def bulit_gru(self,input_var=None,mask_var=None,input_root=None,input_e1=None,input_e2=None):
"""
Bulit the GRU network
"""
#inputlayer
l_in=lasagne.layers.InputLayer(shape=self.input_shape,input_var=input_var,name="l_in")
#mask layer
l_mask=lasagne.layers.InputLayer(shape=self.mask_shape,input_var=mask_var,name="l_mask")
#inpute dropout
l_input_drop=lasagne.layers.DropoutLayer(l_in,p=self.keep_prob_input)
"""Input attention entity and root"""
l_in_root=lasagne.layers.InputLayer(shape=self.input_shape_att,input_var=input_root,name="l_in_root")
l_in_e1=lasagne.layers.InputLayer(shape=self.input_shape_att,input_var=input_e1,name="l_in_e1")
l_in_e2=lasagne.layers.InputLayer(shape=self.input_shape_att,input_var=input_e2,name="l_in_e2")
#Two GRU forward
l_gru_forward=lasagne.layers.GRULayer(\
l_input_drop,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,name="l_gru_forward1")
l_gru_forward=lasagne.layers.GRULayer(\
l_gru_forward,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,name="l_gru_forward2")
#Two GRU backward
l_gru_backward=lasagne.layers.GRULayer(\
l_input_drop,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,backwards=True,name="l_gru_backward1")
l_gru_backward=lasagne.layers.GRULayer(\
l_gru_backward,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,backwards=True,name="l_gru_backward2")
#Merge forward layers and backward layers
l_merge=lasagne.layers.ElemwiseSumLayer([l_gru_forward,l_gru_backward])
#Keyword-Attention
l_self_att=SelfAttEntRootLayer((l_in,l_in_e1,l_in_e2,l_in_root,l_merge),attention_size2=self.attention_size2)
## #when you need order in [Fe1,Fe2,Froot],you can use this line.
## l_kas=SelfAttEntRootLayer3((l_self_att,l_merge))
#output dropout
l_merge_drop=lasagne.layers.DropoutLayer(l_self_att,p=self.keep_prob_gru_output)
l_merge_fc = lasagne.layers.batch_norm(lasagne.layers.DenseLayer(
l_merge_drop,
num_units=250,
W=lasagne.init.GlorotUniform(gain=1.0), b=lasagne.init.Constant(1.),
nonlinearity=lasagne.nonlinearities.selu))
# #Keyword-attention(variant)
# l_self_att_var=SelfAttEntRootLayerVariant((l_in,l_in_e1,l_in_e2,l_in_root,l_merge),attention_size2=self.attention_size2)
#
# #output dropout
# l_merge_drop_var=lasagne.layers.DropoutLayer(l_self_att_var,p=self.keep_prob_gru_output)
#
# l_merge_fc_var = lasagne.layers.batch_norm(lasagne.layers.DenseLayer(
# l_merge_drop_var,
# num_units=250,
# W=lasagne.init.GlorotUniform(gain=1.0), b=lasagne.init.Constant(0.),
# nonlinearity=lasagne.nonlinearities.selu))
#
# #Merge keyword-attention
# l_key_merge=lasagne.layers.ElemwiseSumLayer([l_merge_fc,l_merge_fc_var])
#Margin loss
#init w_classes: it's format is (class_number,network_output[1])
w_classes_init=np.sqrt(6.0/(19+200))
l_out_margin=MarginLossLayer(l_merge_fc,w_classes=lasagne.init.Uniform(w_classes_init), class_number=19)
return l_out_margin,l_in,l_mask,l_out_margin,l_out_margin
def bulit_gru(self,
input_var=None,
mask_var=None,
input_root=None,
input_e1=None,
input_e2=None,
sen_length=None):
"""
Bulit the GRU network
"""
#inputlayer
l_in = lasagne.layers.InputLayer(shape=self.input_shape,
input_var=input_var,
name="l_in")
#mask layer
l_mask = lasagne.layers.InputLayer(shape=self.mask_shape,
input_var=mask_var,
name="l_mask")
#inpute dropout
l_input_drop = lasagne.layers.DropoutLayer(l_in,
p=self.keep_prob_input)
"""Input attention entity and root"""
l_in_root = lasagne.layers.InputLayer(shape=self.input_shape_att,
input_var=input_root,
name="l_in_root")
l_in_e1 = lasagne.layers.InputLayer(shape=self.input_shape_att,
input_var=input_e1,
name="l_in_e1")
l_in_e2 = lasagne.layers.InputLayer(shape=self.input_shape_att,
input_var=input_e2,
name="l_in_e2")
# #input attention
# l_att_alpha=InputAttEntRootLayer((l_in,l_in_e1,l_in_e2,l_in_root))
"""GRU"""
#Two GRU forward
l_gru_forward=lasagne.layers.GRULayer(\
l_input_drop,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,name="l_gru_forward1")
l_gru_forward=lasagne.layers.GRULayer(\
l_gru_forward,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,name="l_gru_forward2")
# #Attention
# l_att_forward=InputAttentionDotLayer((l_gru_forward,l_att_alpha))
#Two GRU backward
l_gru_backward=lasagne.layers.GRULayer(\
l_input_drop,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,backwards=True,name="l_gru_backward1")
l_gru_backward=lasagne.layers.GRULayer(\
l_gru_backward,num_units=self.gru_size,grad_clipping=self.grad_clip,
resetgate=self.gate_parameters,updategate=self.gate_parameters,
hidden_update=self.cell_parameters,learn_init=True,
only_return_final=False,backwards=True,name="l_gru_backward2")
# #Attention
# l_att_backward=InputAttentionDotLayer((l_gru_backward,l_att_alpha))
#Merge forward layers and backward layers
l_merge = lasagne.layers.ElemwiseSumLayer(
(l_gru_forward, l_gru_backward))
# l_merge=lasagne.layers.ConcatLayer((l_gru_forward,l_gru_backward),axis=2)
#Self-Attention
# l_self_att=SelfAttentionDotLayer((l_merge,l_att_alpha))
l_self_att = SelfAttEntRootLayer(
(l_in, l_in_e1, l_in_e2, l_in_root, l_merge),
attention_size2=self.attention_size2)
#output dropout
l_merge_drop = lasagne.layers.DropoutLayer(l_self_att,
p=self.keep_prob_gru_output)
l_merge_fc = lasagne.layers.DenseLayer(
l_merge_drop,
num_units=50,
W=lasagne.init.GlorotUniform(gain=1.0),
b=lasagne.init.Constant(0.),
nonlinearity=lasagne.nonlinearities.selu)
# #Margin loss
# #init w_classes: it's format is (class_number,network_output[1])
# w_classes_init=np.sqrt(6.0/(19+self.gru_size))
# l_out_margin=MarginLossLayer(l_merge_drop,w_classes=lasagne.init.Uniform(w_classes_init), class_number=19)
# Finally, we'll add the fully-connected output layer, of 10 softmax units:
l_out_margin = lasagne.layers.DenseLayer(\
l_merge_fc, num_units=self.num_classes,\
nonlinearity=lasagne.nonlinearities.softmax)
return l_out_margin, l_in, l_mask, l_self_att, l_gru_forward