def build_model(self):
self.ctsm_model = ContinuousLSTM()
with tf.variable_scope('ShortTermIntentEncoder', reuse=tf.AUTO_REUSE):
time_aware_gru_net_input = tf.concat([
self.behavior_list_embedding_dense,
],
axis=2)
self.short_term_intent_temp = self.ctsm_model.my_lstm_net(
hidden_units=self.num_units,
input_data=time_aware_gru_net_input,
input_length=tf.add(self.seq_length,
-1)) # batch_size, max_len, num_units
emb = gather_indexes(batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=self.num_units,
sequence_tensor=self.short_term_intent_temp,
positions=self.mask_index -
1) #batch_size, num_units
self.predict_behavior_emb = layer_norm(
emb) # batch_size, num_units
self.output()
def build_model(self):
self.ctsm_model = ContinuousLSTM()
last_time = tf.squeeze(gather_indexes(
batch_size=self.now_bacth_data_size,
seq_length=self.max_len,
width=1,
sequence_tensor=self.time_list,
positions=self.mask_index - 1),
axis=1)
h_i, h_i_bar, delta_i = self.get_state()
predict_target_lambda_emb = self.cal_ht(h_i, h_i_bar, delta_i,
self.target[2] - last_time)
self.predict_behavior_emb = predict_target_lambda_emb
self.output()
def build_model(self):
self.ctsm_model = ContinuousLSTM()
self.transformer_model = transformer_encoder()
last_time = tf.squeeze(gather_indexes(batch_size=self.now_batch_size,
seq_length=self.max_seq_len,
width=1,
sequence_tensor=self.time_lst,
positions=self.mask_index - 1),
axis=1)
o_i, c_i, c_i_bar, delta_i, h_i_minus = self.get_state(
self.target_time_last_lst)
predict_target_lambda_emb = self.cal_ht(o_i, c_i, c_i_bar, delta_i,
self.target_time - last_time)
# sims_time_lst: batch_size, sims_len
predict_sims_emb = tf.zeros([self.now_batch_size, 1])
self.test = tf.split(self.sims_time_lst, self.sims_len, 1)
sims_time = tf.squeeze(tf.split(self.sims_time_lst, self.sims_len, 1),
2)
for i in range(self.sims_len):
# 第i个时间 batch_size, num_units
cur_sims_emb = self.cal_ht(o_i, c_i, c_i_bar, delta_i,
sims_time[i] - last_time)
predict_sims_emb = tf.concat([predict_sims_emb, cur_sims_emb],
axis=1)
predict_sims_emb = predict_sims_emb[:,
1:] # batch_size, sims_len * num_units
predict_sims_emb = tf.reshape(predict_sims_emb,
[-1, self.sims_len, self.num_units
]) # batch_size, sims_len , num_units
self.predict_target_emb = predict_target_lambda_emb #
self.predict_sims_emb = predict_sims_emb
with tf.variable_scope('prepare_emb'):
emb_for_type = self.predict_target_emb
emb_for_time = h_i_minus
with tf.variable_scope('intensity_calculation', reuse=tf.AUTO_REUSE):
intensity_model = nhp_intensity_calculation()
self.target_lambda = intensity_model.cal_target_intensity(
hidden_emb=self.predict_target_emb, type_num=self.type_num)
self.sims_lambda = intensity_model.cal_sims_intensity(
hidden_emb=self.predict_sims_emb,
sims_len=self.sims_len,
type_num=self.type_num)
with tf.variable_scope('type_time_calculation', reuse=tf.AUTO_REUSE):
time_predictor = thp_time_predictor()
self.predict_time = time_predictor.predict_time(
emb=emb_for_time,
num_units=self.num_units,
)
type_predictor = thp_type_predictor()
self.predict_type_prob = type_predictor.predict_type(
emb=emb_for_type,
num_units=self.num_units,
type_num=self.type_num)
# self.predict_type_prob = tf.matmul(self.predict_type_prob, self.embedding.type_emb_lookup_table[:-3, :],
# transpose_b=True)
self.output()