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_model(self, l, tN, N=100000, d=10, K=5, H=1000, m=0.05, reuse=False):
'''
N = 1000000 (Paper)
d = Unknown
'''
with tf.variable_scope('TagSpace', reuse=reuse):
lr = tf.placeholder('float32', shape=[1], name='lr')
doc = tf.placeholder('float32', shape=[None, l], name='doc')
tag_flag = tf.placeholder('float32', shape=[None, tN], name='tag_flag')
doc_embed = tflearn.embedding(doc, input_dim=N, output_dim=d)
self.lt_embed = lt_embed = tf.Variable(tf.random_normal([tN, d], stddev=0.1))
net = tflearn.conv_1d(doc_embed, K, H, activation='tanh')
net = tflearn.max_pool_1d(net, l)
net = tflearn.tanh(net)
self.logit = logit = tflearn.fully_connected(net, d, activation=None)
zero_vector = tf.zeros(shape=(1,1), dtype=tf.float32)
logit = tf.expand_dims(logit, 1)
logit_set = tf.concat([logit for i in range(tN)], axis=1)
tag_flag_ex = tf.expand_dims(tag_flag, 2)
tg = tf.concat([tag_flag_ex for i in range(d)], axis=2)
self.tag_logit = tf.reduce_sum(tf.multiply(logit_set, tf.multiply(tf.ones_like(tg), lt_embed)), axis=2)
self.positive_logit = positive_logit = tf.reduce_sum(tf.multiply(logit_set, tf.multiply(tg, lt_embed)), axis=2)
with tf.device('/cpu:0'):
self.f_positive = f_positive = tf.map_fn(lambda x: (tf.boolean_mask(x[0], x[1]), True), (positive_logit, tf.not_equal(positive_logit, zero_vector)))
positive = tf.reduce_min(f_positive[0], axis=1)
self.positive = positive
tag_flag_ex = tf.expand_dims(1-tag_flag, 2)
tg = tf.concat([tag_flag_ex for i in range(d)], axis=2)
negative_logit = tf.reduce_sum(tf.multiply(logit_set, tf.multiply(tg, lt_embed)), axis=2)
with tf.device('/cpu:0'):
self.f_negative = f_negative = tf.map_fn(lambda x: (tf.boolean_mask(x[0], x[1]), True), (negative_logit, tf.not_equal(negative_logit, zero_vector)))
self.negative = negative = tf.reduce_max(f_negative[0], axis=1)
self.f_loss = f_loss = tf.reduce_mean(tf.reduce_max([tf.reduce_min([tf.expand_dims(m - positive + negative,1), tf.expand_dims(tf.fill([tf.shape(doc)[0]], 10e7),1)], axis=0), tf.zeros([tf.shape(doc)[0], 1])], axis=0))
opt = tf.train.AdamOptimizer(learning_rate=lr[0])
self.op = opt.minimize(f_loss)
from tflearn.helpers.evaluator import Evaluator from tflearn.helpers.summarizer import summarize from tflearn.helpers.regularizer import add_weights_regularizer from tensorflow.contrib.slim import dataset from tensorflow.contrib.slim import dataset tflearn.input_data() tflearn.variable() tflearn.conv_2d() tflearn.single_unit() tflearn.lstm() tflearn.embedding() tflearn.batch_normalization() tflearn.merge() tflearn.regression() tflearn.tanh() tflearn.softmax_categorical_crossentropy() tflearn.SGD() tflearn.initializations.uniform() tflearn.losses.L1() tflearn.add_weights_regularizer() tflearn.metrics.Accuracy() tflearn.summaries() tflearn.ImagePreprocessing() tflearn.ImageAugmentation() tflearn.init_graph()
l = word_pad_length
tN = tag_size
N=100000
d=10
K=5
H=1000
m=0.05
lr = tf.placeholder('float32', shape=[1], name='lr')
doc = tf.placeholder('float32', shape=[None, l], name='doc')
tag_flag = tf.placeholder('float32', shape=[None, tN], name='tag_flag')
doc_embed = tflearn.embedding(doc, input_dim=N, output_dim=d)
lt_embed = tf.Variable(tf.random_normal([tN, d], stddev=0.1))
net = tflearn.conv_1d(doc_embed, H, K, activation='tanh')
net = tflearn.max_pool_1d(net, K)
net = tflearn.tanh(net)
logit = tflearn.fully_connected(net, d, activation=None)
zero_vector = tf.zeros(shape=(1,1), dtype=tf.float32)
logit = tf.expand_dims(logit, 1)
logit_set = tf.concat([logit for i in range(tN)], axis=1)
tag_flag_ex = tf.expand_dims(tag_flag, 2)
tg = tf.concat([tag_flag_ex for i in range(d)], axis=2)
tag_logit = tf.reduce_sum(tf.multiply(logit_set, tf.multiply(tf.ones_like(tg), lt_embed)), axis=2)
positive_logit = tf.reduce_sum(tf.multiply(logit_set, tf.multiply(tg, lt_embed)), axis=2)
random_sample = tf.random_uniform([batch_size],minval=0,maxval=1,dtype=tf.float32)
#f_test = tf.not_equal(positive_logit, zero_vector)
