def ops(self):
"""
operation
"""
gru = DynamicGRU(size=self.gru_dim,
param_attr=attr.ParamAttr(name="%s.w" % self.name),
bias_attr=attr.ParamAttr(name="%s.b" % self.name))
return gru
def ops(self):
"""
operation
"""
fc = FC(size=self.fc_dim,
param_attr=attr.ParamAttr(name="%s.w" % self.name),
bias_attr=attr.ParamAttr(name="%s.b" % self.name),
act=self.act)
return fc
def ops(self, input):
"""
operation
"""
proj = fluid.layers.fc(input=input, size=self.gru_dim * 3)
gru = fluid.layers.dynamic_gru(
input=proj,
size=self.gru_dim,
param_attr=attr.ParamAttr(name="%s.w" % self.name),
bias_attr=attr.ParamAttr(name="%s.b" % self.name))
return gru
def ops(self, input):
"""
operation
"""
proj = fluid.layers.fc(input=input, size=self.lstm_dim * 4)
lstm, _ = fluid.layers.dynamic_lstm(
input=proj,
size=self.lstm_dim * 4,
param_attr=attr.ParamAttr(name="%s.w" % self.name),
bias_attr=attr.ParamAttr(name="%s.b" % self.name))
return lstm
def ops(self, input):
"""
operation
"""
fc = fluid.layers.fc(input=input,
size=self.fc_dim,
param_attr=attr.ParamAttr(name="%s.w" %
self.name),
bias_attr=attr.ParamAttr(name="%s.b" % self.name),
act=self.act,
name=self.name)
return fc
def lstm(left_emb, right_emb, hid_dim=128):
"""
LSTM network.
"""
left_proj = fluid.layers.fc(
input=left_emb,
size=hid_dim * 4,
param_attr=attr.ParamAttr(name="matching_lstm_fc.w"),
bias_attr=attr.ParamAttr(name="matching_lstm_fc.b"))
right_proj = fluid.layers.fc(
input=right_emb,
size=hid_dim * 4,
param_attr=attr.ParamAttr(name="matching_lstm_fc.w"),
bias_attr=attr.ParamAttr(name="matching_lstm_fc.b"))
left_lstm, cell = fluid.layers.dynamic_lstm(
input=left_proj,
size=hid_dim * 4,
param_attr=attr.ParamAttr(name="matching_lstm.w"),
bias_attr=attr.ParamAttr(name="matching_lstm.b"),
is_reverse=False)
right_lstm, _ = fluid.layers.dynamic_lstm(
input=right_proj,
size=hid_dim * 4,
param_attr=attr.ParamAttr(name="matching_lstm.w"),
bias_attr=attr.ParamAttr(name="matching_lstm.b"),
is_reverse=False)
left_last = fluid.layers.sequence_last_step(left_lstm)
right_last = fluid.layers.sequence_last_step(right_lstm)
return left_last, right_last
def gru(left_emb, right_emb, hid_dim=128):
"""
GRU network.
"""
left_proj = fluid.layers.fc(
input=left_emb,
size=hid_dim * 3,
param_attr=attr.ParamAttr(name="matching_gru_fc.w"),
bias_attr=attr.ParamAttr(name="matching_gru_fc.b"))
right_proj = fluid.layers.fc(
input=right_emb,
size=hid_dim * 3,
param_attr=attr.ParamAttr(name="matching_gru_fc.w"),
bias_attr=attr.ParamAttr(name="matching_gru_fc.b"))
left_gru = fluid.layers.dynamic_gru(
input=left_proj,
size=hid_dim,
param_attr=attr.ParamAttr(name="matching_gru.w"),
bias_attr=attr.ParamAttr(name="matching_gru.b"))
right_gru = fluid.layers.dynamic_gru(
input=right_proj,
size=hid_dim,
param_attr=attr.ParamAttr(name="matching_gru.w"),
bias_attr=attr.ParamAttr(name="matching_gru.b"))
left_last = fluid.layers.sequence_last_step(left_gru)
right_last = fluid.layers.sequence_last_step(right_gru)
return left_last, right_last
def ops(self, input):
"""
:param input:
"""
if self.init_w2v:
#weight_data = EmbeddingLayer.v2np(self)
w_param_attrs = fluid.ParamAttr(
name=self.name,
learning_rate=1,
initializer=fluid.initializer.NumpyArrayInitializer(
self.weight_data),
trainable=True)
emb = fluid.layers.embedding(input=input,
size=[self.dict_dim, self.emb_dim],
is_sparse=False,
param_attr=w_param_attrs,
dtype='float32')
else:
emb = fluid.layers.embedding(
input=input,
size=[self.dict_dim, self.emb_dim],
is_sparse=False,
param_attr=attr.ParamAttr(name=self.name),
dtype='float32')
return emb
def ops(self, input):
"""
operation
"""
emb = fluid.layers.embedding(input=input,
size=[self.dict_size, self.emb_dim],
is_sparse=True,
param_attr=attr.ParamAttr(name=self.name))
return emb
def cnn(left_emb, right_emb, hid_dim=128, filter_size=3, num_filters=256):
"""
CNN network.
"""
# cnn layer
left_cnn = fluid.nets.sequence_conv_pool(
input=left_emb,
num_filters=num_filters,
filter_size=filter_size,
param_attr=attr.ParamAttr(name="matching_cnn"),
act="relu",
pool_type="max")
right_cnn = fluid.nets.sequence_conv_pool(
input=right_emb,
num_filters=num_filters,
filter_size=filter_size,
param_attr=attr.ParamAttr(name="matching_cnn"),
act="relu",
pool_type="max")
return left_cnn, right_cnn
def ops(self, input):
"""
operation
"""
conv = fluid.nets.sequence_conv_pool(
input=input,
filter_size=self.filter_size,
num_filters=self.num_filters,
param_attr=attr.ParamAttr(name=self.name),
act="relu")
return conv
def ops(self):
"""
operation
"""
emb = Embedding(size=[self.dict_size, self.emb_dim],
is_sparse=True,
padding_idx=self.padding_idx,
param_attr=attr.ParamAttr(
name=self.name,
initializer=fluid.initializer.Xavier()))
return emb
def ops(self):
"""
operation
"""
# TODO(huihuangzheng): The original code set the is_sparse=True, but it
# causes crush in dy2stat. Set it to True after fixing it.
emb = Embedding(size=[self.dict_size, self.emb_dim],
is_sparse=True,
padding_idx=self.padding_idx,
param_attr=attr.ParamAttr(
name=self.name,
initializer=fluid.initializer.Xavier()))
return emb
