.dim_role_type + self.arg_type_and_word_dim
self.hidden_dim = int(self.arg_type_and_word_dim / 2)
embed_init = chainer.initializers.Normal()
self.embed_positiontype = L.EmbedID(self.max_pos,
self.dim_pos,
initialW=embed_init,
ignore_label=-1)
self.embed_wordtype = L.EmbedID(n_word_types,
self.dim_embed,
initialW=embed_init,
ignore_label=-1)
self.embed_trigtype = L.EmbedID(n_trig_types,
self.dim_arg_type,
initialW=embed_init,
ignore_label=-1)
self.embed_roletype = L.EmbedID(n_role_types,
self.dim_role_type,
initialW=embed_init,
ignore_label=-1)
self.embed_enttype = L.EmbedID(n_entity_types,
self.dim_arg_type,
initialW=embed_init,
ignore_label=-1)
self.embed_action = L.EmbedID(self.action_dim,
self.action_dim,
initialW=embed_init,
ignore_label=-1)
self.bilstm = L.NStepBiLSTM(1, self.dim_embed, self.dim_bilstm, 0)
self.linear_structure = L.Linear(
None, self.relation_dim + self.action_dim)
self.linear_buffer = L.Linear(None,
self.relation_dim + self.action_dim)
self.state_representation = L.Linear(None,
self.arg_type_and_word_dim)
self.linear1 = L.Linear(None, self.hidden_dim)
self.linear2 = L.Linear(None, self.hidden_dim)
self.linear = L.Linear(None, 1)
def __init__(self,
vocab_size: int,
word_size: int,
hidden_size: int,
dropout: float = 0.5,
**kwargs):
"""Initialize the Relevant Sentences Detector model.
Parameters
----------
vocab_size : int
The number of items in the vocabulary of the input.
word_size : int
The number of dimensions used to represent the vocabulary.
hidden_size : int
The number of dimensions used to represent the hidden state of the LSTM.
dropout : float, optional
The dropout used for the LSTM (default: 0.5).
kwargs
Optional keyword arguments passed to the super class.
"""
super().__init__(**kwargs)
with self.init_scope():
self.embed = links.EmbedID(vocab_size, word_size)
self.rnn = links.NStepBiLSTM(1, word_size, hidden_size, dropout)
self.l_out = links.Linear(2)
def __init__(self, inputDim, outDimPdf, outDimPho):
super(LSTM, self).__init__()
with self.init_scope():
global args
self.nl = L.NStepBiLSTM(args.layer, inputDim, args.hiddenNode,
args.dropout)
self.bn = L.BatchNormalization(args.hiddenNode * 2, 0.95)
#for param in self.params():
# param.array[...] = np.random.uniform(-0.1, 0.1, param.shape)
initializerW = None #chainer.initializers.HeNormal()
initializerBias = None #chainer.initializers.Zero()
self.ln1 = L.Linear(None,
outDimPdf,
initialW=initializerW,
initial_bias=initializerBias)
self.ln2 = L.Linear(None,
outDimPho,
initialW=initializerW,
initial_bias=initializerBias)
def __init__(self, database, n_layers:int, in_size:int, out_size:int, initialW=None,
spatial_edge_model: SpatialEdgeMode = SpatialEdgeMode.all_edge,
recurrent_block_type: RecurrentType = RecurrentType.rnn, attn_heads=8, bi_lstm=False):
super(SpaceTimeRNN, self).__init__()
self.neg_pos_ratio = 3
self.database = database
self.spatial_edge_mode = spatial_edge_model
self.out_size = out_size
self.in_size = in_size
