def create_citation2(in_dim, em_dim):
# Building 'penNet'
input_layer = tflearn.input_data(shape=[
None, in_dim + G_NUM + SUBC_NUM + C_NUM + SUBS_NUM + Citation_Dim
],
name='input')
input_p = input_layer[:, 0:in_dim]
input_ss = input_layer[:, in_dim:in_dim + SUBS_NUM]
input_c = input_layer[:, in_dim + SUBS_NUM:in_dim + SUBS_NUM + C_NUM]
input_sc = input_layer[:, in_dim + SUBS_NUM + C_NUM:in_dim + SUBS_NUM +
C_NUM + SUBC_NUM]
input_g = input_layer[:, in_dim + SUBS_NUM + C_NUM + SUBC_NUM:in_dim +
SUBS_NUM + C_NUM + SUBC_NUM + G_NUM]
input_citation = input_layer[:,
in_dim + SUBS_NUM + C_NUM + SUBC_NUM + G_NUM:]
group_embedding = tf.Variable(tf.random_normal([G_NUM, em_dim]),
name='group_embedding')
subclass_embedding = tf.Variable(tf.random_normal([SUBC_NUM, em_dim]),
name='subclass_embedding')
class_embedding = tf.Variable(tf.random_normal([C_NUM, em_dim]),
name='class_embedding')
subsection_embedding = tf.Variable(tf.random_normal([SUBS_NUM, em_dim]),
name='subsection_embedding')
section_embedding = tf.Variable(tf.random_normal([SEC_NUM, em_dim]),
name='section_embedding')
# SG_output = _cat_weighted(X, subgroup_embedding, W_SG)
textual_inf = tflearn.embedding(input_p,
input_dim=110240,
output_dim=128,
name='word_embedding')
textual_embedding = tflearn.lstm(textual_inf,
em_dim,
dropout=0.8,
name='lstm_weight')
network = tf.concat([textual_embedding, input_citation],
1) # dimension=256=(d(t) + d(c))
network = tflearn.fully_connected(network, em_dim, activation='softmax')
# network = tflearn.bidirectional_rnn(network, tflearn.BasicLSTMCell(128), tflearn.BasicLSTMCell(128),
# name='bilstm_weight')
# network = tflearn.dropout(network, 0.8)
G_output = _cat_weighted2(network, group_embedding, input_g)
SC_output = _cat_weighted2(G_output, subclass_embedding, input_sc)
C_output = _cat_weighted2(SC_output, class_embedding, input_c)
SS_output = _cat_weighted2(C_output, subsection_embedding, input_ss)
network = tf.matmul(SS_output,
tf.transpose(section_embedding),
name='section_weight')
# network = tflearn.fully_connected(network, SEC_NUM, activation='softmax') # not sure
network = tflearn.softmax(network)
network = tflearn.regression(network,
optimizer='adam',
learning_rate=0.001,
loss='categorical_crossentropy')
return network
def create_critic_network(self, Scope):
inputs = tf.placeholder(shape=[1, self.max_lenth], dtype=tf.int32, name="inputs")
action = tf.placeholder(shape=[1, self.max_lenth], dtype=tf.int32, name="action")
action_pos = tf.placeholder(shape=[1, None], dtype=tf.int32, name="action_pos")
lenth = tf.placeholder(shape=[1], dtype=tf.int32, name="lenth")
lenth_up = tf.placeholder(shape=[1], dtype=tf.int32, name="lenth_up")
#Lower network
if Scope[-1] == 'e':
vec = tf.nn.embedding_lookup(self.wordvector, inputs)
print "active"
else:
vec = tf.nn.embedding_lookup(self.target_wordvector, inputs)
print "target"
cell = LSTMCell(self.dim, initializer=self.init, state_is_tuple=False)
self.state_size = cell.state_size
actions = tf.to_float(action)
h = cell.zero_state(1, tf.float32)
embedding = []
for step in range(self.max_lenth):
with tf.variable_scope("Lower/"+Scope, reuse=True):
o, h = cell(vec[:,step,:], h)
embedding.append(o[0])
h = h *(1.0 - actions[0,step])
#Upper network
embedding = tf.stack(embedding)
embedding = tf.gather(embedding, action_pos, name="Upper_input")
with tf.variable_scope("Upper", reuse=True):
out, _ = tf.nn.bidirectional_dynamic_rnn(cell, cell, embedding, lenth_up, dtype=tf.float32, scope=Scope)
if self.isAttention:
out = tf.concat(out, 2)
out = out[0,:,:]
tmp = tflearn.fully_connected(out, self.dim, scope=Scope, name="att")
tmp = tflearn.tanh(tmp)
with tf.variable_scope(Scope):
v_T = tf.get_variable("v_T", dtype=tf.float32, shape=[self.dim, 1], trainable=True)
a = tflearn.softmax(tf.matmul(tmp,v_T))
out = tf.reduce_sum(out * a, 0)
out = tf.expand_dims(out, 0)
else:
#out = embedding[:, -1, :]
out = tf.concat((out[0][:,-1,:], out[1][:,0,:]), 1)
out = tflearn.dropout(out, self.keep_prob)
out = tflearn.fully_connected(out, self.grained, scope=Scope+"/pred", name="get_pred")
return inputs, action, action_pos, lenth, lenth_up, out
def create_citation(in_dim, em_dim):
# Building 'citation'
input_layer = tflearn.input_data(
shape=[None, in_dim + G_NUM + SUBC_NUM + C_NUM + SUBS_NUM],
name='input')
input_p = input_layer[:, 0:in_dim]
input_ss = input_layer[:, in_dim:in_dim + SUBS_NUM]
input_c = input_layer[:, in_dim + SUBS_NUM:in_dim + SUBS_NUM + C_NUM]
input_sc = input_layer[:, in_dim + SUBS_NUM + C_NUM:in_dim + SUBS_NUM +
C_NUM + SUBC_NUM]
input_g = input_layer[:, in_dim + SUBS_NUM + C_NUM + SUBC_NUM:]
group_embedding = tf.Variable(tf.random_normal([G_NUM, em_dim]),
name='group_embedding')
subclass_embedding = tf.Variable(tf.random_normal([SUBC_NUM, em_dim]),
name='subclass_embedding')
class_embedding = tf.Variable(tf.random_normal([C_NUM, em_dim]),
name='class_embedding')
subsection_embedding = tf.Variable(tf.random_normal([SUBS_NUM, em_dim]),
name='subsection_embedding')
section_embedding = tf.Variable(tf.random_normal([SEC_NUM, em_dim]),
name='section_embedding')
# SG_output = _cat_weighted(X, subgroup_embedding, W_SG)
network = tflearn.embedding(input_p,
input_dim=110240,
output_dim=128,
name='word_embedding')
network = tflearn.lstm(network, em_dim, dropout=0.8, name='lstm_weight')
G_output = _cat_weighted1(network, group_embedding, input_g)
SC_output = _cat_weighted1(G_output, subclass_embedding, input_sc)
C_output = _cat_weighted1(SC_output, class_embedding, input_c)
SS_output = _cat_weighted1(C_output, subsection_embedding, input_ss)
network = tf.matmul(SS_output,
tf.transpose(section_embedding),
name='section_weight')
# network = tflearn.fully_connected(network, SEC_NUM, activation='softmax') # not sure
network = tflearn.softmax(network)
network = tflearn.regression(network,
optimizer='adam',
learning_rate=0.001,
loss='categorical_crossentropy')
return network