self.frame_node_num = config.BOX_NUM[self.database]
self.mid_size = 1024
NodeRecurrentModule = AttentionBlock if recurrent_block_type == RecurrentType.attention_block else TemporalRNN
if recurrent_block_type == RecurrentType.no_temporal:
NodeRecurrentModule = PositionwiseFeedForwardLayer
with self.init_scope():
if not initialW:
initialW = initializers.HeNormal()
self.top = dict()
for i in range(self.frame_node_num):
if recurrent_block_type == RecurrentType.rnn:
self.add_link("Node_{}".format(i),
TemporalRNN(n_layers, self.in_size, self.mid_size, use_bi_lstm=bi_lstm))
else:
self.add_link("Node_{}".format(i),
NodeRecurrentModule(n_layers, self.in_size, self.mid_size))
self.top[str(i)] = getattr(self, "Node_{}".format(i))
fc_in_len = self.mid_size
if spatial_edge_model != SpatialEdgeMode.no_edge:
self.space_lstm = L.NStepBiLSTM(n_layers, self.in_size, self.mid_size//2, dropout=0.1, initialW=initialW)
fc_in_len = self.mid_size * 2
self.fc = L.Linear(fc_in_len, self.out_size, initialW=initialW)
def __init__(self,
idim,
elayers,
cdim,
hdim,
subsample,
dropout,
typ="lstm"):
super(BRNNP, self).__init__()
with self.init_scope():
for i in six.moves.range(elayers):
if i == 0:
inputdim = idim
else:
inputdim = hdim
setattr(
self, "birnn%d" % i,
L.NStepBiLSTM(1, inputdim, cdim, dropout) if typ == "lstm"
else L.NStepBiGRU(1, inputdim, cdim, dropout))
# bottleneck layer to merge
setattr(self, "bt%d" % i, L.Linear(2 * cdim, hdim))
self.elayers = elayers
self.cdim = cdim
self.subsample = subsample
self.typ = typ
def setUp(self):
shape = (self.n_layer * 2, len(self.lengths), self.out_size)
if self.hidden_none:
self.h = self.c = numpy.zeros(shape, 'f')
else:
self.h = numpy.random.uniform(-1, 1, shape).astype('f')
self.c = numpy.random.uniform(-1, 1, shape).astype('f')
self.xs = [
numpy.random.uniform(-1, 1, (l, self.in_size)).astype('f')
for l in self.lengths
]
self.gh = numpy.random.uniform(-1, 1, shape).astype('f')
self.gc = numpy.random.uniform(-1, 1, shape).astype('f')
self.gys = [
numpy.random.uniform(-1, 1, (l, self.out_size * 2)).astype('f')
for l in self.lengths
]
self.rnn = links.NStepBiLSTM(self.n_layer, self.in_size, self.out_size,
self.dropout)
for layer in self.rnn:
for p in layer.params():
p.array[...] = numpy.random.uniform(-1, 1, p.shape)
self.rnn.cleargrads()
def __init__(self, n_vocab, embed, hidden, dropout, initialW):
n_layers = 1
super(WordEncoder, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab, embed, initialW=initialW)
self.Nlstm = L.NStepBiLSTM(n_layers, embed, hidden, dropout)
self.hidden = hidden
def __init__(self,
n_layers=consts.LAYERS,
n_in=consts.IN * 2,
n_units=consts.UNITS * 2,
n_tag=9,
dropout=consts.DROPOUT):
super(RNN_TOP, self).__init__()
# パラメータを持つ層の登録
with self.init_scope():
self.xh = L.NStepBiLSTM(n_layers, n_in, n_units, dropout)
self.h_100_self = L.Linear(n_units * 2, 100 * 2)
self.h_1_self = L.Linear(100 * 2, 1)
self.h_100_qyn = L.Linear(n_units * 2, 100 * 2)
self.h_1_qyn = L.Linear(100 * 2, 1)
self.h_100_qw = L.Linear(n_units * 2, 100 * 2)
self.h_1_qw = L.Linear(100 * 2, 1)
self.h_100_ayn = L.Linear(n_units * 2, 100 * 2)
self.h_1_ayn = L.Linear(100 * 2, 1)
self.h_100_aw = L.Linear(n_units * 2, 100 * 2)
self.h_1_aw = L.Linear(100 * 2, 1)
self.h_100_res = L.Linear(n_units * 2, 100 * 2)
self.h_1_res = L.Linear(100 * 2, 1)
self.h_100_fil = L.Linear(n_units * 2, 100 * 2)
self.h_1_fil = L.Linear(100 * 2, 1)
self.h_100_con = L.Linear(n_units * 2, 100 * 2)
self.h_1_con = L.Linear(100 * 2, 1)
self.h_100_req = L.Linear(n_units * 2, 100 * 2)
self.h_1_req = L.Linear(100 * 2, 1)
def __init__(self, num_classes):
super(Model, self).__init__()
with self.init_scope():
self.conv1 = L.ConvolutionND(ndim=1,
in_channels=3,
out_channels=48,
ksize=5,
stride=1,
pad=2)
self.batch2 = L.BatchNormalization(48)
self.conv2 = L.ConvolutionND(ndim=1,
in_channels=48,
out_channels=64,
ksize=5,
stride=1,
pad=2)
self.batch3 = L.BatchNormalization(64)
self.conv3 = L.ConvolutionND(ndim=1,
in_channels=64,
out_channels=96,
ksize=3,
stride=1,
pad=1)
self.lstm = L.NStepBiLSTM(n_layers=3,
in_size=96,
out_size=128,
dropout=0.3)
self.fc = L.Linear(in_size=256, out_size=num_classes)
def __init__(self,
n_vocab=None,
emb_dim=100,
hidden_dim=200,
init_emb=None,
add_dim=0,
use_dropout=0.33,
n_layers=1,
pos_dim=0,
n_pos=0):
feature_dim = emb_dim + add_dim + pos_dim
super(FastBiLSTM,
self).__init__(word_embed=L.EmbedID(n_vocab,
emb_dim,
ignore_label=-1),
bi_lstm=L.NStepBiLSTM(n_layers=n_layers,
in_size=feature_dim,
out_size=hidden_dim,
dropout=use_dropout,
use_cudnn=True))
if n_pos:
pos_embed = L.EmbedID(n_pos, pos_dim, ignore_label=-1)
self.add_link('pos_embed', pos_embed)
self.n_pos = n_pos
self.hidden_dim = hidden_dim
self.train = True
self.use_dropout = use_dropout
self.n_layers = n_layers
# Forget gate bias => 1.0
# MEMO: Values 1 and 5 reference the forget gate.
for w in self.bi_lstm:
w.b1.data[:] = 1.0
w.b5.data[:] = 1.0
def __init__(self,
n_layers=consts.LAYERS,
n_in=consts.IN * 2,
n_units=consts.UNITS * 2,
dropout=consts.DROPOUT):
super(RNN_CONNECT_AT, self).__init__()
with self.init_scope():
self.xh = L.NStepBiLSTM(n_layers, n_in, n_units, dropout)
self.hy1 = L.Linear(n_units * 2, 100 * 2)
self.h_100_self = L.Linear(n_units * 2, 100 * 2)
self.h_100_qyn = L.Linear(n_units * 2, 100 * 2)
self.h_100_qw = L.Linear(n_units * 2, 100 * 2)
self.h_100_ayn = L.Linear(n_units * 2, 100 * 2)
self.h_100_aw = L.Linear(n_units * 2, 100 * 2)
self.h_100_res = L.Linear(n_units * 2, 100 * 2)
self.h_100_fil = L.Linear(n_units * 2, 100 * 2)
self.h_100_con = L.Linear(n_units * 2, 100 * 2)
self.h_100_req = L.Linear(n_units * 2, 100 * 2)
self.at = L.Linear(100 * 2, 1)
# self.at1 = L.Linear(100*2, 100) # 200->100->1じゃなくて200->1でもいいかも
# self.at2 = L.Linear(100, 1)
self.out = L.Linear(100 * 2, 9) # 200->9じゃなくて200->100->9でもいいかも
def __init__(self, n_layers, n_source_vocab, trans_data, n_units,
v_eos_src, n_maxsize):
super(Seq2Tree_Flatten, self).__init__()
# for each nodetype, for each move, the result array.
self.trans_data = trans_data
self.embed_idx = []
ns = 0
def inc():
nonlocal ns
ns += 1
return ns - 1
self.embed_idx = [[[inc() for v in vs] for vs in moves]
for moves in self.trans_data]
self.embed_root_idx = ns
self.embed_y_size = ns + 1
self.choicerange = []
self.choice_idx = []
self.is_trivial = []
s = 0
for d in self.trans_data:
ist = len(d) <= 1
self.is_trivial.append(ist)
#if ist:
# self.choicerange.append(None)
# self.choice_idx.append([0])
# continue
b = s
s += len(d)
self.choicerange.append((b, s))
self.choice_idx.append(list(range(b, s)))
#self.choice_num_sum = sum(list(map(lambda d: len(d),self.trans_data)))
self.type_size = len(self.embed_idx)
self.n_all_choice = sum(map(lambda x: len(x), self.trans_data))
with self.init_scope():
self.embed_x = L.EmbedID(n_source_vocab, n_units)
#self.embed_y = L.EmbedID(self.embed_y_size, n_units) # maybe mergable
self.embed_y_0 = L.EmbedID(self.embed_y_size,
n_units) # maybe mergable
self.embed_y_1 = L.EmbedID(self.type_size,
n_units) # maybe mergable
self.encoder = L.NStepBiLSTM(n_layers, n_units, n_units, 0.1)
self.decoder = L.NStepLSTM(n_layers, n_units * 2, n_units * 2, 0.1)
self.Wc = L.Linear(n_units * 4, n_units)
self.Ws = L.Linear(n_units, self.n_all_choice)
#self.att = Attention(n_units)
self.att = GlobalGeneralAttention(n_units)
self.n_layers = n_layers
self.n_units = n_units
self.v_eos_src = v_eos_src
self.n_maxsize = n_maxsize
self.rootidx = len(trans_data) - 1
def __init__(self, ch):
super(Link_NStepBiLSTM, self).__init__(L.NStepBiLSTM(1, 1, 1, 0))
# code.InteractiveConsole({'ch': ch}).interact()
hd = ch.children().__next__()
if not (hd.w0 is None):
self.n_in = hd.w0.shape[1]
else:
self.n_in = None
self.out_size = ch.out_size
self.n_layers = ch.n_layers
self.dropout = ch.dropout
self.ws = []
self.bs = []
for i in range(self.n_layers * 2):
ws = []
bs = []
for j in range(8):
ws.append(
helper.make_tensor_value_info(('/%d/w%d' % (i, j)),
TensorProto.FLOAT, ["TODO"]))
bs.append(
helper.make_tensor_value_info(('/%d/b%d' % (i, j)),
TensorProto.FLOAT, ["TODO"]))
self.ws.append(ws)
self.bs.append(bs)
def __init__(self, n_layers=consts.LAYERS, n_in=consts.IN*2, n_units=consts.UNITS*2, n_tag=9, dropout=consts.DROPOUT):
super(RNN_FINETUNING, self).__init__()
# パラメータを持つ層の登録
with self.init_scope():
self.xh = L.NStepBiLSTM(n_layers, n_in, n_units, dropout)
self.hy1 = L.Linear(n_units*2, 100*2)
self.hy2 = L.Linear(100*2, n_tag)
def __init__(self, n_vocab=30000, n_units=200, n_layers=2, dropout=0.5):
super(SentRepRNN, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
n_units) # word embedding（入力は単一のリスト)
self.encoder = L.NStepBiLSTM(
n_layers, n_units, n_units,
dropout) # (embed時と異なり、)ラベル内の各文ごとに文ベクトルを作成
def __init__(self, vocab_size, embed_size, hidden_size, dropout):
n_layers = 1
super(Encoder, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(vocab_size, embed_size)
self.Nlstm = L.NStepBiLSTM(n_layers, embed_size, hidden_size,
dropout)
self.hidden_size = hidden_size
def __init__(self, input_size, hidden_size, num_layers, num_classes):
super(BiRNN, self).__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
with self.init_scope():
self.nsteplstm = L.NStepBiLSTM(num_layers, input_size, hidden_size,
dropout=0) # chainer.Links.NStepLSTM and torch.nn.LSTM are alike.
self.fc = L.Linear(2 * hidden_size, num_classes)
def __init__(self, vocab_size):
super(LSTMChain, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(vocab_size, 200)
self.bi_lstm = L.NStepBiLSTM(1, 200, 300, 0.2)
self.mid = L.Linear(600, 200)
self.out = L.Linear(200, 11)
self.dropout = 0.2
def __init__(self, n_word_types, n_trig_types, n_role_types,
n_entity_types, trigger_type2id, entity_type2id, DIM_EMBED,
DIM_EVENT, DIM_BILSTM, DIM_TRIG_TYPE, DIM_ROLE_TYPE,
DIM_ARG_TYPE, DIM_IO, DROPOUT, REPLACE_TYPE, GENERALISATION,
THRESHOLD):
super(Loader, self).__init__()
with self.init_scope():
self.DIM_EMBED = DIM_EMBED
self.DIM_EVENT = DIM_EVENT
self.DIM_BILSTM = DIM_BILSTM
self.DIM_TRIG_TYPE = DIM_TRIG_TYPE
self.DIM_ROLE_TYPE = DIM_ROLE_TYPE
self.DIM_ARG_TYPE = DIM_ARG_TYPE
self.DIM_IO = DIM_IO
self.DROPOUT = DROPOUT
self.THRESHOLD = THRESHOLD
self.GENERALISATION = GENERALISATION
self.REPLACE_TYPE = REPLACE_TYPE
self.DROPOUT = DROPOUT
self.DIM_TREE_LSTM_INPUT = self.DIM_TRIG_TYPE + (
self.DIM_BILSTM *
2) + self.DIM_ROLE_TYPE + self.DIM_TRIG_TYPE + (
self.DIM_BILSTM * 2) + self.DIM_IO
self.DIM_ARG = self.DIM_TRIG_TYPE + (self.DIM_BILSTM * 2)
self.id2triggertype = {v: k for k, v in trigger_type2id.items()}
self.id2entitytype = {v: k for k, v in entity_type2id.items()}
self.embed = L.EmbedID(n_word_types,
self.DIM_EMBED,
ignore_label=-1)
self.bilstm = L.NStepBiLSTM(1, self.DIM_EMBED, self.DIM_BILSTM, 0)
self.embed_trigtype = L.EmbedID(n_trig_types,
self.DIM_TRIG_TYPE,
ignore_label=-1)
self.embed_roletype = L.EmbedID(n_role_types,
self.DIM_ROLE_TYPE,
ignore_label=-1)
self.embed_argtype = L.EmbedID(n_entity_types,
self.DIM_ARG_TYPE,
ignore_label=-1)
self.embed_io = L.EmbedID(2, self.DIM_IO, ignore_label=-1)
self.treelstm = L.ChildSumTreeLSTM(self.DIM_TREE_LSTM_INPUT,
self.DIM_EVENT)
self.l1 = L.Linear(None, self.DIM_EVENT)
self.y = L.Linear(None, self.DIM_EVENT)
self.final = L.Linear(None, 1) # event or non-event
self.reducedEvent = L.Linear(None, self.DIM_ARG)
self.len_type_and_arg = self.DIM_TRIG_TYPE + (self.DIM_BILSTM * 2)
self.len_relation = self.DIM_TRIG_TYPE + (self.DIM_BILSTM * 2) + self.DIM_ROLE_TYPE + self.DIM_ARG_TYPE + \
(self.DIM_BILSTM * 2) + self.DIM_IO
self.trigger_type2id = trigger_type2id
self.entity_type2id = entity_type2id
def __init__(self, inSize, hidSize, outSize):
# out size is usually class size
super().__init__()
bilstm = L.NStepBiLSTM(1, inSize, hidSize, dropout=0.3)
linear = L.Linear(hidSize * 2, outSize)
self.add_link('bilstm', bilstm)
self.add_link('linear', linear)
def __init__(self, train=True):
self.train = train
super(Mynet, self).__init__()
with self.init_scope():
self.h = L.NStepBiLSTM(n_layers=1,
in_size=num_classes,
out_size=128,
dropout=0)
self.out = L.Linear(None, num_classes)